Intel® 64 and IA-32 Architectures

Software Developer’s Manual

Volume 1:
Basic Architecture

NOTE: The Intel® 64 and IA-32 Architectures Software Developer's

Manual consists of five volumes: Basic Architecture, Order Number
253665; Instruction Set Reference A-M, Order Number 253666;
Instruction Set Reference N-Z, Order Number 253667; System
Programming Guide, Part 1, Order Number 253668; System Programming
Guide, Part 2, Order Number 253669. Refer to all five volumes when
evaluating your design needs.

Order Number: 253665-032US

September 2009
INFORMATION IN THIS DOCUMENT IS PROVIDED IN CONNECTION WITH INTEL PRODUCTS. NO LICENSE,
EXPRESS OR IMPLIED, BY ESTOPPEL OR OTHERWISE, TO ANY INTELLECTUAL PROPERTY RIGHTS IS GRANT-
ED BY THIS DOCUMENT. EXCEPT AS PROVIDED IN INTEL'S TERMS AND CONDITIONS OF SALE FOR SUCH
PRODUCTS, INTEL ASSUMES NO LIABILITY WHATSOEVER AND INTEL DISCLAIMS ANY EXPRESS OR IMPLIED
WARRANTY, RELATING TO SALE AND/OR USE OF INTEL PRODUCTS INCLUDING LIABILITY OR WARRANTIES
RELATING TO FITNESS FOR A PARTICULAR PURPOSE, MERCHANTABILITY, OR INFRINGEMENT OF ANY
PATENT, COPYRIGHT OR OTHER INTELLECTUAL PROPERTY RIGHT.

UNLESS OTHERWISE AGREED IN WRITING BY INTEL, THE INTEL PRODUCTS ARE NOT DESIGNED NOR IN-
TENDED FOR ANY APPLICATION IN WHICH THE FAILURE OF THE INTEL PRODUCT COULD CREATE A SITUA-
TION WHERE PERSONAL INJURY OR DEATH MAY OCCUR.
Intel may make changes to specifications and product descriptions at any time, without notice. Designers
must not rely on the absence or characteristics of any features or instructions marked "reserved" or "unde-
fined." Intel reserves these for future definition and shall have no responsibility whatsoever for conflicts or
incompatibilities arising from future changes to them. The information here is subject to change without no-
tice. Do not finalize a design with this information.
The Intel® 64 architecture processors may contain design defects or errors known as errata. Current char-
acterized errata are available on request.
Intel® Hyper-Threading Technology requires a computer system with an Intel® processor supporting Intel
Hyper-Threading Technology and an Intel® HT Technology enabled chipset, BIOS and operating system.
Performance will vary depending on the specific hardware and software you use. For more information, see
http://www.intel.com/technology/hyperthread/index.htm; including details on which processors support Intel HT
Technology.
Intel® Virtualization Technology requires a computer system with an enabled Intel® processor, BIOS, virtual
machine monitor (VMM) and for some uses, certain platform software enabled for it. Functionality, perfor-
mance or other benefits will vary depending on hardware and software configurations. Intel® Virtualization
Technology-enabled BIOS and VMM applications are currently in development.
64-bit computing on Intel architecture requires a computer system with a processor, chipset, BIOS, oper-
ating system, device drivers and applications enabled for Intel® 64 architecture. Processors will not operate
(including 32-bit operation) without an Intel® 64 architecture-enabled BIOS. Performance will vary de-
pending on your hardware and software configurations. Consult with your system vendor for more infor-
mation.
Enabling Execute Disable Bit functionality requires a PC with a processor with Execute Disable Bit capability
and a supporting operating system. Check with your PC manufacturer on whether your system delivers Ex-
ecute Disable Bit functionality.
Intel, Pentium, Intel Xeon, Intel NetBurst, Intel Core, Intel Core Solo, Intel Core Duo, Intel Core 2 Duo,
Intel Core 2 Extreme, Intel Pentium D, Itanium, Intel SpeedStep, MMX, Intel Atom, and VTune are trade-
marks or registered trademarks of Intel Corporation or its subsidiaries in the United States and other coun-
tries.
*Other names and brands may be claimed as the property of others.
Contact your local Intel sales office or your distributor to obtain the latest specifications and before placing
your product order.
Copies of documents which have an ordering number and are referenced in this document, or other Intel
literature, may be obtained by calling 1-800-548-4725, or by visiting Intel’s website at http://www.intel.com

Copyright © 1997-2009 Intel Corporation

 CONTENTS
PAGE
CHAPTER 1
ABOUT THIS MANUAL
1.1 INTEL® 64 AND IA-32 PROCESSORS COVERED IN THIS MANUAL . . . . . . . . . . . . . . . . . . . . . . 1-1
1.2 OVERVIEW OF VOLUME 1: BASIC ARCHITECTURE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-3
1.3 NOTATIONAL CONVENTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-5
1.3.1 Bit and Byte Order . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-5
1.3.2 Reserved Bits and Software Compatibility. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-5
1.3.2.1 Instruction Operands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-6
1.3.3 Hexadecimal and Binary Numbers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-7
1.3.4 Segmented Addressing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-7
1.3.5 A New Syntax for CPUID, CR, and MSR Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-7
1.3.6 Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-8
1.4 RELATED LITERATURE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-9

CHAPTER 2
INTEL® 64 AND IA-32 ARCHITECTURES
2.1 BRIEF HISTORY OF INTEL® 64 AND IA-32 ARCHITECTURE. . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-1
2.1.1 16-bit Processors and Segmentation (1978) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-1
2.1.2 The Intel® 286 Processor (1982) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-1
2.1.3 The Intel386™ Processor (1985) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-2
2.1.4 The Intel486™ Processor (1989) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-2
2.1.5 The Intel® Pentium® Processor (1993) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-2
2.1.6 The P6 Family of Processors (1995-1999) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-3
2.1.7 The Intel® Pentium® 4 Processor Family (2000-2006) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-4
2.1.8 The Intel® Xeon® Processor (2001- 2007) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-4
2.1.9 The Intel® Pentium® M Processor (2003-Current). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-5
2.1.10 The Intel® Pentium® Processor Extreme Edition (2005-2007). . . . . . . . . . . . . . . . . . . . . 2-5
2.1.11 The Intel® Core™ Duo and Intel® Core™ Solo Processors (2006-2007). . . . . . . . . . . . . 2-5
2.1.12 The Intel® Xeon® Processor 5100, 5300 Series and
Intel® Core™2 Processor Family (2006-Current) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-6
2.1.13 The Intel® Xeon® Processor 5200, 5400, 7400 Series and
Intel® Core™2 Processor Family (2007-Current) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-6
2.1.14 The Intel® Atom™ Processor Family (2008-Current) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-7
2.1.15 The Intel® Core™i7 Processor Family (2008-Current) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-7
2.2 MORE ON SPECIFIC ADVANCES. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-8
2.2.1 P6 Family Microarchitecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-8
2.2.2 Intel NetBurst® Microarchitecture. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-10
2.2.2.1 The Front End Pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-12
2.2.2.2 Out-Of-Order Execution Core . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-13
2.2.2.3 Retirement Unit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-13
2.2.3 Intel® Core™ Microarchitecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-13
2.2.3.1 The Front End . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-15
2.2.3.2 Execution Core . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-16
2.2.4 Intel® Atom™ Microarchitecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-16

Vol. 1 iii
CONTENTS

PAGE
2.2.5 Intel Microarchitecture (Nehalem) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-17
2.2.6 SIMD Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-18
2.2.7 Intel® Hyper-Threading Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-21
2.2.7.1 Some Implementation Notes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-22
2.2.8 Multi-Core Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-22
2.2.9 Intel® 64 Architecture. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-26
2.2.10 Intel® Virtualization Technology (Intel® VT) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-27
2.3 INTEL® 64 AND IA-32 PROCESSOR GENERATIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-27

CHAPTER 3
BASIC EXECUTION ENVIRONMENT
3.1 MODES OF OPERATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-1
3.1.1 Intel® 64 Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-2
3.2 OVERVIEW OF THE BASIC EXECUTION ENVIRONMENT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-2
3.2.1 64-Bit Mode Execution Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-6
3.3 MEMORY ORGANIZATION. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-8
3.3.1 IA-32 Memory Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-8
3.3.2 Paging and Virtual Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-10
3.3.3 Memory Organization in 64-Bit Mode. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-10
3.3.4 Modes of Operation vs. Memory Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-10
3.3.5 32-Bit and 16-Bit Address and Operand Sizes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-11
3.3.6 Extended Physical Addressing in Protected Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-12
3.3.7 Address Calculations in 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-12
3.3.7.1 Canonical Addressing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-13
3.4 BASIC PROGRAM EXECUTION REGISTERS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-13
3.4.1 General-Purpose Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-14
3.4.1.1 General-Purpose Registers in 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-16
3.4.2 Segment Registers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-17
3.4.2.1 Segment Registers in 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-20
3.4.3 EFLAGS Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-20
3.4.3.1 Status Flags . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-21
3.4.3.2 DF Flag . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-22
3.4.3.3 System Flags and IOPL Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-23
3.4.3.4 RFLAGS Register in 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-24
3.5 INSTRUCTION POINTER. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-24
3.5.1 Instruction Pointer in 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-24
3.6 OPERAND-SIZE AND ADDRESS-SIZE ATTRIBUTES. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-24
3.6.1 Operand Size and Address Size in 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-25
3.7 OPERAND ADDRESSING . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-26
3.7.1 Immediate Operands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-27
3.7.2 Register Operands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-27
3.7.2.1 Register Operands in 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-28
3.7.3 Memory Operands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-28
3.7.3.1 Memory Operands in 64-Bit Mode. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-29
3.7.4 Specifying a Segment Selector. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-29
3.7.4.1 Segmentation in 64-Bit Mode. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-30
3.7.5 Specifying an Offset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-30

iv Vol. 1
 CONTENTS

PAGE
3.7.5.1 Specifying an Offset in 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3-32
3.7.6 Assembler and Compiler Addressing Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3-32
3.7.7 I/O Port Addressing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3-33

CHAPTER 4
DATA TYPES
4.1 FUNDAMENTAL DATA TYPES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-1
4.1.1 Alignment of Words, Doublewords, Quadwords, and Double Quadwords . . . . . . . . . . . . 4-2
4.2 NUMERIC DATA TYPES. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-3
4.2.1 Integers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-4
4.2.1.1 Unsigned Integers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-5
4.2.1.2 Signed Integers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-5
4.2.2 Floating-Point Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-6
4.3 POINTER DATA TYPES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-9
4.3.1 Pointer Data Types in 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-9
4.4 BIT FIELD DATA TYPE. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-10
4.5 STRING DATA TYPES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-11
4.6 PACKED SIMD DATA TYPES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-11
4.6.1 64-Bit SIMD Packed Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-11
4.6.2 128-Bit Packed SIMD Data Types. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-12
4.7 BCD AND PACKED BCD INTEGERS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-13
4.8 REAL NUMBERS AND FLOATING-POINT FORMATS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-15
4.8.1 Real Number System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-16
4.8.2 Floating-Point Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-16
4.8.2.1 Normalized Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-18
4.8.2.2 Biased Exponent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-18
4.8.3 Real Number and Non-number Encodings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-19
4.8.3.1 Signed Zeros . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-20
4.8.3.2 Normalized and Denormalized Finite Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-20
4.8.3.3 Signed Infinities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-21
4.8.3.4 NaNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-21
4.8.3.5 Operating on SNaNs and QNaNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-22
4.8.3.6 Using SNaNs and QNaNs in Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-23
4.8.3.7 QNaN Floating-Point Indefinite . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-24
4.8.4 Rounding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-24
4.8.4.1 Rounding Control (RC) Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-25
4.8.4.2 Truncation with SSE and SSE2 Conversion Instructions . . . . . . . . . . . . . . . . . . . . . . . .4-26
4.9 OVERVIEW OF FLOATING-POINT EXCEPTIONS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-26
4.9.1 Floating-Point Exception Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-28
4.9.1.1 Invalid Operation Exception (#I) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-28
4.9.1.2 Denormal Operand Exception (#D). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-28
4.9.1.3 Divide-By-Zero Exception (#Z) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-29
4.9.1.4 Numeric Overflow Exception (#O) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-29
4.9.1.5 Numeric Underflow Exception (#U) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-30
4.9.1.6 Inexact-Result (Precision) Exception (#P) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-31
4.9.2 Floating-Point Exception Priority . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-32
4.9.3 Typical Actions of a Floating-Point Exception Handler . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-33

Vol. 1 v
CONTENTS

PAGE
CHAPTER 5
INSTRUCTION SET SUMMARY
5.1 GENERAL-PURPOSE INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-3
5.1.1 Data Transfer Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-3
5.1.2 Binary Arithmetic Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-4
5.1.3 Decimal Arithmetic Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-5
5.1.4 Logical Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-5
5.1.5 Shift and Rotate Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-5
5.1.6 Bit and Byte Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-5
5.1.7 Control Transfer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-6
5.1.8 String Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-7
5.1.9 I/O Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-8
5.1.10 Enter and Leave Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-8
5.1.11 Flag Control (EFLAG) Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-8
5.1.12 Segment Register Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-9
5.1.13 Miscellaneous Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-9
5.2 X87 FPU INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-9
5.2.1 x87 FPU Data Transfer Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-10
5.2.2 x87 FPU Basic Arithmetic Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-10
5.2.3 x87 FPU Comparison Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-11
5.2.4 x87 FPU Transcendental Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-12
5.2.5 x87 FPU Load Constants Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-12
5.2.6 x87 FPU Control Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-12
5.3 X87 FPU AND SIMD STATE MANAGEMENT INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . 5-13
5.4 MMX™ INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-13
5.4.1 MMX Data Transfer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-14
5.4.2 MMX Conversion Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-14
5.4.3 MMX Packed Arithmetic Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-14
5.4.4 MMX Comparison Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-15
5.4.5 MMX Logical Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-15
5.4.6 MMX Shift and Rotate Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-15
5.4.7 MMX State Management Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-16
5.5 SSE INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-16
5.5.1 SSE SIMD Single-Precision Floating-Point Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-16
5.5.1.1 SSE Data Transfer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-16
5.5.1.2 SSE Packed Arithmetic Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-17
5.5.1.3 SSE Comparison Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-18
5.5.1.4 SSE Logical Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-18
5.5.1.5 SSE Shuffle and Unpack Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-18
5.5.1.6 SSE Conversion Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-19
5.5.2 SSE MXCSR State Management Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-19
5.5.3 SSE 64-Bit SIMD Integer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-19
5.5.4 SSE Cacheability Control, Prefetch, and Instruction Ordering Instructions. . . . . . . . . . 5-20
5.6 SSE2 INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-20
5.6.1 SSE2 Packed and Scalar Double-Precision Floating-Point Instructions . . . . . . . . . . . . . 5-21
5.6.1.1 SSE2 Data Movement Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-21
5.6.1.2 SSE2 Packed Arithmetic Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-21

vi Vol. 1
 CONTENTS

PAGE
5.6.1.3 SSE2 Logical Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-22
5.6.1.4 SSE2 Compare Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-22
5.6.1.5 SSE2 Shuffle and Unpack Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-22
5.6.1.6 SSE2 Conversion Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-23
5.6.2 SSE2 Packed Single-Precision Floating-Point Instructions . . . . . . . . . . . . . . . . . . . . . . . . .5-23
5.6.3 SSE2 128-Bit SIMD Integer Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-24
5.6.4 SSE2 Cacheability Control and Ordering Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-24
5.7 SSE3 INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-25
5.7.1 SSE3 x87-FP Integer Conversion Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-25
5.7.2 SSE3 Specialized 128-bit Unaligned Data Load Instruction . . . . . . . . . . . . . . . . . . . . . . . .5-25
5.7.3 SSE3 SIMD Floating-Point Packed ADD/SUB Instructions . . . . . . . . . . . . . . . . . . . . . . . . . .5-25
5.7.4 SSE3 SIMD Floating-Point Horizontal ADD/SUB Instructions . . . . . . . . . . . . . . . . . . . . . . .5-26
5.7.5 SSE3 SIMD Floating-Point LOAD/MOVE/DUPLICATE Instructions. . . . . . . . . . . . . . . . . . .5-26
5.7.6 SSE3 Agent Synchronization Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-27
5.8 SUPPLEMENTAL STREAMING SIMD EXTENSIONS 3 (SSSE3) INSTRUCTIONS . . . . . . . . . . 5-27
5.8.1 Horizontal Addition/Subtraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-27
5.8.2 Packed Absolute Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-28
5.8.3 Multiply and Add Packed Signed and Unsigned Bytes . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-28
5.8.4 Packed Multiply High with Round and Scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-28
5.8.5 Packed Shuffle Bytes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-29
5.8.6 Packed Sign . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-29
5.8.7 Packed Align Right. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-29
5.9 SSE4 INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-29
5.10 SSE4.1 INSTRUCTIONS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-30
5.10.1 Dword Multiply Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-30
5.10.2 Floating-Point Dot Product Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-31
5.10.3 Streaming Load Hint Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-31
5.10.4 Packed Blending Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-31
5.10.5 Packed Integer MIN/MAX Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-31
5.10.6 Floating-Point Round Instructions with Selectable Rounding Mode . . . . . . . . . . . . . . . .5-32
5.10.7 Insertion and Extractions from XMM Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-32
5.10.8 Packed Integer Format Conversions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-33
5.10.9 Improved Sums of Absolute Differences (SAD) for 4-Byte Blocks . . . . . . . . . . . . . . . . . .5-33
5.10.10 Horizontal Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-33
5.10.11 Packed Test. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-34
5.10.12 Packed Qword Equality Comparisons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-34
5.10.13 Dword Packing With Unsigned Saturation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-34
5.11 SSE4.2 INSTRUCTION SET. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-34
5.11.1 String and Text Processing Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-34
5.11.2 Packed Comparison SIMD integer Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-34
5.11.3 Application-Targeted Accelerator Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-35
5.12 SYSTEM INSTRUCTIONS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-35
5.13 64-BIT MODE INSTRUCTIONS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-36
5.14 VIRTUAL-MACHINE EXTENSIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-36
5.15 SAFER MODE EXTENSIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-37

Vol. 1 vii
CONTENTS

PAGE
CHAPTER 6
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS
6.1 PROCEDURE CALL TYPES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-1
6.2 STACKS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-1
6.2.1 Setting Up a Stack . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-2
6.2.2 Stack Alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-3
6.2.3 Address-Size Attributes for Stack Accesses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-3
6.2.4 Procedure Linking Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-4
6.2.4.1 Stack-Frame Base Pointer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-4
6.2.4.2 Return Instruction Pointer. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-4
6.2.5 Stack Behavior in 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-5
6.3 CALLING PROCEDURES USING CALL AND RET. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-5
6.3.1 Near CALL and RET Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-5
6.3.2 Far CALL and RET Operation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-6
6.3.3 Parameter Passing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-7
6.3.3.1 Passing Parameters Through the General-Purpose Registers . . . . . . . . . . . . . . . . . . . 6-7
6.3.3.2 Passing Parameters on the Stack . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-7
6.3.3.3 Passing Parameters in an Argument List . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-8
6.3.4 Saving Procedure State Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-8
6.3.5 Calls to Other Privilege Levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-8
6.3.6 CALL and RET Operation Between Privilege Levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-10
6.3.7 Branch Functions in 64-Bit Mode. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-11
6.4 INTERRUPTS AND EXCEPTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-13
6.4.1 Call and Return Operation for Interrupt or Exception Handling Procedures . . . . . . . . 6-14
6.4.2 Calls to Interrupt or Exception Handler Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-17
6.4.3 Interrupt and Exception Handling in Real-Address Mode. . . . . . . . . . . . . . . . . . . . . . . . . . 6-17
6.4.4 INT n, INTO, INT 3, and BOUND Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-18
6.4.5 Handling Floating-Point Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-18
6.4.6 Interrupt and Exception Behavior in 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-19
6.5 PROCEDURE CALLS FOR BLOCK-STRUCTURED LANGUAGES . . . . . . . . . . . . . . . . . . . . . . . . . 6-19
6.5.1 ENTER Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-20
6.5.2 LEAVE Instruction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-25

CHAPTER 7
PROGRAMMING WITH
GENERAL-PURPOSE INSTRUCTIONS
7.1 PROGRAMMING ENVIRONMENT FOR GP INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-1
7.2 PROGRAMMING ENVIRONMENT FOR GP INSTRUCTIONS IN 64-BIT MODE . . . . . . . . . . . . . . 7-2
7.3 SUMMARY OF GP INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-3
7.3.1 Data Transfer Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-3
7.3.1.1 General Data Movement Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-4
7.3.1.2 Exchange Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-5
7.3.1.3 Exchange Instructions in 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-7
7.3.1.4 Stack Manipulation Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-7
7.3.1.5 Stack Manipulation Instructions in 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-9
7.3.1.6 Type Conversion Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-10
7.3.1.7 Type Conversion Instructions in 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-11

viii Vol. 1
 CONTENTS

PAGE
7.3.2 Binary Arithmetic Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-12
7.3.2.1 Addition and Subtraction Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-12
7.3.2.2 Increment and Decrement Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-12
7.3.2.3 Increment and Decrement Instructions in 64-Bit Mode. . . . . . . . . . . . . . . . . . . . . . . . .7-12
7.3.2.4 Comparison and Sign Change Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-12
7.3.2.5 Multiplication and Divide Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-13
7.3.3 Decimal Arithmetic Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-13
7.3.3.1 Packed BCD Adjustment Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-14
7.3.3.2 Unpacked BCD Adjustment Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-14
7.3.4 Decimal Arithmetic Instructions in 64-Bit Mode. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-15
7.3.5 Logical Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-15
7.3.6 Shift and Rotate Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-15
7.3.6.1 Shift Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-15
7.3.6.2 Double-Shift Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-17
7.3.6.3 Rotate Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-18
7.3.7 Bit and Byte Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-20
7.3.7.1 Bit Test and Modify Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-20
7.3.7.2 Bit Scan Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-20
7.3.7.3 Byte Set on Condition Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-20
7.3.7.4 Test Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-21
7.3.8 Control Transfer Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-21
7.3.8.1 Unconditional Transfer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-21
7.3.8.2 Conditional Transfer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-23
7.3.8.3 Control Transfer Instructions in 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-25
7.3.8.4 Software Interrupt Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-25
7.3.8.5 Software Interrupt Instructions in 64-bit Mode and Compatibility Mode . . . . . . . .7-26
7.3.9 String Operations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-26
7.3.9.1 Repeating String Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-27
7.3.10 String Operations in 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-28
7.3.10.1 Repeating String Operations in 64-bit Mode. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-28
7.3.11 I/O Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-28
7.3.12 I/O Instructions in 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-29
7.3.13 Enter and Leave Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-29
7.3.14 Flag Control (EFLAG) Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-29
7.3.14.1 Carry and Direction Flag Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-29
7.3.14.2 EFLAGS Transfer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-30
7.3.14.3 Interrupt Flag Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-31
7.3.15 Flag Control (RFLAG) Instructions in 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-31
7.3.16 Segment Register Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-31
7.3.16.1 Segment-Register Load and Store Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-31
7.3.16.2 Far Control Transfer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-32
7.3.16.3 Software Interrupt Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-32
7.3.16.4 Load Far Pointer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-32
7.3.17 Miscellaneous Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-32
7.3.17.1 Address Computation Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-33
7.3.17.2 Table Lookup Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-33
7.3.17.3 Processor Identification Instruction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-33

Vol. 1 ix
CONTENTS

PAGE
7.3.17.4 No-Operation and Undefined Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-33

CHAPTER 8
PROGRAMMING WITH THE X87 FPU
8.1 X87 FPU EXECUTION ENVIRONMENT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-1
8.1.1 x87 FPU in 64-Bit Mode and Compatibility Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-2
8.1.2 x87 FPU Data Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-2
8.1.2.1 Parameter Passing With the x87 FPU Register Stack . . . . . . . . . . . . . . . . . . . . . . . . . . 8-5
8.1.3 x87 FPU Status Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-6
8.1.3.1 Top of Stack (TOP) Pointer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-6
8.1.3.2 Condition Code Flags . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-6
8.1.3.3 x87 FPU Floating-Point Exception Flags. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-7
8.1.3.4 Stack Fault Flag. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-9
8.1.4 Branching and Conditional Moves on Condition Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-9
8.1.5 x87 FPU Control Word . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-10
8.1.5.1 x87 FPU Floating-Point Exception Mask Bits. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-11
8.1.5.2 Precision Control Field. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-11
8.1.5.3 Rounding Control Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-12
8.1.6 Infinity Control Flag . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-12
8.1.7 x87 FPU Tag Word . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-12
8.1.8 x87 FPU Instruction and Data (Operand) Pointers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-13
8.1.9 Last Instruction Opcode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-14
8.1.9.1 Fopcode Compatibility Sub-mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-14
8.1.10 Saving the x87 FPU’s State with FSTENV/FNSTENV and FSAVE/FNSAVE . . . . . . . . . 8-15
8.1.11 Saving the x87 FPU’s State with FXSAVE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-17
8.2 X87 FPU DATA TYPES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-17
8.2.1 Indefinites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-19
8.2.2 Unsupported Double Extended-Precision
Floating-Point Encodings and Pseudo-Denormals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-20
8.3 X86 FPU INSTRUCTION SET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-21
8.3.1 Escape (ESC) Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-22
8.3.2 x87 FPU Instruction Operands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-22
8.3.3 Data Transfer Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-22
8.3.4 Load Constant Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-24
8.3.5 Basic Arithmetic Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-24
8.3.6 Comparison and Classification Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-26
8.3.6.1 Branching on the x87 FPU Condition Codes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-28
8.3.7 Trigonometric Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-29
8.3.8 Pi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-30
8.3.9 Logarithmic, Exponential, and Scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-31
8.3.10 Transcendental Instruction Accuracy. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-31
8.3.11 x87 FPU Control Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-32
8.3.12 Waiting vs. Non-waiting Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-33
8.3.13 Unsupported x87 FPU Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-34
8.4 X87 FPU FLOATING-POINT EXCEPTION HANDLING. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-34
8.4.1 Arithmetic vs. Non-arithmetic Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-35
8.5 X87 FPU FLOATING-POINT EXCEPTION CONDITIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-36

x Vol. 1
 CONTENTS

PAGE
8.5.1 Invalid Operation Exception . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-36
8.5.1.1 Stack Overflow or Underflow Exception (#IS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-37
8.5.1.2 Invalid Arithmetic Operand Exception (#IA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-38
8.5.2 Denormal Operand Exception (#D). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-39
8.5.3 Divide-By-Zero Exception (#Z) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-40
8.5.4 Numeric Overflow Exception (#O) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-40
8.5.5 Numeric Underflow Exception (#U) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-41
8.5.6 Inexact-Result (Precision) Exception (#P) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-42
8.6 X87 FPU EXCEPTION SYNCHRONIZATION. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-43
8.7 HANDLING X87 FPU EXCEPTIONS IN SOFTWARE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-45
8.7.1 Native Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-45
8.7.2 MS-DOS* Compatibility Sub-mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-45
8.7.3 Handling x87 FPU Exceptions in Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-47

CHAPTER 9
PROGRAMMING WITH INTEL® MMX™ TECHNOLOGY
9.1 OVERVIEW OF MMX TECHNOLOGY. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-1
9.2 THE MMX TECHNOLOGY PROGRAMMING ENVIRONMENT . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-2
9.2.1 MMX Technology in 64-Bit Mode and Compatibility Mode . . . . . . . . . . . . . . . . . . . . . . . . . . 9-2
9.2.2 MMX Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-3
9.2.3 MMX Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-3
9.2.4 Memory Data Formats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-4
9.2.5 Single Instruction, Multiple Data (SIMD) Execution Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-4
9.3 SATURATION AND WRAPAROUND MODES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-5
9.4 MMX INSTRUCTIONS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-6
9.4.1 Data Transfer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-8
9.4.2 Arithmetic Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-8
9.4.3 Comparison Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-9
9.4.4 Conversion Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-9
9.4.5 Unpack Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-9
9.4.6 Logical Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-10
9.4.7 Shift Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-10
9.4.8 EMMS Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-10
9.5 COMPATIBILITY WITH X87 FPU ARCHITECTURE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-10
9.5.1 MMX Instructions and the x87 FPU Tag Word . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-11
9.6 WRITING APPLICATIONS WITH MMX CODE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-11
9.6.1 Checking for MMX Technology Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-11
9.6.2 Transitions Between x87 FPU and MMX Code. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-12
9.6.3 Using the EMMS Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-12
9.6.4 Mixing MMX and x87 FPU Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-13
9.6.5 Interfacing with MMX Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-13
9.6.6 Using MMX Code in a Multitasking Operating System Environment . . . . . . . . . . . . . . . .9-14
9.6.7 Exception Handling in MMX Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-14
9.6.8 Register Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-14
9.6.9 Effect of Instruction Prefixes on MMX Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-14

Vol. 1 xi
CONTENTS

PAGE
CHAPTER 10
PROGRAMMING WITH
STREAMING SIMD EXTENSIONS (SSE)
10.1 OVERVIEW OF SSE EXTENSIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-1
10.2 SSE PROGRAMMING ENVIRONMENT. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-3
10.2.1 SSE in 64-Bit Mode and Compatibility Mode. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-4
10.2.2 XMM Registers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-4
10.2.3 MXCSR Control and Status Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-5
10.2.3.1 SIMD Floating-Point Mask and Flag Bits. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-6
10.2.3.2 SIMD Floating-Point Rounding Control Field. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-7
10.2.3.3 Flush-To-Zero . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-7
10.2.3.4 Denormals-Are-Zeros . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-7
10.2.4 Compatibility of SSE Extensions with SSE2/SSE3/MMX and the x87 FPU . . . . . . . . . . 10-8
10.3 SSE DATA TYPES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-8
10.4 SSE INSTRUCTION SET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-9
10.4.1 SSE Packed and Scalar Floating-Point Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-9
10.4.1.1 SSE Data Movement Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-11
10.4.1.2 SSE Arithmetic Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-11
10.4.2 SSE Logical Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-13
10.4.2.1 SSE Comparison Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-13
10.4.2.2 SSE Shuffle and Unpack Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-14
10.4.3 SSE Conversion Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-15
10.4.4 SSE 64-Bit SIMD Integer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-16
10.4.5 MXCSR State Management Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-17
10.4.6 Cacheability Control, Prefetch, and Memory Ordering Instructions . . . . . . . . . . . . . . . 10-18
10.4.6.1 Cacheability Control Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-18
10.4.6.2 Caching of Temporal vs. Non-Temporal Data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-18
10.4.6.3 PREFETCHh Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-19
10.4.6.4 SFENCE Instruction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-20
10.5 FXSAVE AND FXRSTOR INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-20
10.6 HANDLING SSE INSTRUCTION EXCEPTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-21
10.7 WRITING APPLICATIONS WITH THE SSE EXTENSIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-21

CHAPTER 11
PROGRAMMING WITH
STREAMING SIMD EXTENSIONS 2 (SSE2)
11.1 OVERVIEW OF SSE2 EXTENSIONS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-1
11.2 SSE2 PROGRAMMING ENVIRONMENT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-3
11.2.1 SSE2 in 64-Bit Mode and Compatibility Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-4
11.2.2 Compatibility of SSE2 Extensions with SSE, MMX
Technology and x87 FPU Programming Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-4
11.2.3 Denormals-Are-Zeros Flag . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-4
11.3 SSE2 DATA TYPES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-5
11.4 SSE2 INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-6
11.4.1 Packed and Scalar Double-Precision Floating-Point Instructions . . . . . . . . . . . . . . . . . . . 11-6
11.4.1.1 Data Movement Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-7
11.4.1.2 SSE2 Arithmetic Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-8

xii Vol. 1
 CONTENTS

PAGE
11.4.1.3 SSE2 Logical Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .11-9
11.4.1.4 SSE2 Comparison Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .11-9
11.4.1.5 SSE2 Shuffle and Unpack Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-10
11.4.1.6 SSE2 Conversion Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-12
11.4.2 SSE2 64-Bit and 128-Bit SIMD Integer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-15
11.4.3 128-Bit SIMD Integer Instruction Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-16
11.4.4 Cacheability Control and Memory Ordering Instructions . . . . . . . . . . . . . . . . . . . . . . . . . 11-16
11.4.4.1 FLUSH Cache Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-17
11.4.4.2 Cacheability Control Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-17
11.4.4.3 Memory Ordering Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-17
11.4.4.4 Pause. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-18
11.4.5 Branch Hints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-18
11.5 SSE, SSE2, AND SSE3 EXCEPTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-18
11.5.1 SIMD Floating-Point Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-19
11.5.2 SIMD Floating-Point Exception Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-19
11.5.2.1 Invalid Operation Exception (#I) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-20
11.5.2.2 Denormal-Operand Exception (#D) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-21
11.5.2.3 Divide-By-Zero Exception (#Z) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-22
11.5.2.4 Numeric Overflow Exception (#O) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-22
11.5.2.5 Numeric Underflow Exception (#U) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-22
11.5.2.6 Inexact-Result (Precision) Exception (#P) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-23
11.5.3 Generating SIMD Floating-Point Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-23
11.5.3.1 Handling Masked Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-23
11.5.3.2 Handling Unmasked Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-25
11.5.3.3 Handling Combinations of Masked and Unmasked Exceptions . . . . . . . . . . . . . . . . 11-26
11.5.4 Handling SIMD Floating-Point Exceptions in Software. . . . . . . . . . . . . . . . . . . . . . . . . . . 11-26
11.5.5 Interaction of SIMD and x87 FPU Floating-Point Exceptions. . . . . . . . . . . . . . . . . . . . . 11-26
11.6 WRITING APPLICATIONS WITH SSE/SSE2 EXTENSIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-27
11.6.1 General Guidelines for Using SSE/SSE2 Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-27
11.6.2 Checking for SSE/SSE2 Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-28
11.6.3 Checking for the DAZ Flag in the MXCSR Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-28
11.6.4 Initialization of SSE/SE2 Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-29
11.6.5 Saving and Restoring the SSE/SSE2 State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-30
11.6.6 Guidelines for Writing to the MXCSR Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-30
11.6.7 Interaction of SSE/SSE2 Instructions with x87 FPU and MMX Instructions . . . . . . . 11-31
11.6.8 Compatibility of SIMD and x87 FPU Floating-Point Data Types . . . . . . . . . . . . . . . . . . 11-32
11.6.9 Mixing Packed and Scalar Floating-Point and 128-Bit SIMD Integer Instructions and Data
11-32
11.6.10 Interfacing with SSE/SSE2 Procedures and Functions. . . . . . . . . . . . . . . . . . . . . . . . . . . 11-34
11.6.10.1 Passing Parameters in XMM Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-34
11.6.10.2 Saving XMM Register State on a Procedure or Function Call. . . . . . . . . . . . . . . . . . 11-34
11.6.10.3 Caller-Save Requirement for Procedure and Function Calls. . . . . . . . . . . . . . . . . . . 11-35
11.6.11 Updating Existing MMX Technology Routines
Using 128-Bit SIMD Integer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-35
11.6.12 Branching on Arithmetic Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-36
11.6.13 Cacheability Hint Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-36
11.6.14 Effect of Instruction Prefixes on the SSE/SSE2 Instructions. . . . . . . . . . . . . . . . . . . . . 11-37

Vol. 1 xiii
CONTENTS

PAGE
CHAPTER 12
PROGRAMMING WITH SSE3, SSSE3, AND SSE4
12.1 PROGRAMMING ENVIRONMENT AND DATA TYPES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-1
12.1.1 SSE3, SSSE3, SSE4 in 64-Bit Mode and Compatibility Mode . . . . . . . . . . . . . . . . . . . . . . . 12-1
12.1.2 Compatibility of SSE3/SSSE3 with MMX Technology, the x87 FPU Environment, and
SSE/SSE2 Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-2
12.1.3 Horizontal and Asymmetric Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-2
12.2 OVERVIEW OF SSE3 INSTRUCTIONS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-3
12.3 SSE3 INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-3
12.3.1 x87 FPU Instruction for Integer Conversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-4
12.3.2 SIMD Integer Instruction for Specialized 128-bit Unaligned Data Load. . . . . . . . . . . . . 12-4
12.3.3 SIMD Floating-Point Instructions That Enhance LOAD/MOVE/DUPLICATE Performance . .
12-4
12.3.4 SIMD Floating-Point Instructions Provide Packed Addition/Subtraction . . . . . . . . . . . . 12-5
12.3.5 SIMD Floating-Point Instructions Provide Horizontal Addition/Subtraction . . . . . . . . . 12-5
12.3.6 Two Thread Synchronization Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-7
12.4 WRITING APPLICATIONS WITH SSE3 EXTENSIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-7
12.4.1 Guidelines for Using SSE3 Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-7
12.4.2 Checking for SSE3 Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-7
12.4.3 Enable FTZ and DAZ for SIMD Floating-Point Computation. . . . . . . . . . . . . . . . . . . . . . . . 12-8
12.4.4 Programming SSE3 with SSE/SSE2 Extensions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-8
12.5 OVERVIEW OF SSSE3 INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-8
12.6 SSSE3 INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-9
12.6.1 Horizontal Addition/Subtraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-9
12.6.2 Packed Absolute Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-11
12.6.3 Multiply and Add Packed Signed and Unsigned Bytes. . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-11
12.6.4 Packed Multiply High with Round and Scale. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-11
12.6.5 Packed Shuffle Bytes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-12
12.6.6 Packed Sign . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-12
12.6.7 Packed Align Right . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-12
12.7 WRITING APPLICATIONS WITH SSSE3 EXTENSIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-12
12.7.1 Guidelines for Using SSSE3 Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-12
12.7.2 Checking for SSSE3 Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-13
12.8 SSE3/SSSE3 AND SSE4 EXCEPTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-13
12.8.1 Device Not Available (DNA) Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-13
12.8.2 Numeric Error flag and IGNNE# . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-14
12.8.3 Emulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-14
12.8.4 IEEE 754 Compliance of SSE4.1 Floating-Point Instructions . . . . . . . . . . . . . . . . . . . . . . 12-14
12.9 SSE4 OVERVIEW . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-15
12.10 SSE4.1 INSTRUCTION SET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-16
12.10.1 Dword Multiply Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-16
12.10.2 Floating-Point Dot Product Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-16
12.10.3 Streaming Load Hint Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-17
12.10.4 Packed Blending Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-21
12.10.5 Packed Integer MIN/MAX Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-22
12.10.6 Floating-Point Round Instructions with Selectable Rounding Mode . . . . . . . . . . . . . . . 12-23
12.10.7 Insertion and Extractions from XMM Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-23

xiv Vol. 1
 CONTENTS

PAGE
12.10.8 Packed Integer Format Conversions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-23
12.10.9 Improved Sums of Absolute Differences (SAD) for 4-Byte Blocks . . . . . . . . . . . . . . . . 12-24
12.10.10 Horizontal Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-25
12.10.11 Packed Test. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-25
12.10.12 Packed Qword Equality Comparisons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-26
12.10.13 Dword Packing With Unsigned Saturation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-26
12.11 SSE4.2 INSTRUCTION SET. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-26
12.11.1 String and Text Processing Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-26
12.11.1.1 Memory Operand Alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-27
12.11.2 Packed Comparison SIMD Integer Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-28
12.11.3 Application-Targeted Accelerator Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-28
12.12 WRITING APPLICATIONS WITH SSE4 EXTENSIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-28
12.12.1 Guidelines for Using SSE4 Extensions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-28
12.12.2 Checking for SSE4.1 Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-28
12.12.3 Checking for SSE4.2 Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-29

CHAPTER 13
INPUT/OUTPUT
13.1 I/O PORT ADDRESSING. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-1
13.2 I/O PORT HARDWARE. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-1
13.3 I/O ADDRESS SPACE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-2
13.3.1 Memory-Mapped I/O . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .13-2
13.4 I/O INSTRUCTIONS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-3
13.5 PROTECTED-MODE I/O . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-4
13.5.1 I/O Privilege Level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .13-4
13.5.2 I/O Permission Bit Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .13-5
13.6 ORDERING I/O . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-7

CHAPTER 14
PROCESSOR IDENTIFICATION AND FEATURE DETERMINATION
14.1 USING THE CPUID INSTRUCTION. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-1
14.1.1 Notes on Where to Start . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .14-1
14.1.2 Identification of Earlier IA-32 Processors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .14-2

APPENDIX A
EFLAGS CROSS-REFERENCE
A.1 EFLAGS AND INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A-1

APPENDIX B
EFLAGS CONDITION CODES
B.1 CONDITION CODES. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-1

APPENDIX C
FLOATING-POINT EXCEPTIONS SUMMARY
C.1 OVERVIEW . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C-1
C.2 X87 FPU INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C-2
C.3 SSE INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C-4

Vol. 1 xv
CONTENTS

PAGE
C.4 SSE2 INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C-7
C.5 SSE3 INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C-11
C.6 SSSE3 INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C-12
C.7 SSE4 INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C-12

APPENDIX D
GUIDELINES FOR WRITING X87 FPU
EXCEPTION HANDLERS
D.1 MS-DOS COMPATIBILITY SUB-MODE FOR HANDLING X87 FPU EXCEPTIONS . . . . . . . . . . . D-1
D.2 IMPLEMENTATION OF THE MS-DOS* COMPATIBILITY SUB-MODE IN THE INTEL486™,
PENTIUM®, AND P6 PROCESSOR FAMILY, AND PENTIUM® 4 PROCESSORS . . . . . . . . . . . . . D-3
D.2.1 MS-DOS* Compatibility Sub-mode in the Intel486™ and Pentium® Processors . . . . . . . D-3
D.2.1.1 Basic Rules: When FERR# Is Generated. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . D-4
D.2.1.2 Recommended External Hardware to Support the MS-DOS* Compatibility Sub-mode
D-5
D.2.1.3 No-Wait x87 FPU Instructions Can Get x87 FPU Interrupt in Window . . . . . . . . . . . D-8
D.2.2 MS-DOS* Compatibility Sub-mode in the P6 Family
and Pentium® 4 Processors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . D-10
D.3 RECOMMENDED PROTOCOL FOR MS-DOS* COMPATIBILITY HANDLERS. . . . . . . . . . . . . . D-11
D.3.1 Floating-Point Exceptions and Their Defaults . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . D-12
D.3.2 Two Options for Handling Numeric Exceptions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . D-12
D.3.2.1 Automatic Exception Handling: Using Masked Exceptions. . . . . . . . . . . . . . . . . . . . . . D-12
D.3.2.2 Software Exception Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . D-14
D.3.3 Synchronization Required for Use of x87 FPU Exception Handlers . . . . . . . . . . . . . . . . D-15
D.3.3.1 Exception Synchronization: What, Why, and When . . . . . . . . . . . . . . . . . . . . . . . . . . . . D-16
D.3.3.2 Exception Synchronization Examples. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . D-17
D.3.3.3 Proper Exception Synchronization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . D-18
D.3.4 x87 FPU Exception Handling Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . D-18
D.3.5 Need for Storing State of IGNNE# Circuit If Using x87 FPU and SMM. . . . . . . . . . . . . . D-22
D.3.6 Considerations When x87 FPU Shared Between Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . D-23
D.3.6.1 Speculatively Deferring x87 FPU Saves, General Overview . . . . . . . . . . . . . . . . . . . . D-23
D.3.6.2 Tracking x87 FPU Ownership . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . D-24
D.3.6.3 Interaction of x87 FPU State Saves and Floating-Point Exception Association. . D-25
D.3.6.4 Interrupt Routing From the Kernel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . D-28
D.3.6.5 Special Considerations for Operating Systems that Support Streaming SIMD
Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . D-28
D.4 DIFFERENCES FOR HANDLERS USING NATIVE MODE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . D-29
D.4.1 Origin with the Intel 286 and Intel 287, and Intel386 and Intel 387 Processors . . . D-29
D.4.2 Changes with Intel486, Pentium and Pentium Pro Processors with CR0.NE[bit 5] = 1 .D-
30
D.4.3 Considerations When x87 FPU Shared Between Tasks Using Native Mode . . . . . . . . D-30

APPENDIX E
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS
E.1 TWO OPTIONS FOR HANDLING FLOATING-POINT EXCEPTIONS . . . . . . . . . . . . . . . . . . . . . . . E-1
E.2 SOFTWARE EXCEPTION HANDLING . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-1
E.3 EXCEPTION SYNCHRONIZATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-3

xvi Vol. 1
 CONTENTS

PAGE
E.4 SIMD FLOATING-POINT EXCEPTIONS AND THE IEEE STANDARD 754 . . . . . . . . . . . . . . . . . . E-4
E.4.1 Floating-Point Emulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-4
E.4.2 SSE/SSE2/SSE3 Response To Floating-Point Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . E-6
E.4.2.1 Numeric Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-7
E.4.2.2 Results of Operations with NaN Operands or a NaN Result for SSE/SSE2/SSE3
Numeric Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-7
E.4.2.3 Condition Codes, Exception Flags, and Response for Masked and Unmasked Numeric
Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-12
E.4.3 Example SIMD Floating-Point Emulation Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . E-22

Vol. 1 xvii
CONTENTS

PAGE

FIGURES
Figure 1-1. Bit and Byte Order . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-5
Figure 1-2. Syntax for CPUID, CR, and MSR Data Presentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-8
Figure 2-1. The P6 Processor Microarchitecture with Advanced Transfer Cache Enhancement 2-
9
Figure 2-2. The Intel NetBurst Microarchitecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-12
Figure 2-3. The Intel Core Microarchitecture Pipeline Functionality. . . . . . . . . . . . . . . . . . . . . . . . 2-15
Figure 2-4. SIMD Extensions, Register Layouts, and Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . 2-20
Figure 2-5. Comparison of an IA-32 Processor Supporting Hyper-Threading Technology and a
Traditional Dual Processor System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-21
Figure 2-6. Intel 64 and IA-32 Processors that Support Dual-Core . . . . . . . . . . . . . . . . . . . . . . . . 2-24
Figure 2-7. Intel 64 Processors that Support Quad-Core. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-25
Figure 2-8. Intel Core i7 Processor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-26
Figure 3-1. IA-32 Basic Execution Environment for Non-64-bit Modes. . . . . . . . . . . . . . . . . . . . . . 3-4
Figure 3-2. 64-Bit Mode Execution Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-7
Figure 3-3. Three Memory Management Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-9
Figure 3-4. General System and Application Programming Registers . . . . . . . . . . . . . . . . . . . . . . 3-15
Figure 3-5. Alternate General-Purpose Register Names . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-16
Figure 3-6. Use of Segment Registers for Flat Memory Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-18
Figure 3-7. Use of Segment Registers in Segmented Memory Model . . . . . . . . . . . . . . . . . . . . . . 3-19
Figure 3-8. EFLAGS Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-21
Figure 3-9. Memory Operand Address . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-28
Figure 3-10. Memory Operand Address in 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-29
Figure 3-11. Offset (or Effective Address) Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-31
Figure 4-1. Fundamental Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-1
Figure 4-2. Bytes, Words, Doublewords, Quadwords, and Double Quadwords in Memory . . . . 4-2
Figure 4-3. Numeric Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-4
Figure 4-4. Pointer Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-9
Figure 4-5. Pointers in 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-10
Figure 4-6. Bit Field Data Type. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-10
Figure 4-7. 64-Bit Packed SIMD Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-12
Figure 4-8. 128-Bit Packed SIMD Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-13
Figure 4-9. BCD Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-14
Figure 4-10. Binary Real Number System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-17
Figure 4-11. Binary Floating-Point Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-17
Figure 4-12. Real Numbers and NaNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-19
Figure 6-1. Stack Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-2
Figure 6-2. Stack on Near and Far Calls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-7
Figure 6-3. Protection Rings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-9
Figure 6-4. Stack Switch on a Call to a Different Privilege Level. . . . . . . . . . . . . . . . . . . . . . . . . . . 6-10
Figure 6-5. Stack Usage on Transfers to Interrupt and Exception Handling Routines . . . . . . . 6-16
Figure 6-6. Nested Procedures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-22
Figure 6-7. Stack Frame After Entering the MAIN Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-23
Figure 6-8. Stack Frame After Entering Procedure A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-23
Figure 6-9. Stack Frame After Entering Procedure B . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-24
Figure 6-10. Stack Frame After Entering Procedure C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-25

xviii Vol. 1
 CONTENTS

PAGE
Figure 7-1. Operation of the PUSH Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-8
Figure 7-2. Operation of the PUSHA Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-8
Figure 7-3. Operation of the POP Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-9
Figure 7-4. Operation of the POPA Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-9
Figure 7-5. Sign Extension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-11
Figure 7-7. SHR Instruction Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-16
Figure 7-6. SHL/SAL Instruction Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-16
Figure 7-8. SAR Instruction Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-17
Figure 7-9. SHLD and SHRD Instruction Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-18
Figure 7-10. ROL, ROR, RCL, and RCR Instruction Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-19
Figure 7-11. Flags Affected by the PUSHF, POPF, PUSHFD, and POPFD Instructions . . . . . . . . .7-30
Figure 8-1. x87 FPU Execution Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-3
Figure 8-2. x87 FPU Data Register Stack . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-4
Figure 8-3. Example x87 FPU Dot Product Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-5
Figure 8-4. x87 FPU Status Word. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-6
Figure 8-5. Moving the Condition Codes to the EFLAGS Register . . . . . . . . . . . . . . . . . . . . . . . . . .8-10
Figure 8-6. x87 FPU Control Word . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-11
Figure 8-7. x87 FPU Tag Word . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-13
Figure 8-8. Contents of x87 FPU Opcode Registers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-15
Figure 8-10. Real Mode x87 FPU State Image in Memory, 32-Bit Format . . . . . . . . . . . . . . . . . . . .8-16
Figure 8-9. Protected Mode x87 FPU State Image in Memory, 32-Bit Format . . . . . . . . . . . . . .8-16
Figure 8-12. Real Mode x87 FPU State Image in Memory, 16-Bit Format . . . . . . . . . . . . . . . . . . . .8-17
Figure 8-11. Protected Mode x87 FPU State Image in Memory, 16-Bit Format . . . . . . . . . . . . . .8-17
Figure 8-13. x87 FPU Data Type Formats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-19
Figure 9-1. MMX Technology Execution Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-2
Figure 9-2. MMX Register Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-3
Figure 9-3. Data Types Introduced with the MMX Technology. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-4
Figure 9-4. SIMD Execution Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-5
Figure 10-1. SSE Execution Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10-3
Figure 10-2. XMM Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10-4
Figure 10-3. MXCSR Control/Status Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10-6
Figure 10-4. 128-Bit Packed Single-Precision Floating-Point Data Type . . . . . . . . . . . . . . . . . . . . .10-8
Figure 10-5. Packed Single-Precision Floating-Point Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-10
Figure 10-6. Scalar Single-Precision Floating-Point Operation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-10
Figure 10-7. SHUFPS Instruction, Packed Shuffle Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-14
Figure 10-8. UNPCKHPS Instruction, High Unpack and Interleave Operation . . . . . . . . . . . . . . . 10-15
Figure 10-9. UNPCKLPS Instruction, Low Unpack and Interleave Operation. . . . . . . . . . . . . . . . 10-15
Figure 11-1. Steaming SIMD Extensions 2 Execution Environment . . . . . . . . . . . . . . . . . . . . . . . . . .11-3
Figure 11-2. Data Types Introduced with the SSE2 Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . .11-5
Figure 11-3. Packed Double-Precision Floating-Point Operations. . . . . . . . . . . . . . . . . . . . . . . . . . . .11-6
Figure 11-4. Scalar Double-Precision Floating-Point Operations. . . . . . . . . . . . . . . . . . . . . . . . . . . . .11-7
Figure 11-5. SHUFPD Instruction, Packed Shuffle Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-11
Figure 11-6. UNPCKHPD Instruction, High Unpack and Interleave Operation . . . . . . . . . . . . . . . 11-11
Figure 11-7. UNPCKLPD Instruction, Low Unpack and Interleave Operation . . . . . . . . . . . . . . . 11-12
Figure 11-8. SSE and SSE2 Conversion Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-13
Figure 11-9. Example Masked Response for Packed Operations . . . . . . . . . . . . . . . . . . . . . . . . . . 11-24
Figure 12-1. Asymmetric Processing in ADDSUBPD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .12-2

Vol. 1 xix
CONTENTS

PAGE
Figure 12-2. Horizontal Data Movement in HADDPD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-3
Figure 12-3. Horizontal Data Movement in PHADDD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-10
Figure 12-4. MPSADBW Operation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-25
Figure 13-1. Memory-Mapped I/O. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-3
Figure 13-2. I/O Permission Bit Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-6
Figure D-1. Recommended Circuit for MS-DOS Compatibility x87 FPU
Exception Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . D-7
Figure D-2. Behavior of Signals During x87 FPU Exception Handling . . . . . . . . . . . . . . . . . . . . . . . D-8
Figure D-3. Timing of Receipt of External Interrupt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . D-9
Figure D-4. Arithmetic Example Using Infinity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . D-13
Figure D-5. General Program Flow for DNA Exception Handler . . . . . . . . . . . . . . . . . . . . . . . . . . . . D-26
Figure D-6. Program Flow for a Numeric Exception Dispatch Routine. . . . . . . . . . . . . . . . . . . . . . D-27
Figure E-1. Control Flow for Handling Unmasked Floating-Point Exceptions . . . . . . . . . . . . . . . . .E-6

xx Vol. 1
 CONTENTS

PAGE

TABLES
Table 2-1. Key Features of Most Recent IA-32 Processors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2-28
Table 2-2. Key Features of Most Recent Intel 64 Processors . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2-28
Table 2-3. Key Features of Previous Generations of IA-32 Processors . . . . . . . . . . . . . . . . . . . .2-32
Table 3-1. Instruction Pointer Sizes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3-12
Table 3-2. Addressable General Purpose Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3-17
Table 3-3. Effective Operand- and Address-Size Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3-25
Table 3-4. Effective Operand- and Address-Size Attributes in 64-Bit Mode. . . . . . . . . . . . . . . .3-26
Table 3-5. Default Segment Selection Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3-29
Table 4-1. Signed Integer Encodings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-6
Table 4-2. Length, Precision, and Range of Floating-Point Data Types . . . . . . . . . . . . . . . . . . . . . 4-7
Table 4-3. Floating-Point Number and NaN Encodings. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-8
Table 4-4. Packed Decimal Integer Encodings. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-15
Table 4-5. Real and Floating-Point Number Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-18
Table 4-6. Denormalization Process. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-21
Table 4-7. Rules for Handling NaNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-23
Table 4-8. Rounding Modes and Encoding of Rounding Control (RC) Field . . . . . . . . . . . . . . . . . .4-25
Table 4-10. Masked Responses to Numeric Overflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-30
Table 4-9. Numeric Overflow Thresholds. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-30
Table 4-11. Numeric Underflow (Normalized) Thresholds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-31
Table 5-1. Instruction Groups and IA-32 Processors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-1
Table 6-1. Exceptions and Interrupts. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-14
Table 7-1. Move Instruction Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-4
Table 7-2. Conditional Move Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-6
Table 7-3. Bit Test and Modify Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-20
Table 7-4. Conditional Jump Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-23
Table 8-1. Condition Code Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-8
Table 8-2. Precision Control Field (PC). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-12
Table 8-3. Unsupported Double Extended-Precision Floating-Point Encodings and Pseudo-
Denormals8-21
Table 8-4. Data Transfer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-23
Table 8-5. Floating-Point Conditional Move Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-23
Table 8-6. Setting of x87 FPU Condition Code Flags for Floating-Point Number Comparisons. 8-
27
Table 8-7. Setting of EFLAGS Status Flags for Floating-Point Number Comparisons. . . . . . . .8-28
Table 8-8. TEST Instruction Constants for Conditional Branching . . . . . . . . . . . . . . . . . . . . . . . . .8-29
Table 8-9. Arithmetic and Non-arithmetic Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-35
Table 8-10. Invalid Arithmetic Operations and the
Masked Responses to Them 8-38
Table 8-11. Divide-By-Zero Conditions and the Masked Responses to Them . . . . . . . . . . . . . . . .8-40
Table 9-1. Data Range Limits for Saturation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-6
Table 9-2. MMX Instruction Set Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-7
Table 9-3. Effect of Prefixes on MMX Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-15
Table 10-1. PREFETCHh Instructions Caching Hints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-20
Table 11-1. Masked Responses of SSE/SSE2/SSE3 Instructions to Invalid Arithmetic Operations
11-20

Vol. 1 xxi
CONTENTS

PAGE
Table 11-2. SSE and SSE2 State Following a Power-up/Reset or INIT . . . . . . . . . . . . . . . . . . . . . 11-30
Table 11-3. Effect of Prefixes on SSE, SSE2, and SSE3 Instructions . . . . . . . . . . . . . . . . . . . . . . 11-37
Table 12-1. SIMD numeric exceptions signaled by SSE4.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-15
Table 12-2. Enhanced 32-bit SIMD Multiply Supported by SSE4.1. . . . . . . . . . . . . . . . . . . . . . . . . 12-16
Table 12-3. Blend Field Size and Control Modes Supported by SSE4.1 . . . . . . . . . . . . . . . . . . . . 12-22
Table 12-4. Enhanced SIMD Integer MIN/MAX Instructions Supported by SSE4.1 . . . . . . . . . . 12-22
Table 12-5. New SIMD Integer conversions supported by SSE4.1 . . . . . . . . . . . . . . . . . . . . . . . . . 12-24
Table 12-6. New SIMD Integer Conversions Supported by SSE4.1 . . . . . . . . . . . . . . . . . . . . . . . . 12-24
Table 12-7. Enhanced SIMD Pack support by SSE4.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-26
Table 13-1. I/O Instruction Serialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-8
Table A-1. Codes Describing Flags . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A-1
Table A-2. EFLAGS Cross-Reference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A-1
Table B-1. EFLAGS Condition Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-1
Table C-1. x87 FPU and SIMD Floating-Point Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .C-1
Table C-2. Exceptions Generated with x87 FPU Floating-Point Instructions. . . . . . . . . . . . . . . . .C-2
Table C-3. Exceptions Generated with SSE Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .C-4
Table C-4. Exceptions Generated with SSE2 Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .C-7
Table C-5. Exceptions Generated with SSE3 Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C-11
Table C-6. Exceptions Generated with SSE4 Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C-13
Table E-1. ADDPS, ADDSS, SUBPS, SUBSS, MULPS, MULSS, DIVPS, DIVSS, ADDPD, ADDSD,
SUBPD, SUBSD, MULPD, MULSD, DIVPD, DIVSD, ADDSUBPS, ADDSUBPD, HADDPS,
HADDPD, HSUBPS, HSUBPDE-8
Table E-2. CMPPS.EQ, CMPSS.EQ, CMPPS.ORD, CMPSS.ORD,
CMPPD.EQ, CMPSD.EQ, CMPPD.ORD, CMPSD.ORDE-9
Table E-3. CMPPS.NEQ, CMPSS.NEQ, CMPPS.UNORD, CMPSS.UNORD, CMPPD.NEQ, CMPSD.NEQ,
CMPPD.UNORD, CMPSD.UNORDE-9
Table E-4. CMPPS.LT, CMPSS.LT, CMPPS.LE, CMPSS.LE, CMPPD.LT, CMPSD.LT, CMPPD.LE,
CMPSD.LEE-9
Table E-5. CMPPS.NLT, CMPSS.NLT, CMPPS.NLE, CMPSS.NLE, CMPPD.NLT, CMPSD.NLT,
CMPPD.NLE, CMPSD.NLEE-10
Table E-6. COMISS, COMISD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-10
Table E-7. UCOMISS, UCOMISD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-10
Table E-8. CVTPS2PI, CVTSS2SI, CVTTPS2PI, CVTTSS2SI, CVTPD2PI, CVTSD2SI, CVTTPD2PI,
CVTTSD2SI, CVTPS2DQ, CVTTPS2DQ, CVTPD2DQ, CVTTPD2DQE-11
Table E-9. MAXPS, MAXSS, MINPS, MINSS, MAXPD, MAXSD, MINPD, MINSD . . . . . . . . . . . . . . . E-11
Table E-10. SQRTPS, SQRTSS, SQRTPD, SQRTSD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-11
Table E-11. CVTPS2PD, CVTSS2SD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-12
Table E-12. CVTPD2PS, CVTSD2SS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-12
Table E-13. #I - Invalid Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-13
Table E-14. #Z - Divide-by-Zero . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-16
Table E-15. #D - Denormal Operand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-17
Table E-16. #O - Numeric Overflow. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-18
Table E-17. #U - Numeric Underflow. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-20
Table E-18. #P - Inexact Result (Precision) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-21

xxii Vol. 1
 CHAPTER 1
ABOUT THIS MANUAL

The Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volume 1:

Basic Architecture (order number 253665) is part of a set that describes the architec-
ture and programming environment of Intel® 64 and IA-32 architecture processors.
Other volumes in this set are:
• The Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volumes
2A & 2B: Instruction Set Reference (order numbers 253666 and 253667).
• The Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volumes
3A & 3B: System Programming Guide (order number 253668 and 253669).
The Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volume 1,
describes the basic architecture and programming environment of Intel 64 and IA-32
processors. The Intel® 64 and IA-32 Architectures Software Developer’s Manual,
Volumes 2A & 2B, describe the instruction set of the processor and the opcode struc-
ture. These volumes apply to application programmers and to programmers who
write operating systems or executives. The Intel® 64 and IA-32 Architectures Soft-
ware Developer’s Manual, Volumes 3A & 3B, describe the operating-system support
environment of Intel 64 and IA-32 processors. These volumes target operating-
system and BIOS designers. In addition, the Intel® 64 and IA-32 Architectures Soft-
ware Developer’s Manual, Volume 3B, addresses the programming environment for
classes of software that host operating systems.

1.1 INTEL® 64 AND IA-32 PROCESSORS COVERED IN

THIS MANUAL
This manual set includes information pertaining primarily to the most recent Intel 64
and IA-32 processors, which include:
• Pentium® processors
• P6 family processors
• Pentium® 4 processors
• Pentium® M processors
• Intel® Xeon® processors
• Pentium® D processors
• Pentium® processor Extreme Editions
• 64-bit Intel® Xeon® processors
• Intel® CoreTM Duo processor
• Intel® CoreTM Solo processor

Vol. 1 1-1
ABOUT THIS MANUAL

• Dual-Core Intel® Xeon® processor LV

• Intel® CoreTM2 Duo processor
• Intel® CoreTM2 Quad processor Q6000 series
• Intel® Xeon® processor 3000, 3200 series
• Intel® Xeon® processor 5000 series
• Intel® Xeon® processor 5100, 5300 series
• Intel® CoreTM2 Extreme processor X7000 and X6800 series
• Intel® CoreTM2 Extreme processor QX6000 series
• Intel® Xeon® processor 7100 series
• Intel® Pentium® Dual-Core processor
• Intel® Xeon® processor 7200, 7300 series
• Intel® Xeon® processor 5200, 5400, 7400 series
• Intel® CoreTM2 Extreme processor QX9000 and X9000 series
• Intel® CoreTM2 Quad processor Q9000 series
• Intel® CoreTM2 Duo processor E8000, T9000 series
• Intel® AtomTM processor family
• Intel® CoreTM i7 processor
• Intel® CoreTM i5 processor
P6 family processors are IA-32 processors based on the P6 family microarchitecture.
This includes the Pentium® Pro, Pentium® II, Pentium® III, and Pentium® III Xeon®
processors.
The Pentium® 4, Pentium® D, and Pentium® processor Extreme Editions are based
on the Intel NetBurst® microarchitecture. Most early Intel® Xeon® processors are
based on the Intel NetBurst® microarchitecture. Intel Xeon processor 5000, 7100
series are based on the Intel NetBurst® microarchitecture.
The Intel® CoreTM Duo, Intel® CoreTM Solo and dual-core Intel® Xeon® processor LV
are based on an improved Pentium® M processor microarchitecture.
The Intel® Xeon® processor 3000, 3200, 5100, 5300, 7200 and 7300 series, Intel®
Pentium® dual-core, Intel® CoreTM2 Duo, Intel® CoreTM2 Quad, and Intel® CoreTM2
Extreme processors are based on Intel® CoreTM microarchitecture.
The Intel® Xeon® processor 5200, 5400, 7400 series, Intel® CoreTM2 Quad processor
Q9000 series, and Intel® CoreTM2 Extreme processor QX9000, X9000 series, Intel®
CoreTM2 processor E8000 series are based on Enhanced Intel® CoreTM microarchitec-
ture.
The Intel® AtomTM processor family is based on the Intel® AtomTM microarchitecture
and supports Intel 64 architecture.
The Intel® CoreTM i7 processor and the Intel® CoreTM i5 processor are based on the
Intel® microarchitecture (Nehalem) and support Intel 64 architecture.

1-2 Vol. 1
 ABOUT THIS MANUAL

Processors based on the Next Generation Intel Processor, codenamed Westmere,

support Intel 64 architecture.
P6 family, Pentium® M, Intel® CoreTM Solo, Intel® CoreTM Duo processors, dual-core
Intel® Xeon® processor LV, and early generations of Pentium 4 and Intel Xeon
processors support IA-32 architecture. The Intel® AtomTM processor Z5xx series
support IA-32 architecture.
The Intel® Xeon® processor 3000, 3200, 5000, 5100, 5200, 5300, 5400, 7100,
7200, 7300, 7400 series, Intel® CoreTM2 Duo, Intel® CoreTM2 Extreme processors,
Intel Core 2 Quad processors, Pentium® D processors, Pentium® Dual-Core
processor, newer generations of Pentium 4 and Intel Xeon processor family support
Intel® 64 architecture.
IA-32 architecture is the instruction set architecture and programming environment
for Intel's 32-bit microprocessors.
Intel® 64 architecture is the instruction set architecture and programming environ-
ment which is the superset of Intel’s 32-bit and 64-bit architectures. It is compatible
with the IA-32 architecture.

1.2 OVERVIEW OF VOLUME 1: BASIC ARCHITECTURE

A description of this manual’s content follows:
Chapter 1 — About This Manual. Gives an overview of all five volumes of the
Intel® 64 and IA-32 Architectures Software Developer’s Manual. It also describes
the notational conventions in these manuals and lists related Intel manuals and
documentation of interest to programmers and hardware designers.
Chapter 2 — Intel® 64 and IA-32 Architectures. Introduces the Intel 64 and
IA-32 architectures along with the families of Intel processors that are based on
these architectures. It also gives an overview of the common features found in these
processors and brief history of the Intel 64 and IA-32 architectures.
Chapter 3 — Basic Execution Environment. Introduces the models of memory
organization and describes the register set used by applications.
Chapter 4 — Data Types. Describes the data types and addressing modes recog-
nized by the processor; provides an overview of real numbers and floating-point
formats and of floating-point exceptions.
Chapter 5 — Instruction Set Summary. Lists all Intel 64 and IA-32 instructions,
divided into technology groups.
Chapter 6 — Procedure Calls, Interrupts, and Exceptions. Describes the proce-
dure stack and mechanisms provided for making procedure calls and for servicing
interrupts and exceptions.
Chapter 7 — Programming with General-Purpose Instructions. Describes
basic load and store, program control, arithmetic, and string instructions that

Vol. 1 1-3
ABOUT THIS MANUAL

operate on basic data types, general-purpose and segment registers; also describes
system instructions that are executed in protected mode.
Chapter 8 — Programming with the x87 FPU. Describes the x87 floating-point
unit (FPU), including floating-point registers and data types; gives an overview of the
floating-point instruction set and describes the processor's floating-point exception
conditions.
Chapter 9 — Programming with Intel® MMX™ Technology. Describes Intel
MMX technology, including MMX registers and data types; also provides an overview
of the MMX instruction set.
Chapter 10 — Programming with Streaming SIMD Extensions (SSE).
Describes SSE extensions, including XMM registers, the MXCSR register, and packed
single-precision floating-point data types; provides an overview of the SSE instruc-
tion set and gives guidelines for writing code that accesses the SSE extensions.
Chapter 11 — Programming with Streaming SIMD Extensions 2 (SSE2).
Describes SSE2 extensions, including XMM registers and packed double-precision
floating-point data types; provides an overview of the SSE2 instruction set and gives
guidelines for writing code that accesses SSE2 extensions. This chapter also
describes SIMD floating-point exceptions that can be generated with SSE and SSE2
instructions. It also provides general guidelines for incorporating support for SSE and
SSE2 extensions into operating system and applications code.
Chapter 12 — Programming with SSE3, SSSE3 and SSE4. Provides an overview
of the SSE3 instruction set, Supplemental SSE3, SSE4, and guidelines for writing
code that accesses these extensions.
Chapter 13 — Input/Output. Describes the processor’s I/O mechanism, including
I/O port addressing, I/O instructions, and I/O protection mechanisms.
Chapter 14 — Processor Identification and Feature Determination. Describes
how to determine the CPU type and features available in the processor.
Appendix A — EFLAGS Cross-Reference. Summarizes how the IA-32 instructions
affect the flags in the EFLAGS register.
Appendix B — EFLAGS Condition Codes. Summarizes how conditional jump,
move, and ‘byte set on condition code’ instructions use condition code flags (OF, CF,
ZF, SF, and PF) in the EFLAGS register.
Appendix C — Floating-Point Exceptions Summary. Summarizes exceptions
raised by the x87 FPU floating-point and SSE/SSE2/SSE3 floating-point instructions.
Appendix D — Guidelines for Writing x87 FPU Exception Handlers. Describes
how to design and write MS-DOS* compatible exception handling facilities for FPU
exceptions (includes software and hardware requirements and assembly-language
code examples). This appendix also describes general techniques for writing robust
FPU exception handlers.
Appendix E — Guidelines for Writing SIMD Floating-Point Exception
Handlers. Gives guidelines for writing exception handlers for exceptions generated
by SSE/SSE2/SSE3 floating-point instructions.

1-4 Vol. 1
 ABOUT THIS MANUAL

1.3 NOTATIONAL CONVENTIONS

This manual uses specific notation for data-structure formats, for symbolic represen-
tation of instructions, and for hexadecimal and binary numbers. This notation is
described below.

1.3.1 Bit and Byte Order

In illustrations of data structures in memory, smaller addresses appear toward the
bottom of the figure; addresses increase toward the top. Bit positions are numbered
from right to left. The numerical value of a set bit is equal to two raised to the power
of the bit position. Intel 64 and IA-32 processors are “little endian” machines; this
means the bytes of a word are numbered starting from the least significant byte. See
Figure 1-1.

Data Structure
Highest
Address 32 24 23 16 15 8 7 0 Bit offset
28
24
20
16
12
8
4
Byte 3 Byte 2 Byte 1 Byte 0 0
Lowest
Address

Byte Offset

Figure 1-1. Bit and Byte Order

1.3.2 Reserved Bits and Software Compatibility

In many register and memory layout descriptions, certain bits are marked as
reserved. When bits are marked as reserved, it is essential for compatibility with
future processors that software treat these bits as having a future, though unknown,
effect. The behavior of reserved bits should be regarded as not only undefined, but
unpredictable.
Software should follow these guidelines in dealing with reserved bits:
• Do not depend on the states of any reserved bits when testing the values of
registers that contain such bits. Mask out the reserved bits before testing.
• Do not depend on the states of any reserved bits when storing to memory or to a
register.

Vol. 1 1-5
ABOUT THIS MANUAL

• Do not depend on the ability to retain information written into any reserved bits.
• When loading a register, always load the reserved bits with the values indicated
in the documentation, if any, or reload them with values previously read from the
same register.

NOTE
Avoid any software dependence upon the state of reserved bits in
Intel 64 and IA-32 registers. Depending upon the values of reserved
register bits will make software dependent upon the unspecified
manner in which the processor handles these bits. Programs that
depend upon reserved values risk incompatibility with future
processors.

1.3.2.1 Instruction Operands

When instructions are represented symbolically, a subset of the IA-32 assembly
language is used. In this subset, an instruction has the following format:

label: mnemonic argument1, argument2, argument3

where:
• A label is an identifier which is followed by a colon.
• A mnemonic is a reserved name for a class of instruction opcodes which have
the same function.
• The operands argument1, argument2, and argument3 are optional. There
may be from zero to three operands, depending on the opcode. When present,
they take the form of either literals or identifiers for data items. Operand
identifiers are either reserved names of registers or are assumed to be assigned
to data items declared in another part of the program (which may not be shown
in the example).
When two operands are present in an arithmetic or logical instruction, the right
operand is the source and the left operand is the destination.
For example:

LOADREG: MOV EAX, SUBTOTAL

In this example, LOADREG is a label, MOV is the mnemonic identifier of an opcode,
EAX is the destination operand, and SUBTOTAL is the source operand. Some
assembly languages put the source and destination in reverse order.

1-6 Vol. 1
 ABOUT THIS MANUAL

1.3.3 Hexadecimal and Binary Numbers

Base 16 (hexadecimal) numbers are represented by a string of hexadecimal digits
followed by the character H (for example, 0F82EH). A hexadecimal digit is a char-
acter from the following set: 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, A, B, C, D, E, and F.
Base 2 (binary) numbers are represented by a string of 1s and 0s, sometimes
followed by the character B (for example, 1010B). The “B” designation is only used in
situations where confusion as to the type of number might arise.

1.3.4 Segmented Addressing

The processor uses byte addressing. This means memory is organized and accessed
as a sequence of bytes. Whether one or more bytes are being accessed, a byte
address is used to locate the byte or bytes memory. The range of memory that can
be addressed is called an address space.
The processor also supports segmented addressing. This is a form of addressing
where a program may have many independent address spaces, called segments.
For example, a program can keep its code (instructions) and stack in separate
segments. Code addresses would always refer to the code space, and stack
addresses would always refer to the stack space. The following notation is used to
specify a byte address within a segment:

Segment-register:Byte-address
For example, the following segment address identifies the byte at address FF79H in
the segment pointed by the DS register:

DS:FF79H
The following segment address identifies an instruction address in the code segment.
The CS register points to the code segment and the EIP register contains the address
of the instruction.

CS:EIP

1.3.5 A New Syntax for CPUID, CR, and MSR Values

Obtain feature flags, status, and system information by using the CPUID instruction,
by checking control register bits, and by reading model-specific registers. We are
moving toward a new syntax to represent this information. See Figure 1-2.

Vol. 1 1-7
ABOUT THIS MANUAL

&38,',QSXWDQG2XWSXW
&38,'+(&;66(>ELW@ 

,QSXWYDOXHVIRU($; (&;UHJLVWHUV
,IRQO\RQHYDOXH($;LVLPSOLHG

2XWSXWUHJLVWHUDQGIHDWXUHIODJRUILHOG
QDPHZLWKELWSRVLWLRQV
9DOXHRUUDQJHRIRXWSXW

&RQWURO5HJLVWHU9DOXHV
&526);65>ELW@ 

([DPSOH&5QDPH
)HDWXUHIODJRUILHOGQDPH
ZLWKELWSRVLWLRQV
9DOXHRUUDQJHRIRXWSXW

0RGHO6SHFLILF5HJLVWHU9DOXHV

,$B0,6&B(1$%/(6(1$%/()23&2'(>ELW@ 

([DPSOH065QDPH
)HDWXUHIODJRUILHOGQDPHZLWKELWSRVLWLRQV
9DOXHRUUDQJHRIRXWSXW

20

Figure 1-2. Syntax for CPUID, CR, and MSR Data Presentation

1.3.6 Exceptions
An exception is an event that typically occurs when an instruction causes an error.
For example, an attempt to divide by zero generates an exception. However, some
exceptions, such as breakpoints, occur under other conditions. Some types of excep-
tions may provide error codes. An error code reports additional information about the
error. An example of the notation used to show an exception and error code is shown
below:

#PF(fault code)

1-8 Vol. 1
 ABOUT THIS MANUAL

This example refers to a page-fault exception under conditions where an error code
naming a type of fault is reported. Under some conditions, exceptions that produce
error codes may not be able to report an accurate code. In this case, the error code
is zero, as shown below for a general-protection exception:

#GP(0)

1.4 RELATED LITERATURE

Literature related to Intel 64 and IA-32 processors is listed on-line at:
http://developer.intel.com/products/processor/manuals/index.htm
Some of the documents listed at this web site can be viewed on-line; others can be
ordered. The literature available is listed by Intel processor and then by the following
literature types: applications notes, data sheets, manuals, papers, and specification
updates.
See also:
• The data sheet for a particular Intel 64 or IA-32 processor
• The specification update for a particular Intel 64 or IA-32 processor
• Intel® C++ Compiler documentation and online help
http://www.intel.com/cd/software/products/asmo-na/eng/index.htm
• Intel® Fortran Compiler documentation and online help
http://www.intel.com/cd/software/products/asmo-na/end/index.htm
• Intel® VTune™ Performance Analyzer documentation and online help
http://www.intel.com/cd/software/products/asmo-na/eng/index.htm
• Intel® 64 and IA-32 Architectures Software Developer’s Manual (in five volumes)
http://developer.intel.com/products/processor/manuals/index.htm
• Intel® 64 and IA-32 Architectures Optimization Reference Manual
http://developer.intel.com/products/processor/manuals/index.htm
• Intel® Processor Identification with the CPUID Instruction, AP-485
http://www.intel.com/support/processors/sb/cs-009861.htm
• TLBs, Paging-Structure Caches, and Their Invalidation,
http://developer.intel.com/products/processor/manuals/index.htm
• Intel 64 Architecture x2APIC Specification:
http://developer.intel.com/products/processor/manuals/index.htm
• Intel 64 Architecture Processor Topology Enumeration:
http://softwarecommunity.intel.com/articles/eng/3887.htm
• Intel® Trusted Execution Technology Measured Launched Environment
Programming Guide, http://www.intel.com/technology/security/index.htm

Vol. 1 1-9
ABOUT THIS MANUAL

• Intel® SSE4 Programming Reference,

http://developer.intel.com/products/processor/manuals/index.htm
• Developing Multi-threaded Applications: A Platform Consistent Approach
http://cache-
www.intel.com/cd/00/00/05/15/51534_developing_multithreaded_applications.
pdf
• Using Spin-Loops on Intel Pentium 4 Processor and Intel Xeon Processor MP
http://www3.intel.com/cd/ids/developer/asmo-
na/eng/dc/threading/knowledgebase/19083.htm
More relevant links are:
• Software network link:
http://softwarecommunity.intel.com/isn/home/
• Developer centers:
http://www.intel.com/cd/ids/developer/asmo-na/eng/dc/index.htm
• Processor support general link:
http://www.intel.com/support/processors/
• Software products and packages:
http://www.intel.com/cd/software/products/asmo-na/eng/index.htm
• Intel 64 and IA-32 processor manuals (printed or PDF downloads):
http://developer.intel.com/products/processor/manuals/index.htm
• Intel® Multi-Core Technology:
http://developer.intel.com/multi-core/index.htm
• Intel® Hyper-Threading Technology (Intel® HT Technology):
http://developer.intel.com/technology/hyperthread/

1-10 Vol. 1
 CHAPTER 2
®
INTEL 64 AND IA-32 ARCHITECTURES

The exponential growth of computing power and ownership has made the computer
one of the most important forces shaping business and society. Intel 64 and IA-32
architectures have been at the forefront of the computer revolution and is today the
preferred computer architecture, as measured by computers in use and the total
computing power available in the world.

2.1 BRIEF HISTORY OF INTEL® 64 AND IA-32

ARCHITECTURE
The following sections provide a summary of the major technical evolutions from
IA-32 to Intel 64 architecture: starting from the Intel 8086 processor to the latest
Intel® Core® 2 Duo, Core 2 Quad and Intel Xeon processor 5300 and 7300 series.
Object code created for processors released as early as 1978 still executes on the
latest processors in the Intel 64 and IA-32 architecture families.

2.1.1 16-bit Processors and Segmentation (1978)

The IA-32 architecture family was preceded by 16-bit processors, the 8086 and
8088. The 8086 has 16-bit registers and a 16-bit external data bus, with 20-bit
addressing giving a 1-MByte address space. The 8088 is similar to the 8086 except it
has an 8-bit external data bus.
The 8086/8088 introduced segmentation to the IA-32 architecture. With segmenta-
tion, a 16-bit segment register contains a pointer to a memory segment of up to
64 KBytes. Using four segment registers at a time, 8086/8088 processors are able to
address up to 256 KBytes without switching between segments. The 20-bit
addresses that can be formed using a segment register and an additional 16-bit
pointer provide a total address range of 1 MByte.

2.1.2 The Intel® 286 Processor (1982)

The Intel 286 processor introduced protected mode operation into the IA-32 archi-
tecture. Protected mode uses the segment register content as selectors or pointers
into descriptor tables. Descriptors provide 24-bit base addresses with a physical
memory size of up to 16 MBytes, support for virtual memory management on a
segment swapping basis, and a number of protection mechanisms. These mecha-
nisms include:
• Segment limit checking

Vol. 1 2-1
INTEL® 64 AND IA-32 ARCHITECTURES

• Read-only and execute-only segment options

• Four privilege levels

2.1.3 The Intel386™ Processor (1985)

The Intel386 processor was the first 32-bit processor in the IA-32 architecture family.
It introduced 32-bit registers for use both to hold operands and for addressing. The
lower half of each 32-bit Intel386 register retains the properties of the 16-bit regis-
ters of earlier generations, permitting backward compatibility. The processor also
provides a virtual-8086 mode that allows for even greater efficiency when executing
programs created for 8086/8088 processors.
In addition, the Intel386 processor has support for:
• A 32-bit address bus that supports up to 4-GBytes of physical memory
• A segmented-memory model and a flat memory model
• Paging, with a fixed 4-KByte page size providing a method for virtual memory
management
• Support for parallel stages

2.1.4 The Intel486™ Processor (1989)

The Intel486™ processor added more parallel execution capability by expanding the
Intel386 processor’s instruction decode and execution units into five pipelined
stages. Each stage operates in parallel with the others on up to five instructions in
different stages of execution.
In addition, the processor added:
• An 8-KByte on-chip first-level cache that increased the percent of instructions
that could execute at the scalar rate of one per clock
• An integrated x87 FPU
• Power saving and system management capabilities

2.1.5 The Intel® Pentium® Processor (1993)

The introduction of the Intel Pentium processor added a second execution pipeline to
achieve superscalar performance (two pipelines, known as u and v, together can
execute two instructions per clock). The on-chip first-level cache doubled, with 8
KBytes devoted to code and another 8 KBytes devoted to data. The data cache uses
the MESI protocol to support more efficient write-back cache in addition to the write-
through cache previously used by the Intel486 processor. Branch prediction with an
on-chip branch table was added to increase performance in looping constructs.
In addition, the processor added:

2-2 Vol. 1
 INTEL® 64 AND IA-32 ARCHITECTURES

• Extensions to make the virtual-8086 mode more efficient and allow for 4-MByte
as well as 4-KByte pages
• Internal data paths of 128 and 256 bits add speed to internal data transfers
• Burstable external data bus was increased to 64 bits
• An APIC to support systems with multiple processors
• A dual processor mode to support glueless two processor systems
A subsequent stepping of the Pentium family introduced Intel MMX technology (the
Pentium Processor with MMX technology). Intel MMX technology uses the single-
instruction, multiple-data (SIMD) execution model to perform parallel computations
on packed integer data contained in 64-bit registers.
See Section 2.2.6, “SIMD Instructions.”

2.1.6 The P6 Family of Processors (1995-1999)

The P6 family of processors was based on a superscalar microarchitecture that set
new performance standards; see also Section 2.2.1, “P6 Family Microarchitecture.”
One of the goals in the design of the P6 family microarchitecture was to exceed the
performance of the Pentium processor significantly while using the same 0.6-
micrometer, four-layer, metal BICMOS manufacturing process. Members of this
family include the following:
• The Intel Pentium Pro processor is three-way superscalar. Using parallel
processing techniques, the processor is able on average to decode, dispatch, and
complete execution of (retire) three instructions per clock cycle. The Pentium Pro
introduced the dynamic execution (micro-data flow analysis, out-of-order
execution, superior branch prediction, and speculative execution) in a
superscalar implementation. The processor was further enhanced by its caches.
It has the same two on-chip 8-KByte 1st-Level caches as the Pentium processor
and an additional 256-KByte Level 2 cache in the same package as the processor.
• The Intel Pentium II processor added Intel MMX technology to the P6 family
processors along with new packaging and several hardware enhancements. The
processor core is packaged in the single edge contact cartridge (SECC). The Level
l data and instruction caches were enlarged to 16 KBytes each, and Level 2 cache
sizes of 256 KBytes, 512 KBytes, and 1 MByte are supported. A half-clock speed
backside bus connects the Level 2 cache to the processor. Multiple low-power
states such as AutoHALT, Stop-Grant, Sleep, and Deep Sleep are supported to
conserve power when idling.
• The Pentium II Xeon processor combined the premium characteristics of
previous generations of Intel processors. This includes: 4-way, 8-way (and up)
scalability and a 2 MByte 2nd-Level cache running on a full-clock speed backside
bus.
• The Intel Celeron processor family focused on the value PC market segment.
Its introduction offers an integrated 128 KBytes of Level 2 cache and a plastic pin
grid array (P.P.G.A.) form factor to lower system design cost.

Vol. 1 2-3
INTEL® 64 AND IA-32 ARCHITECTURES

• The Intel Pentium III processor introduced the Streaming SIMD Extensions
(SSE) to the IA-32 architecture. SSE extensions expand the SIMD execution
model introduced with the Intel MMX technology by providing a new set of 128-
bit registers and the ability to perform SIMD operations on packed single-
precision floating-point values. See Section 2.2.6, “SIMD Instructions.”
• The Pentium III Xeon processor extended the performance levels of the IA-32
processors with the enhancement of a full-speed, on-die, and Advanced Transfer
Cache.

2.1.7 The Intel® Pentium® 4 Processor Family (2000-2006)

The Intel Pentium 4 processor family is based on Intel NetBurst microarchitecture;
see Section 2.2.2, “Intel NetBurst® Microarchitecture.”
The Intel Pentium 4 processor introduced Streaming SIMD Extensions 2 (SSE2); see
Section 2.2.6, “SIMD Instructions.” The Intel Pentium 4 processor 3.40 GHz,
supporting Hyper-Threading Technology introduced Streaming SIMD Extensions 3
(SSE3); see Section 2.2.6, “SIMD Instructions.”
Intel 64 architecture was introduced in the Intel Pentium 4 Processor Extreme Edition
supporting Hyper-Threading Technology and in the Intel Pentium 4 Processor 6xx and
5xx sequences.
Intel® Virtualization Technology (Intel® VT) was introduced in the Intel Pentium 4
processor 672 and 662.

2.1.8 The Intel® Xeon® Processor (2001- 2007)

Intel Xeon processors (with exception for dual-core Intel Xeon processor LV, Intel
Xeon processor 5100 series) are based on the Intel NetBurst microarchitecture; see
Section 2.2.2, “Intel NetBurst® Microarchitecture.” As a family, this group of IA-32
processors (more recently Intel 64 processors) is designed for use in multi-processor
server systems and high-performance workstations.
The Intel Xeon processor MP introduced support for Intel® Hyper-Threading Tech-
nology; see Section 2.2.7, “Intel® Hyper-Threading Technology.”
The 64-bit Intel Xeon processor 3.60 GHz (with an 800 MHz System Bus) was used to
introduce Intel 64 architecture. The Dual-Core Intel Xeon processor includes dual
core technology. The Intel Xeon processor 70xx series includes Intel Virtualization
Technology.
The Intel Xeon processor 5100 series introduces power-efficient, high performance
Intel Core microarchitecture. This processor is based on Intel 64 architecture; it
includes Intel Virtualization Technology and dual-core technology. The Intel Xeon
processor 3000 series are also based on Intel Core microarchitecture. The Intel Xeon
processor 5300 series introduces four processor cores in a physical package, they are
also based on Intel Core microarchitecture.

2-4 Vol. 1
 INTEL® 64 AND IA-32 ARCHITECTURES

2.1.9 The Intel® Pentium® M Processor (2003-Current)

The Intel Pentium M processor family is a high performance, low power mobile
processor family with microarchitectural enhancements over previous generations of
IA-32 Intel mobile processors. This family is designed for extending battery life and
seamless integration with platform innovations that enable new usage models (such
as extended mobility, ultra thin form-factors, and integrated wireless networking).
Its enhanced microarchitecture includes:
• Support for Intel Architecture with Dynamic Execution
• A high performance, low-power core manufactured using Intel’s advanced
process technology with copper interconnect
• On-die, primary 32-KByte instruction cache and 32-KByte write-back data cache
• On-die, second-level cache (up to 2 MByte) with Advanced Transfer Cache Archi-
tecture
• Advanced Branch Prediction and Data Prefetch Logic
• Support for MMX technology, Streaming SIMD instructions, and the SSE2
instruction set
• A 400 or 533 MHz, Source-Synchronous Processor System Bus
• Advanced power management using Enhanced Intel SpeedStep® technology

2.1.10 The Intel® Pentium® Processor Extreme Edition (2005-2007)

The Intel Pentium processor Extreme Edition introduced dual-core technology. This
technology provides advanced hardware multi-threading support. The processor is
based on Intel NetBurst microarchitecture and supports SSE, SSE2, SSE3, Hyper-
Threading Technology, and Intel 64 architecture.
See also:
• Section 2.2.2, “Intel NetBurst® Microarchitecture”
• Section 2.2.3, “Intel® Core™ Microarchitecture”
• Section 2.2.6, “SIMD Instructions”
• Section 2.2.7, “Intel® Hyper-Threading Technology”
• Section 2.2.8, “Multi-Core Technology”
• Section 2.2.9, “Intel® 64 Architecture”

2.1.11 The Intel® Core™ Duo and Intel® Core™ Solo Processors
(2006-2007)
The Intel Core Duo processor offers power-efficient, dual-core performance with a
low-power design that extends battery life. This family and the single-core Intel Core

Vol. 1 2-5
INTEL® 64 AND IA-32 ARCHITECTURES

Solo processor offer microarchitectural enhancements over Pentium M processor

family.
Its enhanced microarchitecture includes:
• Intel® Smart Cache which allows for efficient data sharing between two
processor cores
• Improved decoding and SIMD execution
• Intel® Dynamic Power Coordination and Enhanced Intel® Deeper Sleep to reduce
power consumption
• Intel® Advanced Thermal Manager which features digital thermal sensor
interfaces
• Support for power-optimized 667 MHz bus
The dual-core Intel Xeon processor LV is based on the same microarchitecture as
Intel Core Duo processor, and supports IA-32 architecture.

2.1.12 The Intel® Xeon® Processor 5100, 5300 Series and

Intel® Core™2 Processor Family (2006-Current)
The Intel Xeon processor 3000, 3200, 5100, 5300, and 7300 series, Intel Pentium
Dual-Core, Intel Core 2 Extreme, Intel Core 2 Quad processors, and Intel Core 2 Duo
processor family support Intel 64 architecture; they are based on the high-perfor-
mance, power-efficient Intel® Core microarchitecture built on 65 nm process tech-
nology. The Intel Core microarchitecture includes the following innovative features:
• Intel® Wide Dynamic Execution to increase performance and execution
throughput
• Intel® Intelligent Power Capability to reduce power consumption
• Intel® Advanced Smart Cache which allows for efficient data sharing between
two processor cores
• Intel® Smart Memory Access to increase data bandwidth and hide latency of
memory accesses
• Intel® Advanced Digital Media Boost which improves application performance
using multiple generations of Streaming SIMD extensions
The Intel Xeon processor 5300 series, Intel Core 2 Extreme processor QX6800 series,
and Intel Core 2 Quad processors support Intel quad-core technology.

2.1.13 The Intel® Xeon® Processor 5200, 5400, 7400 Series and
Intel® Core™2 Processor Family (2007-Current)
The Intel Xeon processor 5200, 5400, and 7400 series, Intel Core 2 Quad processor
Q9000 Series, Intel Core 2 Duo processor E8000 series support Intel 64 architecture;
they are based on the Enhanced Intel® Core microarchitecture using 45 nm process

2-6 Vol. 1
 INTEL® 64 AND IA-32 ARCHITECTURES

technology. The Enhanced Intel Core microarchitecture provides the following

improved features:
• A radix-16 divider, faster OS primitives further increases the performance of
Intel® Wide Dynamic Execution.
• Improves Intel® Advanced Smart Cache with Up to 50% larger level-two cache
and up to 50% increase in way-set associativity.
• A 128-bit shuffler engine significantly improves the performance of Intel®
Advanced Digital Media Boost and SSE4.
Intel Xeon processor 5400 series and Intel Core 2 Quad processor Q9000 Series
support Intel quad-core technology. Intel Xeon processor 7400 series offers up to six
processor cores and an L3 cache up to 16 MBytes.

2.1.14 The Intel® Atom™ Processor Family (2008-Current)

The Intel® AtomTM processors are built on 45 nm process technology. They are based
on a new microarchitecture, Intel® AtomTM microarchitecture, which is optimized for
ultra low power devices. The Intel® AtomTM microarchitecture features two in-order
execution pipelines that minimize power consumption, increase battery life, and
enable ultra-small form factors. It provides the following features:
• Enhanced Intel® SpeedStep® Technology
• Intel® Hyper-Threading Technology
• Deep Power Down Technology with Dynamic Cache Sizing
• Support for new instructions up to and including Supplemental Streaming SIMD
Extensions 3 (SSSE3).
• Support for Intel® Virtualization Technology
• Support for Intel® 64 Architecture (excluding Intel Atom processor Z5xx Series)

2.1.15 The Intel® Core™i7 Processor Family (2008-Current)

The Intel Core i7 processor support Intel 64 architecture; they are based on Intel
microarchitecture (Nehalem) using 45 nm process technology. The Intel Core i7
processor includes the following innovative features:
• Intel® Turbo Boost Technology converts thermal headroom into higher perfor-
mance.
• Intel® HyperThreading Technology in conjunction with Quadcore to provide four
cores and eight threads.
• Dedicated power control unit to reduce active and idle power consumption.
• Integrated memory controller on the processor supporting three channel of DDR3
memory.
• 8 MB inclusive Intel® Smart Cache.

Vol. 1 2-7
INTEL® 64 AND IA-32 ARCHITECTURES

• Intel® QuickPath interconnect (QPI) providing point-to-point link to chipset.

• Support for SSE4.2 and SSE4.1 instruction sets.
• Second generation Intel Virtualization Technology.

2.2 MORE ON SPECIFIC ADVANCES

The following sections provide more information on major innovations.

2.2.1 P6 Family Microarchitecture

The Pentium Pro processor introduced a new microarchitecture commonly referred to
as P6 processor microarchitecture. The P6 processor microarchitecture was later
enhanced with an on-die, Level 2 cache, called Advanced Transfer Cache.
The microarchitecture is a three-way superscalar, pipelined architecture. Three-way
superscalar means that by using parallel processing techniques, the processor is able
on average to decode, dispatch, and complete execution of (retire) three instructions
per clock cycle. To handle this level of instruction throughput, the P6 processor family
uses a decoupled, 12-stage superpipeline that supports out-of-order instruction
execution.
Figure 2-1 shows a conceptual view of the P6 processor microarchitecture pipeline
with the Advanced Transfer Cache enhancement.

2-8 Vol. 1
 INTEL® 64 AND IA-32 ARCHITECTURES

System Bus

Frequently used
Bus Unit Less frequently used

2nd Level Cache 1st Level Cache

On-die, 8-way 4-way, low latency

Front End

Execution
Instruction Execution
Fetch/
Cache Out-of-Order Retirement
Decode
Microcode Core
ROM

Branch History Update

BTSs/Branch Prediction

OM16520

Figure 2-1. The P6 Processor Microarchitecture with Advanced Transfer Cache

Enhancement

To ensure a steady supply of instructions and data for the instruction execution pipe-
line, the P6 processor microarchitecture incorporates two cache levels. The Level 1
cache provides an 8-KByte instruction cache and an 8-KByte data cache, both closely
coupled to the pipeline. The Level 2 cache provides 256-KByte, 512-KByte, or
1-MByte static RAM that is coupled to the core processor through a full clock-speed
64-bit cache bus.
The centerpiece of the P6 processor microarchitecture is an out-of-order execution
mechanism called dynamic execution. Dynamic execution incorporates three data-
processing concepts:
• Deep branch prediction allows the processor to decode instructions beyond
branches to keep the instruction pipeline full. The P6 processor family
implements highly optimized branch prediction algorithms to predict the direction
of the instruction.
• Dynamic data flow analysis requires real-time analysis of the flow of data
through the processor to determine dependencies and to detect opportunities for
out-of-order instruction execution. The out-of-order execution core can monitor

Vol. 1 2-9
INTEL® 64 AND IA-32 ARCHITECTURES

many instructions and execute these instructions in the order that best optimizes
the use of the processor’s multiple execution units, while maintaining the data
integrity.
• Speculative execution refers to the processor’s ability to execute instructions
that lie beyond a conditional branch that has not yet been resolved, and
ultimately to commit the results in the order of the original instruction stream. To
make speculative execution possible, the P6 processor microarchitecture
decouples the dispatch and execution of instructions from the commitment of
results. The processor’s out-of-order execution core uses data-flow analysis to
execute all available instructions in the instruction pool and store the results in
temporary registers. The retirement unit then linearly searches the instruction
pool for completed instructions that no longer have data dependencies with other
instructions or unresolved branch predictions. When completed instructions are
found, the retirement unit commits the results of these instructions to memory
and/or the IA-32 registers (the processor’s eight general-purpose registers and
eight x87 FPU data registers) in the order they were originally issued and retires
the instructions from the instruction pool.

2.2.2 Intel NetBurst® Microarchitecture

The Intel NetBurst microarchitecture provides:
• The Rapid Execution Engine
— Arithmetic Logic Units (ALUs) run at twice the processor frequency
— Basic integer operations can dispatch in 1/2 processor clock tick
• Hyper-Pipelined Technology
— Deep pipeline to enable industry-leading clock rates for desktop PCs and
servers
— Frequency headroom and scalability to continue leadership into the future
• Advanced Dynamic Execution
— Deep, out-of-order, speculative execution engine
• Up to 126 instructions in flight
• Up to 48 loads and 24 stores in pipeline1
— Enhanced branch prediction capability
• Reduces the misprediction penalty associated with deeper pipelines
• Advanced branch prediction algorithm
• 4K-entry branch target array

1. Intel 64 and IA-32 processors based on the Intel NetBurst microarchitecture at 90 nm process
can handle more than 24 stores in flight.

2-10 Vol. 1
 INTEL® 64 AND IA-32 ARCHITECTURES

• New cache subsystem

— First level caches
• Advanced Execution Trace Cache stores decoded instructions
• Execution Trace Cache removes decoder latency from main execution
loops
• Execution Trace Cache integrates path of program execution flow into a
single line
• Low latency data cache
— Second level cache
• Full-speed, unified 8-way Level 2 on-die Advance Transfer Cache
• Bandwidth and performance increases with processor frequency
• High-performance, quad-pumped bus interface to the Intel NetBurst microarchi-
tecture system bus
— Supports quad-pumped, scalable bus clock to achieve up to 4X effective
speed
— Capable of delivering up to 8.5 GBytes of bandwidth per second
• Superscalar issue to enable parallelism
• Expanded hardware registers with renaming to avoid register name space
limitations
• 64-byte cache line size (transfers data up to two lines per sector)
Figure 2-2 is an overview of the Intel NetBurst microarchitecture. This microarchitec-
ture pipeline is made up of three sections: (1) the front end pipeline, (2) the out-of-
order execution core, and (3) the retirement unit.

Vol. 1 2-11
INTEL® 64 AND IA-32 ARCHITECTURES

System Bus
Frequently used paths

Less frequently used

paths
Bus Unit

3rd Level Cache

Optional

2nd Level Cache 1st Level Cache

8-Way 4-way

Front End

Execution
Trace Cache
Fetch/Decode Out-Of-Order Retirement
Microcode ROM
Core

Branch History Update

BTBs/Branch Prediction

OM16521

Figure 2-2. The Intel NetBurst Microarchitecture

2.2.2.1 The Front End Pipeline

The front end supplies instructions in program order to the out-of-order execution
core. It performs a number of functions:
• Prefetches instructions that are likely to be executed
• Fetches instructions that have not already been prefetched
• Decodes instructions into micro-operations
• Generates microcode for complex instructions and special-purpose code
• Delivers decoded instructions from the execution trace cache
• Predicts branches using highly advanced algorithm
The pipeline is designed to address common problems in high-speed, pipelined
microprocessors. Two of these problems contribute to major sources of delays:
• time to decode instructions fetched from the target

2-12 Vol. 1
 INTEL® 64 AND IA-32 ARCHITECTURES

• wasted decode bandwidth due to branches or branch target in the middle of

cache lines
The operation of the pipeline’s trace cache addresses these issues. Instructions are
constantly being fetched and decoded by the translation engine (part of the
fetch/decode logic) and built into sequences of µops called traces. At any time,
multiple traces (representing prefetched branches) are being stored in the trace
cache. The trace cache is searched for the instruction that follows the active branch.
If the instruction also appears as the first instruction in a pre-fetched branch, the
fetch and decode of instructions from the memory hierarchy ceases and the pre-
fetched branch becomes the new source of instructions (see Figure 2-2).
The trace cache and the translation engine have cooperating branch prediction hard-
ware. Branch targets are predicted based on their linear addresses using branch
target buffers (BTBs) and fetched as soon as possible.

2.2.2.2 Out-Of-Order Execution Core

The out-of-order execution core’s ability to execute instructions out of order is a key
factor in enabling parallelism. This feature enables the processor to reorder instruc-
tions so that if one µop is delayed, other µops may proceed around it. The processor
employs several buffers to smooth the flow of µops.
The core is designed to facilitate parallel execution. It can dispatch up to six µops per
cycle (this exceeds trace cache and retirement µop bandwidth). Most pipelines can
start executing a new µop every cycle, so several instructions can be in flight at a
time for each pipeline. A number of arithmetic logical unit (ALU) instructions can
start at two per cycle; many floating-point instructions can start once every two
cycles.

2.2.2.3 Retirement Unit

The retirement unit receives the results of the executed µops from the out-of-order
execution core and processes the results so that the architectural state updates
according to the original program order.
When a µop completes and writes its result, it is retired. Up to three µops may be
retired per cycle. The Reorder Buffer (ROB) is the unit in the processor which buffers
completed µops, updates the architectural state in order, and manages the ordering
of exceptions. The retirement section also keeps track of branches and sends
updated branch target information to the BTB. The BTB then purges pre-fetched
traces that are no longer needed.

2.2.3 Intel® Core™ Microarchitecture

Intel Core microarchitecture introduces the following features that enable high
performance and power-efficient performance for single-threaded as well as multi-
threaded workloads:

Vol. 1 2-13
INTEL® 64 AND IA-32 ARCHITECTURES

• Intel® Wide Dynamic Execution enable each processor core to fetch,

dispatch, execute in high bandwidths to support retirement of up to four instruc-
tions per cycle.
— Fourteen-stage efficient pipeline
— Three arithmetic logical units
— Four decoders to decode up to five instruction per cycle
— Macro-fusion and micro-fusion to improve front-end throughput
— Peak issue rate of dispatching up to six micro-ops per cycle
— Peak retirement bandwidth of up to 4 micro-ops per cycle
— Advanced branch prediction
— Stack pointer tracker to improve efficiency of executing function/procedure
entries and exits
• Intel® Advanced Smart Cache delivers higher bandwidth from the second
level cache to the core, and optimal performance and flexibility for single-
threaded and multi-threaded applications.
— Large second level cache up to 4 MB and 16-way associativity
— Optimized for multicore and single-threaded execution environments
— 256 bit internal data path to improve bandwidth from L2 to first-level data
cache
• Intel® Smart Memory Access prefetches data from memory in response to
data access patterns and reduces cache-miss exposure of out-of-order
execution.
— Hardware prefetchers to reduce effective latency of second-level cache
misses
— Hardware prefetchers to reduce effective latency of first-level data cache
misses
— Memory disambiguation to improve efficiency of speculative execution
execution engine
• Intel® Advanced Digital Media Boost improves most 128-bit SIMD instruction
with single-cycle throughput and floating-point operations.
— Single-cycle throughput of most 128-bit SIMD instructions
— Up to eight floating-point operation per cycle
— Three issue ports available to dispatching SIMD instructions for execution
Intel Core 2 Extreme, Intel Core 2 Duo processors and Intel Xeon processor 5100
series implement two processor cores based on the Intel Core microarchitecture, the
functionality of the subsystems in each core are depicted in Figure 2-3.

2-14 Vol. 1
 INTEL® 64 AND IA-32 ARCHITECTURES

Instruction Fetch and P reD ecode

Instruction Q ueue

M icro-
code D ecode
ROM

S hared L2 C ache
R enam e/A lloc U p to 10.7 G B /s
FS B

R etirem ent U nit

(R e-O rder B uffer)

S cheduler

A LU A LU A LU
B ranch FA dd FM ul Load S tore
M M X /S S E /FP M M X /S S E M M X/S S E
M ove

L1D C ache and D T LB

Figure 2-3. The Intel Core Microarchitecture Pipeline Functionality

2.2.3.1 The Front End

The front end of Intel Core microarchitecture provides several enhancements to feed
the Intel Wide Dynamic Execution engine:
• Instruction fetch unit prefetches instructions into an instruction queue to
maintain steady supply of instruction to the decode units.
• Four-wide decode unit can decode 4 instructions per cycle or 5 instructions per
cycle with Macrofusion.
• Macrofusion fuses common sequence of two instructions as one decoded
instruction (micro-ops) to increase decoding throughput.
• Microfusion fuses common sequence of two micro-ops as one micro-ops to
improve retirement throughput.
• Instruction queue provides caching of short loops to improve efficiency.
• Stack pointer tracker improves efficiency of executing procedure/function entries
and exits.

Vol. 1 2-15
INTEL® 64 AND IA-32 ARCHITECTURES

• Branch prediction unit employs dedicated hardware to handle different types of

branches for improved branch prediction.
• Advanced branch prediction algorithm directs instruction fetch unit to fetch
instructions likely in the architectural code path for decoding.

2.2.3.2 Execution Core

The execution core of the Intel Core microarchitecture is superscalar and can process
instructions out of order to increase the overall rate of instructions executed per cycle
(IPC). The execution core employs the following feature to improve execution
throughput and efficiency:
• Up to six micro-ops can be dispatched to execute per cycle
• Up to four instructions can be retired per cycle
• Three full arithmetic logical units
• SIMD instructions can be dispatched through three issue ports
• Most SIMD instructions have 1-cycle throughput (including 128-bit SIMD instruc-
tions)
• Up to eight floating-point operation per cycle
• Many long-latency computation operation are pipelined in hardware to increase
overall throughput
• Reduced exposure to data access delays using Intel Smart Memory Access

2.2.4 Intel® Atom™ Microarchitecture

Intel Atom microarchitecture maximizes power-efficient performance for single-
threaded and multi-threaded workloads by providing:
• Advanced Micro-Ops Execution
— Single-micro-op instruction execution from decode to retirement, including
instructions with register-only, load, and store semantics.
— Sixteen-stage, in-order pipeline optimized for throughput and reduced power
consumption.
— Dual pipelines to enable decode, issue, execution and retirement of two
instructions per cycle.
— Advanced stack pointer to improve efficiency of executing function
entry/returns.
• Intel® Smart Cache
— Second level cache is 512 KB and 8-way associativity.
— Optimized for multi-threaded and single-threaded execution environments

2-16 Vol. 1
 INTEL® 64 AND IA-32 ARCHITECTURES

— 256 bit internal data path between L2 and L1 data cache improves high
bandwidth.
• Efficient Memory Access
— Efficient hardware prefetchers to L1 and L2, speculatively loading data likely
to be requested by processor to reduce cache miss impact.
• Intel® Digital Media Boost
— Two issue ports for dispatching SIMD instructions to execution units.
— Single-cycle throughput for most 128-bit integer SIMD instructions
— Up to six floating-point operations per cycle
— Up to two 128-bit SIMD integer operations per cycle
— Safe Instruction Recognition (SIR) to allow long-latency floating-point
operations to retire out of order with respect to integer instructions.

2.2.5 Intel Microarchitecture (Nehalem)

Intel microarchitecture (Nehalem) provides the foundation for many innovative
features of Intel Core i7 processors. It builds on the success of 45nm Intel Core
microarchitecture and provides the following feature enhancements:
• Enhanced processor core
— Improved branch prediction and recovery from misprediction.
— Enhanced loop streaming to improve front end performance and reduce
power consumption.
— Deeper buffering in out-of-order engine to extract parallelism.
— Enhanced execution units to provide acceleration in CRC, string/text
processing and data shuffling.
• Smart Memory Access
— Integrated memory controller provides low-latency access to system memory
and scalable memory bandwidth
— New cache hierarchy organization with shared, inclusive L3 to reduce snoop
traffic
— Two level TLBs and increased TLB size.
— Fast unaligned memory access.
• HyperThreading Technology
— Provides two hardware threads (logical processors) per core.
— Takes advantage of 4-wide execution engine, large L3, and massive memory
bandwidth.
• Dedicated Power management Innovations

Vol. 1 2-17
INTEL® 64 AND IA-32 ARCHITECTURES

— Integrated microcontroller with optimized embedded firmware to manage

power consumption.
— Embedded real-time sensors for temperature, current, and power.
— Integrated power gate to turn off/on per-core power consumption
— Versatility to reduce power consumption of memory, link subsystems.

2.2.6 SIMD Instructions

Beginning with the Pentium II and Pentium with Intel MMX technology processor
families, six extensions have been introduced into the Intel 64 and IA-32 architec-
tures to perform single-instruction multiple-data (SIMD) operations. These exten-
sions include the MMX technology, SSE extensions, SSE2 extensions, SSE3
extensions, Supplemental Streaming SIMD Extensions 3, and SSE4. Each of these
extensions provides a group of instructions that perform SIMD operations on packed
integer and/or packed floating-point data elements.
SIMD integer operations can use the 64-bit MMX or the 128-bit XMM registers. SIMD
floating-point operations use 128-bit XMM registers. Figure 2-4 shows a summary of
the various SIMD extensions (MMX technology, SSE, SSE2, SSE3, SSSE3, and SSE4),
the data types they operate on, and how the data types are packed into MMX and
XMM registers.
The Intel MMX technology was introduced in the Pentium II and Pentium with MMX
technology processor families. MMX instructions perform SIMD operations on packed
byte, word, or doubleword integers located in MMX registers. These instructions are
useful in applications that operate on integer arrays and streams of integer data that
lend themselves to SIMD processing.
SSE extensions were introduced in the Pentium III processor family. SSE instructions
operate on packed single-precision floating-point values contained in XMM registers
and on packed integers contained in MMX registers. Several SSE instructions provide
state management, cache control, and memory ordering operations. Other SSE
instructions are targeted at applications that operate on arrays of single-precision
floating-point data elements (3-D geometry, 3-D rendering, and video encoding and
decoding applications).
SSE2 extensions were introduced in Pentium 4 and Intel Xeon processors. SSE2
instructions operate on packed double-precision floating-point values contained in
XMM registers and on packed integers contained in MMX and XMM registers. SSE2
integer instructions extend IA-32 SIMD operations by adding new 128-bit SIMD
integer operations and by expanding existing 64-bit SIMD integer operations to
128-bit XMM capability. SSE2 instructions also provide new cache control and
memory ordering operations.
SSE3 extensions were introduced with the Pentium 4 processor supporting Hyper-
Threading Technology (built on 90 nm process technology). SSE3 offers 13 instruc-
tions that accelerate performance of Streaming SIMD Extensions technology,
Streaming SIMD Extensions 2 technology, and x87-FP math capabilities.

2-18 Vol. 1
 INTEL® 64 AND IA-32 ARCHITECTURES

SSSE3 extensions were introduced with the Intel Xeon processor 5100 series and
Intel Core 2 processor family. SSSE3 offer 32 instructions to accelerate processing of
SIMD integer data.
SSE4 extensions offer 54 instructions. 47 of them are referred to as SSE4.1 instruc-
tions. SSE4.1 are introduced with Intel Xeon processor 5400 series and Intel Core 2
Extreme processor QX9650. The other 7 SSE4 instructions are referred to as SSE4.2
instructions.
Intel 64 architecture allows four generations of 128-bit SIMD extensions to access up
to 16 XMM registers. IA-32 architecture provides 8 XMM registers.
See also:
• Section 5.4, “MMX™ Instructions,” and Chapter 9, “Programming with Intel®
MMX™ Technology”
• Section 5.5, “SSE Instructions,” and Chapter 10, “Programming with Streaming
SIMD Extensions (SSE)”
• Section 5.6, “SSE2 Instructions,” and Chapter 11, “Programming with Streaming
SIMD Extensions 2 (SSE2)”
• Section 5.7, “SSE3 Instructions”, Section 5.8, “Supplemental Streaming SIMD
Extensions 3 (SSSE3) Instructions”, Section 5.9, “SSE4 Instructions”, and
Chapter 12, “Programming with SSE3, SSSE3, AND SSE4”

Vol. 1 2-19
INTEL® 64 AND IA-32 ARCHITECTURES

SIMD Extension Register Layout Data Type

MMX Registers
MMX Technology 8 Packed Byte Integers

4 Packed Word Integers

2 Packed Doubleword Integers

Quadword

MMX Registers
SSE 8 Packed Byte Integers

4 Packed Word Integers

2 Packed Doubleword Integers

Quadword

XMM Registers
4 Packed Single-Precision
Floating-Point Values

MMX Registers
SSE2/SSE3/SSSE3 2 Packed Doubleword Integers

Quadword

XMM Registers
2 Packed Double-Precision
Floating-Point Values

16 Packed Byte Integers

8 Packed Word Integers

4 Packed Doubleword
Integers
2 Quadword Integers

Double Quadword

Figure 2-4. SIMD Extensions, Register Layouts, and Data Types

2-20 Vol. 1
 INTEL® 64 AND IA-32 ARCHITECTURES

2.2.7 Intel® Hyper-Threading Technology

Intel Hyper-Threading Technology (Intel HT Technology) was developed to improve
the performance of IA-32 processors when executing multi-threaded operating
system and application code or single-threaded applications under multi-tasking
environments. The technology enables a single physical processor to execute two or
more separate code streams (threads) concurrently using shared execution
resources.
Intel HT Technology is one form of hardware multi-threading capability in IA-32
processor families. It differs from multi-processor capability using separate physi-
cally distinct packages with each physical processor package mated with a physical
socket. Intel HT Technology provides hardware multi-threading capability with a
single physical package by using shared execution resources in a processor core.
Architecturally, an IA-32 processor that supports Intel HT Technology consists of two
or more logical processors, each of which has its own IA-32 architectural state. Each
logical processor consists of a full set of IA-32 data registers, segment registers,
control registers, debug registers, and most of the MSRs. Each also has its own
advanced programmable interrupt controller (APIC).
Figure 2-5 shows a comparison of a processor that supports Intel HT Technology
(implemented with two logical processors) and a traditional dual processor system.

IA-32 Processor Supporting

Traditional Multiple Processor (MP) System
Hyper-Threading Technology

AS AS AS AS

Processor Core Processor Core Processor Core

IA-32 processor IA-32 processor IA-32 processor

Two logical Each processor is a

processors that share separate physical
a single core package

AS = IA-32 Architectural State

OM16522

Figure 2-5. Comparison of an IA-32 Processor Supporting Hyper-Threading

Technology and a Traditional Dual Processor System
Unlike a traditional MP system configuration that uses two or more separate physical
IA-32 processors, the logical processors in an IA-32 processor supporting Intel HT
Technology share the core resources of the physical processor. This includes the

Vol. 1 2-21
INTEL® 64 AND IA-32 ARCHITECTURES

execution engine and the system bus interface. After power up and initialization,
each logical processor can be independently directed to execute a specified thread,
interrupted, or halted.
Intel HT Technology leverages the process and thread-level parallelism found in
contemporary operating systems and high-performance applications by providing
two or more logical processors on a single chip. This configuration allows two or more
threads1 to be executed simultaneously on each a physical processor. Each logical
processor executes instructions from an application thread using the resources in the
processor core. The core executes these threads concurrently, using out-of-order
instruction scheduling to maximize the use of execution units during each clock cycle.

2.2.7.1 Some Implementation Notes

All Intel HT Technology configurations require:
• A processor that supports Intel HT Technology
• A chipset and BIOS that utilize the technology
• Operating system optimizations
See http://www.intel.com/products/ht/hyperthreading_more.htm for information.
At the firmware (BIOS) level, the basic procedures to initialize the logical processors
in a processor supporting Intel HT Technology are the same as those for a traditional
DP or MP platform. The mechanisms that are described in the Multiprocessor Specifi-
cation, Version 1.4 to power-up and initialize physical processors in an MP system
also apply to logical processors in a processor that supports Intel HT Technology.
An operating system designed to run on a traditional DP or MP platform may use
CPUID to determine the presence of hardware multi-threading support feature and
the number of logical processors they provide.
Although existing operating system and application code should run correctly on a
processor that supports Intel HT Technology, some code modifications are recom-
mended to get the optimum benefit. These modifications are discussed in Chapter 7,
“Multiple-Processor Management,” Intel® 64 and IA-32 Architectures Software
Developer’s Manual, Volume 3A.

2.2.8 Multi-Core Technology

Multi-core technology is another form of hardware multi-threading capability in IA-32
processor families. Multi-core technology enhances hardware multi-threading capa-
bility by providing two or more execution cores in a physical package.
The Intel Pentium processor Extreme Edition is the first member in the IA-32
processor family to introduce multi-core technology. The processor provides hard-

1. In the remainder of this document, the term “thread” will be used as a general term for the terms
“process” and “thread.”

2-22 Vol. 1
 INTEL® 64 AND IA-32 ARCHITECTURES

ware multi-threading support with both two processor cores and Intel Hyper-
Threading Technology. This means that the Intel Pentium processor Extreme Edition
provides four logical processors in a physical package (two logical processors for
each processor core). The Dual-Core Intel Xeon processor features multi-core, Intel
Hyper-Threading Technology and supports multi-processor platforms.
The Intel Pentium D processor also features multi-core technology. This processor
provides hardware multi-threading support with two processor cores but does not
offer Intel Hyper-Threading Technology. This means that the Intel Pentium D
processor provides two logical processors in a physical package, with each logical
processor owning the complete execution resources of a processor core.
The Intel Core 2 processor family, Intel Xeon processor 3000 series, Intel Xeon
processor 5100 series, and Intel Core Duo processor offer power-efficient multi-core
technology. The processor contains two cores that share a smart second level cache.
The Level 2 cache enables efficient data sharing between two cores to reduce
memory traffic to the system bus.

Vol. 1 2-23
INTEL® 64 AND IA-32 ARCHITECTURES

Intel Core Duo Processor

Intel Core 2 Duo Processor
Intel Pentium dual-core Processor Pentium D Processor
Architectual State Architectual State Architectual State Architectual State
Execution Engine Execution Engine
Execution Engine Execution Engine
Local APIC Local APIC
Local APIC Local APIC
Second Level Cache
Bus Interface Bus Interface
Bus Interface

System Bus System Bus

Pentium Processor Extreme Edition

Architectual Architectual Architectual Architectual
State State State State

Execution Engine Execution Engine

Local APIC Local APIC Local APIC Local APIC

Bus Interface Bus Interface

OM19809
System Bus

Figure 2-6. Intel 64 and IA-32 Processors that Support Dual-Core

The Pentium® dual-core processor is based on the same technology as the Intel Core
2 Duo processor family.
The Intel Xeon processor 7300, 5300 and 3200 series, Intel Core 2 Extreme Quad-
Core processor, and Intel Core 2 Quad processors support Intel quad-core tech-
nology. The Quad-core Intel Xeon processors and the Quad-Core Intel Core 2
processor family are also in Figure 2-7.

2-24 Vol. 1
 INTEL® 64 AND IA-32 ARCHITECTURES

Intel Core 2 Extreme Quad-core Processor

Intel Core 2 Quad Processor
Intel Xeon Processor 3200 Series
Intel Xeon Processor 5300 Series

Architectual State Architectual State Architectual State Architectual State

Execution Engine Execution Engine Execution Engine Execution Engine

Local APIC Local APIC Local APIC Local APIC

Second Level Cache Second Level Cache

Bus Interface Bus Interface

System Bus

OM19810

Figure 2-7. Intel 64 Processors that Support Quad-Core

Intel Core i7 processors support Intel quad-core technology, Intel HyperThreading

Technology, provides Intel QuickPath interconnect link to the chipset and have inte-
grated memory controller supporting three channel to DDR3 memory.

Vol. 1 2-25
INTEL® 64 AND IA-32 ARCHITECTURES

Intel Core i7 Processor

Logical Logical Logical Logical Logical Logical Logical Logical

Proces Proces Proces Proces Proces Proces Proces Proces
sor sor sor sor sor sor sor sor

L1 and L2 L1 and L2 L1 and L2 L1 and L2

Execution Engine Execution Engine Execution Engine Execution Engine

Third Level Cache

QuickPath Interconnect (QPI) Interface, Integrated Memory Controller

IMC
QPI
DDR3

Chipset
OM19810b

Figure 2-8. Intel Core i7 Processor

2.2.9 Intel® 64 Architecture

Intel 64 architecture increases the linear address space for software to 64 bits and
supports physical address space up to 40 bits. The technology also introduces a new
operating mode referred to as IA-32e mode.
IA-32e mode operates in one of two sub-modes: (1) compatibility mode enables a
64-bit operating system to run most legacy 32-bit software unmodified, (2) 64-bit
mode enables a 64-bit operating system to run applications written to access 64-bit
address space.
In the 64-bit mode, applications may access:
• 64-bit flat linear addressing
• 8 additional general-purpose registers (GPRs)
• 8 additional registers for streaming SIMD extensions (SSE, SSE2, SSE3 and
SSSE3)
• 64-bit-wide GPRs and instruction pointers
• uniform byte-register addressing
• fast interrupt-prioritization mechanism

2-26 Vol. 1
 INTEL® 64 AND IA-32 ARCHITECTURES

• a new instruction-pointer relative-addressing mode

An Intel 64 architecture processor supports existing IA-32 software because it is able
to run all non-64-bit legacy modes supported by IA-32 architecture. Most existing
IA-32 applications also run in compatibility mode.

2.2.10 Intel® Virtualization Technology (Intel® VT)

Intel® Virtualization Technology for Intel 64 and IA-32 architectures provide exten-
sions that support virtualization. The extensions are referred to as Virtual Machine
Extensions (VMX). An Intel 64 or IA-32 platform with VMX can function as multiple
virtual systems (or virtual machines). Each virtual machine can run operating
systems and applications in separate partitions.
VMX also provides programming interface for a new layer of system software (called
the Virtual Machine Monitor (VMM)) used to manage the operation of virtual
machines. Information on VMX and on the programming of VMMs is in Intel® 64 and
IA-32 Architectures Software Developer’s Manual, Volume 3B. Chapter 5, “VMX
Instruction Reference,” in the Intel® 64 and IA-32 Architectures Software Devel-
oper’s Manual, Volume 2B, provides information on VMX instructions.
Intel Core i7 processor provides the following enhancements to Intel Virtualization
Technology:
• Virtual processor ID (VPID) to reduce the cost of VMM managing transitions.
• Extended page table (EPT) to reduce the number of transitions for VMM to
manage memory virtualization.
• Reduced latency of VM transitions.

2.3 INTEL® 64 AND IA-32 PROCESSOR GENERATIONS

In the mid-1960s, Intel cofounder and Chairman Emeritus Gordon Moore had this
observation: “... the number of transistors that would be incorporated on a silicon die
would double every 18 months for the next several years.” Over the past three and
half decades, this prediction known as “Moore's Law” has continued to hold true.
The computing power and the complexity (or roughly, the number of transistors per
processor) of Intel architecture processors has grown in close relation to Moore's law.
By taking advantage of new process technology and new microarchitecture designs,
each new generation of IA-32 processors has demonstrated frequency-scaling head-
room and new performance levels over the previous generation processors.
The key features of the Intel Pentium 4 processor, Intel Xeon processor, Intel Xeon
processor MP, Pentium III processor, and Pentium III Xeon processor with advanced

Vol. 1 2-27
INTEL® 64 AND IA-32 ARCHITECTURES

transfer cache are shown in Table 2-1. Older generation IA-32 processors, which do
not employ on-die Level 2 cache, are shown in Table 2-2.
Table 2-1. Key Features of Most Recent IA-32 Processors
Intel Date Micro- Top-Bin Tran- Register Syste Max. On-Die
Processor Intro- architecture Clock Fre- sistors Sizes1 m Bus Extern. Caches2
duced quency at Band- Addr.
Intro- width Space
duction
Intel Pentium M 2004 Intel Pentium M 2.00 GHz 140 M GP: 32 3.2 GB/s 4 GB L1: 64 KB
Processor 7553 Processor FPU: 80 L2: 2 MB
MMX: 64
XMM: 128

Intel Core Duo 2006 Improved Intel Pentium 2.16 GHz 152M GP: 32 5.3 GB/s 4 GB L1: 64 KB
Processor M Processor FPU: 80 L2: 2 MB (2MB
T26003 Microarchitecture; Dual MMX: 64 Total)
Core; XMM: 128
Intel Smart Cache,
Advanced Thermal
Manager

Intel Atom 2008 Intel Atom 1.86 GHz - 800 47M GP: 32 Up to 4.2 4 GB L1: 56 KB4
Processor Z5xx Microarchitecture; MHz FPU: 80 GB/s L2: 512KB
series Intel Virtualization MMX: 64
Technology. XMM: 128

NOTES:
1. The register size and external data bus size are given in bits.
2. First level cache is denoted using the abbreviation L1, 2nd level cache is denoted as L2. The size
of L1 includes the first-level data cache and the instruction cache where applicable, but
does not include the trace cache.
3. Intel processor numbers are not a measure of performance. Processor numbers differentiate
features within each processor family, not across different processor families.
See http://www.intel.com/products/processor_number for details.
4. In Intel Atom Processor, the size of L1 instruction cache is 32 KBytes, L1 data cache is 24 KBytes.

Table 2-2. Key Features of Most Recent Intel 64 Processors

Intel Date Micro- Top-Bin Tran- Register System Max. On-Die
Processor Intro- architec-ture Fre- sistor Sizes Bus/QP Extern Caches
duced quency s I Link . Addr.
at Intro- Speed Space
duction
64-bit Intel Xeon 2004 Intel NetBurst 3.60 GHz 125 M GP: 32, 64 6.4 GB/s 64 GB 12K µop
Processor with Microarchitecture; FPU: 80 Execution
800 MHz Intel Hyper-Threading MMX: 64 Trace Cache;
System Bus Technology; Intel 64 XMM: 128 16 KB L1;
Architecture 1 MB L2

64-bit Intel Xeon 2005 Intel NetBurst 3.33 GHz 675M GP: 32, 64 5.3 GB/s 1 1024 GB 12K µop
Processor MP Microarchitecture; FPU: 80 (1 TB) Execution
with 8MB L3 Intel Hyper-Threading MMX: 64 Trace Cache;
Technology; Intel 64 XMM: 128 16 KB L1;
Architecture 1 MB L2,
8 MB L3

2-28 Vol. 1
 INTEL® 64 AND IA-32 ARCHITECTURES

Table 2-2. Key Features of Most Recent Intel 64 Processors (Contd.)

Intel Date Micro- Top-Bin Tran- Register System Max. On-Die
Processor Intro- architec-ture Fre- sistor Sizes Bus/QP Extern Caches
duced quency s I Link . Addr.
at Intro- Speed Space
duction
Intel Pentium 4 2005 Intel NetBurst 3.73 GHz 164 M GP: 32, 64 8.5 GB/s 64 GB 12K µop
Processor Microarchitecture; FPU: 80 Execution
Extreme Edition Intel Hyper-Threading MMX: 64 Trace Cache;
Supporting Technology; Intel 64 XMM: 128 16 KB L1;
Hyper-Threading Architecture 2 MB L2
Technology

Intel Pentium 2005 Intel NetBurst 3.20 GHz 230 M GP: 32, 64 6.4 GB/s 64 GB 12K µop
Processor Microarchitecture; FPU: 80 Execution
Extreme Edition Intel Hyper-Threading MMX: 64 Trace Cache;
840 Technology; Intel 64 XMM: 128 16 KB L1;
Architecture; 1MB L2 (2MB
Dual-core 2 Total)

Dual-Core Intel 2005 Intel NetBurst 3.00 GHz 321M GP: 32, 64 6.4 GB/s 64 GB 12K µop
Xeon Microarchitecture; FPU: 80 Execution
Processor 7041 Intel Hyper-Threading MMX: 64 Trace Cache;
Technology; Intel 64 XMM: 128 16 KB L1;
Architecture; 2MB L2 (4MB
Dual-core 3 Total)

Intel Pentium 4 2005 Intel NetBurst 3.80 GHz 164 M GP: 32, 64 6.4 GB/s 64 GB 12K µop
Processor 672 Microarchitecture; FPU: 80 Execution
Intel Hyper-Threading MMX: 64 Trace Cache;
Technology; Intel 64 XMM: 128 16 KB L1;
Architecture; 2MB L2
Intel Virtualization
Technology.

Intel Pentium 2006 Intel NetBurst 3.46 GHz 376M GP: 32, 64 8.5 GB/s 64 GB 12K µop
Processor Microarchitecture; FPU: 80 Execution
Extreme Edition Intel 64 Architecture; MMX: 64 Trace Cache;
955 Dual Core; XMM: 128 16 KB L1;
Intel Virtualization 2MB L2
Technology. (4MB Total)

Intel Core 2 2006 Intel Core 2.93 GHz 291M GP: 32,64 8.5 GB/s 64 GB L1: 64 KB
Extreme Microarchitecture; FPU: 80 L2: 4MB (4MB
Processor Dual Core; MMX: 64 Total)
X6800 Intel 64 Architecture; XMM: 128
Intel Virtualization
Technology.

Intel Xeon 2006 Intel Core 3.00 GHz 291M GP: 32, 64 10.6 GB/s 64 GB L1: 64 KB
Processor 5160 Microarchitecture; FPU: 80 L2: 4MB (4MB
Dual Core; MMX: 64 Total)
Intel 64 Architecture; XMM: 128
Intel Virtualization
Technology.

Intel Xeon 2006 Intel NetBurst 3.40 GHz 1.3 B GP: 32, 64 12.8 GB/s 64 GB L1: 64 KB
Processor 7140 Microarchitecture; FPU: 80 L2: 1MB (2MB
Dual Core; MMX: 64 Total)
Intel 64 Architecture; XMM: 128 L3: 16 MB
Intel Virtualization (16MB Total)
Technology.

Intel Core 2 2006 Intel Core 2.66 GHz 582M GP: 32,64 8.5 GB/s 64 GB L1: 64 KB
Extreme Microarchitecture; FPU: 80 L2: 4MB (4MB
Processor Quad Core; MMX: 64 Total)
QX6700 Intel 64 Architecture; XMM: 128
Intel Virtualization
Technology.

Vol. 1 2-29
INTEL® 64 AND IA-32 ARCHITECTURES

Table 2-2. Key Features of Most Recent Intel 64 Processors (Contd.)

Intel Date Micro- Top-Bin Tran- Register System Max. On-Die
Processor Intro- architec-ture Fre- sistor Sizes Bus/QP Extern Caches
duced quency s I Link . Addr.
at Intro- Speed Space
duction
Quad-core Intel 2006 Intel Core 2.66 GHz 582 M GP: 32, 64 10.6 GB/s 256 GB L1: 64 KB
Xeon Microarchitecture; FPU: 80 L2: 4MB (8 MB
Processor 5355 Quad Core; MMX: 64 Total)
Intel 64 Architecture; XMM: 128
Intel Virtualization
Technology.

Intel Core 2 Duo 2007 Intel Core 3.00 GHz 291 M GP: 32, 64 10.6 GB/s 64 GB L1: 64 KB
Processor Microarchitecture; FPU: 80 L2: 4MB (4MB
E6850 Dual Core; MMX: 64 Total)
Intel 64 Architecture; XMM: 128
Intel Virtualization
Technology;
Intel Trusted
Execution Technology

Intel Xeon 2007 Intel Core 2.93 GHz 582 M GP: 32, 64 8.5 GB/s 1024 GB L1: 64 KB
Processor 7350 Microarchitecture; FPU: 80 L2: 4MB (8MB
Quad Core; MMX: 64 Total)
Intel 64 Architecture; XMM: 128
Intel Virtualization
Technology.

Intel Xeon 2007 Enhanced Intel Core 3.00 GHz 820 M GP: 32, 64 12.8 GB/s 256 GB L1: 64 KB
Processor 5472 Microarchitecture; FPU: 80 L2: 6MB
Quad Core; MMX: 64 (12MB Total)
Intel 64 Architecture; XMM: 128
Intel Virtualization
Technology.

Intel Atom 2008 Intel Atom 2.0 - 1.60 GHz 47 M GP: 32, 64 Up to 4.2 Up to 64GB L1: 56 KB4
Processor Microarchitecture; FPU: 80 GB/s L2: 512KB
Intel 64 Architecture; MMX: 64
Intel Virtualization XMM: 128
Technology.

Intel Xeon 2008 Enhanced Intel Core 2.67 GHz 1.9 B GP: 32, 64 8.5 GB/s 1024 GB L1: 64 KB
Processor 7460 Microarchitecture; Six FPU: 80 L2: 3MB (9MB
Cores; MMX: 64 Total)
Intel 64 Architecture; XMM: 128 L3: 16MB
Intel Virtualization
Technology.

Intel Atom 2008 Intel Atom 1.60 GHz 94 M GP: 32, 64 Up to 4.2 Up to 64GB L1: 56 KB5
Processor 330 Microarchitecture; FPU: 80 GB/s L2: 512KB
Intel 64 Architecture; MMX: 64 (1MB Total)
Dual core; XMM: 128
Intel Virtualization
Technology.

Intel Core i7-965 2008 Intel microarchitecture 3.20 GHz 731 M GP: 32, 64 QPI: 6.4 64 GB L1: 64 KB
Processor (Nehalem); Quadcore; FPU: 80 GT/s; L2: 256KB
Extreme Edition HyperThreading MMX: 64 Memory: L3: 8MB
Technology; Intel QPI; XMM: 128 25.6 GB/s
Intel 64 Architecture;
Intel Virtualization
Technology.

NOTES:
1. The 64-bit Intel Xeon Processor MP with an 8-MByte L3 supports a multi-processor platform with a
dual system bus; this creates a platform bandwidth with 10.6 GBytes.
2. In Intel Pentium Processor Extreme Edition 840, the size of on-die cache is listed for each core. The
total size of L2 in the physical package in 2 MBytes.

2-30 Vol. 1
 INTEL® 64 AND IA-32 ARCHITECTURES

3. In Dual-Core Intel Xeon Processor 7041, the size of on-die cache is listed for each core. The total
size of L2 in the physical package in 4 MBytes.
4. In Intel Atom Processor, the size of L1 instruction cache is 32 KBytes, L1 data cache is 24 KBytes.
5. In Intel Atom Processor, the size of L1 instruction cache is 32 KBytes, L1 data cache is 24 KBytes.

Vol. 1 2-31
INTEL® 64 AND IA-32 ARCHITECTURES

Table 2-3. Key Features of Previous Generations of IA-32 Processors

Intel Date Max. Clock Tran- Register Ext. Data Max. Caches
Processor Intro- Frequency/ sistors Sizes1 Bus Extern.
duced Technology at Size2 Addr.
Introduction Space
8086 1978 8 MHz 29 K 16 GP 16 1 MB None

Intel 286 1982 12.5 MHz 134 K 16 GP 16 16 MB Note 3

Intel386 DX Processor 1985 20 MHz 275 K 32 GP 32 4 GB Note 3

Intel486 DX Processor 1989 25 MHz 1.2 M 32 GP 32 4 GB L1: 8 KB

80 FPU

Pentium Processor 1993 60 MHz 3.1 M 32 GP 64 4 GB L1:16 KB

80 FPU

Pentium Pro Processor 1995 200 MHz 5.5 M 32 GP 64 64 GB L1: 16 KB

80 FPU L2: 256 KB or
512 KB

Pentium II Processor 1997 266 MHz 7M 32 GP 64 64 GB L1: 32 KB

80 FPU L2: 256 KB or
64 MMX 512 KB

Pentium III Processor 1999 500 MHz 8.2 M 32 GP 64 64 GB L1: 32 KB

80 FPU L2: 512 KB
64 MMX
128 XMM

Pentium III and Pentium 1999 700 MHz 28 M 32 GP 64 64 GB L1: 32 KB

III Xeon Processors 80 FPU L2: 256 KB
64 MMX
128 XMM

Pentium 4 Processor 2000 1.50 GHz, Intel NetBurst 42 M 32 GP 64 64 GB 12K µop

Microarchitecture 80 FPU Execution Trace
64 MMX Cache; L1: 8KB
128 XMM L2: 256 KB

Intel Xeon Processor 2001 1.70 GHz, Intel NetBurst 42 M 32 GP 64 64 GB 12K µop
Microarchitecture 80 FPU Execution Trace
64 MMX Cache; L1: 8KB
128 XMM L2: 512KB

Intel Xeon Processor 2002 2.20 GHz, Intel NetBurst 55 M 32 GP 64 64 GB 12K µop
Microarchitecture, 80 FPU Execution Trace
HyperThreading 64 MMX Cache; L1: 8KB
Technology 128 XMM L2: 512KB

Pentium M Processor 2003 1.60 GHz, Intel NetBurst 77 M 32 GP 64 4 GB L1: 64KB

Microarchitecture 80 FPU L2: 1 MB
64 MMX
128 XMM

Intel Pentium 4 2004 3.40 GHz, Intel NetBurst 125 M 32 GP 64 64 GB 12K µop
Processor Supporting Microarchitecture, 80 FPU Execution Trace
Hyper-Threading HyperThreading 64 MMX Cache; L1: 16KB
Technology at 90 nm Technology 128 XMM L2: 1 MB
process

NOTE:
1. The register size and external data bus size are given in bits. Note also that each 32-bit general-
purpose (GP) registers can be addressed as an 8- or a 16-bit data registers in all of the processors.
2. Internal data paths are 2 to 4 times wider than the external data bus for each processor.

2-32 Vol. 1
 CHAPTER 3
BASIC EXECUTION ENVIRONMENT

This chapter describes the basic execution environment of an Intel 64 or IA-32

processor as seen by assembly-language programmers. It describes how the
processor executes instructions and how it stores and manipulates data. The execu-
tion environment described here includes memory (the address space), general-
purpose data registers, segment registers, the flag register, and the instruction
pointer register.

3.1 MODES OF OPERATION

The IA-32 architecture supports three basic operating modes: protected mode, real-
address mode, and system management mode. The operating mode determines
which instructions and architectural features are accessible:
• Protected mode — This mode is the native state of the processor. Among the
capabilities of protected mode is the ability to directly execute “real-address
mode” 8086 software in a protected, multi-tasking environment. This feature is
called virtual-8086 mode, although it is not actually a processor mode. Virtual-
8086 mode is actually a protected mode attribute that can be enabled for any
task.
• Real-address mode — This mode implements the programming environment of
the Intel 8086 processor with extensions (such as the ability to switch to
protected or system management mode). The processor is placed in real-address
mode following power-up or a reset.
• System management mode (SMM) — This mode provides an operating
system or executive with a transparent mechanism for implementing platform-
specific functions such as power management and system security. The
processor enters SMM when the external SMM interrupt pin (SMI#) is activated
or an SMI is received from the advanced programmable interrupt controller
(APIC).
In SMM, the processor switches to a separate address space while saving the
basic context of the currently running program or task. SMM-specific code may
then be executed transparently. Upon returning from SMM, the processor is
placed back into its state prior to the system management interrupt. SMM was
introduced with the Intel386™ SL and Intel486™ SL processors and became a
standard IA-32 feature with the Pentium processor family.

Vol. 1 3-1
BASIC EXECUTION ENVIRONMENT

3.1.1 Intel® 64 Architecture

Intel 64 architecture adds IA-32e mode. IA-32e mode has two sub-modes.
These are:
• Compatibility mode (sub-mode of IA-32e mode) — Compatibility mode
permits most legacy 16-bit and 32-bit applications to run without re-compilation
under a 64-bit operating system. For brevity, the compatibility sub-mode is
referred to as compatibility mode in IA-32 architecture. The execution
environment of compatibility mode is the same as described in Section 3.2.
Compatibility mode also supports all of the privilege levels that are supported in
64-bit and protected modes. Legacy applications that run in Virtual 8086 mode or
use hardware task management will not work in this mode.
Compatibility mode is enabled by the operating system (OS) on a code segment
basis. This means that a single 64-bit OS can support 64-bit applications running
in 64-bit mode and support legacy 32-bit applications (not recompiled for
64-bits) running in compatibility mode.
Compatibility mode is similar to 32-bit protected mode. Applications access only
the first 4 GByte of linear-address space. Compatibility mode uses 16-bit and 32-
bit address and operand sizes. Like protected mode, this mode allows applica-
tions to access physical memory greater than 4 GByte using PAE (Physical
Address Extensions).
• 64-bit mode (sub-mode of IA-32e mode) — This mode enables a 64-bit
operating system to run applications written to access 64-bit linear address
space. For brevity, the 64-bit sub-mode is referred to as 64-bit mode in IA-32
architecture.
64-bit mode extends the number of general purpose registers and SIMD
extension registers from 8 to 16. General purpose registers are widened to 64
bits. The mode also introduces a new opcode prefix (REX) to access the register
extensions. See Section 3.2.1 for a detailed description.
64-bit mode is enabled by the operating system on a code-segment basis. Its
default address size is 64 bits and its default operand size is 32 bits. The default
operand size can be overridden on an instruction-by-instruction basis using a REX
opcode prefix in conjunction with an operand size override prefix.
REX prefixes allow a 64-bit operand to be specified when operating in 64-bit
mode. By using this mechanism, many existing instructions have been promoted
to allow the use of 64-bit registers and 64-bit addresses.

3.2 OVERVIEW OF THE BASIC EXECUTION

ENVIRONMENT
Any program or task running on an IA-32 processor is given a set of resources for
executing instructions and for storing code, data, and state information. These

3-2 Vol. 1
 BASIC EXECUTION ENVIRONMENT

resources (described briefly in the following paragraphs and shown in Figure 3-1)
make up the basic execution environment for an IA-32 processor.
An Intel 64 processor supports the basic execution environment of an IA-32
processor, and a similar environment under IA-32e mode that can execute 64-bit
programs (64-bit sub-mode) and 32-bit programs (compatibility sub-mode).
The basic execution environment is used jointly by the application programs and the
operating system or executive running on the processor.
• Address space — Any task or program running on an IA-32 processor can
address a linear address space of up to 4 GBytes (232 bytes) and a physical
address space of up to 64 GBytes (236 bytes). See Section 3.3.6, “Extended
Physical Addressing in Protected Mode,” for more information about addressing
an address space greater than 4 GBytes.
• Basic program execution registers — The eight general-purpose registers,
the six segment registers, the EFLAGS register, and the EIP (instruction pointer)
register comprise a basic execution environment in which to execute a set of
general-purpose instructions. These instructions perform basic integer arithmetic
on byte, word, and doubleword integers, handle program flow control, operate on
bit and byte strings, and address memory. See Section 3.4, “Basic Program
Execution Registers,” for more information about these registers.
• x87 FPU registers — The eight x87 FPU data registers, the x87 FPU control
register, the status register, the x87 FPU instruction pointer register, the x87 FPU
operand (data) pointer register, the x87 FPU tag register, and the x87 FPU opcode
register provide an execution environment for operating on single-precision,
double-precision, and double extended-precision floating-point values, word
integers, doubleword integers, quadword integers, and binary coded decimal
(BCD) values. See Section 8.1, “x87 FPU Execution Environment,” for more
information about these registers.
• MMX registers — The eight MMX registers support execution of single-
instruction, multiple-data (SIMD) operations on 64-bit packed byte, word, and
doubleword integers. See Section 9.2, “The MMX Technology Programming
Environment,” for more information about these registers.
• XMM registers — The eight XMM data registers and the MXCSR register support
execution of SIMD operations on 128-bit packed single-precision and double-
precision floating-point values and on 128-bit packed byte, word, doubleword,
and quadword integers. See Section 10.2, “SSE Programming Environment,” for
more information about these registers.

Vol. 1 3-3
BASIC EXECUTION ENVIRONMENT

Basic Program Execution Registers Address Space*

2^32 -1
Eight 32-bit General-Purpose Registers
Registers

Six 16-bit
Segment Registers
Registers

32-bits EFLAGS Register

32-bits EIP (Instruction Pointer Register)

FPU Registers

Eight 80-bit Floating-Point

Registers Data Registers 0

*The address space can be

flat or segmented. Using
16 bits Control Register
the physical address
16 bits Status Register extension mechanism, a
physical address space of
16 bits Tag Register
2^36 - 1 can be addressed.
Opcode Register (11-bits)
48 bits FPU Instruction Pointer Register
48 bits FPU Data (Operand) Pointer Register

MMX Registers

Eight 64-bit
Registers MMX Registers

XMM Registers

Eight 128-bit XMM Registers

Registers

32-bits MXCSR Register

Figure 3-1. IA-32 Basic Execution Environment for Non-64-bit Modes

3-4 Vol. 1
 BASIC EXECUTION ENVIRONMENT

• Stack — To support procedure or subroutine calls and the passing of parameters

between procedures or subroutines, a stack and stack management resources
are included in the execution environment. The stack (not shown in Figure 3-1) is
located in memory. See Section 6.2, “Stacks,” for more information about stack
structure.
In addition to the resources provided in the basic execution environment, the IA-32
architecture provides the following resources as part of its system-level architecture.
They provide extensive support for operating-system and system-development soft-
ware. Except for the I/O ports, the system resources are described in detail in the
Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volumes 3A & 3B.
• I/O ports — The IA-32 architecture supports a transfers of data to and from
input/output (I/O) ports. See Chapter 13, “Input/Output,” in this volume.
• Control registers — The five control registers (CR0 through CR4) determine the
operating mode of the processor and the characteristics of the currently
executing task. See Chapter 2, “System Architecture Overview,” in the Intel® 64
and IA-32 Architectures Software Developer’s Manual, Volume 3A.
• Memory management registers — The GDTR, IDTR, task register, and LDTR
specify the locations of data structures used in protected mode memory
management. See Chapter 2, “System Architecture Overview,” in the.
• Debug registers — The debug registers (DR0 through DR7) control and allow
monitoring of the processor’s debugging operations. See in the Intel® 64 and
IA-32 Architectures Software Developer’s Manual, Volume 3B.
• Memory type range registers (MTRRs) — The MTRRs are used to assign
memory types to regions of memory. See the sections on MTRRs in the .
• Machine specific registers (MSRs) — The processor provides a variety of
machine specific registers that are used to control and report on processor
performance. Virtually all MSRs handle system related functions and are not
accessible to an application program. One exception to this rule is the time-
stamp counter. The MSRs are described in Appendix B, “Model-Specific Registers
(MSRs),” of the Intel® 64 and IA-32 Architectures Software Developer’s Manual,
Volume 3B.
• Machine check registers — The machine check registers consist of a set of
control, status, and error-reporting MSRs that are used to detect and report on
hardware (machine) errors. See Chapter 14, “Machine-Check Architecture,” of
the Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volume 3A.
• Performance monitoring counters — The performance monitoring counters
allow processor performance events to be monitored. See Chapter 20, “Intro-
duction to Virtual-Machine Extensions,” in the Intel® 64 and IA-32 Architectures
Software Developer’s Manual, Volume 3B.
The remainder of this chapter describes the organization of memory and the address
space, the basic program execution registers, and addressing modes. Refer to the
following chapters in this volume for descriptions of the other program execution
resources shown in Figure 3-1:

Vol. 1 3-5
BASIC EXECUTION ENVIRONMENT

• x87 FPU registers — See Chapter 8, “Programming with the x87 FPU.”
• MMX Registers — See Chapter 9, “Programming with Intel® MMX™
Technology.”
• XMM registers — See Chapter 10, “Programming with Streaming SIMD
Extensions (SSE),” Chapter 11, “Programming with Streaming SIMD Extensions 2
(SSE2),” and Chapter 12, “Programming with SSE3, SSSE3, AND SSE4.”
• Stack implementation and procedure calls — See Chapter 6, “Procedure
Calls, Interrupts, and Exceptions.”

3.2.1 64-Bit Mode Execution Environment

The execution environment for 64-bit mode is similar to that described in Section
3.2. The following paragraphs describe the differences that apply.
• Address space — A task or program running in 64-bit mode on an IA-32
processor can address linear address space of up to 264 bytes (subject to the
canonical addressing requirement described in Section 3.3.7.1) and physical
address space of up to 240 bytes. Software can query CPUID for the physical
address size supported by a processor.
• Basic program execution registers — The number of general-purpose
registers (GPRs) available is 16. GPRs are 64-bits wide and they support
operations on byte, word, doubleword and quadword integers. Accessing byte
registers is done uniformly to the lowest 8 bits. The instruction pointer register
becomes 64 bits. The EFLAGS register is extended to 64 bits wide, and is referred
to as the RFLAGS register. The upper 32 bits of RFLAGS is reserved. The lower 32
bits of RFLAGS is the same as EFLAGS. See Figure 3-2.
• XMM registers — There are 16 XMM data registers for SIMD operations. See
Section 10.2, “SSE Programming Environment,” for more information about
these registers.
• Stack — The stack pointer size is 64 bits. Stack size is not controlled by a bit in
the SS descriptor (as it is in non-64-bit modes) nor can the pointer size be
overridden by an instruction prefix.
• Control registers — Control registers expand to 64 bits. A new control register
(the task priority register: CR8 or TPR) has been added. See Chapter 2, “Intel®
64 and IA-32 Architectures,” in the Intel® 64 and IA-32 Architectures Software
Developer’s Manual, Volume 3A.
• Debug registers — Debug registers expand to 64 bits. See Chapter 18,
“Debugging and Performance Monitoring,” in the Intel® 64 and IA-32 Architec-
tures Software Developer’s Manual, Volume 3B.
• Descriptor table registers — The global descriptor table register (GDTR) and
interrupt descriptor table register (IDTR) expand to 10 bytes so that they can

3-6 Vol. 1
 BASIC EXECUTION ENVIRONMENT

hold a full 64-bit base address. The local descriptor table register (LDTR) and the
task register (TR) also expand to hold a full 64-bit base address.

Basic Program Execution Registers Address Space

2^64 -1
Sixteen 64-bit General-Purpose Registers
Registers

Six 16-bit
Segment Registers
Registers

64-bits RFLAGS Register

64-bits RIP (Instruction Pointer Register)

FPU Registers

Eight 80-bit Floating-Point

Registers Data Registers 0

16 bits Control Register

16 bits Status Register
16 bits Tag Register
Opcode Register (11-bits)
64 bits FPU Instruction Pointer Register
64 bits FPU Data (Operand) Pointer Register

MMX Registers

Eight 64-bit MMX Registers

Registers

XMM Registers

Sixteen 128-bit
Registers XMM Registers

32-bits MXCSR Register

Figure 3-2. 64-Bit Mode Execution Environment

Vol. 1 3-7
BASIC EXECUTION ENVIRONMENT

3.3 MEMORY ORGANIZATION

The memory that the processor addresses on its bus is called physical memory.
Physical memory is organized as a sequence of 8-bit bytes. Each byte is assigned a
unique address, called a physical address. The physical address space ranges
from zero to a maximum of 236 − 1 (64 GBytes) if the processor does not support
Intel 64 architecture. Intel 64 architecture introduces a changes in physical and
linear address space; these are described in Section 3.3.3, Section 3.3.4, and
Section 3.3.7.
Virtually any operating system or executive designed to work with an IA-32 or Intel
64 processor will use the processor’s memory management facilities to access
memory. These facilities provide features such as segmentation and paging, which
allow memory to be managed efficiently and reliably. Memory management is
described in detail in Chapter 3, “Protected-Mode Memory Management,” in the
Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volume 3A. The
following paragraphs describe the basic methods of addressing memory when
memory management is used.

3.3.1 IA-32 Memory Models

When employing the processor’s memory management facilities, programs do not
directly address physical memory. Instead, they access memory using one of three
memory models: flat, segmented, or real address mode:
• Flat memory model — Memory appears to a program as a single, continuous
address space (Figure 3-3). This space is called a linear address space. Code,
data, and stacks are all contained in this address space. Linear address space is
byte addressable, with addresses running contiguously from 0 to 232 - 1 (if not in
64-bit mode). An address for any byte in linear address space is called a linear
address.
• Segmented memory model — Memory appears to a program as a group of
independent address spaces called segments. Code, data, and stacks are
typically contained in separate segments. To address a byte in a segment, a
program issues a logical address. This consists of a segment selector and an
offset (logical addresses are often referred to as far pointers). The segment
selector identifies the segment to be accessed and the offset identifies a byte in
the address space of the segment. Programs running on an IA-32 processor can
address up to 16,383 segments of different sizes and types, and each segment
can be as large as 232 bytes.
Internally, all the segments that are defined for a system are mapped into the
processor’s linear address space. To access a memory location, the processor
thus translates each logical address into a linear address. This translation is
transparent to the application program.
The primary reason for using segmented memory is to increase the reliability of
programs and systems. For example, placing a program’s stack in a separate

3-8 Vol. 1
 BASIC EXECUTION ENVIRONMENT

segment prevents the stack from growing into the code or data space and
overwriting instructions or data, respectively.
• Real-address mode memory model — This is the memory model for the Intel
8086 processor. It is supported to provide compatibility with existing programs
written to run on the Intel 8086 processor. The real-address mode uses a specific
implementation of segmented memory in which the linear address space for the
program and the operating system/executive consists of an array of segments of
up to 64 KBytes in size each. The maximum size of the linear address space in
real-address mode is 220 bytes.
See also: Chapter 15, “8086 Emulation,” Intel® 64 and IA-32 Architectures
Software Developer’s Manual, Volume 3A.

Flat Model
Linear Address

Linear
Address
Space*

Segmented Model

Segments

Offset (effective address) Linear

Address
Space*
Logical
Address Segment Selector

Real-Address Mode Model

Linear Address
Offset Space Divided
Into Equal
Sized Segments
Logical
Address Segment Selector

* The linear address space

can be paged when using the
flat or segmented model.

Figure 3-3. Three Memory Management Models

Vol. 1 3-9
BASIC EXECUTION ENVIRONMENT

3.3.2 Paging and Virtual Memory

With the flat or the segmented memory model, linear address space is mapped into
the processor’s physical address space either directly or through paging. When using
direct mapping (paging disabled), each linear address has a one-to-one correspon-
dence with a physical address. Linear addresses are sent out on the processor’s
address lines without translation.
When using the IA-32 architecture’s paging mechanism (paging enabled), linear
address space is divided into pages which are mapped to virtual memory. The pages
of virtual memory are then mapped as needed into physical memory. When an oper-
ating system or executive uses paging, the paging mechanism is transparent to an
application program. All that the application sees is linear address space.
In addition, IA-32 architecture’s paging mechanism includes extensions that
support:
• Page Address Extensions (PAE) to address physical address space greater than
4 GBytes.
• Page Size Extensions (PSE) to map linear address to physical address in
4-MBytes pages.
See also: Chapter 3, “Protected-Mode Memory Management,” in the Intel® 64 and
IA-32 Architectures Software Developer’s Manual, Volume 3A.

3.3.3 Memory Organization in 64-Bit Mode

Intel 64 architecture supports physical address space greater than 64 GBytes; the
actual physical address size of IA-32 processors is implementation specific. In 64-bit
mode, there is architectural support for 64-bit linear address space. However,
processors supporting Intel 64 architecture may implement less than 64-bits (see
Section 3.3.7.1). The linear address space is mapped into the processor physical
address space through the PAE paging mechanism.

3.3.4 Modes of Operation vs. Memory Model

When writing code for an IA-32 or Intel 64 processor, a programmer needs to know
the operating mode the processor is going to be in when executing the code and the
memory model being used. The relationship between operating modes and memory
models is as follows:
• Protected mode — When in protected mode, the processor can use any of the
memory models described in this section. (The real-addressing mode memory
model is ordinarily used only when the processor is in the virtual-8086 mode.)
The memory model used depends on the design of the operating system or
executive. When multitasking is implemented, individual tasks can use different
memory models.

3-10 Vol. 1
 BASIC EXECUTION ENVIRONMENT

• Real-address mode — When in real-address mode, the processor only supports

the real-address mode memory model.
• System management mode — When in SMM, the processor switches to a
separate address space, called the system management RAM (SMRAM). The
memory model used to address bytes in this address space is similar to the real-
address mode model. See Chapter 24, “System Management,” in the Intel® 64
and IA-32 Architectures Software Developer’s Manual, Volume 3B, for more
information on the memory model used in SMM.
• Compatibility mode — Software that needs to run in compatibility mode should
observe the same memory model as those targeted to run in 32-bit protected
mode. The effect of segmentation is the same as it is in 32-bit protected mode
semantics.
• 64-bit mode — Segmentation is generally (but not completely) disabled,
creating a flat 64-bit linear-address space. Specifically, the processor treats the
segment base of CS, DS, ES, and SS as zero in 64-bit mode (this makes a linear
address equal an effective address). Segmented and real address modes are not
available in 64-bit mode.

3.3.5 32-Bit and 16-Bit Address and Operand Sizes

IA-32 processors in protected mode can be configured for 32-bit or 16-bit address
and operand sizes. With 32-bit address and operand sizes, the maximum linear
address or segment offset is FFFFFFFFH (232-1); operand sizes are typically 8 bits or
32 bits. With 16-bit address and operand sizes, the maximum linear address or
segment offset is FFFFH (216-1); operand sizes are typically 8 bits or 16 bits.
When using 32-bit addressing, a logical address (or far pointer) consists of a 16-bit
segment selector and a 32-bit offset; when using 16-bit addressing, an address
consists of a 16-bit segment selector and a 16-bit offset.
Instruction prefixes allow temporary overrides of the default address and/or operand
sizes from within a program.
When operating in protected mode, the segment descriptor for the currently
executing code segment defines the default address and operand size. A segment
descriptor is a system data structure not normally visible to application code. Assem-
bler directives allow the default addressing and operand size to be chosen for a
program. The assembler and other tools then set up the segment descriptor for the
code segment appropriately.
When operating in real-address mode, the default addressing and operand size is 16
bits. An address-size override can be used in real-address mode to enable 32-bit
addressing. However, the maximum allowable 32-bit linear address is still 000FFFFFH
(220-1).

Vol. 1 3-11
BASIC EXECUTION ENVIRONMENT

3.3.6 Extended Physical Addressing in Protected Mode

Beginning with P6 family processors, the IA-32 architecture supports addressing of
up to 64 GBytes (236 bytes) of physical memory. A program or task could not
address locations in this address space directly. Instead, it addresses individual linear
address spaces of up to 4 GBytes that mapped to 64-GByte physical address space
through a virtual memory management mechanism. Using this mechanism, an oper-
ating system can enable a program to switch 4-GByte linear address spaces within
64-GByte physical address space.
The use of extended physical addressing requires the processor to operate in
protected mode and the operating system to provide a virtual memory management
system. See “36-Bit Physical Addressing Using the PAE Paging Mechanism” in
Chapter 3, “Protected-Mode Memory Management,” of the Intel® 64 and IA-32
Architectures Software Developer’s Manual, Volume 3A.

3.3.7 Address Calculations in 64-Bit Mode

In most cases, 64-bit mode uses flat address space for code, data, and stacks. In
64-bit mode (if there is no address-size override), the size of effective address calcu-
lations is 64 bits. An effective-address calculation uses a 64-bit base and index regis-
ters and sign-extend displacements to 64 bits.
In the flat address space of 64-bit mode, linear addresses are equal to effective
addresses because the base address is zero. In the event that FS or GS segments are
used with a non-zero base, this rule does not hold. In 64-bit mode, the effective
address components are added and the effective address is truncated (See for
example the instruction LEA) before adding the full 64-bit segment base. The base is
never truncated, regardless of addressing mode in 64-bit mode.
The instruction pointer is extended to 64 bits to support 64-bit code offsets. The
64-bit instruction pointer is called the RIP. Table 3-1 shows the relationship between
RIP, EIP, and IP.

Table 3-1. Instruction Pointer Sizes

Bits 63:32 Bits 31:16 Bits 15:0
16-bit instruction pointer Not Modified IP
32-bit instruction pointer Zero Extension EIP
64-bit instruction pointer RIP

Generally, displacements and immediates in 64-bit mode are not extended to 64 bits.
They are still limited to 32 bits and sign-extended during effective-address calcula-
tions. In 64-bit mode, however, support is provided for 64-bit displacement and
immediate forms of the MOV instruction.
All 16-bit and 32-bit address calculations are zero-extended in IA-32e mode to form
64-bit addresses. Address calculations are first truncated to the effective address

3-12 Vol. 1
 BASIC EXECUTION ENVIRONMENT

size of the current mode (64-bit mode or compatibility mode), as overridden by any
address-size prefix. The result is then zero-extended to the full 64-bit address width.
Because of this, 16-bit and 32-bit applications running in compatibility mode can
access only the low 4 GBytes of the 64-bit mode effective addresses. Likewise, a
32-bit address generated in 64-bit mode can access only the low 4 GBytes of the
64-bit mode effective addresses.

3.3.7.1 Canonical Addressing

In 64-bit mode, an address is considered to be in canonical form if address bits 63
through to the most-significant implemented bit by the microarchitecture are set to
either all ones or all zeros.
Intel 64 architecture defines a 64-bit linear address. Implementations can support
less. The first implementation of IA-32 processors with Intel 64 architecture supports
a 48-bit linear address. This means a canonical address must have bits 63 through 48
set to zeros or ones (depending on whether bit 47 is a zero or one).
Although implementations may not use all 64 bits of the linear address, they should
check bits 63 through the most-significant implemented bit to see if the address is in
canonical form. If a linear-memory reference is not in canonical form, the implemen-
tation should generate an exception. In most cases, a general-protection exception
(#GP) is generated. However, in the case of explicit or implied stack references, a
stack fault (#SS) is generated.
Instructions that have implied stack references, by default, use the SS segment
register. These include PUSH/POP-related instructions and instructions using
RSP/RBP as base registers. In these cases, the canonical fault is #SF.
If an instruction uses base registers RSP/RBP and uses a segment override prefix to
specify a non-SS segment, a canonical fault generates a #GP (instead of an #SF). In
64-bit mode, only FS and GS segment-overrides are applicable in this situation.
Other segment override prefixes (CS, DS, ES and SS) are ignored. Note that this also
means that an SS segment-override applied to a “non-stack” register reference is
ignored. Such a sequence still produces a #GP for a canonical fault (and not an #SF).

3.4 BASIC PROGRAM EXECUTION REGISTERS

IA-32 architecture provides 16 basic program execution registers for use in general
system and application programing (see Figure 3-4). These registers can be grouped
as follows:
• General-purpose registers. These eight registers are available for storing
operands and pointers.
• Segment registers. These registers hold up to six segment selectors.

Vol. 1 3-13
BASIC EXECUTION ENVIRONMENT

• EFLAGS (program status and control) register. The EFLAGS register report
on the status of the program being executed and allows limited (application-
program level) control of the processor.
• EIP (instruction pointer) register. The EIP register contains a 32-bit pointer
to the next instruction to be executed.

3.4.1 General-Purpose Registers

The 32-bit general-purpose registers EAX, EBX, ECX, EDX, ESI, EDI, EBP, and ESP
are provided for holding the following items:
• Operands for logical and arithmetic operations
• Operands for address calculations
• Memory pointers
Although all of these registers are available for general storage of operands, results,
and pointers, caution should be used when referencing the ESP register. The ESP
register holds the stack pointer and as a general rule should not be used for another
purpose.
Many instructions assign specific registers to hold operands. For example, string
instructions use the contents of the ECX, ESI, and EDI registers as operands. When
using a segmented memory model, some instructions assume that pointers in certain
registers are relative to specific segments. For instance, some instructions assume
that a pointer in the EBX register points to a memory location in the DS segment.

3-14 Vol. 1
 BASIC EXECUTION ENVIRONMENT

General-Purpose Registers
31 0
EAX
EBX
ECX
EDX
ESI
EDI
EBP
ESP

Segment Registers
15 0
CS
DS
SS
ES
FS
GS

Program Status and Control Register

31 0
EFLAGS

Instruction Pointer 0
31
EIP

Figure 3-4. General System and Application Programming Registers

The special uses of general-purpose registers by instructions are described in

Chapter 5, “Instruction Set Summary,” in this volume. See also: Chapter 3 and
Chapter 4 of Intel® 64 and IA-32 Architectures Software Developer’s Manual,
Volumes 2A & 2B. The following is a summary of special uses:
• EAX — Accumulator for operands and results data
• EBX — Pointer to data in the DS segment
• ECX — Counter for string and loop operations
• EDX — I/O pointer
• ESI — Pointer to data in the segment pointed to by the DS register; source
pointer for string operations
• EDI — Pointer to data (or destination) in the segment pointed to by the ES
register; destination pointer for string operations
• ESP — Stack pointer (in the SS segment)

Vol. 1 3-15
BASIC EXECUTION ENVIRONMENT

• EBP — Pointer to data on the stack (in the SS segment)

As shown in Figure 3-5, the lower 16 bits of the general-purpose registers map
directly to the register set found in the 8086 and Intel 286 processors and can be
referenced with the names AX, BX, CX, DX, BP, SI, DI, and SP. Each of the lower two
bytes of the EAX, EBX, ECX, and EDX registers can be referenced by the names AH,
BH, CH, and DH (high bytes) and AL, BL, CL, and DL (low bytes).

General-Purpose Registers
31 16 15 8 7 0 16-bit 32-bit
AH AL AX EAX
BH BL BX EBX
CH CL CX ECX
DH DL DX EDX
BP EBP
SI ESI
DI EDI
SP ESP

Figure 3-5. Alternate General-Purpose Register Names

3.4.1.1 General-Purpose Registers in 64-Bit Mode

In 64-bit mode, there are 16 general purpose registers and the default operand size
is 32 bits. However, general-purpose registers are able to work with either 32-bit or
64-bit operands. If a 32-bit operand size is specified: EAX, EBX, ECX, EDX, EDI, ESI,
EBP, ESP, R8D - R15D are available. If a 64-bit operand size is specified: RAX, RBX,
RCX, RDX, RDI, RSI, RBP, RSP, R8-R15 are available. R8D-R15D/R8-R15 represent
eight new general-purpose registers. All of these registers can be accessed at the
byte, word, dword, and qword level. REX prefixes are used to generate 64-bit
operand sizes or to reference registers R8-R15.
Registers only available in 64-bit mode (R8-R15 and XMM8-XMM15) are preserved
across transitions from 64-bit mode into compatibility mode then back into 64-bit
mode. However, values of R8-R15 and XMM8-XMM15 are undefined after transitions
from 64-bit mode through compatibility mode to legacy or real mode and then back
through compatibility mode to 64-bit mode.

3-16 Vol. 1
 BASIC EXECUTION ENVIRONMENT

Table 3-2. Addressable General Purpose Registers

Register Type Without REX With REX
Byte Registers AL, BL, CL, DL, AH, BH, CH, AL, BL, CL, DL, DIL, SIL, BPL, SPL,
DH R8L - R15L
Word Registers AX, BX, CX, DX, DI, SI, BP, SP AX, BX, CX, DX, DI, SI, BP, SP, R8W -
R15W
Doubleword Registers EAX, EBX, ECX, EDX, EDI, ESI, EAX, EBX, ECX, EDX, EDI, ESI, EBP,
EBP, ESP ESP, R8D - R15D
Quadword Registers N.A. RAX, RBX, RCX, RDX, RDI, RSI,
RBP, RSP, R8 - R15

In 64-bit mode, there are limitations on accessing byte registers. An instruction

cannot reference legacy high-bytes (for example: AH, BH, CH, DH) and one of the
new byte registers at the same time (for example: the low byte of the RAX register).
However, instructions may reference legacy low-bytes (for example: AL, BL, CL or
DL) and new byte registers at the same time (for example: the low byte of the R8
register, or RBP). The architecture enforces this limitation by changing high-byte
references (AH, BH, CH, DH) to low byte references (BPL, SPL, DIL, SIL: the low 8
bits for RBP, RSP, RDI and RSI) for instructions using a REX prefix.
When in 64-bit mode, operand size determines the number of valid bits in the desti-
nation general-purpose register:
• 64-bit operands generate a 64-bit result in the destination general-purpose
register.
• 32-bit operands generate a 32-bit result, zero-extended to a 64-bit result in the
destination general-purpose register.
• 8-bit and 16-bit operands generate an 8-bit or 16-bit result. The upper 56 bits or
48 bits (respectively) of the destination general-purpose register are not be
modified by the operation. If the result of an 8-bit or 16-bit operation is intended
for 64-bit address calculation, explicitly sign-extend the register to the full
64-bits.
Because the upper 32 bits of 64-bit general-purpose registers are undefined in 32-bit
modes, the upper 32 bits of any general-purpose register are not preserved when
switching from 64-bit mode to a 32-bit mode (to protected mode or compatibility
mode). Software must not depend on these bits to maintain a value after a 64-bit to
32-bit mode switch.

3.4.2 Segment Registers

The segment registers (CS, DS, SS, ES, FS, and GS) hold 16-bit segment selectors.
A segment selector is a special pointer that identifies a segment in memory. To
access a particular segment in memory, the segment selector for that segment must
be present in the appropriate segment register.

Vol. 1 3-17
BASIC EXECUTION ENVIRONMENT

When writing application code, programmers generally create segment selectors

with assembler directives and symbols. The assembler and other tools then create
the actual segment selector values associated with these directives and symbols. If
writing system code, programmers may need to create segment selectors directly.
See Chapter 3, “Protected-Mode Memory Management,” in the Intel® 64 and IA-32
Architectures Software Developer’s Manual, Volume 3A.
How segment registers are used depends on the type of memory management model
that the operating system or executive is using. When using the flat (unsegmented)
memory model, segment registers are loaded with segment selectors that point to
overlapping segments, each of which begins at address 0 of the linear address space
(see Figure 3-6). These overlapping segments then comprise the linear address
space for the program. Typically, two overlapping segments are defined: one for code
and another for data and stacks. The CS segment register points to the code
segment and all the other segment registers point to the data and stack segment.
When using the segmented memory model, each segment register is ordinarily
loaded with a different segment selector so that each segment register points to a
different segment within the linear address space (see Figure 3-7). At any time, a
program can thus access up to six segments in the linear address space. To access a
segment not pointed to by one of the segment registers, a program must first load
the segment selector for the segment to be accessed into a segment register.

Linear Address
Space for Program

Segment Registers Overlapping

Segments
CS of up to
4 GBytes
DS
Beginning at
SS
Address 0
ES
FS
GS
The segment selector in
each segment register
points to an overlapping
segment in the linear
address space.

Figure 3-6. Use of Segment Registers for Flat Memory Model

3-18 Vol. 1
 BASIC EXECUTION ENVIRONMENT

Code
Segment
Segment Registers
Data
CS Segment
DS Stack
SS Segment
ES All segments
FS are mapped
GS to the same
linear-address
space
Data
Segment
Data
Segment
Data
Segment

Figure 3-7. Use of Segment Registers in Segmented Memory Model

Each of the segment registers is associated with one of three types of storage: code,
data, or stack. For example, the CS register contains the segment selector for the
code segment, where the instructions being executed are stored. The processor
fetches instructions from the code segment, using a logical address that consists of
the segment selector in the CS register and the contents of the EIP register. The EIP
register contains the offset within the code segment of the next instruction to be
executed. The CS register cannot be loaded explicitly by an application program.
Instead, it is loaded implicitly by instructions or internal processor operations that
change program control (such as, procedure calls, interrupt handling, or task
switching).
The DS, ES, FS, and GS registers point to four data segments. The availability of
four data segments permits efficient and secure access to different types of data
structures. For example, four separate data segments might be created: one for the
data structures of the current module, another for the data exported from a higher-
level module, a third for a dynamically created data structure, and a fourth for data
shared with another program. To access additional data segments, the application
program must load segment selectors for these segments into the DS, ES, FS, and
GS registers, as needed.
The SS register contains the segment selector for the stack segment, where the
procedure stack is stored for the program, task, or handler currently being executed.
All stack operations use the SS register to find the stack segment. Unlike the CS
register, the SS register can be loaded explicitly, which permits application programs
to set up multiple stacks and switch among them.

Vol. 1 3-19
BASIC EXECUTION ENVIRONMENT

See Section 3.3, “Memory Organization,” for an overview of how the segment regis-
ters are used in real-address mode.
The four segment registers CS, DS, SS, and ES are the same as the segment regis-
ters found in the Intel 8086 and Intel 286 processors and the FS and GS registers
were introduced into the IA-32 Architecture with the Intel386™ family of processors.

3.4.2.1 Segment Registers in 64-Bit Mode

In 64-bit mode: CS, DS, ES, SS are treated as if each segment base is 0, regardless
of the value of the associated segment descriptor base. This creates a flat address
space for code, data, and stack. FS and GS are exceptions. Both segment registers
may be used as additional base registers in linear address calculations (in the
addressing of local data and certain operating system data structures).
Even though segmentation is generally disabled, segment register loads may cause
the processor to perform segment access assists. During these activities, enabled
processors will still perform most of the legacy checks on loaded values (even if the
checks are not applicable in 64-bit mode). Such checks are needed because a
segment register loaded in 64-bit mode may be used by an application running in
compatibility mode.
Limit checks for CS, DS, ES, SS, FS, and GS are disabled in 64-bit mode.

3.4.3 EFLAGS Register

The 32-bit EFLAGS register contains a group of status flags, a control flag, and a
group of system flags. Figure 3-8 defines the flags within this register. Following
initialization of the processor (either by asserting the RESET pin or the INIT pin), the
state of the EFLAGS register is 00000002H. Bits 1, 3, 5, 15, and 22 through 31 of this
register are reserved. Software should not use or depend on the states of any of
these bits.
Some of the flags in the EFLAGS register can be modified directly, using special-
purpose instructions (described in the following sections). There are no instructions
that allow the whole register to be examined or modified directly.
The following instructions can be used to move groups of flags to and from the proce-
dure stack or the EAX register: LAHF, SAHF, PUSHF, PUSHFD, POPF, and POPFD. After
the contents of the EFLAGS register have been transferred to the procedure stack or
EAX register, the flags can be examined and modified using the processor’s bit
manipulation instructions (BT, BTS, BTR, and BTC).
When suspending a task (using the processor’s multitasking facilities), the processor
automatically saves the state of the EFLAGS register in the task state segment (TSS)
for the task being suspended. When binding itself to a new task, the processor loads
the EFLAGS register with data from the new task’s TSS.
When a call is made to an interrupt or exception handler procedure, the processor
automatically saves the state of the EFLAGS registers on the procedure stack. When

3-20 Vol. 1
 BASIC EXECUTION ENVIRONMENT

an interrupt or exception is handled with a task switch, the state of the EFLAGS
register is saved in the TSS for the task being suspended.

31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0
I
V V O
I I I A V R 0 N O D I T S Z A P C
0 0 0 0 0 0 0 0 0 0 T F F F F F F 0 F 0 F 1 F
D C M F P
P F
L

X ID Flag (ID)
X Virtual Interrupt Pending (VIP)
X Virtual Interrupt Flag (VIF)
X Alignment Check (AC)
X Virtual-8086 Mode (VM)
X Resume Flag (RF)
X Nested Task (NT)
X I/O Privilege Level (IOPL)
S Overflow Flag (OF)
C Direction Flag (DF)
X Interrupt Enable Flag (IF)
X Trap Flag (TF)
S Sign Flag (SF)
S Zero Flag (ZF)
S Auxiliary Carry Flag (AF)
S Parity Flag (PF)
S Carry Flag (CF)

S Indicates a Status Flag

C Indicates a Control Flag
X Indicates a System Flag

Reserved bit positions. DO NOT USE.

Always set to values previously read.

Figure 3-8. EFLAGS Register

As the IA-32 Architecture has evolved, flags have been added to the EFLAGS register,
but the function and placement of existing flags have remained the same from one
family of the IA-32 processors to the next. As a result, code that accesses or modifies
these flags for one family of IA-32 processors works as expected when run on later
families of processors.

3.4.3.1 Status Flags

The status flags (bits 0, 2, 4, 6, 7, and 11) of the EFLAGS register indicate the results
of arithmetic instructions, such as the ADD, SUB, MUL, and DIV instructions. The
status flag functions are:
CF (bit 0) Carry flag — Set if an arithmetic operation generates a carry or
a borrow out of the most-significant bit of the result; cleared

Vol. 1 3-21
BASIC EXECUTION ENVIRONMENT

otherwise. This flag indicates an overflow condition for

unsigned-integer arithmetic. It is also used in multiple-precision
arithmetic.
PF (bit 2) Parity flag — Set if the least-significant byte of the result
contains an even number of 1 bits; cleared otherwise.
AF (bit 4) Adjust flag — Set if an arithmetic operation generates a carry
or a borrow out of bit 3 of the result; cleared otherwise. This flag
is used in binary-coded decimal (BCD) arithmetic.
ZF (bit 6) Zero flag — Set if the result is zero; cleared otherwise.
SF (bit 7) Sign flag — Set equal to the most-significant bit of the result,
which is the sign bit of a signed integer. (0 indicates a positive
value and 1 indicates a negative value.)
OF (bit 11) Overflow flag — Set if the integer result is too large a positive
number or too small a negative number (excluding the sign-bit)
to fit in the destination operand; cleared otherwise. This flag
indicates an overflow condition for signed-integer (two’s
complement) arithmetic.
Of these status flags, only the CF flag can be modified directly, using the STC, CLC,
and CMC instructions. Also the bit instructions (BT, BTS, BTR, and BTC) copy a spec-
ified bit into the CF flag.
The status flags allow a single arithmetic operation to produce results for three
different data types: unsigned integers, signed integers, and BCD integers. If the
result of an arithmetic operation is treated as an unsigned integer, the CF flag indi-
cates an out-of-range condition (carry or a borrow); if treated as a signed integer
(two’s complement number), the OF flag indicates a carry or borrow; and if treated
as a BCD digit, the AF flag indicates a carry or borrow. The SF flag indicates the sign
of a signed integer. The ZF flag indicates either a signed- or an unsigned-integer
zero.
When performing multiple-precision arithmetic on integers, the CF flag is used in
conjunction with the add with carry (ADC) and subtract with borrow (SBB) instruc-
tions to propagate a carry or borrow from one computation to the next.
The condition instructions Jcc (jump on condition code cc), SETcc (byte set on condi-
tion code cc), LOOPcc, and CMOVcc (conditional move) use one or more of the status
flags as condition codes and test them for branch, set-byte, or end-loop conditions.

3.4.3.2 DF Flag
The direction flag (DF, located in bit 10 of the EFLAGS register) controls string
instructions (MOVS, CMPS, SCAS, LODS, and STOS). Setting the DF flag causes the
string instructions to auto-decrement (to process strings from high addresses to low
addresses). Clearing the DF flag causes the string instructions to auto-increment
(process strings from low addresses to high addresses).
The STD and CLD instructions set and clear the DF flag, respectively.

3-22 Vol. 1
 BASIC EXECUTION ENVIRONMENT

3.4.3.3 System Flags and IOPL Field

The system flags and IOPL field in the EFLAGS register control operating-system or
executive operations. They should not be modified by application programs.
The functions of the system flags are as follows:
TF (bit 8) Trap flag — Set to enable single-step mode for debugging;
clear to disable single-step mode.
IF (bit 9) Interrupt enable flag — Controls the response of the
processor to maskable interrupt requests. Set to respond to
maskable interrupts; cleared to inhibit maskable interrupts.
IOPL (bits 12 and 13)
I/O privilege level field — Indicates the I/O privilege level of
the currently running program or task. The current privilege
level (CPL) of the currently running program or task must be
less than or equal to the I/O privilege level to access the I/O
address space. This field can only be modified by the POPF and
IRET instructions when operating at a CPL of 0.
NT (bit 14) Nested task flag — Controls the chaining of interrupted and
called tasks. Set when the current task is linked to the previ-
ously executed task; cleared when the current task is not linked
to another task.
RF (bit 16) Resume flag — Controls the processor’s response to debug
exceptions.
VM (bit 17) Virtual-8086 mode flag — Set to enable virtual-8086 mode;
clear to return to protected mode without virtual-8086 mode
semantics.
AC (bit 18) Alignment check flag — Set this flag and the AM bit in the CR0
register to enable alignment checking of memory references;
clear the AC flag and/or the AM bit to disable alignment
checking.
VIF (bit 19) Virtual interrupt flag — Virtual image of the IF flag. Used in
conjunction with the VIP flag. (To use this flag and the VIP flag
the virtual mode extensions are enabled by setting the VME flag
in control register CR4.)
VIP (bit 20) Virtual interrupt pending flag — Set to indicate that an inter-
rupt is pending; clear when no interrupt is pending. (Software
sets and clears this flag; the processor only reads it.) Used in
conjunction with the VIF flag.
ID (bit 21) Identification flag — The ability of a program to set or clear
this flag indicates support for the CPUID instruction.
For a detailed description of these flags: see Chapter 3, “Protected-Mode Memory
Management,” in the Intel® 64 and IA-32 Architectures Software Developer’s
Manual, Volume 3A.

Vol. 1 3-23
BASIC EXECUTION ENVIRONMENT

3.4.3.4 RFLAGS Register in 64-Bit Mode

In 64-bit mode, EFLAGS is extended to 64 bits and called RFLAGS. The upper 32 bits
of RFLAGS register is reserved. The lower 32 bits of RFLAGS is the same as EFLAGS.

3.5 INSTRUCTION POINTER

The instruction pointer (EIP) register contains the offset in the current code segment
for the next instruction to be executed. It is advanced from one instruction boundary
to the next in straight-line code or it is moved ahead or backwards by a number of
instructions when executing JMP, Jcc, CALL, RET, and IRET instructions.
The EIP register cannot be accessed directly by software; it is controlled implicitly by
control-transfer instructions (such as JMP, Jcc, CALL, and RET), interrupts, and
exceptions. The only way to read the EIP register is to execute a CALL instruction and
then read the value of the return instruction pointer from the procedure stack. The
EIP register can be loaded indirectly by modifying the value of a return instruction
pointer on the procedure stack and executing a return instruction (RET or IRET). See
Section 6.2.4.2, “Return Instruction Pointer.”
All IA-32 processors prefetch instructions. Because of instruction prefetching, an
instruction address read from the bus during an instruction load does not match the
value in the EIP register. Even though different processor generations use different
prefetching mechanisms, the function of the EIP register to direct program flow
remains fully compatible with all software written to run on IA-32 processors.

3.5.1 Instruction Pointer in 64-Bit Mode

In 64-bit mode, the RIP register becomes the instruction pointer. This register holds
the 64-bit offset of the next instruction to be executed. 64-bit mode also supports a
technique called RIP-relative addressing. Using this technique, the effective address
is determined by adding a displacement to the RIP of the next instruction.

3.6 OPERAND-SIZE AND ADDRESS-SIZE ATTRIBUTES

When the processor is executing in protected mode, every code segment has a
default operand-size attribute and address-size attribute. These attributes are
selected with the D (default size) flag in the segment descriptor for the code segment
(see Chapter 3, “Protected-Mode Memory Management,” in the Intel® 64 and IA-32
Architectures Software Developer’s Manual, Volume 3A). When the D flag is set, the
32-bit operand-size and address-size attributes are selected; when the flag is clear,
the 16-bit size attributes are selected. When the processor is executing in real-
address mode, virtual-8086 mode, or SMM, the default operand-size and address-
size attributes are always 16 bits.

3-24 Vol. 1
 BASIC EXECUTION ENVIRONMENT

The operand-size attribute selects the size of operands. When the 16-bit operand-
size attribute is in force, operands can generally be either 8 bits or 16 bits, and when
the 32-bit operand-size attribute is in force, operands can generally be 8 bits or 32
bits.
The address-size attribute selects the sizes of addresses used to address memory:
16 bits or 32 bits. When the 16-bit address-size attribute is in force, segment offsets
and displacements are 16 bits. This restriction limits the size of a segment to 64
KBytes. When the 32-bit address-size attribute is in force, segment offsets and
displacements are 32 bits, allowing up to 4 GBytes to be addressed.
The default operand-size attribute and/or address-size attribute can be overridden
for a particular instruction by adding an operand-size and/or address-size prefix to
an instruction. See Chapter 2, “Instruction Format,” in the Intel® 64 and IA-32 Archi-
tectures Software Developer’s Manual, Volume 3A. The effect of this prefix applies
only to the targeted instruction.
Table 3-4 shows effective operand size and address size (when executing in
protected mode or compatibility mode) depending on the settings of the D flag and
the operand-size and address-size prefixes.

Table 3-3. Effective Operand- and Address-Size Attributes

D Flag in Code Segment Descriptor 0 0 0 0 1 1 1 1
Operand-Size Prefix 66H N N Y Y N N Y Y
Address-Size Prefix 67H N Y N Y N Y N Y
Effective Operand Size 16 16 32 32 32 32 16 16
Effective Address Size 16 32 16 32 32 16 32 16
NOTES:
Y: Yes - this instruction prefix is present.
N: No - this instruction prefix is not present.

3.6.1 Operand Size and Address Size in 64-Bit Mode

In 64-bit mode, the default address size is 64 bits and the default operand size is 32
bits. Defaults can be overridden using prefixes. Address-size and operand-size
prefixes allow mixing of 32/64-bit data and 32/64-bit addresses on an instruction-
by-instruction basis. Table 3-4 shows valid combinations of the 66H instruction prefix
and the REX.W prefix that may be used to specify operand-size overrides in 64-bit
mode. Note that 16-bit addresses are not supported in 64-bit mode.
REX prefixes consist of 4-bit fields that form 16 different values. The W-bit field in the
REX prefixes is referred to as REX.W. If the REX.W field is properly set, the prefix
specifies an operand size override to 64 bits. Note that software can still use the
operand-size 66H prefix to toggle to a 16-bit operand size. However, setting REX.W
takes precedence over the operand-size prefix (66H) when both are used.

Vol. 1 3-25
BASIC EXECUTION ENVIRONMENT

In the case of SSE/SSE2/SSE3/SSSE3 SIMD instructions: the 66H, F2H, and F3H
prefixes are mandatory for opcode extensions. In such a case, there is no interaction
between a valid REX.W prefix and a 66H opcode extension prefix.
See Chapter 2, “Instruction Format,” in the Intel® 64 and IA-32 Architectures Soft-
ware Developer’s Manual, Volume 3A.

Table 3-4. Effective Operand- and Address-Size Attributes in 64-Bit Mode

L Flag in Code Segment
Descriptor 1 1 1 1 1 1 1 1
REX.W Prefix 0 0 0 0 1 1 1 1
Operand-Size Prefix 66H N N Y Y N N Y Y
Address-Size Prefix 67H N Y N Y N Y N Y
Effective Operand Size 32 32 16 16 64 64 64 64
Effective Address Size 64 32 64 32 64 32 64 32
NOTES:
Y: Yes - this instruction prefix is present.
N: No - this instruction prefix is not present.

3.7 OPERAND ADDRESSING

IA-32 machine-instructions act on zero or more operands. Some operands are spec-
ified explicitly and others are implicit. The data for a source operand can be located
in:
• the instruction itself (an immediate operand)
• a register
• a memory location
• an I/O port
When an instruction returns data to a destination operand, it can be returned to:
• a register
• a memory location
• an I/O port

3-26 Vol. 1
 BASIC EXECUTION ENVIRONMENT

3.7.1 Immediate Operands

Some instructions use data encoded in the instruction itself as a source operand.
These operands are called immediate operands (or simply immediates). For
example, the following ADD instruction adds an immediate value of 14 to the
contents of the EAX register:

ADD EAX, 14
All arithmetic instructions (except the DIV and IDIV instructions) allow the source
operand to be an immediate value. The maximum value allowed for an immediate
operand varies among instructions, but can never be greater than the maximum
value of an unsigned doubleword integer (232).

3.7.2 Register Operands

Source and destination operands can be any of the following registers, depending on
the instruction being executed:
• 32-bit general-purpose registers (EAX, EBX, ECX, EDX, ESI, EDI, ESP, or EBP)
• 16-bit general-purpose registers (AX, BX, CX, DX, SI, DI, SP, or BP)
• 8-bit general-purpose registers (AH, BH, CH, DH, AL, BL, CL, or DL)
• segment registers (CS, DS, SS, ES, FS, and GS)
• EFLAGS register
• x87 FPU registers (ST0 through ST7, status word, control word, tag word, data
operand pointer, and instruction pointer)
• MMX registers (MM0 through MM7)
• XMM registers (XMM0 through XMM7) and the MXCSR register
• control registers (CR0, CR2, CR3, and CR4) and system table pointer registers
(GDTR, LDTR, IDTR, and task register)
• debug registers (DR0, DR1, DR2, DR3, DR6, and DR7)
• MSR registers
Some instructions (such as the DIV and MUL instructions) use quadword operands
contained in a pair of 32-bit registers. Register pairs are represented with a colon
separating them. For example, in the register pair EDX:EAX, EDX contains the high
order bits and EAX contains the low order bits of a quadword operand.
Several instructions (such as the PUSHFD and POPFD instructions) are provided to
load and store the contents of the EFLAGS register or to set or clear individual flags
in this register. Other instructions (such as the Jcc instructions) use the state of the
status flags in the EFLAGS register as condition codes for branching or other decision
making operations.
The processor contains a selection of system registers that are used to control
memory management, interrupt and exception handling, task management,

Vol. 1 3-27
BASIC EXECUTION ENVIRONMENT

processor management, and debugging activities. Some of these system registers

are accessible by an application program, the operating system, or the executive
through a set of system instructions. When accessing a system register with a
system instruction, the register is generally an implied operand of the instruction.

3.7.2.1 Register Operands in 64-Bit Mode

Register operands in 64-bit mode can be any of the following:
• 64-bit general-purpose registers (RAX, RBX, RCX, RDX, RSI, RDI, RSP, RBP, or
R8-R15)
• 32-bit general-purpose registers (EAX, EBX, ECX, EDX, ESI, EDI, ESP, EBP, or
R8D-R15D)
• 16-bit general-purpose registers (AX, BX, CX, DX, SI, DI, SP, BP, or R8W-R15W)
• 8-bit general-purpose registers: AL, BL, CL, DL, SIL, DIL, SPL, BPL, and R8L-
R15L are available using REX prefixes; AL, BL, CL, DL, AH, BH, CH, DH are
available without using REX prefixes.
• Segment registers (CS, DS, SS, ES, FS, and GS)
• RFLAGS register
• x87 FPU registers (ST0 through ST7, status word, control word, tag word, data
operand pointer, and instruction pointer)
• MMX registers (MM0 through MM7)
• XMM registers (XMM0 through XMM15) and the MXCSR register
• Control registers (CR0, CR2, CR3, CR4, and CR8) and system table pointer
registers (GDTR, LDTR, IDTR, and task register)
• Debug registers (DR0, DR1, DR2, DR3, DR6, and DR7)
• MSR registers
• RDX:RAX register pair representing a 128-bit operand

3.7.3 Memory Operands

Source and destination operands in memory are referenced by means of a segment
selector and an offset (see Figure 3-9). Segment selectors specify the segment
containing the operand. Offsets specify the linear or effective address of the operand.
Offsets can be 32 bits (represented by the notation m16:32) or 16 bits (represented
by the notation m16:16).

15 0 31 0
Segment Offset (or Linear Address)
Selector

Figure 3-9. Memory Operand Address

3-28 Vol. 1
 BASIC EXECUTION ENVIRONMENT

3.7.3.1 Memory Operands in 64-Bit Mode

In 64-bit mode, a memory operand can be referenced by a segment selector and an
offset. The offset can be 16 bits, 32 bits or 64 bits (see Figure 3-10).

15 0 63 0
Segment Offset (or Linear Address)
Selector

Figure 3-10. Memory Operand Address in 64-Bit Mode

3.7.4 Specifying a Segment Selector

The segment selector can be specified either implicitly or explicitly. The most
common method of specifying a segment selector is to load it in a segment register
and then allow the processor to select the register implicitly, depending on the type
of operation being performed. The processor automatically chooses a segment
according to the rules given in Table 3-5.
When storing data in memory or loading data from memory, the DS segment default
can be overridden to allow other segments to be accessed. Within an assembler, the
segment override is generally handled with a colon “:” operator. For example, the
following MOV instruction moves a value from register EAX into the segment pointed
to by the ES register. The offset into the segment is contained in the EBX register:

MOV ES:[EBX], EAX;

Table 3-5. Default Segment Selection Rules

Reference Register Segment
Type Used Used Default Selection Rule
Instructions CS Code Segment All instruction fetches.
Stack SS Stack Segment All stack pushes and pops.
Any memory reference which uses the ESP or EBP
register as a base register.
Local Data DS Data Segment All data references, except when relative to stack or
string destination.
Destination ES Data Segment Destination of string instructions.
Strings pointed to with
the ES register

At the machine level, a segment override is specified with a segment-override prefix,

which is a byte placed at the beginning of an instruction. The following default
segment selections cannot be overridden:
• Instruction fetches must be made from the code segment.

Vol. 1 3-29
BASIC EXECUTION ENVIRONMENT

• Destination strings in string instructions must be stored in the data segment

pointed to by the ES register.
• Push and pop operations must always reference the SS segment.
Some instructions require a segment selector to be specified explicitly. In these
cases, the 16-bit segment selector can be located in a memory location or in a 16-bit
register. For example, the following MOV instruction moves a segment selector
located in register BX into segment register DS:

MOV DS, BX
Segment selectors can also be specified explicitly as part of a 48-bit far pointer in
memory. Here, the first doubleword in memory contains the offset and the next word
contains the segment selector.

3.7.4.1 Segmentation in 64-Bit Mode

In IA-32e mode, the effects of segmentation depend on whether the processor is
running in compatibility mode or 64-bit mode. In compatibility mode, segmentation
functions just as it does in legacy IA-32 mode, using the 16-bit or 32-bit protected
mode semantics described above.
In 64-bit mode, segmentation is generally (but not completely) disabled, creating a
flat 64-bit linear-address space. The processor treats the segment base of CS, DS,
ES, SS as zero, creating a linear address that is equal to the effective address. The
exceptions are the FS and GS segments, whose segment registers (which hold the
segment base) can be used as additional base registers in some linear address calcu-
lations.

3.7.5 Specifying an Offset

The offset part of a memory address can be specified directly as a static value (called
a displacement) or through an address computation made up of one or more of the
following components:
• Displacement — An 8-, 16-, or 32-bit value.
• Base — The value in a general-purpose register.
• Index — The value in a general-purpose register.
• Scale factor — A value of 2, 4, or 8 that is multiplied by the index value.
The offset which results from adding these components is called an effective
address. Each of these components can have either a positive or negative (2s
complement) value, with the exception of the scaling factor. Figure 3-11 shows all
the possible ways that these components can be combined to create an effective
address in the selected segment.

3-30 Vol. 1
 BASIC EXECUTION ENVIRONMENT

Base Index Scale Displacement

EAX
EAX None
EBX 1
EBX
ECX
ECX 2 8-bit
EDX
+ EDX +
ESP * 16-bit
EBP 4
EBP
ESI
ESI 8 32-bit
EDI
EDI

Offset = Base + (Index * Scale) + Displacement

Figure 3-11. Offset (or Effective Address) Computation

The uses of general-purpose registers as base or index components are restricted in

the following manner:
• The ESP register cannot be used as an index register.
• When the ESP or EBP register is used as the base, the SS segment is the default
segment. In all other cases, the DS segment is the default segment.
The base, index, and displacement components can be used in any combination, and
any of these components can be NULL. A scale factor may be used only when an
index also is used. Each possible combination is useful for data structures commonly
used by programmers in high-level languages and assembly language.
The following addressing modes suggest uses for common combinations of address
components.
• Displacement ⎯ A displacement alone represents a direct (uncomputed) offset
to the operand. Because the displacement is encoded in the instruction, this form
of an address is sometimes called an absolute or static address. It is commonly
used to access a statically allocated scalar operand.
• Base ⎯ A base alone represents an indirect offset to the operand. Since the
value in the base register can change, it can be used for dynamic storage of
variables and data structures.
• Base + Displacement ⎯ A base register and a displacement can be used
together for two distinct purposes:
— As an index into an array when the element size is not 2, 4, or 8 bytes—The
displacement component encodes the static offset to the beginning of the
array. The base register holds the results of a calculation to determine the
offset to a specific element within the array.
— To access a field of a record: the base register holds the address of the
beginning of the record, while the displacement is a static offset to the field.
An important special case of this combination is access to parameters in a
procedure activation record. A procedure activation record is the stack frame

Vol. 1 3-31
BASIC EXECUTION ENVIRONMENT

created when a procedure is entered. Here, the EBP register is the best choice for
the base register, because it automatically selects the stack segment. This is a
compact encoding for this common function.
• (Index ∗ Scale) + Displacement ⎯ This address mode offers an efficient way
to index into a static array when the element size is 2, 4, or 8 bytes. The
displacement locates the beginning of the array, the index register holds the
subscript of the desired array element, and the processor automatically converts
the subscript into an index by applying the scaling factor.
• Base + Index + Displacement ⎯ Using two registers together supports either
a two-dimensional array (the displacement holds the address of the beginning of
the array) or one of several instances of an array of records (the displacement is
an offset to a field within the record).
• Base + (Index ∗ Scale) + Displacement ⎯ Using all the addressing
components together allows efficient indexing of a two-dimensional array when
the elements of the array are 2, 4, or 8 bytes in size.

3.7.5.1 Specifying an Offset in 64-Bit Mode

The offset part of a memory address in 64-bit mode can be specified directly as a
static value or through an address computation made up of one or more of the
following components:
• Displacement — An 8-bit, 16-bit, or 32-bit value.
• Base — The value in a 32-bit (or 64-bit if REX.W is set) general-purpose register.
• Index — The value in a 32-bit (or 64-bit if REX.W is set) general-purpose
register.
• Scale factor — A value of 2, 4, or 8 that is multiplied by the index value.
The base and index value can be specified in one of sixteen available general-purpose
registers in most cases. See Chapter 2, “Instruction Format,” in the Intel® 64 and
IA-32 Architectures Software Developer’s Manual, Volume 3A.
The following unique combination of address components is also available.
• RIP + Displacement ⎯ In 64-bit mode, RIP-relative addressing uses a signed
32-bit displacement to calculate the effective address of the next instruction by
sign-extend the 32-bit value and add to the 64-bit value in RIP.

3.7.6 Assembler and Compiler Addressing Modes

At the machine-code level, the selected combination of displacement, base register,
index register, and scale factor is encoded in an instruction. All assemblers permit a
programmer to use any of the allowable combinations of these addressing compo-
nents to address operands. High-level language compilers will select an appropriate
combination of these components based on the language construct a programmer
defines.

3-32 Vol. 1
 BASIC EXECUTION ENVIRONMENT

3.7.7 I/O Port Addressing

The processor supports an I/O address space that contains up to 65,536 8-bit I/O
ports. Ports that are 16-bit and 32-bit may also be defined in the I/O address space.
An I/O port can be addressed with either an immediate operand or a value in the DX
register. See Chapter 13, “Input/Output,” for more information about I/O port
addressing.

Vol. 1 3-33
BASIC EXECUTION ENVIRONMENT

3-34 Vol. 1
 CHAPTER 4
DATA TYPES

This chapter introduces data types defined for the Intel 64 and IA-32 architectures.
A section at the end of this chapter describes the real-number and floating-point
concepts used in x87 FPU, SSE, SSE2, SSE3 and SSSE3 extensions.

4.1 FUNDAMENTAL DATA TYPES

The fundamental data types are bytes, words, doublewords, quadwords, and double
quadwords (see Figure 4-1). A byte is eight bits, a word is 2 bytes (16 bits), a
doubleword is 4 bytes (32 bits), a quadword is 8 bytes (64 bits), and a double quad-
word is 16 bytes (128 bits). A subset of the IA-32 architecture instructions operates
on these fundamental data types without any additional operand typing.

7 0
Byte

N
15 8 7 0
High Low
Byte Byte Word
N+1 N
31 16 15 0
High Word Low Word Doubleword
N+2 N
63 32 31 0
High Doubleword Low Doubleword Quadword
N+4 N
127 64 63 0
High Quadword Low Quadword Double
Quadword
N+8 N

Figure 4-1. Fundamental Data Types

The quadword data type was introduced into the IA-32 architecture in the Intel486
processor; the double quadword data type was introduced in the Pentium III
processor with the SSE extensions.
Figure 4-2 shows the byte order of each of the fundamental data types when refer-
enced as operands in memory. The low byte (bits 0 through 7) of each data type
occupies the lowest address in memory and that address is also the address of the
operand.

Vol. 1 4-1
DATA TYPES

4EH FH
12H EH
7AH DH

Word at Address BH FEH CH

Doubleword at Address AH
Contains FE06H 06H BH Contains 7AFE0636H

36H AH
Byte at Address 9H
1FH 9H
Contains 1FH Quadword at Address 6H
A4H 8H Contains
7AFE06361FA4230BH
Word at Address 6H 23H 7H
Contains 230BH 0BH 6H

45H 5H
67H 4H
Word at Address 2H
Contains 74CBH 74H 3H
CBH 2H Double quadword at Address 0H
Word at Address 1H Contains
Contains CB31H 31H 1H 4E127AFE06361FA4230B456774CB3112
12H 0H

Figure 4-2. Bytes, Words, Doublewords, Quadwords, and Double Quadwords in

Memory

4.1.1 Alignment of Words, Doublewords, Quadwords, and Double

Quadwords
Words, doublewords, and quadwords do not need to be aligned in memory on natural
boundaries. The natural boundaries for words, double words, and quadwords are
even-numbered addresses, addresses evenly divisible by four, and addresses evenly
divisible by eight, respectively. However, to improve the performance of programs,
data structures (especially stacks) should be aligned on natural boundaries when-
ever possible. The reason for this is that the processor requires two memory
accesses to make an unaligned memory access; aligned accesses require only one
memory access. A word or doubleword operand that crosses a 4-byte boundary or a
quadword operand that crosses an 8-byte boundary is considered unaligned and
requires two separate memory bus cycles for access.
Some instructions that operate on double quadwords require memory operands to be
aligned on a natural boundary. These instructions generate a general-protection
exception (#GP) if an unaligned operand is specified. A natural boundary for a double
quadword is any address evenly divisible by 16. Other instructions that operate on
double quadwords permit unaligned access (without generating a general-protection

4-2 Vol. 1
 DATA TYPES

exception). However, additional memory bus cycles are required to access unaligned
data from memory.

4.2 NUMERIC DATA TYPES

Although bytes, words, and doublewords are fundamental data types, some instruc-
tions support additional interpretations of these data types to allow operations to be
performed on numeric data types (signed and unsigned integers, and floating-point
numbers). See Figure 4-3.

Vol. 1 4-3
DATA TYPES

Byte Unsigned Integer

7 0

Word Unsigned Integer

15 0

Doubleword Unsigned Integer

31 0

Quadword Unsigned Integer

63 0
Sign
Byte Signed Integer
76 0

Sign
Word Signed Integer
15 14 0
Sign
Doubleword Signed Integer
31 30 0

Sign
Quadword Signed Integer
63 62 0
Sign
Single Precision
Floating Point
31 30 23 22 0

Sign
Double Precision
Floating Point
63 62 52 51 0

Sign Integer Bit

Double Extended Precision
Floating Point
79 78 64 63 62 0

Figure 4-3. Numeric Data Types

4.2.1 Integers
The Intel 64 and IA-32 architectures define two types of integers: unsigned and
signed. Unsigned integers are ordinary binary values ranging from 0 to the maximum
positive number that can be encoded in the selected operand size. Signed integers

4-4 Vol. 1
 DATA TYPES

are two’s complement binary values that can be used to represent both positive and
negative integer values.
Some integer instructions (such as the ADD, SUB, PADDB, and PSUBB instructions)
operate on either unsigned or signed integer operands. Other integer instructions
(such as IMUL, MUL, IDIV, DIV, FIADD, and FISUB) operate on only one integer type.
The following sections describe the encodings and ranges of the two types of
integers.

4.2.1.1 Unsigned Integers

Unsigned integers are unsigned binary numbers contained in a byte, word, double-
word, and quadword. Their values range from 0 to 255 for an unsigned byte integer,
from 0 to 65,535 for an unsigned word integer, from 0 to 232 – 1 for an unsigned
doubleword integer, and from 0 to 264 – 1 for an unsigned quadword integer.
Unsigned integers are sometimes referred to as ordinals.

4.2.1.2 Signed Integers

Signed integers are signed binary numbers held in a byte, word, doubleword, or
quadword. All operations on signed integers assume a two's complement representa-
tion. The sign bit is located in bit 7 in a byte integer, bit 15 in a word integer, bit 31 in
a doubleword integer, and bit 63 in a quadword integer (see the signed integer
encodings in Table 4-1).

Vol. 1 4-5
DATA TYPES

Table 4-1. Signed Integer Encodings

Class Two’s Complement Encoding
Sign
Positive Largest 0 11..11
. .
. .
Smallest 0 00..01
Zero 0 00..00
Negative Smallest 1 11..11
. .
. .
Largest 1 00..00
Integer indefinite 1 00..00
Signed Byte Integer: ← 7 bits →
Signed Word Integer: ← 15 bits →
Signed Doubleword Integer: ← 31 bits →
Signed Quadword Integer: ← 63 bits →
The sign bit is set for negative integers and cleared for positive integers and zero.
Integer values range from –128 to +127 for a byte integer, from –32,768 to +32,767
for a word integer, from –231 to +231 – 1 for a doubleword integer, and from –263 to
+263 – 1 for a quadword integer.
When storing integer values in memory, word integers are stored in 2 consecutive
bytes; doubleword integers are stored in 4 consecutive bytes; and quadword inte-
gers are stored in 8 consecutive bytes.
The integer indefinite is a special value that is sometimes returned by the x87 FPU
when operating on integer values. For more information, see Section 8.2.1, “Indefi-
nites.”

4.2.2 Floating-Point Data Types

The IA-32 architecture defines and operates on three floating-point data types:
single-precision floating-point, double-precision floating-point, and double-extended
precision floating-point (see Figure 4-3). The data formats for these data types
correspond directly to formats specified in the IEEE Standard 754 for Binary Floating-
Point Arithmetic.

4-6 Vol. 1
 DATA TYPES

Table 4-2 gives the length, precision, and approximate normalized range that can be
represented by each of these data types. Denormal values are also supported in each
of these types.

Table 4-2. Length, Precision, and Range of Floating-Point Data Types

Data Type Length Precision Approximate Normalized Range
(Bits) Binary Decimal
Single Precision 32 24 2–126 to 2127 1.18 × 10–38 to 3.40 × 1038
Double Precision 64 53 2–1022 to 21023 2.23 × 10–308 to 1.79 × 10308
Double Extended 80 64 2–16382 to 216383 3.37 × 10–4932 to 1.18 × 104932
Precision

NOTE
Section 4.8, “Real Numbers and Floating-Point Formats,” gives an
overview of the IEEE Standard 754 floating-point formats and defines
the terms integer bit, QNaN, SNaN, and denormal value.

Table 4-3 shows the floating-point encodings for zeros, denormalized finite numbers,
normalized finite numbers, infinites, and NaNs for each of the three floating-point
data types. It also gives the format for the QNaN floating-point indefinite value. (See
Section 4.8.3.7, “QNaN Floating-Point Indefinite,” for a discussion of the use of the
QNaN floating-point indefinite value.)
For the single-precision and double-precision formats, only the fraction part of the
significand is encoded. The integer is assumed to be 1 for all numbers except 0 and
denormalized finite numbers. For the double extended-precision format, the integer
is contained in bit 63, and the most-significant fraction bit is bit 62. Here, the integer
is explicitly set to 1 for normalized numbers, infinities, and NaNs, and to 0 for zero
and denormalized numbers.

Vol. 1 4-7
DATA TYPES

Table 4-3. Floating-Point Number and NaN Encodings

Class Sign Biased Exponent Significand
1
Integer Fraction
Positive +∞ 0 11..11 1 00..00
+Normals 0 11..10 1 11..11
. . . .
. . . .
0 00..01 1 00..00
+Denormals 0 00..00 0 11.11
. . . .
. . . .
0 00..00 0 00..01
+Zero 0 00..00 0 00..00
Negative −Zero 1 00..00 0 00..00
−Denormals 1 00..00 0 00..01
. . . .
. . . .
1 00..00 0 11..11
−Normals 1 00..01 1 00..00
. . . .
. . . .
1 11..10 1 11..11
-∞ 1 11..11 1 00..00
NaNs SNaN X 11..11 1 0X..XX2
QNaN X 11..11 1 1X..XX
QNaN 1 11..11 1 10..00
Floating-Point
Indefinite
Single-Precision: ← 8 Bits → ← 23 Bits →
Double-Precision: ← 11 Bits → ← 52 Bits →
Double Extended-Precision: ← 15 Bits → ← 63 Bits →
NOTES:
1. Integer bit is implied and not stored for single-precision and double-precision formats.
2. The fraction for SNaN encodings must be non-zero with the most-significant bit 0.

The exponent of each floating-point data type is encoded in biased format; see
Section 4.8.2.2, “Biased Exponent.” The biasing constant is 127 for the single-
precision format, 1023 for the double-precision format, and 16,383 for the double
extended-precision format.

4-8 Vol. 1
 DATA TYPES

When storing floating-point values in memory, single-precision values are stored in 4

consecutive bytes in memory; double-precision values are stored in 8 consecutive
bytes; and double extended-precision values are stored in 10 consecutive bytes.
The single-precision and double-precision floating-point data types are operated on
by x87 FPU, and SSE/SSE2/SSE3 instructions. The double-extended-precision
floating-point format is only operated on by the x87 FPU. See Section 11.6.8,
“Compatibility of SIMD and x87 FPU Floating-Point Data Types,” for a discussion of
the compatibility of single-precision and double-precision floating-point data types
between the x87 FPU and SSE/SSE2/SSE3 extensions.

4.3 POINTER DATA TYPES

Pointers are addresses of locations in memory.
In non-64-bit modes, the architecture defines two types of pointers: a near pointer
and a far pointer. A near pointer is a 32-bit (or 16-bit) offset (also called an effec-
tive address) within a segment. Near pointers are used for all memory references in
a flat memory model or for references in a segmented model where the identity of
the segment being accessed is implied.
A far pointer is a logical address, consisting of a 16-bit segment selector and a 32-bit
(or 16-bit) offset. Far pointers are used for memory references in a segmented
memory model where the identity of a segment being accessed must be specified
explicitly. Near and far pointers with 32-bit offsets are shown in Figure 4-4.

Near Pointer
Offset
31 0

Far Pointer or Logical Address

Segment Selector Offset
47 32 31 0

Figure 4-4. Pointer Data Types

4.3.1 Pointer Data Types in 64-Bit Mode

In 64-bit mode (a sub-mode of IA-32e mode), a near pointer is 64 bits. This
equates to an effective address. Far pointers in 64-bit mode can be one of three
forms:
• 16-bit segment selector, 16-bit offset if the operand size is 32 bits
• 16-bit segment selector, 32-bit offset if the operand size is 32 bits
• 16-bit segment selector, 64-bit offset if the operand size is 64 bits
See Figure 4-5.

Vol. 1 4-9
DATA TYPES

Near Pointer

64-bit Offset

63 0

Far Pointer with 64-bit Operand Size

16-bit Segment Selector 64-bit Offset

79 64 63 0

Far Pointer with 32-bit Operand Size

16-bit Segment Selector 32-bit Offset

47 32 31 0

Far Pointer with 32-bit Operand Size

16-bit Segment Selector 16-bit Offset

31 16 15 0

Figure 4-5. Pointers in 64-Bit Mode

4.4 BIT FIELD DATA TYPE

A bit field (see Figure 4-6) is a contiguous sequence of bits. It can begin at any bit
position of any byte in memory and can contain up to 32 bits.

Bit Field

Field Length
Least
Significant
Bit

Figure 4-6. Bit Field Data Type

4-10 Vol. 1
 DATA TYPES

4.5 STRING DATA TYPES

Strings are continuous sequences of bits, bytes, words, or doublewords. A bit string
can begin at any bit position of any byte and can contain up to 232 – 1 bits. A byte
string can contain bytes, words, or doublewords and can range from zero to 232 – 1
bytes (4 GBytes).

4.6 PACKED SIMD DATA TYPES

Intel 64 and IA-32 architectures define and operate on a set of 64-bit and 128-bit
packed data type for use in SIMD operations. These data types consist of funda-
mental data types (packed bytes, words, doublewords, and quadwords) and numeric
interpretations of fundamental types for use in packed integer and packed floating-
point operations.

4.6.1 64-Bit SIMD Packed Data Types

The 64-bit packed SIMD data types were introduced into the IA-32 architecture in the
Intel MMX technology. They are operated on in MMX registers. The fundamental
64-bit packed data types are packed bytes, packed words, and packed doublewords
(see Figure 4-7). When performing numeric SIMD operations on these data types,
these data types are interpreted as containing byte, word, or doubleword integer
values.

Vol. 1 4-11
DATA TYPES

Fundamental 64-Bit Packed SIMD Data Types

Packed Bytes

63 0

Packed Words

63 0

Packed Doublewords

63 0

64-Bit Packed Integer Data Types

Packed Byte Integers

63 0

Packed Word Integers

63 0

Packed Doubleword Integers

63 0

Figure 4-7. 64-Bit Packed SIMD Data Types

4.6.2 128-Bit Packed SIMD Data Types

The 128-bit packed SIMD data types were introduced into the IA-32 architecture in
the SSE extensions and used with SSE2, SSE3 and SSSE3 extensions. They are oper-
ated on primarily in the 128-bit XMM registers and memory. The fundamental 128-bit
packed data types are packed bytes, packed words, packed doublewords, and
packed quadwords (see Figure 4-8). When performing SIMD operations on these
fundamental data types in XMM registers, these data types are interpreted as
containing packed or scalar single-precision floating-point or double-precision
floating-point values, or as containing packed byte, word, doubleword, or quadword
integer values.

4-12 Vol. 1
 DATA TYPES

Fundamental 128-Bit Packed SIMD Data Types

Packed Bytes
127 0

Packed Words
127 0
Packed Doublewords
127 0
Packed Quadwords
127 0

128-Bit Packed Floating-Point and Integer Data Types

Packed Single Precision

Floating Point
127 0

Packed Double Precision

Floating Point
127 0

Packed Byte Integers

127 0

Packed Word Integers

127 0

Packed Doubleword Integers

127 0

Packed Quadword Integers

127 0

Figure 4-8. 128-Bit Packed SIMD Data Types

4.7 BCD AND PACKED BCD INTEGERS

Binary-coded decimal integers (BCD integers) are unsigned 4-bit integers with valid
values ranging from 0 to 9. IA-32 architecture defines operations on BCD integers
located in one or more general-purpose registers or in one or more x87 FPU registers
(see Figure 4-9).

Vol. 1 4-13
DATA TYPES

BCD Integers
X BCD
7 43 0
Packed BCD Integers
BCD BCD
7 43 0
Sign 80-Bit Packed BCD Decimal Integers
X D17 D16 D15 D14 D13 D12 D11 D10 D9 D8 D7 D6 D5 D4 D3 D2 D1 D0
79 78 72 71 0
4 Bits = 1 BCD Digit

Figure 4-9. BCD Data Types

When operating on BCD integers in general-purpose registers, the BCD values can be
unpacked (one BCD digit per byte) or packed (two BCD digits per byte). The value of
an unpacked BCD integer is the binary value of the low half-byte (bits 0 through 3).
The high half-byte (bits 4 through 7) can be any value during addition and subtrac-
tion, but must be zero during multiplication and division. Packed BCD integers allow
two BCD digits to be contained in one byte. Here, the digit in the high half-byte is
more significant than the digit in the low half-byte.
When operating on BCD integers in x87 FPU data registers, BCD values are packed in
an 80-bit format and referred to as decimal integers. In this format, the first 9 bytes
hold 18 BCD digits, 2 digits per byte. The least-significant digit is contained in the
lower half-byte of byte 0 and the most-significant digit is contained in the upper half-
byte of byte 9. The most significant bit of byte 10 contains the sign bit (0 = positive
and 1 = negative; bits 0 through 6 of byte 10 are don’t care bits). Negative decimal
integers are not stored in two's complement form; they are distinguished from posi-
tive decimal integers only by the sign bit. The range of decimal integers that can be
encoded in this format is –1018 + 1 to 1018 – 1.
The decimal integer format exists in memory only. When a decimal integer is loaded
in an x87 FPU data register, it is automatically converted to the double-extended-
precision floating-point format. All decimal integers are exactly representable in
double extended-precision format.
Table 4-4 gives the possible encodings of value in the decimal integer data type.

4-14 Vol. 1
 DATA TYPES

Table 4-4. Packed Decimal Integer Encodings

Magnitude
Class Sign digit digit digit digit ... digit
Positive
Largest 0 0000000 1001 1001 1001 1001 ... 1001

. . .
. . .
Smallest 0 0000000 0000 0000 0000 0000 ... 0001
Zero 0 0000000 0000 0000 0000 0000 ... 0000
Negative
Zero 1 0000000 0000 0000 0000 0000 ... 0000

Smallest 1 0000000 0000 0000 0000 0000 ... 0001

. . .
. . .
Largest 1 0000000 1001 1001 1001 1001 ... 1001
Packed 1 1111111 1111 1111 1100 0000 ... 0000
BCD
Integer
Indefinit
e
← 1 byte → ← 9 bytes →

The packed BCD integer indefinite encoding (FFFFC000000000000000H) is stored by

the FBSTP instruction in response to a masked floating-point invalid-operation
exception. Attempting to load this value with the FBLD instruction produces an unde-
fined result.

4.8 REAL NUMBERS AND FLOATING-POINT FORMATS

This section describes how real numbers are represented in floating-point format in
x87 FPU and SSE/SSE2/SSE3 floating-point instructions. It also introduces terms
such as normalized numbers, denormalized numbers, biased exponents, signed
zeros, and NaNs. Readers who are already familiar with floating-point processing
techniques and the IEEE Standard 754 for Binary Floating-Point Arithmetic may wish
to skip this section.

Vol. 1 4-15
DATA TYPES

4.8.1 Real Number System

As shown in Figure 4-10, the real-number system comprises the continuum of real
numbers from minus infinity (− ∞) to plus infinity (+ ∞).
Because the size and number of registers that any computer can have is limited, only
a subset of the real-number continuum can be used in real-number (floating-point)
calculations. As shown at the bottom of Figure 4-10, the subset of real numbers that
the IA-32 architecture supports represents an approximation of the real number
system. The range and precision of this real-number subset is determined by the
IEEE Standard 754 floating-point formats.

4.8.2 Floating-Point Format

To increase the speed and efficiency of real-number computations, computers and
microprocessors typically represent real numbers in a binary floating-point format.
In this format, a real number has three parts: a sign, a significand, and an exponent
(see Figure 4-11).
The sign is a binary value that indicates whether the number is positive (0) or nega-
tive (1). The significand has two parts: a 1-bit binary integer (also referred to as
the J-bit) and a binary fraction. The integer-bit is often not represented, but instead
is an implied value. The exponent is a binary integer that represents the base-2
power by which the significand is multiplied.
Table 4-5 shows how the real number 178.125 (in ordinary decimal format) is stored
in IEEE Standard 754 floating-point format. The table lists a progression of real
number notations that leads to the single-precision, 32-bit floating-point format. In
this format, the significand is normalized (see Section 4.8.2.1, “Normalized
Numbers”) and the exponent is biased (see Section 4.8.2.2, “Biased Exponent”). For
the single-precision floating-point format, the biasing constant is +127.

4-16 Vol. 1
 DATA TYPES

Binary Real Number System

-100 -10 -1 0 1 10 100
ςς ςς

Subset of binary real numbers that can be represented with

IEEE single-precision (32-bit) floating-point format
-100 -10 -1 0 1 10 100
ςς ςς

+10

10.0000000000000000000000
1.11111111111111111111111
Precision 24 Binary Digits

Numbers within this range

cannot be represented.

Figure 4-10. Binary Real Number System

Sign
Exponent Significand

Fraction

Integer or J-Bit

Figure 4-11. Binary Floating-Point Format

Vol. 1 4-17
DATA TYPES

Table 4-5. Real and Floating-Point Number Notation

Notation Value
Ordinary Decimal 178.125
Scientific Decimal 1.78125E10
2
Scientific Binary 1.0110010001E2111
Scientific Binary 1.0110010001E210000110
(Biased Exponent)
IEEE Single-Precision Format Sign Biased Exponent Normalized Significand
0 10000110 0110010001000000000000
0
1. (Implied)

4.8.2.1 Normalized Numbers

In most cases, floating-point numbers are encoded in normalized form. This means
that except for zero, the significand is always made up of an integer of 1 and the
following fraction:
1.fff...ff
For values less than 1, leading zeros are eliminated. (For each leading zero elimi-
nated, the exponent is decremented by one.)
Representing numbers in normalized form maximizes the number of significant digits
that can be accommodated in a significand of a given width. To summarize, a normal-
ized real number consists of a normalized significand that represents a real number
between 1 and 2 and an exponent that specifies the number’s binary point.

4.8.2.2 Biased Exponent

In the IA-32 architecture, the exponents of floating-point numbers are encoded in a
biased form. This means that a constant is added to the actual exponent so that the
biased exponent is always a positive number. The value of the biasing constant
depends on the number of bits available for representing exponents in the floating-
point format being used. The biasing constant is chosen so that the smallest normal-
ized number can be reciprocated without overflow.
See Section 4.2.2, “Floating-Point Data Types,” for a list of the biasing constants that
the IA-32 architecture uses for the various sizes of floating-point data-types.

4-18 Vol. 1
 DATA TYPES

4.8.3 Real Number and Non-number Encodings

A variety of real numbers and special values can be encoded in the IEEE Standard
754 floating-point format. These numbers and values are generally divided into the
following classes:
• Signed zeros
• Denormalized finite numbers
• Normalized finite numbers
• Signed infinities
• NaNs
• Indefinite numbers
(The term NaN stands for “Not a Number.”)
Figure 4-12 shows how the encodings for these numbers and non-numbers fit into
the real number continuum. The encodings shown here are for the IEEE single-preci-
sion floating-point format. The term “S” indicates the sign bit, “E” the biased expo-
nent, and “Sig” the significand. The exponent values are given in decimal. The
integer bit is shown for the significands, even though the integer bit is implied in
single-precision floating-point format.

NaN NaN
− Denormalized Finite + Denormalized Finite
−∞ − Normalized Finite − 0+ 0 + Normalized Finite + ∞

Real Number and NaN Encodings For 32-Bit Floating-Point Format

S E Sig1 S E Sig1
1 0 0.000... −0 +0 0 0 0.000...

0.XXX...2 − Denormalized +Denormalized 0.XXX...2

1 0 Finite Finite 0 0
− Normalized +Normalized
1 1...254 1.XXX... Finite 0 1...254 1.XXX...
Finite
1 255 1.000... −∞ +∞ 0 255 1.000...

X3 255 1.0XX...2 SNaN SNaN X3 255 1.0XX...2

X3 255 1.1XX... QNaN QNaN X3 255 1.1XX...

NOTES:
1. Integer bit of fraction implied for
single-precision floating-point format.
2. Fraction must be non-zero.
3. Sign bit ignored.

Figure 4-12. Real Numbers and NaNs

Vol. 1 4-19
DATA TYPES

An IA-32 processor can operate on and/or return any of these values, depending on
the type of computation being performed. The following sections describe these
number and non-number classes.

4.8.3.1 Signed Zeros

Zero can be represented as a +0 or a −0 depending on the sign bit. Both encodings
are equal in value. The sign of a zero result depends on the operation being
performed and the rounding mode being used. Signed zeros have been provided to
aid in implementing interval arithmetic. The sign of a zero may indicate the direction
from which underflow occurred, or it may indicate the sign of an ∞ that has been
reciprocated.

4.8.3.2 Normalized and Denormalized Finite Numbers

Non-zero, finite numbers are divided into two classes: normalized and denormalized.
The normalized finite numbers comprise all the non-zero finite values that can be
encoded in a normalized real number format between zero and ∞. In the single-preci-
sion floating-point format shown in Figure 4-12, this group of numbers includes all
the numbers with biased exponents ranging from 1 to 25410 (unbiased, the exponent
range is from −12610 to +12710).

When floating-point numbers become very close to zero, the normalized-number

format can no longer be used to represent the numbers. This is because the range of
the exponent is not large enough to compensate for shifting the binary point to the
right to eliminate leading zeros.
When the biased exponent is zero, smaller numbers can only be represented by
making the integer bit (and perhaps other leading bits) of the significand zero. The
numbers in this range are called denormalized (or tiny) numbers. The use of
leading zeros with denormalized numbers allows smaller numbers to be represented.
However, this denormalization causes a loss of precision (the number of significant
bits in the fraction is reduced by the leading zeros).
When performing normalized floating-point computations, an IA-32 processor
normally operates on normalized numbers and produces normalized numbers as
results. Denormalized numbers represent an underflow condition. The exact condi-
tions are specified in Section 4.9.1.5, “Numeric Underflow Exception (#U).”
A denormalized number is computed through a technique called gradual underflow.
Table 4-6 gives an example of gradual underflow in the denormalization process.
Here the single-precision format is being used, so the minimum exponent (unbiased)
is −12610. The true result in this example requires an exponent of −12910 in order to
have a normalized number. Since −12910 is beyond the allowable exponent range,
the result is denormalized by inserting leading zeros until the minimum exponent of
−12610 is reached.

4-20 Vol. 1
 DATA TYPES

Table 4-6. Denormalization Process

Operation Sign Exponent* Significand
True Result 0 −129 1.01011100000...00
Denormalize 0 −128 0.10101110000...00
Denormalize 0 −127 0.01010111000...00
Denormalize 0 −126 0.00101011100...00
Denormal Result 0 −126 0.00101011100...00
* Expressed as an unbiased, decimal number.

In the extreme case, all the significant bits are shifted out to the right by leading
zeros, creating a zero result.
The Intel 64 and IA-32 architectures deal with denormal values in the following ways:
• It avoids creating denormals by normalizing numbers whenever possible.
• It provides the floating-point underflow exception to permit programmers to
detect cases when denormals are created.
• It provides the floating-point denormal-operand exception to permit procedures
or programs to detect when denormals are being used as source operands for
computations.

4.8.3.3 Signed Infinities

The two infinities, + ∞ and − ∞, represent the maximum positive and negative real
numbers, respectively, that can be represented in the floating-point format. Infinity
is always represented by a significand of 1.00...00 (the integer bit may be implied)
and the maximum biased exponent allowed in the specified format (for example,
25510 for the single-precision format).
The signs of infinities are observed, and comparisons are possible. Infinities are
always interpreted in the affine sense; that is, –∞ is less than any finite number and
+∞ is greater than any finite number. Arithmetic on infinities is always exact. Excep-
tions are generated only when the use of an infinity as a source operand constitutes
an invalid operation.
Whereas denormalized numbers may represent an underflow condition, the two ∞
numbers may represent the result of an overflow condition. Here, the normalized
result of a computation has a biased exponent greater than the largest allowable
exponent for the selected result format.

4.8.3.4 NaNs
Since NaNs are non-numbers, they are not part of the real number line. In
Figure 4-12, the encoding space for NaNs in the floating-point formats is shown

Vol. 1 4-21
DATA TYPES

above the ends of the real number line. This space includes any value with the
maximum allowable biased exponent and a non-zero fraction (the sign bit is ignored
for NaNs).
The IA-32 architecture defines two classes of NaNs: quiet NaNs (QNaNs) and
signaling NaNs (SNaNs). A QNaN is a NaN with the most significant fraction bit set;
an SNaN is a NaN with the most significant fraction bit clear. QNaNs are allowed to
propagate through most arithmetic operations without signaling an exception.
SNaNs generally signal a floating-point invalid-operation exception whenever they
appear as operands in arithmetic operations.
SNaNs are typically used to trap or invoke an exception handler. They must be
inserted by software; that is, the processor never generates an SNaN as a result of a
floating-point operation.

4.8.3.5 Operating on SNaNs and QNaNs

When a floating-point operation is performed on an SNaN and/or a QNaN, the result
of the operation is either a QNaN delivered to the destination operand or the genera-
tion of a floating-point invalid operating exception, depending on the following rules:
• If one of the source operands is an SNaN and the floating-point invalid-operating
exception is not masked (see Section 4.9.1.1, “Invalid Operation Exception
(#I)”), the a floating-point invalid-operation exception is signaled and no result is
stored in the destination operand.
• If either or both of the source operands are NaNs and floating-point invalid-
operation exception is masked, the result is as shown in Table 4-7. When an
SNaN is converted to a QNaN, the conversion is handled by setting the most-
significant fraction bit of the SNaN to 1. Also, when one of the source operands is
an SNaN, the floating-point invalid-operation exception flag it set. Note that for
some combinations of source operands, the result is different for x87 FPU
operations and for SSE/SSE2/SSE3 operations.
• When neither of the source operands is a NaN, but the operation generates a
floating-point invalid-operation exception (see Tables 8-10 and 11-1), the result
is commonly an SNaN source operand converted to a QNaN or the QNaN floating-
point indefinite value.
Any exceptions to the behavior described in Table 4-7 are described in Section
8.5.1.2, “Invalid Arithmetic Operand Exception (#IA),” and Section 11.5.2.1, “Invalid
Operation Exception (#I).”

4-22 Vol. 1
 DATA TYPES

Table 4-7. Rules for Handling NaNs

Source Operands Result1
SNaN and QNaN x87 FPU — QNaN source operand.
SSE/SSE2/SSE3 — First operand (if this operand is
an SNaN, it is converted to a QNaN)
Two SNaNs x87 FPU—SNaN source operand with the larger
significand, converted into a QNaN
SSE/SSE2/SSE3 — First operand converted to a
QNaN
Two QNaNs x87 FPU — QNaN source operand with the larger
significand
SSE/SSE2/SSE3 — First operand
SNaN and a floating-point value SNaN source operand, converted into a QNaN
QNaN and a floating-point value QNaN source operand
SNaN (for instructions that take only one SNaN source operand, converted into a QNaN
operand)
QNaN (for instructions that take only one QNaN source operand
operand)
NOTE:
1. For SSE/SSE2/SSE3 instructions, the first operand is generally a source operand that becomes
the destination operand. Within the Result column, the x87 FPU notation also applies to the
FISTTP instruction in SSE3; the SSE3 notation applies to the SIMD floating-point instructions.

4.8.3.6 Using SNaNs and QNaNs in Applications

Except for the rules given at the beginning of Section 4.8.3.4, “NaNs,” for encoding
SNaNs and QNaNs, software is free to use the bits in the significand of a NaN for any
purpose. Both SNaNs and QNaNs can be encoded to carry and store data, such as
diagnostic information.
By unmasking the invalid operation exception, the programmer can use signaling
NaNs to trap to the exception handler. The generality of this approach and the large
number of NaN values that are available provide the sophisticated programmer with
a tool that can be applied to a variety of special situations.
For example, a compiler can use signaling NaNs as references to uninitialized (real)
array elements. The compiler can preinitialize each array element with a signaling
NaN whose significand contained the index (relative position) of the element. Then,
if an application program attempts to access an element that it had not initialized, it
can use the NaN placed there by the compiler. If the invalid operation exception is
unmasked, an interrupt will occur, and the exception handler will be invoked. The
exception handler can determine which element has been accessed, since the

Vol. 1 4-23
DATA TYPES

operand address field of the exception pointer will point to the NaN, and the NaN will
contain the index number of the array element.
Quiet NaNs are often used to speed up debugging. In its early testing phase, a
program often contains multiple errors. An exception handler can be written to save
diagnostic information in memory whenever it was invoked. After storing the diag-
nostic data, it can supply a quiet NaN as the result of the erroneous instruction, and
that NaN can point to its associated diagnostic area in memory. The program will
then continue, creating a different NaN for each error. When the program ends, the
NaN results can be used to access the diagnostic data saved at the time the errors
occurred. Many errors can thus be diagnosed and corrected in one test run.
In embedded applications that use computed results in further computations, an
undetected QNaN can invalidate all subsequent results. Such applications should
therefore periodically check for QNaNs and provide a recovery mechanism to be used
if a QNaN result is detected.

4.8.3.7 QNaN Floating-Point Indefinite

For the floating-point data type encodings (single-precision, double-precision, and
double-extended-precision), one unique encoding (a QNaN) is reserved for repre-
senting the special value QNaN floating-point indefinite. The x87 FPU and the
SSE/SSE2/SSE3 extensions return these indefinite values as responses to some
masked floating-point exceptions. Table 4-3 shows the encoding used for the QNaN
floating-point indefinite.

4.8.4 Rounding
When performing floating-point operations, the processor produces an infinitely
precise floating-point result in the destination format (single-precision, double-preci-
sion, or double extended-precision floating-point) whenever possible. However,
because only a subset of the numbers in the real number continuum can be repre-
sented in IEEE Standard 754 floating-point formats, it is often the case that an infi-
nitely precise result cannot be encoded exactly in the format of the destination
operand.
For example, the following value (a) has a 24-bit fraction. The least-significant bit of
this fraction (the underlined bit) cannot be encoded exactly in the single-precision
format (which has only a 23-bit fraction):
(a) 1.0001 0000 1000 0011 1001 0111E2 101
To round this result (a), the processor first selects two representable fractions b and
c that most closely bracket a in value (b < a < c).
(b) 1.0001 0000 1000 0011 1001 011E2 101
(c) 1.0001 0000 1000 0011 1001 100E2 101

4-24 Vol. 1
 DATA TYPES

The processor then sets the result to b or to c according to the selected rounding
mode. Rounding introduces an error in a result that is less than one unit in the last
place (the least significant bit position of the floating-point value) to which the result
is rounded.
The IEEE Standard 754 defines four rounding modes (see Table 4-8): round to
nearest, round up, round down, and round toward zero. The default rounding mode
(for the Intel 64 and IA-32 architectures) is round to nearest. This mode provides the
most accurate and statistically unbiased estimate of the true result and is suitable for
most applications.

Table 4-8. Rounding Modes and Encoding of Rounding Control (RC) Field
Rounding RC Field Description
Mode Setting
Round to 00B Rounded result is the closest to the infinitely precise result. If two
nearest (even) values are equally close, the result is the even value (that is, the
one with the least-significant bit of zero). Default
Round down 01B Rounded result is closest to but no greater than the infinitely
(toward −∞) precise result.
Round up 10B Rounded result is closest to but no less than the infinitely precise
(toward +∞) result.
Round toward 11B Rounded result is closest to but no greater in absolute value than
zero (Truncate) the infinitely precise result.

The round up and round down modes are termed directed rounding and can be
used to implement interval arithmetic. Interval arithmetic is used to determine upper
and lower bounds for the true result of a multistep computation, when the interme-
diate results of the computation are subject to rounding.
The round toward zero mode (sometimes called the “chop” mode) is commonly used
when performing integer arithmetic with the x87 FPU.
The rounded result is called the inexact result. When the processor produces an
inexact result, the floating-point precision (inexact) flag (PE) is set (see Section
4.9.1.6, “Inexact-Result (Precision) Exception (#P)”).
The rounding modes have no effect on comparison operations, operations that
produce exact results, or operations that produce NaN results.

4.8.4.1 Rounding Control (RC) Fields

In the Intel 64 and IA-32 architectures, the rounding mode is controlled by a 2-bit
rounding-control (RC) field (Table 4-8 shows the encoding of this field). The RC field
is implemented in two different locations:
• x87 FPU control register (bits 10 and 11)

Vol. 1 4-25
DATA TYPES

• The MXCSR register (bits 13 and 14)

Although these two RC fields perform the same function, they control rounding for
different execution environments within the processor. The RC field in the x87 FPU
control register controls rounding for computations performed with the x87 FPU
instructions; the RC field in the MXCSR register controls rounding for SIMD floating-
point computations performed with the SSE/SSE2 instructions.

4.8.4.2 Truncation with SSE and SSE2 Conversion Instructions

The following SSE/SSE2 instructions automatically truncate the results of conver-
sions from floating-point values to integers when the result it inexact: CVTTPD2DQ,
CVTTPS2DQ, CVTTPD2PI, CVTTPS2PI, CVTTSD2SI, CVTTSS2SI. Here, truncation
means the round toward zero mode described in Table 4-8.

4.9 OVERVIEW OF FLOATING-POINT EXCEPTIONS

The following section provides an overview of floating-point exceptions and their
handling in the IA-32 architecture. For information specific to the x87 FPU and to the
SSE/SSE2/SSE3 extensions, refer to the following sections:
• Section 8.4, “x87 FPU Floating-Point Exception Handling”
• Section 11.5, “SSE, SSE2, and SSE3 Exceptions”
When operating on floating-point operands, the IA-32 architecture recognizes and
detects six classes of exception conditions:
• Invalid operation (#I)
• Divide-by-zero (#Z)
• Denormalized operand (#D)
• Numeric overflow (#O)
• Numeric underflow (#U)
• Inexact result (precision) (#P)
The nomenclature of “#” symbol followed by one or two letters (for example, #P) is
used in this manual to indicate exception conditions. It is merely a short-hand form
and is not related to assembler mnemonics.

NOTE
All of the exceptions listed above except the denormal-operand
exception (#D) are defined in IEEE Standard 754.

The invalid-operation, divide-by-zero and denormal-operand exceptions are pre-

computation exceptions (that is, they are detected before any arithmetic operation

4-26 Vol. 1
 DATA TYPES

occurs). The numeric-underflow, numeric-overflow and precision exceptions are

post-computation exceptions.
Each of the six exception classes has a corresponding flag bit (IE, ZE, OE, UE, DE, or
PE) and mask bit (IM, ZM, OM, UM, DM, or PM). When one or more floating-point
exception conditions are detected, the processor sets the appropriate flag bits, then
takes one of two possible courses of action, depending on the settings of the corre-
sponding mask bits:
• Mask bit set. Handles the exception automatically, producing a predefined (and
often times usable) result, while allowing program execution to continue undis-
turbed.
• Mask bit clear. Invokes a software exception handler to handle the exception.
The masked (default) responses to exceptions have been chosen to deliver a reason-
able result for each exception condition and are generally satisfactory for most
floating-point applications. By masking or unmasking specific floating-point excep-
tions, programmers can delegate responsibility for most exceptions to the processor
and reserve the most severe exception conditions for software exception handlers.
Because the exception flags are “sticky,” they provide a cumulative record of the
exceptions that have occurred since they were last cleared. A programmer can thus
mask all exceptions, run a calculation, and then inspect the exception flags to see if
any exceptions were detected during the calculation.
In the IA-32 architecture, floating-point exception flag and mask bits are imple-
mented in two different locations:
• x87 FPU status word and control word. The flag bits are located at bits 0 through
5 of the x87 FPU status word and the mask bits are located at bits 0 through 5 of
the x87 FPU control word (see Figures 8-4 and 8-6).
• MXCSR register. The flag bits are located at bits 0 through 5 of the MXCSR
register and the mask bits are located at bits 7 through 12 of the register (see
Figure 10-3).
Although these two sets of flag and mask bits perform the same function, they report
on and control exceptions for different execution environments within the processor.
The flag and mask bits in the x87 FPU status and control words control exception
reporting and masking for computations performed with the x87 FPU instructions;
the companion bits in the MXCSR register control exception reporting and masking
for SIMD floating-point computations performed with the SSE/SSE2/SSE3 instruc-
tions.
Note that when exceptions are masked, the processor may detect multiple excep-
tions in a single instruction, because it continues executing the instruction after
performing its masked response. For example, the processor can detect a denormal-
ized operand, perform its masked response to this exception, and then detect
numeric underflow.
See Section 4.9.2, “Floating-Point Exception Priority,” for a description of the rules for
exception precedence when more than one floating-point exception condition is
detected for an instruction.

Vol. 1 4-27
DATA TYPES

4.9.1 Floating-Point Exception Conditions

The following sections describe the various conditions that cause a floating-point
exception to be generated and the masked response of the processor when these
conditions are detected. The Intel® 64 and IA-32 Architectures Software Developer’s
Manual, Volumes 3A & 3B, list the floating-point exceptions that can be signaled for
each floating-point instruction.

4.9.1.1 Invalid Operation Exception (#I)

The processor reports an invalid operation exception in response to one or more
invalid arithmetic operands. If the invalid operation exception is masked, the
processor sets the IE flag and returns an indefinite value or a QNaN. This value over-
writes the destination register specified by the instruction. If the invalid operation
exception is not masked, the IE flag is set, a software exception handler is invoked,
and the operands remain unaltered.
See Section 4.8.3.6, “Using SNaNs and QNaNs in Applications,” for information about
the result returned when an exception is caused by an SNaN.
The processor can detect a variety of invalid arithmetic operations that can be coded
in a program. These operations generally indicate a programming error, such as
dividing ∞ by ∞ . See the following sections for information regarding the invalid-
operation exception when detected while executing x87 FPU or SSE/SSE2/SSE3
instructions:
• x87 FPU; Section 8.5.1, “Invalid Operation Exception”
• SIMD floating-point exceptions; Section 11.5.2.1, “Invalid Operation Exception
(#I)”

4.9.1.2 Denormal Operand Exception (#D)

The processor reports the denormal-operand exception if an arithmetic instruction
attempts to operate on a denormal operand (see Section 4.8.3.2, “Normalized and
Denormalized Finite Numbers”). When the exception is masked, the processor sets
the DE flag and proceeds with the instruction. Operating on denormal numbers will
produce results at least as good as, and often better than, what can be obtained
when denormal numbers are flushed to zero. Programmers can mask this exception
so that a computation may proceed, then analyze any loss of accuracy when the final
result is delivered.
When a denormal-operand exception is not masked, the DE flag is set, a software
exception handler is invoked, and the operands remain unaltered. When denormal
operands have reduced significance due to loss of low-order bits, it may be advisable
to not operate on them. Precluding denormal operands from computations can be
accomplished by an exception handler that responds to unmasked denormal-
operand exceptions.

4-28 Vol. 1
 DATA TYPES

See the following sections for information regarding the denormal-operand exception
when detected while executing x87 FPU or SSE/SSE2/SSE3 instructions:
• x87 FPU; Section 8.5.2, “Denormal Operand Exception (#D)”
• SIMD floating-point exceptions; Section 11.5.2.2, “Denormal-Operand Exception
(#D)”

4.9.1.3 Divide-By-Zero Exception (#Z)

The processor reports the floating-point divide-by-zero exception whenever an
instruction attempts to divide a finite non-zero operand by 0. The masked response
for the divide-by-zero exception is to set the ZE flag and return an infinity signed with
the exclusive OR of the sign of the operands. If the divide-by-zero exception is not
masked, the ZE flag is set, a software exception handler is invoked, and the operands
remain unaltered.
See the following sections for information regarding the divide-by-zero exception
when detected while executing x87 FPU or SSE/SSE2 instructions:
• x87 FPU; Section 8.5.3, “Divide-By-Zero Exception (#Z)”
• SIMD floating-point exceptions; Section 11.5.2.3, “Divide-By-Zero Exception
(#Z)”

4.9.1.4 Numeric Overflow Exception (#O)

The processor reports a floating-point numeric overflow exception whenever the
rounded result of an instruction exceeds the largest allowable finite value that will fit
into the destination operand. Table 4-9 shows the threshold range for numeric over-
flow for each of the floating-point formats; overflow occurs when a rounded result
falls at or outside this threshold range.

Vol. 1 4-29
DATA TYPES

Table 4-9. Numeric Overflow Thresholds

Floating-Point Format Overflow Thresholds
Single Precision | x | ≥ 1.0 ∗ 2128
Double Precision | x | ≥ 1.0 ∗ 21024
Double Extended Precision | x | ≥ 1.0 ∗ 216384

When a numeric-overflow exception occurs and the exception is masked, the

processor sets the OE flag and returns one of the values shown in Table 4-10,
according to the current rounding mode. See Section 4.8.4, “Rounding.”
When numeric overflow occurs and the numeric-overflow exception is not masked,
the OE flag is set, a software exception handler is invoked, and the source and desti-
nation operands either remain unchanged or a biased result is stored in the destina-
tion operand (depending whether the overflow exception was generated during an
SSE/SSE2/SSE3 floating-point operation or an x87 FPU operation).

Table 4-10. Masked Responses to Numeric Overflow

Rounding Mode Sign of True Result Result
To nearest + +∞
– –∞
Toward –∞ + Largest finite positive number
– –∞
Toward +∞ + +∞
– Largest finite negative number
Toward zero + Largest finite positive number
– Largest finite negative number

See the following sections for information regarding the numeric overflow exception
when detected while executing x87 FPU instructions or while executing
SSE/SSE2/SSE3 instructions:
• x87 FPU; Section 8.5.4, “Numeric Overflow Exception (#O)”
• SIMD floating-point exceptions; Section 11.5.2.4, “Numeric Overflow Exception
(#O)”

4.9.1.5 Numeric Underflow Exception (#U)

The processor detects a floating-point numeric underflow condition whenever the
result of rounding with unbounded exponent (taking into account precision control
for x87) is tiny; that is, less than the smallest possible normalized, finite value that
will fit into the destination operand. Table 4-11 shows the threshold range for

4-30 Vol. 1
 DATA TYPES

numeric underflow for each of the floating-point formats (assuming normalized

results); underflow occurs when a rounded result falls strictly within the threshold
range. The ability to detect and handle underflow is provided to prevent a very small
result from propagating through a computation and causing another exception (such
as overflow during division) to be generated at a later time.

Table 4-11. Numeric Underflow (Normalized) Thresholds

Floating-Point Format Underflow Thresholds*
Single Precision | x | < 1.0 ∗ 2−126
Double Precision | x | < 1.0 ∗ 2−1022
Double Extended Precision | x | < 1.0 ∗ 2−16382
* Where ‘x’ is the result rounded to destination precision with an unbounded exponent range.

How the processor handles an underflow condition, depends on two related condi-
tions:
• creation of a tiny result
• creation of an inexact result; that is, a result that cannot be represented exactly
in the destination format
Which of these events causes an underflow exception to be reported and how the
processor responds to the exception condition depends on whether the underflow
exception is masked:
• Underflow exception masked — The underflow exception is reported (the UE
flag is set) only when the result is both tiny and inexact. The processor returns a
denormalized result to the destination operand, regardless of inexactness.
• Underflow exception not masked — The underflow exception is reported
when the result is tiny, regardless of inexactness. The processor leaves the
source and destination operands unaltered or stores a biased result in the
designating operand (depending whether the underflow exception was generated
during an SSE/SSE2/SSE3 floating-point operation or an x87 FPU operation) and
invokes a software exception handler.
See the following sections for information regarding the numeric underflow exception
when detected while executing x87 FPU instructions or while executing
SSE/SSE2/SSE3 instructions:
• x87 FPU; Section 8.5.5, “Numeric Underflow Exception (#U)”
• SIMD floating-point exceptions; Section 11.5.2.5, “Numeric Underflow Exception
(#U)”

4.9.1.6 Inexact-Result (Precision) Exception (#P)

The inexact-result exception (also called the precision exception) occurs if the result
of an operation is not exactly representable in the destination format. For example,
the fraction 1/3 cannot be precisely represented in binary floating-point form. This

Vol. 1 4-31
DATA TYPES

exception occurs frequently and indicates that some (normally acceptable) accuracy
will be lost due to rounding. The exception is supported for applications that need to
perform exact arithmetic only. Because the rounded result is generally satisfactory
for most applications, this exception is commonly masked.
If the inexact-result exception is masked when an inexact-result condition occurs and
a numeric overflow or underflow condition has not occurred, the processor sets the
PE flag and stores the rounded result in the destination operand. The current
rounding mode determines the method used to round the result. See Section 4.8.4,
“Rounding.”
If the inexact-result exception is not masked when an inexact result occurs and
numeric overflow or underflow has not occurred, the PE flag is set, the rounded result
is stored in the destination operand, and a software exception handler is invoked.
If an inexact result occurs in conjunction with numeric overflow or underflow, one of
the following operations is carried out:
• If an inexact result occurs along with masked overflow or underflow, the OE flag
or UE flag and the PE flag are set and the result is stored as described for the
overflow or underflow exceptions; see Section 4.9.1.4, “Numeric Overflow
Exception (#O),” or Section 4.9.1.5, “Numeric Underflow Exception (#U).” If the
inexact result exception is unmasked, the processor also invokes a software
exception handler.
• If an inexact result occurs along with unmasked overflow or underflow and the
destination operand is a register, the OE or UE flag and the PE flag are set, the
result is stored as described for the overflow or underflow exceptions, and a
software exception handler is invoked.
If an unmasked numeric overflow or underflow exception occurs and the destination
operand is a memory location (which can happen only for a floating-point store), the
inexact-result condition is not reported and the C1 flag is cleared.
See the following sections for information regarding the inexact-result exception
when detected while executing x87 FPU or SSE/SSE2/SSE3 instructions:
• x87 FPU; Section 8.5.6, “Inexact-Result (Precision) Exception (#P)”
• SIMD floating-point exceptions; Section 11.5.2.3, “Divide-By-Zero Exception
(#Z)”

4.9.2 Floating-Point Exception Priority

The processor handles exceptions according to a predetermined precedence. When
an instruction generates two or more exception conditions, the exception precedence
sometimes results in the higher-priority exception being handled and the lower-
priority exceptions being ignored. For example, dividing an SNaN by zero can poten-
tially signal an invalid-operation exception (due to the SNaN operand) and a divide-
by-zero exception. Here, if both exceptions are masked, the processor handles the
higher-priority exception only (the invalid-operation exception), returning a QNaN to
the destination. Alternately, a denormal-operand or inexact-result exception can

4-32 Vol. 1
 DATA TYPES

accompany a numeric underflow or overflow exception with both exceptions being

handled.
The precedence for floating-point exceptions is as follows:
1. Invalid-operation exception, subdivided as follows:
a. stack underflow (occurs with x87 FPU only)
b. stack overflow (occurs with x87 FPU only)
c. operand of unsupported format (occurs with x87 FPU only when using the
double extended-precision floating-point format)
d. SNaN operand
2. QNaN operand. Though this is not an exception, the handling of a QNaN operand
has precedence over lower-priority exceptions. For example, a QNaN divided by
zero results in a QNaN, not a zero-divide exception.
3. Any other invalid-operation exception not mentioned above or a divide-by-zero
exception.
4. Denormal-operand exception. If masked, then instruction execution continues
and a lower-priority exception can occur as well.
5. Numeric overflow and underflow exceptions; possibly in conjunction with the
inexact-result exception.
6. Inexact-result exception.
Invalid operation, zero divide, and denormal operand exceptions are detected before
a floating-point operation begins. Overflow, underflow, and precision exceptions are
not detected until a true result has been computed. When an unmasked pre-opera-
tion exception is detected, the destination operand has not yet been updated, and
appears as if the offending instruction has not been executed. When an unmasked
post-operation exception is detected, the destination operand may be updated with
a result, depending on the nature of the exception (except for SSE/SSE2/SSE3
instructions, which do not update their destination operands in such cases).

4.9.3 Typical Actions of a Floating-Point Exception Handler

After the floating-point exception handler is invoked, the processor handles the
exception in the same manner that it handles non-floating-point exceptions. The
floating-point exception handler is normally part of the operating system or execu-
tive software, and it usually invokes a user-registered floating-point exception
handle.
A typical action of the exception handler is to store state information in memory.
Other typical exception handler actions include:
• Examining the stored state information to determine the nature of the error
• Taking actions to correct the condition that caused the error

Vol. 1 4-33
DATA TYPES

• Clearing the exception flags

• Returning to the interrupted program and resuming normal execution
In lieu of writing recovery procedures, the exception handler can do the following:
• Increment in software an exception counter for later display or printing
• Print or display diagnostic information (such as the state information)
• Halt further program execution

4-34 Vol. 1
 CHAPTER 5
INSTRUCTION SET SUMMARY

This chapter provides an abridged overview of Intel 64 and IA-32 instructions.

Instructions are divided into the following groups:
• General purpose
• x87 FPU
• x87 FPU and SIMD state management
• Intel MMX technology
• SSE extensions
• SSE2 extensions
• SSE3 extensions
• SSSE3 extensions
• SSE4 extensions
• System instructions
• IA-32e mode: 64-bit mode instructions
• VMX instructions
• SMX instructions
Table 5-1 lists the groups and IA-32 processors that support each group. Within
these groups, most instructions are collected into functional subgroups.

Table 5-1. Instruction Groups and IA-32 Processors

Instruction Set
Architecture Intel 64 and IA-32 Processor Support
General Purpose All Intel 64 and IA-32 processors
x87 FPU Intel486, Pentium, Pentium with MMX Technology, Celeron, Pentium
Pro, Pentium II, Pentium II Xeon, Pentium III, Pentium III Xeon,
Pentium 4, Intel Xeon processors, Pentium M, Intel Core Solo, Intel Core
Duo, Intel Core 2 Duo processors, Intel Atom processors
x87 FPU and SIMD State Pentium II, Pentium II Xeon, Pentium III, Pentium III Xeon, Pentium 4,
Management Intel Xeon processors, Pentium M, Intel Core Solo, Intel Core Duo, Intel
Core 2 Duo processors, Intel Atom processors
MMX Technology Pentium with MMX Technology, Celeron, Pentium II, Pentium II Xeon,
Pentium III, Pentium III Xeon, Pentium 4, Intel Xeon processors,
Pentium M, Intel Core Solo, Intel Core Duo, Intel Core 2 Duo processors,
Intel Atom processors

Vol. 1 5-1
INSTRUCTION SET SUMMARY

Table 5-1. Instruction Groups and IA-32 Processors (Contd.)

Instruction Set
Architecture Intel 64 and IA-32 Processor Support
SSE Extensions Pentium III, Pentium III Xeon, Pentium 4, Intel Xeon processors,
Pentium M, Intel Core Solo, Intel Core Duo, Intel Core 2 Duo processors,
Intel Atom processors
SSE2 Extensions Pentium 4, Intel Xeon processors, Pentium M, Intel Core Solo, Intel Core
Duo, Intel Core 2 Duo processors, Intel Atom processors
SSE3 Extensions Pentium 4 supporting HT Technology (built on 90nm process
technology), Intel Core Solo, Intel Core Duo, Intel Core 2 Duo processors,
Intel Xeon processor 3xxxx, 5xxx, 7xxx Series, Intel Atom processors
SSSE3 Extensions Intel Xeon processor 3000, 3200, 5100, 5300, 7300 series, Intel Core
2 Extreme processors QX6000 series, Intel Core 2 Duo, Intel Core 2
Quad processors, Intel Pentium Dual-Core processors, Intel Atom
processors
SSE4.1 Extensions Intel Xeon processor 3100, 3300, 5200, 5400, 7400 series, Intel Core
2 Extreme processors QX9000 series, Intel Core 2 Quad processor
Q9000 series, Intel Core 2 Duo processors 8000 series, T9000 series.
SSE4.2 Extensions Intel Core i7 processor.
IA-32e mode: 64-bit Intel 64 processors
mode instructions
System Instructions Intel 64 and IA-32 processors
VMX Instructions Intel 64 and IA-32 processors supporting Intel Virtualization
Technology
SMX Instructions Intel Core 2 Duo processor E6x50, E8xxx; Intel Core 2 Quad processor
Q9xxx

The following sections list instructions in each major group and subgroup. Given for
each instruction is its mnemonic and descriptive names. When two or more
mnemonics are given (for example, CMOVA/CMOVNBE), they represent different
mnemonics for the same instruction opcode. Assemblers support redundant
mnemonics for some instructions to make it easier to read code listings. For instance,
CMOVA (Conditional move if above) and CMOVNBE (Conditional move if not below or
equal) represent the same condition. For detailed information about specific instruc-
tions, see the Intel® 64 and IA-32 Architectures Software Developer’s Manual,
Volumes 3A & 3B.

5-2 Vol. 1
 INSTRUCTION SET SUMMARY

5.1 GENERAL-PURPOSE INSTRUCTIONS

The general-purpose instructions preform basic data movement, arithmetic, logic,
program flow, and string operations that programmers commonly use to write appli-
cation and system software to run on Intel 64 and IA-32 processors. They operate on
data contained in memory, in the general-purpose registers (EAX, EBX, ECX, EDX,
EDI, ESI, EBP, and ESP) and in the EFLAGS register. They also operate on address
information contained in memory, the general-purpose registers, and the segment
registers (CS, DS, SS, ES, FS, and GS).
This group of instructions includes the data transfer, binary integer arithmetic,
decimal arithmetic, logic operations, shift and rotate, bit and byte operations,
program control, string, flag control, segment register operations, and miscellaneous
subgroups. The sections that following introduce each subgroup.
For more detailed information on general purpose-instructions, see Chapter 7,
“Programming With General-Purpose Instructions.”

5.1.1 Data Transfer Instructions

The data transfer instructions move data between memory and the general-purpose
and segment registers. They also perform specific operations such as conditional
moves, stack access, and data conversion.
MOV Move data between general-purpose registers; move data
between memory and general-purpose or segment registers;
move immediates to general-purpose registers
CMOVE/CMOVZ Conditional move if equal/Conditional move if zero
CMOVNE/CMOVNZ Conditional move if not equal/Conditional move if not zero
CMOVA/CMOVNBE Conditional move if above/Conditional move if not below or
equal
CMOVAE/CMOVNB Conditional move if above or equal/Conditional move if not
below
CMOVB/CMOVNAE Conditional move if below/Conditional move if not above or
equal
CMOVBE/CMOVNA Conditional move if below or equal/Conditional move if not
above
CMOVG/CMOVNLE Conditional move if greater/Conditional move if not less or equal
CMOVGE/CMOVNL Conditional move if greater or equal/Conditional move if not less
CMOVL/CMOVNGE Conditional move if less/Conditional move if not greater or equal
CMOVLE/CMOVNG Conditional move if less or equal/Conditional move if not greater
CMOVC Conditional move if carry
CMOVNC Conditional move if not carry
CMOVO Conditional move if overflow

Vol. 1 5-3
INSTRUCTION SET SUMMARY

CMOVNO Conditional move if not overflow

CMOVS Conditional move if sign (negative)
CMOVNS Conditional move if not sign (non-negative)
CMOVP/CMOVPE Conditional move if parity/Conditional move if parity even
CMOVNP/CMOVPO Conditional move if not parity/Conditional move if parity odd
XCHG Exchange
BSWAP Byte swap
XADD Exchange and add
CMPXCHG Compare and exchange
CMPXCHG8B Compare and exchange 8 bytes
PUSH Push onto stack
POP Pop off of stack
PUSHA/PUSHAD Push general-purpose registers onto stack
POPA/POPAD Pop general-purpose registers from stack
CWD/CDQ Convert word to doubleword/Convert doubleword to quadword
CBW/CWDE Convert byte to word/Convert word to doubleword in EAX
register
MOVSX Move and sign extend
MOVZX Move and zero extend

5.1.2 Binary Arithmetic Instructions

The binary arithmetic instructions perform basic binary integer computations on
byte, word, and doubleword integers located in memory and/or the general purpose
registers.
ADD Integer add
ADC Add with carry
SUB Subtract
SBB Subtract with borrow
IMUL Signed multiply
MUL Unsigned multiply
IDIV Signed divide
DIV Unsigned divide
INC Increment
DEC Decrement
NEG Negate
CMP Compare

5-4 Vol. 1
 INSTRUCTION SET SUMMARY

5.1.3 Decimal Arithmetic Instructions

The decimal arithmetic instructions perform decimal arithmetic on binary coded
decimal (BCD) data.
DAA Decimal adjust after addition
DAS Decimal adjust after subtraction
AAA ASCII adjust after addition
AAS ASCII adjust after subtraction
AAM ASCII adjust after multiplication
AAD ASCII adjust before division

5.1.4 Logical Instructions

The logical instructions perform basic AND, OR, XOR, and NOT logical operations on
byte, word, and doubleword values.
AND Perform bitwise logical AND
OR Perform bitwise logical OR
XOR Perform bitwise logical exclusive OR
NOT Perform bitwise logical NOT

5.1.5 Shift and Rotate Instructions

The shift and rotate instructions shift and rotate the bits in word and doubleword
operands.
SAR Shift arithmetic right
SHR Shift logical right
SAL/SHL Shift arithmetic left/Shift logical left
SHRD Shift right double
SHLD Shift left double
ROR Rotate right
ROL Rotate left
RCR Rotate through carry right
RCL Rotate through carry left

5.1.6 Bit and Byte Instructions

Bit instructions test and modify individual bits in word and doubleword operands.
Byte instructions set the value of a byte operand to indicate the status of flags in the
EFLAGS register.

Vol. 1 5-5
INSTRUCTION SET SUMMARY

BT Bit test
BTS Bit test and set
BTR Bit test and reset
BTC Bit test and complement
BSF Bit scan forward
BSR Bit scan reverse
SETE/SETZ Set byte if equal/Set byte if zero
SETNE/SETNZ Set byte if not equal/Set byte if not zero
SETA/SETNBE Set byte if above/Set byte if not below or equal
SETAE/SETNB/SETNC Set byte if above or equal/Set byte if not below/Set byte if not
carry
SETB/SETNAE/SETC Set byte if below/Set byte if not above or equal/Set byte if carry
SETBE/SETNA Set byte if below or equal/Set byte if not above
SETG/SETNLE Set byte if greater/Set byte if not less or equal
SETGE/SETNL Set byte if greater or equal/Set byte if not less
SETL/SETNGE Set byte if less/Set byte if not greater or equal
SETLE/SETNG Set byte if less or equal/Set byte if not greater
SETS Set byte if sign (negative)
SETNS Set byte if not sign (non-negative)
SETO Set byte if overflow
SETNO Set byte if not overflow
SETPE/SETP Set byte if parity even/Set byte if parity
SETPO/SETNP Set byte if parity odd/Set byte if not parity
TEST Logical compare

5.1.7 Control Transfer Instructions

The control transfer instructions provide jump, conditional jump, loop, and call and
return operations to control program flow.
JMP Jump
JE/JZ Jump if equal/Jump if zero
JNE/JNZ Jump if not equal/Jump if not zero
JA/JNBE Jump if above/Jump if not below or equal
JAE/JNB Jump if above or equal/Jump if not below
JB/JNAE Jump if below/Jump if not above or equal
JBE/JNA Jump if below or equal/Jump if not above
JG/JNLE Jump if greater/Jump if not less or equal
JGE/JNL Jump if greater or equal/Jump if not less

5-6 Vol. 1
 INSTRUCTION SET SUMMARY

JL/JNGE Jump if less/Jump if not greater or equal

JLE/JNG Jump if less or equal/Jump if not greater
JC Jump if carry
JNC Jump if not carry
JO Jump if overflow
JNO Jump if not overflow
JS Jump if sign (negative)
JNS Jump if not sign (non-negative)
JPO/JNP Jump if parity odd/Jump if not parity
JPE/JP Jump if parity even/Jump if parity
JCXZ/JECXZ Jump register CX zero/Jump register ECX zero
LOOP Loop with ECX counter
LOOPZ/LOOPE Loop with ECX and zero/Loop with ECX and equal
LOOPNZ/LOOPNE Loop with ECX and not zero/Loop with ECX and not equal
CALL Call procedure
RET Return
IRET Return from interrupt
INT Software interrupt
INTO Interrupt on overflow
BOUND Detect value out of range
ENTER High-level procedure entry
LEAVE High-level procedure exit

5.1.8 String Instructions

The string instructions operate on strings of bytes, allowing them to be moved to and
from memory.
MOVS/MOVSB Move string/Move byte string
MOVS/MOVSW Move string/Move word string
MOVS/MOVSD Move string/Move doubleword string
CMPS/CMPSB Compare string/Compare byte string
CMPS/CMPSW Compare string/Compare word string
CMPS/CMPSD Compare string/Compare doubleword string
SCAS/SCASB Scan string/Scan byte string
SCAS/SCASW Scan string/Scan word string
SCAS/SCASD Scan string/Scan doubleword string
LODS/LODSB Load string/Load byte string
LODS/LODSW Load string/Load word string

Vol. 1 5-7
INSTRUCTION SET SUMMARY

LODS/LODSD Load string/Load doubleword string

STOS/STOSB Store string/Store byte string
STOS/STOSW Store string/Store word string
STOS/STOSD Store string/Store doubleword string
REP Repeat while ECX not zero
REPE/REPZ Repeat while equal/Repeat while zero
REPNE/REPNZ Repeat while not equal/Repeat while not zero

5.1.9 I/O Instructions

These instructions move data between the processor’s I/O ports and a register or
memory.
IN Read from a port
OUT Write to a port
INS/INSB Input string from port/Input byte string from port
INS/INSW Input string from port/Input word string from port
INS/INSD Input string from port/Input doubleword string from port
OUTS/OUTSB Output string to port/Output byte string to port
OUTS/OUTSW Output string to port/Output word string to port
OUTS/OUTSD Output string to port/Output doubleword string to port

5.1.10 Enter and Leave Instructions

These instructions provide machine-language support for procedure calls in block-
structured languages.
ENTER High-level procedure entry
LEAVE High-level procedure exit

5.1.11 Flag Control (EFLAG) Instructions

The flag control instructions operate on the flags in the EFLAGS register.
STC Set carry flag
CLC Clear the carry flag
CMC Complement the carry flag
CLD Clear the direction flag
STD Set direction flag
LAHF Load flags into AH register
SAHF Store AH register into flags

5-8 Vol. 1
 INSTRUCTION SET SUMMARY

PUSHF/PUSHFD Push EFLAGS onto stack

POPF/POPFD Pop EFLAGS from stack
STI Set interrupt flag
CLI Clear the interrupt flag

5.1.12 Segment Register Instructions

The segment register instructions allow far pointers (segment addresses) to be
loaded into the segment registers.
LDS Load far pointer using DS
LES Load far pointer using ES
LFS Load far pointer using FS
LGS Load far pointer using GS
LSS Load far pointer using SS

5.1.13 Miscellaneous Instructions

The miscellaneous instructions provide such functions as loading an effective
address, executing a “no-operation,” and retrieving processor identification informa-
tion.
LEA Load effective address
NOP No operation
UD2 Undefined instruction
XLAT/XLATB Table lookup translation
CPUID Processor identification
MOVBE Move data after swapping data bytes

5.2 X87 FPU INSTRUCTIONS

The x87 FPU instructions are executed by the processor’s x87 FPU. These instructions
operate on floating-point, integer, and binary-coded decimal (BCD) operands. For
more detail on x87 FPU instructions, see Chapter 8, “Programming with the x87 FPU.”
These instructions are divided into the following subgroups: data transfer, load
constants, and FPU control instructions. The sections that follow introduce each
subgroup.

Vol. 1 5-9
INSTRUCTION SET SUMMARY

5.2.1 x87 FPU Data Transfer Instructions

The data transfer instructions move floating-point, integer, and BCD values between
memory and the x87 FPU registers. They also perform conditional move operations
on floating-point operands.
FLD Load floating-point value
FST Store floating-point value
FSTP Store floating-point value and pop
FILD Load integer
FIST Store integer
FISTP1 Store integer and pop
FBLD Load BCD
FBSTP Store BCD and pop
FXCH Exchange registers
FCMOVE Floating-point conditional move if equal
FCMOVNE Floating-point conditional move if not equal
FCMOVB Floating-point conditional move if below
FCMOVBE Floating-point conditional move if below or equal
FCMOVNB Floating-point conditional move if not below
FCMOVNBE Floating-point conditional move if not below or equal
FCMOVU Floating-point conditional move if unordered
FCMOVNU Floating-point conditional move if not unordered

5.2.2 x87 FPU Basic Arithmetic Instructions

The basic arithmetic instructions perform basic arithmetic operations on floating-
point and integer operands.
FADD Add floating-point
FADDP Add floating-point and pop
FIADD Add integer
FSUB Subtract floating-point
FSUBP Subtract floating-point and pop
FISUB Subtract integer
FSUBR Subtract floating-point reverse
FSUBRP Subtract floating-point reverse and pop
FISUBR Subtract integer reverse
FMUL Multiply floating-point

1. SSE3 provides an instruction FISTTP for integer conversion.

5-10 Vol. 1
 INSTRUCTION SET SUMMARY

FMULP Multiply floating-point and pop

FIMUL Multiply integer
FDIV Divide floating-point
FDIVP Divide floating-point and pop
FIDIV Divide integer
FDIVR Divide floating-point reverse
FDIVRP Divide floating-point reverse and pop
FIDIVR Divide integer reverse
FPREM Partial remainder
FPREM1 IEEE Partial remainder
FABS Absolute value
FCHS Change sign
FRNDINT Round to integer
FSCALE Scale by power of two
FSQRT Square root
FXTRACT Extract exponent and significand

5.2.3 x87 FPU Comparison Instructions

The compare instructions examine or compare floating-point or integer operands.
FCOM Compare floating-point
FCOMP Compare floating-point and pop
FCOMPP Compare floating-point and pop twice
FUCOM Unordered compare floating-point
FUCOMP Unordered compare floating-point and pop
FUCOMPP Unordered compare floating-point and pop twice
FICOM Compare integer
FICOMP Compare integer and pop
FCOMI Compare floating-point and set EFLAGS
FUCOMI Unordered compare floating-point and set EFLAGS
FCOMIP Compare floating-point, set EFLAGS, and pop
FUCOMIP Unordered compare floating-point, set EFLAGS, and pop
FTST Test floating-point (compare with 0.0)
FXAM Examine floating-point

Vol. 1 5-11
INSTRUCTION SET SUMMARY

5.2.4 x87 FPU Transcendental Instructions

The transcendental instructions perform basic trigonometric and logarithmic opera-
tions on floating-point operands.
FSIN Sine
FCOS Cosine
FSINCOS Sine and cosine
FPTAN Partial tangent
FPATAN Partial arctangent
F2XM1 2x − 1
FYL2X y∗log2x
FYL2XP1 y∗log2(x+1)

5.2.5 x87 FPU Load Constants Instructions

The load constants instructions load common constants, such as π, into the x87
floating-point registers.
FLD1 Load +1.0
FLDZ Load +0.0
FLDPI Load π
FLDL2E Load log2e
FLDLN2 Load loge2
FLDL2T Load log210
FLDLG2 Load log102

5.2.6 x87 FPU Control Instructions

The x87 FPU control instructions operate on the x87 FPU register stack and save and
restore the x87 FPU state.
FINCSTP Increment FPU register stack pointer
FDECSTP Decrement FPU register stack pointer
FFREE Free floating-point register
FINIT Initialize FPU after checking error conditions
FNINIT Initialize FPU without checking error conditions
FCLEX Clear floating-point exception flags after checking for error
conditions
FNCLEX Clear floating-point exception flags without checking for error
conditions
FSTCW Store FPU control word after checking error conditions
FNSTCW Store FPU control word without checking error conditions

5-12 Vol. 1
 INSTRUCTION SET SUMMARY

FLDCW Load FPU control word

FSTENV Store FPU environment after checking error conditions
FNSTENV Store FPU environment without checking error conditions
FLDENV Load FPU environment
FSAVE Save FPU state after checking error conditions
FNSAVE Save FPU state without checking error conditions
FRSTOR Restore FPU state
FSTSW Store FPU status word after checking error conditions
FNSTSW Store FPU status word without checking error conditions
WAIT/FWAIT Wait for FPU
FNOP FPU no operation

5.3 X87 FPU AND SIMD STATE MANAGEMENT

INSTRUCTIONS
Two state management instructions were introduced into the IA-32 architecture with
the Pentium II processor family:
FXSAVE Save x87 FPU and SIMD state
FXRSTOR Restore x87 FPU and SIMD state
Initially, these instructions operated only on the x87 FPU (and MMX) registers to
perform a fast save and restore, respectively, of the x87 FPU and MMX state. With the
introduction of SSE extensions in the Pentium III processor family, these instructions
were expanded to also save and restore the state of the XMM and MXCSR registers.
Intel 64 architecture also supports these instructions.
See Section 10.5, “FXSAVE and FXRSTOR Instructions,” for more detail.

5.4 MMX™ INSTRUCTIONS

Four extensions have been introduced into the IA-32 architecture to permit IA-32
processors to perform single-instruction multiple-data (SIMD) operations. These
extensions include the MMX technology, SSE extensions, SSE2 extensions, and SSE3
extensions. For a discussion that puts SIMD instructions in their historical context,
see Section 2.2.6, “SIMD Instructions.”
MMX instructions operate on packed byte, word, doubleword, or quadword integer
operands contained in memory, in MMX registers, and/or in general-purpose regis-
ters. For more detail on these instructions, see Chapter 9, “Programming with Intel®
MMX™ Technology.”
MMX instructions can only be executed on Intel 64 and IA-32 processors that support
the MMX technology. Support for these instructions can be detected with the CPUID
instruction. See the description of the CPUID instruction in Chapter 3, “Instruction

Vol. 1 5-13
INSTRUCTION SET SUMMARY

Set Reference, A-M,” of the Intel® 64 and IA-32 Architectures Software Developer’s
Manual, Volume 2A.
MMX instructions are divided into the following subgroups: data transfer, conversion,
packed arithmetic, comparison, logical, shift and rotate, and state management
instructions. The sections that follow introduce each subgroup.

5.4.1 MMX Data Transfer Instructions

The data transfer instructions move doubleword and quadword operands between
MMX registers and between MMX registers and memory.
MOVD Move doubleword
MOVQ Move quadword

5.4.2 MMX Conversion Instructions

The conversion instructions pack and unpack bytes, words, and doublewords
PACKSSWB Pack words into bytes with signed saturation
PACKSSDW Pack doublewords into words with signed saturation
PACKUSWB Pack words into bytes with unsigned saturation.
PUNPCKHBW Unpack high-order bytes
PUNPCKHWD Unpack high-order words
PUNPCKHDQ Unpack high-order doublewords
PUNPCKLBW Unpack low-order bytes
PUNPCKLWD Unpack low-order words
PUNPCKLDQ Unpack low-order doublewords

5.4.3 MMX Packed Arithmetic Instructions

The packed arithmetic instructions perform packed integer arithmetic on packed
byte, word, and doubleword integers.
PADDB Add packed byte integers
PADDW Add packed word integers
PADDD Add packed doubleword integers
PADDSB Add packed signed byte integers with signed saturation
PADDSW Add packed signed word integers with signed saturation
PADDUSB Add packed unsigned byte integers with unsigned saturation
PADDUSW Add packed unsigned word integers with unsigned saturation
PSUBB Subtract packed byte integers
PSUBW Subtract packed word integers

5-14 Vol. 1
 INSTRUCTION SET SUMMARY

PSUBD Subtract packed doubleword integers

PSUBSB Subtract packed signed byte integers with signed saturation
PSUBSW Subtract packed signed word integers with signed saturation
PSUBUSB Subtract packed unsigned byte integers with unsigned saturation
PSUBUSW Subtract packed unsigned word integers with unsigned
saturation
PMULHW Multiply packed signed word integers and store high result
PMULLW Multiply packed signed word integers and store low result
PMADDWD Multiply and add packed word integers

5.4.4 MMX Comparison Instructions

The compare instructions compare packed bytes, words, or doublewords.
PCMPEQB Compare packed bytes for equal
PCMPEQW Compare packed words for equal
PCMPEQD Compare packed doublewords for equal
PCMPGTB Compare packed signed byte integers for greater than
PCMPGTW Compare packed signed word integers for greater than
PCMPGTD Compare packed signed doubleword integers for greater than

5.4.5 MMX Logical Instructions

The logical instructions perform AND, AND NOT, OR, and XOR operations on quad-
word operands.
PAND Bitwise logical AND
PANDN Bitwise logical AND NOT
POR Bitwise logical OR
PXOR Bitwise logical exclusive OR

5.4.6 MMX Shift and Rotate Instructions

The shift and rotate instructions shift and rotate packed bytes, words, or double-
words, or quadwords in 64-bit operands.
PSLLW Shift packed words left logical
PSLLD Shift packed doublewords left logical
PSLLQ Shift packed quadword left logical
PSRLW Shift packed words right logical
PSRLD Shift packed doublewords right logical
PSRLQ Shift packed quadword right logical

Vol. 1 5-15
INSTRUCTION SET SUMMARY

PSRAW Shift packed words right arithmetic

PSRAD Shift packed doublewords right arithmetic

5.4.7 MMX State Management Instructions

The EMMS instruction clears the MMX state from the MMX registers.
EMMS Empty MMX state

5.5 SSE INSTRUCTIONS

SSE instructions represent an extension of the SIMD execution model introduced
with the MMX technology. For more detail on these instructions, see Chapter 10,
“Programming with Streaming SIMD Extensions (SSE).”
SSE instructions can only be executed on Intel 64 and IA-32 processors that support
SSE extensions. Support for these instructions can be detected with the CPUID
instruction. See the description of the CPUID instruction in Chapter 3, “Instruction
Set Reference, A-M,” of the Intel® 64 and IA-32 Architectures Software Developer’s
Manual, Volume 2A.
SSE instructions are divided into four subgroups (note that the first subgroup has
subordinate subgroups of its own):
• SIMD single-precision floating-point instructions that operate on the XMM
registers
• MXSCR state management instructions
• 64-bit SIMD integer instructions that operate on the MMX registers
• Cacheability control, prefetch, and instruction ordering instructions
The following sections provide an overview of these groups.

5.5.1 SSE SIMD Single-Precision Floating-Point Instructions

These instructions operate on packed and scalar single-precision floating-point
values located in XMM registers and/or memory. This subgroup is further divided into
the following subordinate subgroups: data transfer, packed arithmetic, comparison,
logical, shuffle and unpack, and conversion instructions.

5.5.1.1 SSE Data Transfer Instructions

SSE data transfer instructions move packed and scalar single-precision floating-point
operands between XMM registers and between XMM registers and memory.

5-16 Vol. 1
 INSTRUCTION SET SUMMARY

MOVAPS Move four aligned packed single-precision floating-point values

between XMM registers or between and XMM register and
memory
MOVUPS Move four unaligned packed single-precision floating-point
values between XMM registers or between and XMM register and
memory
MOVHPS Move two packed single-precision floating-point values to an
from the high quadword of an XMM register and memory
MOVHLPS Move two packed single-precision floating-point values from the
high quadword of an XMM register to the low quadword of
another XMM register
MOVLPS Move two packed single-precision floating-point values to an
from the low quadword of an XMM register and memory
MOVLHPS Move two packed single-precision floating-point values from the
low quadword of an XMM register to the high quadword of
another XMM register
MOVMSKPS Extract sign mask from four packed single-precision floating-
point values
MOVSS Move scalar single-precision floating-point value between XMM
registers or between an XMM register and memory

5.5.1.2 SSE Packed Arithmetic Instructions

SSE packed arithmetic instructions perform packed and scalar arithmetic operations
on packed and scalar single-precision floating-point operands.
ADDPS Add packed single-precision floating-point values
ADDSS Add scalar single-precision floating-point values
SUBPS Subtract packed single-precision floating-point values
SUBSS Subtract scalar single-precision floating-point values
MULPS Multiply packed single-precision floating-point values
MULSS Multiply scalar single-precision floating-point values
DIVPS Divide packed single-precision floating-point values
DIVSS Divide scalar single-precision floating-point values
RCPPS Compute reciprocals of packed single-precision floating-point
values
RCPSS Compute reciprocal of scalar single-precision floating-point
values
SQRTPS Compute square roots of packed single-precision floating-point
values
SQRTSS Compute square root of scalar single-precision floating-point
values

Vol. 1 5-17
INSTRUCTION SET SUMMARY

RSQRTPS Compute reciprocals of square roots of packed single-precision

floating-point values
RSQRTSS Compute reciprocal of square root of scalar single-precision
floating-point values
MAXPS Return maximum packed single-precision floating-point values
MAXSS Return maximum scalar single-precision floating-point values
MINPS Return minimum packed single-precision floating-point values
MINSS Return minimum scalar single-precision floating-point values

5.5.1.3 SSE Comparison Instructions

SSE compare instructions compare packed and scalar single-precision floating-point
operands.
CMPPS Compare packed single-precision floating-point values
CMPSS Compare scalar single-precision floating-point values
COMISS Perform ordered comparison of scalar single-precision floating-
point values and set flags in EFLAGS register
UCOMISS Perform unordered comparison of scalar single-precision
floating-point values and set flags in EFLAGS register

5.5.1.4 SSE Logical Instructions

SSE logical instructions perform bitwise AND, AND NOT, OR, and XOR operations on
packed single-precision floating-point operands.
ANDPS Perform bitwise logical AND of packed single-precision floating-
point values
ANDNPS Perform bitwise logical AND NOT of packed single-precision
floating-point values
ORPS Perform bitwise logical OR of packed single-precision floating-
point values
XORPS Perform bitwise logical XOR of packed single-precision floating-
point values

5.5.1.5 SSE Shuffle and Unpack Instructions

SSE shuffle and unpack instructions shuffle or interleave single-precision floating-
point values in packed single-precision floating-point operands.
SHUFPS Shuffles values in packed single-precision floating-point
operands
UNPCKHPS Unpacks and interleaves the two high-order values from two
single-precision floating-point operands

5-18 Vol. 1
 INSTRUCTION SET SUMMARY

UNPCKLPS Unpacks and interleaves the two low-order values from two
single-precision floating-point operands

5.5.1.6 SSE Conversion Instructions

SSE conversion instructions convert packed and individual doubleword integers into
packed and scalar single-precision floating-point values and vice versa.
CVTPI2PS Convert packed doubleword integers to packed single-precision
floating-point values
CVTSI2SS Convert doubleword integer to scalar single-precision floating-
point value
CVTPS2PI Convert packed single-precision floating-point values to packed
doubleword integers
CVTTPS2PI Convert with truncation packed single-precision floating-point
values to packed doubleword integers
CVTSS2SI Convert a scalar single-precision floating-point value to a
doubleword integer
CVTTSS2SI Convert with truncation a scalar single-precision floating-point
value to a scalar doubleword integer

5.5.2 SSE MXCSR State Management Instructions

MXCSR state management instructions allow saving and restoring the state of the
MXCSR control and status register.
LDMXCSR Load MXCSR register
STMXCSR Save MXCSR register state

5.5.3 SSE 64-Bit SIMD Integer Instructions

These SSE 64-bit SIMD integer instructions perform additional operations on packed
bytes, words, or doublewords contained in MMX registers. They represent enhance-
ments to the MMX instruction set described in Section 5.4, “MMX™ Instructions.”
PAVGB Compute average of packed unsigned byte integers
PAVGW Compute average of packed unsigned word integers
PEXTRW Extract word
PINSRW Insert word
PMAXUB Maximum of packed unsigned byte integers
PMAXSW Maximum of packed signed word integers
PMINUB Minimum of packed unsigned byte integers
PMINSW Minimum of packed signed word integers

Vol. 1 5-19
INSTRUCTION SET SUMMARY

PMOVMSKB Move byte mask

PMULHUW Multiply packed unsigned integers and store high result
PSADBW Compute sum of absolute differences
PSHUFW Shuffle packed integer word in MMX register

5.5.4 SSE Cacheability Control, Prefetch, and Instruction Ordering

Instructions
The cacheability control instructions provide control over the caching of non-
temporal data when storing data from the MMX and XMM registers to memory. The
PREFETCHh allows data to be prefetched to a selected cache level. The SFENCE
instruction controls instruction ordering on store operations.
MASKMOVQ Non-temporal store of selected bytes from an MMX register into
memory
MOVNTQ Non-temporal store of quadword from an MMX register into
memory
MOVNTPS Non-temporal store of four packed single-precision floating-
point values from an XMM register into memory
PREFETCHh Load 32 or more of bytes from memory to a selected level of the
processor’s cache hierarchy
SFENCE Serializes store operations

5.6 SSE2 INSTRUCTIONS

SSE2 extensions represent an extension of the SIMD execution model introduced
with MMX technology and the SSE extensions. SSE2 instructions operate on packed
double-precision floating-point operands and on packed byte, word, doubleword, and
quadword operands located in the XMM registers. For more detail on these instruc-
tions, see Chapter 11, “Programming with Streaming SIMD Extensions 2 (SSE2).”
SSE2 instructions can only be executed on Intel 64 and IA-32 processors that
support the SSE2 extensions. Support for these instructions can be detected with the
CPUID instruction. See the description of the CPUID instruction in Chapter 3,
“Instruction Set Reference, A-M,” of the Intel® 64 and IA-32 Architectures Software
Developer’s Manual, Volume 2A.
These instructions are divided into four subgroups (note that the first subgroup is
further divided into subordinate subgroups):
• Packed and scalar double-precision floating-point instructions
• Packed single-precision floating-point conversion instructions
• 128-bit SIMD integer instructions
• Cacheability-control and instruction ordering instructions

5-20 Vol. 1
 INSTRUCTION SET SUMMARY

The following sections give an overview of each subgroup.

5.6.1 SSE2 Packed and Scalar Double-Precision Floating-Point

Instructions
SSE2 packed and scalar double-precision floating-point instructions are divided into
the following subordinate subgroups: data movement, arithmetic, comparison,
conversion, logical, and shuffle operations on double-precision floating-point oper-
ands. These are introduced in the sections that follow.

5.6.1.1 SSE2 Data Movement Instructions

SSE2 data movement instructions move double-precision floating-point data
between XMM registers and between XMM registers and memory.
MOVAPD Move two aligned packed double-precision floating-point values
between XMM registers or between and XMM register and
memory
MOVUPD Move two unaligned packed double-precision floating-point
values between XMM registers or between and XMM register and
memory
MOVHPD Move high packed double-precision floating-point value to an
from the high quadword of an XMM register and memory
MOVLPD Move low packed single-precision floating-point value to an from
the low quadword of an XMM register and memory
MOVMSKPD Extract sign mask from two packed double-precision floating-
point values
MOVSD Move scalar double-precision floating-point value between XMM
registers or between an XMM register and memory

5.6.1.2 SSE2 Packed Arithmetic Instructions

The arithmetic instructions perform addition, subtraction, multiply, divide, square
root, and maximum/minimum operations on packed and scalar double-precision
floating-point operands.
ADDPD Add packed double-precision floating-point values
ADDSD Add scalar double precision floating-point values
SUBPD Subtract scalar double-precision floating-point values
SUBSD Subtract scalar double-precision floating-point values
MULPD Multiply packed double-precision floating-point values
MULSD Multiply scalar double-precision floating-point values
DIVPD Divide packed double-precision floating-point values
DIVSD Divide scalar double-precision floating-point values

Vol. 1 5-21
INSTRUCTION SET SUMMARY

SQRTPD Compute packed square roots of packed double-precision

floating-point values
SQRTSD Compute scalar square root of scalar double-precision floating-
point values
MAXPD Return maximum packed double-precision floating-point values
MAXSD Return maximum scalar double-precision floating-point values
MINPD Return minimum packed double-precision floating-point values
MINSD Return minimum scalar double-precision floating-point values

5.6.1.3 SSE2 Logical Instructions

SSE2 logical instructions preform AND, AND NOT, OR, and XOR operations on packed
double-precision floating-point values.
ANDPD Perform bitwise logical AND of packed double-precision floating-
point values
ANDNPD Perform bitwise logical AND NOT of packed double-precision
floating-point values
ORPD Perform bitwise logical OR of packed double-precision floating-
point values
XORPD Perform bitwise logical XOR of packed double-precision floating-
point values

5.6.1.4 SSE2 Compare Instructions

SSE2 compare instructions compare packed and scalar double-precision floating-
point values and return the results of the comparison either to the destination
operand or to the EFLAGS register.
CMPPD Compare packed double-precision floating-point values
CMPSD Compare scalar double-precision floating-point values
COMISD Perform ordered comparison of scalar double-precision floating-
point values and set flags in EFLAGS register
UCOMISD Perform unordered comparison of scalar double-precision
floating-point values and set flags in EFLAGS register.

5.6.1.5 SSE2 Shuffle and Unpack Instructions

SSE2 shuffle and unpack instructions shuffle or interleave double-precision floating-
point values in packed double-precision floating-point operands.
SHUFPD Shuffles values in packed double-precision floating-point
operands
UNPCKHPD Unpacks and interleaves the high values from two packed
double-precision floating-point operands

5-22 Vol. 1
 INSTRUCTION SET SUMMARY

UNPCKLPD Unpacks and interleaves the low values from two packed
double-precision floating-point operands

5.6.1.6 SSE2 Conversion Instructions

SSE2 conversion instructions convert packed and individual doubleword integers into
packed and scalar double-precision floating-point values and vice versa. They also
convert between packed and scalar single-precision and double-precision floating-
point values.
CVTPD2PI Convert packed double-precision floating-point values to packed
doubleword integers.
CVTTPD2PI Convert with truncation packed double-precision floating-point
values to packed doubleword integers
CVTPI2PD Convert packed doubleword integers to packed double-precision
floating-point values
CVTPD2DQ Convert packed double-precision floating-point values to packed
doubleword integers
CVTTPD2DQ Convert with truncation packed double-precision floating-point
values to packed doubleword integers
CVTDQ2PD Convert packed doubleword integers to packed double-precision
floating-point values
CVTPS2PD Convert packed single-precision floating-point values to packed
double-precision floating-point values
CVTPD2PS Convert packed double-precision floating-point values to packed
single-precision floating-point values
CVTSS2SD Convert scalar single-precision floating-point values to scalar
double-precision floating-point values
CVTSD2SS Convert scalar double-precision floating-point values to scalar
single-precision floating-point values
CVTSD2SI Convert scalar double-precision floating-point values to a
doubleword integer
CVTTSD2SI Convert with truncation scalar double-precision floating-point
values to scalar doubleword integers
CVTSI2SD Convert doubleword integer to scalar double-precision floating-
point value

5.6.2 SSE2 Packed Single-Precision Floating-Point Instructions

SSE2 packed single-precision floating-point instructions perform conversion opera-
tions on single-precision floating-point and integer operands. These instructions
represent enhancements to the SSE single-precision floating-point instructions.

Vol. 1 5-23
INSTRUCTION SET SUMMARY

CVTDQ2PS Convert packed doubleword integers to packed single-precision

floating-point values
CVTPS2DQ Convert packed single-precision floating-point values to packed
doubleword integers
CVTTPS2DQ Convert with truncation packed single-precision floating-point
values to packed doubleword integers

5.6.3 SSE2 128-Bit SIMD Integer Instructions

SSE2 SIMD integer instructions perform additional operations on packed words,
doublewords, and quadwords contained in XMM and MMX registers.
MOVDQA Move aligned double quadword.
MOVDQU Move unaligned double quadword
MOVQ2DQ Move quadword integer from MMX to XMM registers
MOVDQ2Q Move quadword integer from XMM to MMX registers
PMULUDQ Multiply packed unsigned doubleword integers
PADDQ Add packed quadword integers
PSUBQ Subtract packed quadword integers
PSHUFLW Shuffle packed low words
PSHUFHW Shuffle packed high words
PSHUFD Shuffle packed doublewords
PSLLDQ Shift double quadword left logical
PSRLDQ Shift double quadword right logical
PUNPCKHQDQ Unpack high quadwords
PUNPCKLQDQ Unpack low quadwords

5.6.4 SSE2 Cacheability Control and Ordering Instructions

SSE2 cacheability control instructions provide additional operations for caching of
non-temporal data when storing data from XMM registers to memory. LFENCE and
MFENCE provide additional control of instruction ordering on store operations.
CLFLUSH Flushes and invalidates a memory operand and its associated
cache line from all levels of the processor’s cache hierarchy
LFENCE Serializes load operations
MFENCE Serializes load and store operations
PAUSE Improves the performance of “spin-wait loops”
MASKMOVDQU Non-temporal store of selected bytes from an XMM register into
memory
MOVNTPD Non-temporal store of two packed double-precision floating-
point values from an XMM register into memory

5-24 Vol. 1
 INSTRUCTION SET SUMMARY

MOVNTDQ Non-temporal store of double quadword from an XMM register

into memory
MOVNTI Non-temporal store of a doubleword from a general-purpose
register into memory

5.7 SSE3 INSTRUCTIONS

The SSE3 extensions offers 13 instructions that accelerate performance of Streaming
SIMD Extensions technology, Streaming SIMD Extensions 2 technology, and x87-FP
math capabilities. These instructions can be grouped into the following categories:
• One x87FPU instruction used in integer conversion
• One SIMD integer instruction that addresses unaligned data loads
• Two SIMD floating-point packed ADD/SUB instructions
• Four SIMD floating-point horizontal ADD/SUB instructions
• Three SIMD floating-point LOAD/MOVE/DUPLICATE instructions
• Two thread synchronization instructions
SSE3 instructions can only be executed on Intel 64 and IA-32 processors that
support SSE3 extensions. Support for these instructions can be detected with the
CPUID instruction. See the description of the CPUID instruction in Chapter 3,
“Instruction Set Reference, A-M,” of the Intel® 64 and IA-32 Architectures Software
Developer’s Manual, Volume 2A.
The sections that follow describe each subgroup.

5.7.1 SSE3 x87-FP Integer Conversion Instruction

FISTTP Behaves like the FISTP instruction but uses truncation, irrespec-
tive of the rounding mode specified in the floating-point control
word (FCW)

5.7.2 SSE3 Specialized 128-bit Unaligned Data Load Instruction

LDDQU Special 128-bit unaligned load designed to avoid cache line
splits

5.7.3 SSE3 SIMD Floating-Point Packed ADD/SUB Instructions

ADDSUBPS Performs single-precision addition on the second and fourth
pairs of 32-bit data elements within the operands; single-preci-
sion subtraction on the first and third pairs

Vol. 1 5-25
INSTRUCTION SET SUMMARY

ADDSUBPD Performs double-precision addition on the second pair of quad-

words, and double-precision subtraction on the first pair

5.7.4 SSE3 SIMD Floating-Point Horizontal ADD/SUB Instructions

HADDPS Performs a single-precision addition on contiguous data
elements. The first data element of the result is obtained by
adding the first and second elements of the first operand; the
second element by adding the third and fourth elements of the
first operand; the third by adding the first and second elements
of the second operand; and the fourth by adding the third and
fourth elements of the second operand.
HSUBPS Performs a single-precision subtraction on contiguous data
elements. The first data element of the result is obtained by
subtracting the second element of the first operand from the
first element of the first operand; the second element by
subtracting the fourth element of the first operand from the third
element of the first operand; the third by subtracting the second
element of the second operand from the first element of the
second operand; and the fourth by subtracting the fourth
element of the second operand from the third element of the
second operand.
HADDPD Performs a double-precision addition on contiguous data
elements. The first data element of the result is obtained by
adding the first and second elements of the first operand; the
second element by adding the first and second elements of the
second operand.
HSUBPD Performs a double-precision subtraction on contiguous data
elements. The first data element of the result is obtained by
subtracting the second element of the first operand from the
first element of the first operand; the second element by
subtracting the second element of the second operand from the
first element of the second operand.

5.7.5 SSE3 SIMD Floating-Point LOAD/MOVE/DUPLICATE

Instructions
MOVSHDUP Loads/moves 128 bits; duplicating the second and fourth 32-bit
data elements
MOVSLDUP Loads/moves 128 bits; duplicating the first and third 32-bit data
elements
MOVDDUP Loads/moves 64 bits (bits[63:0] if the source is a register) and
returns the same 64 bits in both the lower and upper halves of

5-26 Vol. 1
 INSTRUCTION SET SUMMARY

the 128-bit result register; duplicates the 64 bits from the

source

5.7.6 SSE3 Agent Synchronization Instructions

MONITOR Sets up an address range used to monitor write-back stores
MWAIT Enables a logical processor to enter into an optimized state while
waiting for a write-back store to the address range set up by the
MONITOR instruction

5.8 SUPPLEMENTAL STREAMING SIMD EXTENSIONS 3

(SSSE3) INSTRUCTIONS
SSSE3 provide 32 instructions (represented by 14 mnemonics) to accelerate compu-
tations on packed integers. These include:
• Twelve instructions that perform horizontal addition or subtraction operations.
• Six instructions that evaluate absolute values.
• Two instructions that perform multiply and add operations and speed up the
evaluation of dot products.
• Two instructions that accelerate packed-integer multiply operations and produce
integer values with scaling.
• Two instructions that perform a byte-wise, in-place shuffle according to the
second shuffle control operand.
• Six instructions that negate packed integers in the destination operand if the
signs of the corresponding element in the source operand is less than zero.
• Two instructions that align data from the composite of two operands.
SSSE3 instructions can only be executed on Intel 64 and IA-32 processors that
support SSSE3 extensions. Support for these instructions can be detected with the
CPUID instruction. See the description of the CPUID instruction in Chapter 3,
“Instruction Set Reference, A-M,” of the Intel® 64 and IA-32 Architectures Software
Developer’s Manual, Volume 2A.
The sections that follow describe each subgroup.

5.8.1 Horizontal Addition/Subtraction

PHADDW Adds two adjacent, signed 16-bit integers horizontally from the
source and destination operands and packs the signed 16-bit
results to the destination operand.

Vol. 1 5-27
INSTRUCTION SET SUMMARY

PHADDSW Adds two adjacent, signed 16-bit integers horizontally from the
source and destination operands and packs the signed, satu-
rated 16-bit results to the destination operand.
PHADDD Adds two adjacent, signed 32-bit integers horizontally from the
source and destination operands and packs the signed 32-bit
results to the destination operand.
PHSUBW Performs horizontal subtraction on each adjacent pair of 16-bit
signed integers by subtracting the most significant word from
the least significant word of each pair in the source and destina-
tion operands. The signed 16-bit results are packed and written
to the destination operand.
PHSUBSW Performs horizontal subtraction on each adjacent pair of 16-bit
signed integers by subtracting the most significant word from
the least significant word of each pair in the source and destina-
tion operands. The signed, saturated 16-bit results are packed
and written to the destination operand.
PHSUBD Performs horizontal subtraction on each adjacent pair of 32-bit
signed integers by subtracting the most significant doubleword
from the least significant double word of each pair in the source
and destination operands. The signed 32-bit results are packed
and written to the destination operand.

5.8.2 Packed Absolute Values

PABSB Computes the absolute value of each signed byte data element.
PABSW Computes the absolute value of each signed 16-bit data
element.
PABSD Computes the absolute value of each signed 32-bit data
element.

5.8.3 Multiply and Add Packed Signed and Unsigned Bytes

PMADDUBSW Multiplies each unsigned byte value with the corresponding
signed byte value to produce an intermediate, 16-bit signed
integer. Each adjacent pair of 16-bit signed values are added
horizontally. The signed, saturated 16-bit results are packed to
the destination operand.

5.8.4 Packed Multiply High with Round and Scale

PMULHRSW Multiplies vertically each signed 16-bit integer from the destina-
tion operand with the corresponding signed 16-bit integer of the
source operand, producing intermediate, signed 32-bit integers.
Each intermediate 32-bit integer is truncated to the 18 most

5-28 Vol. 1
 INSTRUCTION SET SUMMARY

significant bits. Rounding is always performed by adding 1 to the

least significant bit of the 18-bit intermediate result. The final
result is obtained by selecting the 16 bits immediately to the
right of the most significant bit of each 18-bit intermediate
result and packed to the destination operand.

5.8.5 Packed Shuffle Bytes

PSHUFB Permutes each byte in place, according to a shuffle control
mask. The least significant three or four bits of each shuffle
control byte of the control mask form the shuffle index. The
shuffle mask is unaffected. If the most significant bit (bit 7) of a
shuffle control byte is set, the constant zero is written in the
result byte.

5.8.6 Packed Sign

PSIGNB/W/D Negates each signed integer element of the destination operand
if the sign of the corresponding data element in the source
operand is less than zero.

5.8.7 Packed Align Right

PALIGNR Source operand is appended after the destination operand
forming an intermediate value of twice the width of an operand.
The result is extracted from the intermediate value into the
destination operand by selecting the 128 bit or 64 bit value that
are right-aligned to the byte offset specified by the immediate
value.

5.9 SSE4 INSTRUCTIONS

Intel® Streaming SIMD Extensions 4 (SSE4) introduces 54 new instructions. 47 of
the SSE4 instructions are referred to as SSE4.1 in this document, 7 new SSE4
instructions are referred to as SSE4.2.
SSE4.1 is targeted to improve the performance of media, imaging, and 3D work-
loads. SSE4.1 adds instructions that improve compiler vectorization and significantly
increase support for packed dword computation. The technology also provides a hint
that can improve memory throughput when reading from uncacheable WC memory
type.
The 47 SSE4.1 instructions include:
• Two instructions perform packed dword multiplies.

Vol. 1 5-29
INSTRUCTION SET SUMMARY

• Two instructions perform floating-point dot products with input/output selects.

• One instruction performs a load with a streaming hint.
• Six instructions simplify packed blending.
• Eight instructions expand support for packed integer MIN/MAX.
• Four instructions support floating-point round with selectable rounding mode and
precision exception override.
• Seven instructions improve data insertion and extractions from XMM registers
• Twelve instructions improve packed integer format conversions (sign and zero
extensions).
• One instruction improves SAD (sum absolute difference) generation for small
block sizes.
• One instruction aids horizontal searching operations.
• One instruction improves masked comparisons.
• One instruction adds qword packed equality comparisons.
• One instruction adds dword packing with unsigned saturation.
The seven SSE4.2 instructions include:
• String and text processing that can take advantage of single-instruction multiple-
data programming techniques.
• Application-targeted accelerator (ATA) instructions.
• A SIMD integer instruction that enhances the capability of the 128-bit integer
SIMD capability in SSE4.1.

5.10 SSE4.1 INSTRUCTIONS

SSE4.1 instructions can use an XMM register as a source or destination. Program-
ming SSE4.1 is similar to programming 128-bit Integer SIMD and floating-point
SIMD instructions in SSE/SSE2/SSE3/SSSE3. SSE4.1 does not provide any 64-bit
integer SIMD instructions operating on MMX registers. The sections that follow
describe each subgroup.

5.10.1 Dword Multiply Instructions

PMULLD Returns four lower 32-bits of the 64-bit results of signed 32-bit
integer multiplies.
PMULDQ Returns two 64-bit signed result of signed 32-bit integer multi-
plies.

5-30 Vol. 1
 INSTRUCTION SET SUMMARY

5.10.2 Floating-Point Dot Product Instructions

DPPD Perform double-precision dot product for up to 2 elements and
broadcast.
DPPS Perform single-precision dot products for up to 4 elements and
broadcast

5.10.3 Streaming Load Hint Instruction

MOVNTDQA Provides a non-temporal hint that can cause adjacent 16-byte
items within an aligned 64-byte region (a streaming line) to be
fetched and held in a small set of temporary buffers (“streaming
load buffers”). Subsequent streaming loads to other aligned 16-
byte items in the same streaming line may be supplied from the
streaming load buffer and can improve throughput.

5.10.4 Packed Blending Instructions

BLENDPD Conditionally copies specified double-precision floating-point
data elements in the source operand to the corresponding data
elements in the destination, using an immediate byte control.
BLENDPS Conditionally copies specified single-precision floating-point
data elements in the source operand to the corresponding data
elements in the destination, using an immediate byte control.
BLENDVPD Conditionally copies specified double-precision floating-point
data elements in the source operand to the corresponding data
elements in the destination, using an implied mask.
BLENDVPS Conditionally copies specified single-precision floating-point
data elements in the source operand to the corresponding data
elements in the destination, using an implied mask.
PBLENDVB Conditionally copies specified byte elements in the source
operand to the corresponding elements in the destination, using
an implied mask.
PBLENDW Conditionally copies specified word elements in the source
operand to the corresponding elements in the destination, using
an immediate byte control.

5.10.5 Packed Integer MIN/MAX Instructions

PMINUW Compare packed unsigned word integers.
PMINUD Compare packed unsigned dword integers.
PMINSB Compare packed signed byte integers.
PMINSD Compare packed signed dword integers.

Vol. 1 5-31
INSTRUCTION SET SUMMARY

PMAXUW Compare packed unsigned word integers.

PMAXUD Compare packed unsigned dword integers.
PMAXSB Compare packed signed byte integers.
PMAXSD Compare packed signed dword integers.

5.10.6 Floating-Point Round Instructions with Selectable Rounding

Mode
ROUNDPS Round packed single precision floating-point values into integer
values and return rounded floating-point values.
ROUNDPD Round packed double precision floating-point values into integer
values and return rounded floating-point values.
ROUNDSS Round the low packed single precision floating-point value into
an integer value and return a rounded floating-point value.
ROUNDSD Round the low packed double precision floating-point value into
an integer value and return a rounded floating-point value.

5.10.7 Insertion and Extractions from XMM Registers

EXTRACTPS Extracts a single-precision floating-point value from a specified
offset in an XMM register and stores the result to memory or a
general-purpose register
INSERTPS Inserts a single-precision floating-point value from either a 32-
bit memory location or selected from a specified offset in an
XMM register to a specified offset in the destination XMM
register. In addition, INSERTPS allows zeroing out selected data
elements in the destination, using a mask.
PINSRB Insert a byte value from a register or memory into an XMM
register
PINSRD Insert a dword value from 32-bit register or memory into an
XMM register
PINSRQ Insert a qword value from 64-bit register or memory into an
XMM register
PEXTRB Extract a byte from an XMM register and insert the value into a
general-purpose register or memory
PEXTRW Extract a word from an XMM register and insert the value into a
general-purpose register or memory
PEXTRD Extract a dword from an XMM register and insert the value into a
general-purpose register or memory
PEXTRQ Extract a qword from an XMM register and insert the value into a
general-purpose register or memory

5-32 Vol. 1
 INSTRUCTION SET SUMMARY

5.10.8 Packed Integer Format Conversions

PMOVSXBW Sign extend the lower 8-bit integer of each packed word
element into packed signed word integers.
PMOVZXBW Zero extend the lower 8-bit integer of each packed word
element into packed signed word integers.
PMOVSXBD Sign extend the lower 8-bit integer of each packed dword
element into packed signed dword integers.
PMOVZXBD Zero extend the lower 8-bit integer of each packed dword
element into packed signed dword integers.
PMOVSXWD Sign extend the lower 16-bit integer of each packed dword
element into packed signed dword integers.
PMOVZXWD Zero extend the lower 16-bit integer of each packed dword
element into packed signed dword integers..
PMOVSXBQ Sign extend the lower 8-bit integer of each packed qword
element into packed signed qword integers.
PMOVZXBQ Zero extend the lower 8-bit integer of each packed qword
element into packed signed qword integers.
PMOVSXWQ Sign extend the lower 16-bit integer of each packed qword
element into packed signed qword integers.
PMOVZXWQ Zero extend the lower 16-bit integer of each packed qword
element into packed signed qword integers.
PMOVSXDQ Sign extend the lower 32-bit integer of each packed qword
element into packed signed qword integers.
PMOVZXDQ Zero extend the lower 32-bit integer of each packed qword
element into packed signed qword integers.

5.10.9 Improved Sums of Absolute Differences (SAD) for 4-Byte

Blocks
MPSADBW Performs eight 4-byte wide Sum of Absolute Differences opera-
tions to produce eight word integers.

5.10.10 Horizontal Search

PHMINPOSUW Finds the value and location of the minimum unsigned word
from one of 8 horizontally packed unsigned words. The resulting
value and location (offset within the source) are packed into the
low dword of the destination XMM register.

Vol. 1 5-33
INSTRUCTION SET SUMMARY

5.10.11 Packed Test

PTEST Performs a logical AND between the destination with this mask
and sets the ZF flag if the result is zero. The CF flag (zero for
TEST) is set if the inverted mask AND’d with the destination is all
zero

5.10.12 Packed Qword Equality Comparisons

PCMPEQQ 128-bit packed qword equality test

5.10.13 Dword Packing With Unsigned Saturation

PACKUSDW PACKUSDW packs dword to word with unsigned saturation

5.11 SSE4.2 INSTRUCTION SET

Five of the seven SSE4.2 instructions can use an XMM register as a source or desti-
nation. These include four text/string processing instructions and one packed quad-
word compare SIMD instruction. Programming these five SSE4.2 instructions is
similar to programming 128-bit Integer SIMD in SSE2/SSSE3. SSE4.2 does not
provide any 64-bit integer SIMD instructions.
The remaining two SSE4.2 instructions uses general-purpose registers to perform
accelerated processing functions in specific application areas.
The sections that follow describe each subgroup.

5.11.1 String and Text Processing Instructions

PCMPESTRI Packed compare explicit-length strings, return index in ECX/RCX
PCMPESTRM Packed compare explicit-length strings, return mask in XMM0
PCMPISTRI Packed compare implicit-length strings, return index in ECX/RCX
PCMPISTRM Packed compare implicit-length strings, return mask in XMM0

5.11.2 Packed Comparison SIMD integer Instruction

PCMPGTQ Performs logical compare of greater-than on packed integer
quadwords.

5-34 Vol. 1
 INSTRUCTION SET SUMMARY

5.11.3 Application-Targeted Accelerator Instructions

CRC32 Provides hardware acceleration to calculate cyclic redundancy
checks for fast and efficient implementation of data integrity
protocols.
POPCNT This instruction calculates of number of bits set to 1 in the
second operand (source) and returns the count in the first
operand (a destination register)

5.12 SYSTEM INSTRUCTIONS

The following system instructions are used to control those functions of the processor
that are provided to support for operating systems and executives.
LGDT Load global descriptor table (GDT) register
SGDT Store global descriptor table (GDT) register
LLDT Load local descriptor table (LDT) register
SLDT Store local descriptor table (LDT) register
LTR Load task register
STR Store task register
LIDT Load interrupt descriptor table (IDT) register
SIDT Store interrupt descriptor table (IDT) register
MOV Load and store control registers
LMSW Load machine status word
SMSW Store machine status word
CLTS Clear the task-switched flag
ARPL Adjust requested privilege level
LAR Load access rights
LSL Load segment limit
VERR Verify segment for reading
VERW Verify segment for writing
MOV Load and store debug registers
INVD Invalidate cache, no writeback
WBINVD Invalidate cache, with writeback
INVLPG Invalidate TLB Entry
LOCK (prefix) Lock Bus
HLT Halt processor
RSM Return from system management mode (SMM)
RDMSR Read model-specific register
WRMSR Write model-specific register

Vol. 1 5-35
INSTRUCTION SET SUMMARY

RDPMC Read performance monitoring counters

RDTSC Read time stamp counter
RDTSCP Read time stamp counter and processor ID
SYSENTER Fast System Call, transfers to a flat protected mode kernel at
CPL = 0
SYSEXIT Fast System Call, transfers to a flat protected mode kernel at
CPL = 3
XSAVE Save processor extended states to memory
XRSTOR Restore processor extended states from memory
XGETBV Reads the state of an extended control register
XSETBV Writes the state of an extended control register

5.13 64-BIT MODE INSTRUCTIONS

The following instructions are introduced in 64-bit mode. This mode is a sub-mode of
IA-32e mode.
CDQE Convert doubleword to quadword
CMPSQ Compare string operands
CMPXCHG16B Compare RDX:RAX with m128
LODSQ Load qword at address (R)SI into RAX
MOVSQ Move qword from address (R)SI to (R)DI
MOVZX (64-bits) Move doubleword to quadword, zero-extension
STOSQ Store RAX at address RDI
SWAPGS Exchanges current GS base register value with value in MSR
address C0000102H
SYSCALL Fast call to privilege level 0 system procedures
SYSRET Return from fast system call

5.14 VIRTUAL-MACHINE EXTENSIONS

The behavior of the VMCS-maintenance instructions is summarized below:
VMPTRLD Takes a single 64-bit source operand in memory. It makes the
referenced VMCS active and current.
VMPTRST Takes a single 64-bit destination operand that is in memory.
Current-VMCS pointer is stored into the destination operand.
VMCLEAR Takes a single 64-bit operand in memory. The instruction sets
the launch state of the VMCS referenced by the operand to
“clear”, renders that VMCS inactive, and ensures that data for

5-36 Vol. 1
 INSTRUCTION SET SUMMARY

the VMCS have been written to the VMCS-data area in the refer-
enced VMCS region.
VMREAD Reads a component from the VMCS (the encoding of that field is
given in a register operand) and stores it into a destination
operand.
VMWRITE Writes a component to the VMCS (the encoding of that field is
given in a register operand) from a source operand.
The behavior of the VMX management instructions is summarized below:
VMCALL Allows a guest in VMX non-root operation to call the VMM for
service. A VM exit occurs, transferring control to the VMM.
VMLAUNCH Launches a virtual machine managed by the VMCS. A VM entry
occurs, transferring control to the VM.
VMRESUME Resumes a virtual machine managed by the VMCS. A VM entry
occurs, transferring control to the VM.
VMXOFF Causes the processor to leave VMX operation.
VMXON Takes a single 64-bit source operand in memory. It causes a
logical processor to enter VMX root operation and to use the
memory referenced by the operand to support VMX operation.
INVEPT Invalidate cached Extended Page Table (EPT) mappings in the
processor to synchronize address translation in virtual machines
with memory-resident EPT pages.
INVVPID Invalidate cached mappings of address translation based on the
Virtual Processor ID (VPID).

5.15 SAFER MODE EXTENSIONS

The behavior of the GETSEC instruction leaves of the Safer Mode Extensions (SMX)
are summarized below:
GETSEC[CAPABILITIES]Returns the available leaf functions of the GETSEC instruc-
tion.
GETSEC[ENTERACCS] Loads an authenticated code chipset module and enters
authenticated code execution mode.
GETSEC[EXITAC] Exits authenticated code execution mode.
GETSEC[SENTER] Establishes a Measured Launched Environment (MLE) which has
its dynamic root of trust anchored to a chipset supporting Intel
Trusted Execution Technology.
GETSEC[SEXIT] Exits the MLE.
GETSEC[PARAMETERS]Returns SMX related parameter information.
GETSEC[SMCRTL] SMX mode control.
GETSEC[WAKEUP] Wakes up sleeping logical processors inside an MLE.

Vol. 1 5-37
INSTRUCTION SET SUMMARY

5-38 Vol. 1
 CHAPTER 6
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

This chapter describes the facilities in the Intel 64 and IA-32 architectures for
executing calls to procedures or subroutines. It also describes how interrupts and
exceptions are handled from the perspective of an application programmer.

6.1 PROCEDURE CALL TYPES

The processor supports procedure calls in the following two different ways:
• CALL and RET instructions.
• ENTER and LEAVE instructions, in conjunction with the CALL and RET
instructions.
Both of these procedure call mechanisms use the procedure stack, commonly
referred to simply as “the stack,” to save the state of the calling procedure, pass
parameters to the called procedure, and store local variables for the currently
executing procedure.
The processor’s facilities for handling interrupts and exceptions are similar to those
used by the CALL and RET instructions.

6.2 STACKS
The stack (see Figure 6-1) is a contiguous array of memory locations. It is contained
in a segment and identified by the segment selector in the SS register. When using
the flat memory model, the stack can be located anywhere in the linear address
space for the program. A stack can be up to 4 GBytes long, the maximum size of a
segment.
Items are placed on the stack using the PUSH instruction and removed from the
stack using the POP instruction. When an item is pushed onto the stack, the
processor decrements the ESP register, then writes the item at the new top of stack.
When an item is popped off the stack, the processor reads the item from the top of
stack, then increments the ESP register. In this manner, the stack grows down in
memory (towards lesser addresses) when items are pushed on the stack and shrinks
up (towards greater addresses) when the items are popped from the stack.
A program or operating system/executive can set up many stacks. For example, in
multitasking systems, each task can be given its own stack. The number of stacks in
a system is limited by the maximum number of segments and the available physical
memory.

Vol. 2 6-1
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

When a system sets up many stacks, only one stack—the current stack—is avail-
able at a time. The current stack is the one contained in the segment referenced by
the SS register.

Stack Segment
Bottom of Stack
(Initial ESP Value)

Local Variables
for Calling
Procedure The Stack Can Be
16 or 32 Bits Wide

Parameters
Passed to The EBP register is
Called typically set to point
Procedure to the return
instruction pointer.
Frame Boundary
Return Instruction EBP Register
Pointer
ESP Register
Top of Stack

Pushes Move the Pops Move the

Top Of Stack to Top Of Stack to
Lower Addresses Higher Addresses

Figure 6-1. Stack Structure

The processor references the SS register automatically for all stack operations. For
example, when the ESP register is used as a memory address, it automatically points
to an address in the current stack. Also, the CALL, RET, PUSH, POP, ENTER, and
LEAVE instructions all perform operations on the current stack.

6.2.1 Setting Up a Stack

To set a stack and establish it as the current stack, the program or operating
system/executive must do the following:
1. Establish a stack segment.
2. Load the segment selector for the stack segment into the SS register using a
MOV, POP, or LSS instruction.

6-2 Vol. 2
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

3. Load the stack pointer for the stack into the ESP register using a MOV, POP, or
LSS instruction. The LSS instruction can be used to load the SS and ESP registers
in one operation.
See “Segment Descriptors” in of the Intel® 64 and IA-32 Architectures Software
Developer’s Manual, Volume 3A, for information on how to set up a segment
descriptor and segment limits for a stack segment.

6.2.2 Stack Alignment

The stack pointer for a stack segment should be aligned on 16-bit (word) or 32-bit
(double-word) boundaries, depending on the width of the stack segment. The D flag
in the segment descriptor for the current code segment sets the stack-segment width
(see “Segment Descriptors” in Chapter 3, “Protected-Mode Memory Management,” of
the Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volume 3A).
The PUSH and POP instructions use the D flag to determine how much to decrement
or increment the stack pointer on a push or pop operation, respectively. When the
stack width is 16 bits, the stack pointer is incremented or decremented in 16-bit
increments; when the width is 32 bits, the stack pointer is incremented or decre-
mented in 32-bit increments. Pushing a 16-bit value onto a 32-bit wide stack can
result in stack misaligned (that is, the stack pointer is not aligned on a doubleword
boundary). One exception to this rule is when the contents of a segment register (a
16-bit segment selector) are pushed onto a 32-bit wide stack. Here, the processor
automatically aligns the stack pointer to the next 32-bit boundary.
The processor does not check stack pointer alignment. It is the responsibility of the
programs, tasks, and system procedures running on the processor to maintain
proper alignment of stack pointers. Misaligning a stack pointer can cause serious
performance degradation and in some instances program failures.

6.2.3 Address-Size Attributes for Stack Accesses

Instructions that use the stack implicitly (such as the PUSH and POP instructions)
have two address-size attributes each of either 16 or 32 bits. This is because they
always have the implicit address of the top of the stack, and they may also have an
explicit memory address (for example, PUSH Array1[EBX]). The attribute of the
explicit address is determined by the D flag of the current code segment and the
presence or absence of the 67H address-size prefix.
The address-size attribute of the top of the stack determines whether SP or ESP is
used for the stack access. Stack operations with an address-size attribute of 16 use
the 16-bit SP stack pointer register and can use a maximum stack address of FFFFH;
stack operations with an address-size attribute of 32 bits use the 32-bit ESP register
and can use a maximum address of FFFFFFFFH. The default address-size attribute for
data segments used as stacks is controlled by the B flag of the segment’s descriptor.
When this flag is clear, the default address-size attribute is 16; when the flag is set,
the address-size attribute is 32.

Vol. 2 6-3
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

6.2.4 Procedure Linking Information

The processor provides two pointers for linking of procedures: the stack-frame base
pointer and the return instruction pointer. When used in conjunction with a standard
software procedure-call technique, these pointers permit reliable and coherent
linking of procedures.

6.2.4.1 Stack-Frame Base Pointer

The stack is typically divided into frames. Each stack frame can then contain local
variables, parameters to be passed to another procedure, and procedure linking
information. The stack-frame base pointer (contained in the EBP register) identifies a
fixed reference point within the stack frame for the called procedure. To use the
stack-frame base pointer, the called procedure typically copies the contents of the
ESP register into the EBP register prior to pushing any local variables on the stack.
The stack-frame base pointer then permits easy access to data structures passed on
the stack, to the return instruction pointer, and to local variables added to the stack
by the called procedure.
Like the ESP register, the EBP register automatically points to an address in the
current stack segment (that is, the segment specified by the current contents of the
SS register).

6.2.4.2 Return Instruction Pointer

Prior to branching to the first instruction of the called procedure, the CALL instruction
pushes the address in the EIP register onto the current stack. This address is then
called the return-instruction pointer and it points to the instruction where execution
of the calling procedure should resume following a return from the called procedure.
Upon returning from a called procedure, the RET instruction pops the return-instruc-
tion pointer from the stack back into the EIP register. Execution of the calling proce-
dure then resumes.
The processor does not keep track of the location of the return-instruction pointer. It
is thus up to the programmer to insure that stack pointer is pointing to the return-
instruction pointer on the stack, prior to issuing a RET instruction. A common way to
reset the stack pointer to the point to the return-instruction pointer is to move the
contents of the EBP register into the ESP register. If the EBP register is loaded with
the stack pointer immediately following a procedure call, it should point to the return
instruction pointer on the stack.
The processor does not require that the return instruction pointer point back to the
calling procedure. Prior to executing the RET instruction, the return instruction
pointer can be manipulated in software to point to any address in the current code
segment (near return) or another code segment (far return). Performing such an
operation, however, should be undertaken very cautiously, using only well defined
code entry points.

6-4 Vol. 2
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

6.2.5 Stack Behavior in 64-Bit Mode

In 64-bit mode, address calculations that reference SS segments are treated as if the
segment base is zero. Fields (base, limit, and attribute) in segment descriptor regis-
ters are ignored. SS DPL is modified such that it is always equal to CPL. This will be
true even if it is the only field in the SS descriptor that is modified.
Registers E(SP), E(IP) and E(BP) are promoted to 64-bits and are re-named RSP, RIP,
and RBP respectively. Some forms of segment load instructions are invalid (for
example, LDS, POP ES).
PUSH/POP instructions increment/decrement the stack using a 64-bit width. When
the contents of a segment register is pushed onto 64-bit stack, the pointer is auto-
matically aligned to 64 bits (as with a stack that has a 32-bit width).

6.3 CALLING PROCEDURES USING CALL AND RET

The CALL instruction allows control transfers to procedures within the current code
segment (near call) and in a different code segment (far call). Near calls usually
provide access to local procedures within the currently running program or task. Far
calls are usually used to access operating system procedures or procedures in a
different task. See “CALL—Call Procedure” in Chapter 3, “Instruction Set Reference,
A-M,” of the Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volume
2A, for a detailed description of the CALL instruction.
The RET instruction also allows near and far returns to match the near and far
versions of the CALL instruction. In addition, the RET instruction allows a program to
increment the stack pointer on a return to release parameters from the stack. The
number of bytes released from the stack is determined by an optional argument (n)
to the RET instruction. See “RET—Return from Procedure” in Chapter 4, “Instruction
Set Reference, N-Z,” of the Intel® 64 and IA-32 Architectures Software Developer’s
Manual, Volume 2B, for a detailed description of the RET instruction.

6.3.1 Near CALL and RET Operation

When executing a near call, the processor does the following (see Figure 6-2):
1. Pushes the current value of the EIP register on the stack.
2. Loads the offset of the called procedure in the EIP register.
3. Begins execution of the called procedure.
When executing a near return, the processor performs these actions:
1. Pops the top-of-stack value (the return instruction pointer) into the EIP register.
2. If the RET instruction has an optional n argument, increments the stack pointer
by the number of bytes specified with the n operand to release parameters from
the stack.
3. Resumes execution of the calling procedure.

Vol. 2 6-5
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

6.3.2 Far CALL and RET Operation

When executing a far call, the processor performs these actions (see Figure 6-2):
1. Pushes the current value of the CS register on the stack.
2. Pushes the current value of the EIP register on the stack.
3. Loads the segment selector of the segment that contains the called procedure in
the CS register.
4. Loads the offset of the called procedure in the EIP register.
5. Begins execution of the called procedure.
When executing a far return, the processor does the following:
1. Pops the top-of-stack value (the return instruction pointer) into the EIP register.
2. Pops the top-of-stack value (the segment selector for the code segment being
returned to) into the CS register.
3. If the RET instruction has an optional n argument, increments the stack pointer
by the number of bytes specified with the n operand to release parameters from
the stack.
4. Resumes execution of the calling procedure.

6-6 Vol. 2
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

Stack During Stack During

Stack Near Call Far Call
Stack
Frame
Frame
Before
Before
Call Param 1 Param 1
Call
Param 2 Param 2
Param 3 ESP Before Call Param 3 ESP Before Call
Stack Calling EIP ESP After Call Calling CS
Frame Stack Calling EIP ESP After Call
After Frame
Call After
Call
Stack During Stack During
Near Return Far Return

ESP After Return ESP After Return

Param 1 Param 1
Param 2 Param 2
Param 3 Param 3
Calling EIP ESP Before Return Calling CS
Calling EIP ESP Before Return

Note: On a near or far return, parameters are

released from the stack based on the
optional n operand in the RET n instruction.

Figure 6-2. Stack on Near and Far Calls

6.3.3 Parameter Passing

Parameters can be passed between procedures in any of three ways: through
general-purpose registers, in an argument list, or on the stack.

6.3.3.1 Passing Parameters Through the General-Purpose Registers

The processor does not save the state of the general-purpose registers on procedure
calls. A calling procedure can thus pass up to six parameters to the called procedure
by copying the parameters into any of these registers (except the ESP and EBP regis-
ters) prior to executing the CALL instruction. The called procedure can likewise pass
parameters back to the calling procedure through general-purpose registers.

6.3.3.2 Passing Parameters on the Stack

To pass a large number of parameters to the called procedure, the parameters can be
placed on the stack, in the stack frame for the calling procedure. Here, it is useful to

Vol. 2 6-7
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

use the stack-frame base pointer (in the EBP register) to make a frame boundary for
easy access to the parameters.
The stack can also be used to pass parameters back from the called procedure to the
calling procedure.

6.3.3.3 Passing Parameters in an Argument List

An alternate method of passing a larger number of parameters (or a data structure)
to the called procedure is to place the parameters in an argument list in one of the
data segments in memory. A pointer to the argument list can then be passed to the
called procedure through a general-purpose register or the stack. Parameters can
also be passed back to the calling procedure in this same manner.

6.3.4 Saving Procedure State Information

The processor does not save the contents of the general-purpose registers, segment
registers, or the EFLAGS register on a procedure call. A calling procedure should
explicitly save the values in any of the general-purpose registers that it will need
when it resumes execution after a return. These values can be saved on the stack or
in memory in one of the data segments.
The PUSHA and POPA instructions facilitate saving and restoring the contents of the
general-purpose registers. PUSHA pushes the values in all the general-purpose
registers on the stack in the following order: EAX, ECX, EDX, EBX, ESP (the value
prior to executing the PUSHA instruction), EBP, ESI, and EDI. The POPA instruction
pops all the register values saved with a PUSHA instruction (except the ESP value)
from the stack to their respective registers.
If a called procedure changes the state of any of the segment registers explicitly, it
should restore them to their former values before executing a return to the calling
procedure.
If a calling procedure needs to maintain the state of the EFLAGS register, it can save
and restore all or part of the register using the PUSHF/PUSHFD and POPF/POPFD
instructions. The PUSHF instruction pushes the lower word of the EFLAGS register on
the stack, while the PUSHFD instruction pushes the entire register. The POPF instruc-
tion pops a word from the stack into the lower word of the EFLAGS register, while the
POPFD instruction pops a double word from the stack into the register.

6.3.5 Calls to Other Privilege Levels

The IA-32 architecture’s protection mechanism recognizes four privilege levels,
numbered from 0 to 3, where a greater number mean less privilege. The reason to
use privilege levels is to improve the reliability of operating systems. For example,
Figure 6-3 shows how privilege levels can be interpreted as rings of protection.

6-8 Vol. 2
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

Protection Rings

Operating
System
Kernel Level 0
Operating System
Services (Device
Drivers, Etc.) Level 1

Applications Level 2

Level 3
Highest Lowest
0 1 2 3

Privilege Levels

Figure 6-3. Protection Rings

In this example, the highest privilege level 0 (at the center of the diagram) is used for
segments that contain the most critical code modules in the system, usually the
kernel of an operating system. The outer rings (with progressively lower privileges)
are used for segments that contain code modules for less critical software.
Code modules in lower privilege segments can only access modules operating at
higher privilege segments by means of a tightly controlled and protected interface
called a gate. Attempts to access higher privilege segments without going through a
protection gate and without having sufficient access rights causes a general-protec-
tion exception (#GP) to be generated.
If an operating system or executive uses this multilevel protection mechanism, a call
to a procedure that is in a more privileged protection level than the calling procedure
is handled in a similar manner as a far call (see Section 6.3.2, “Far CALL and RET
Operation”). The differences are as follows:
• The segment selector provided in the CALL instruction references a special data
structure called a call gate descriptor. Among other things, the call gate
descriptor provides the following:
— access rights information
— the segment selector for the code segment of the called procedure
— an offset into the code segment (that is, the instruction pointer for the called
procedure)

Vol. 2 6-9
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

• The processor switches to a new stack to execute the called procedure. Each
privilege level has its own stack. The segment selector and stack pointer for the
privilege level 3 stack are stored in the SS and ESP registers, respectively, and
are automatically saved when a call to a more privileged level occurs. The
segment selectors and stack pointers for the privilege level 2, 1, and 0 stacks are
stored in a system segment called the task state segment (TSS).
The use of a call gate and the TSS during a stack switch are transparent to the calling
procedure, except when a general-protection exception is raised.

6.3.6 CALL and RET Operation Between Privilege Levels

When making a call to a more privileged protection level, the processor does the
following (see Figure 6-4):
1. Performs an access rights check (privilege check).
2. Temporarily saves (internally) the current contents of the SS, ESP, CS, and EIP
registers.

Stack for Stack for

Calling Procedure Called Procedure

Calling SS
Calling ESP
Param 1 Param 1
Stack Frame
Param 2 Param 2 Stack Frame
Before Call
Param 3 ESP Before Call Param 3 After Call
Calling CS
ESP After Call Calling EIP

Calling SS
ESP After Return Calling ESP
Param 1 Param 1
Param 2 Param 2
Param 3 Param 3
Calling CS
ESP Before Return Calling EIP

Note: On a return, parameters are

released on both stacks based on the
optional n operand in the RET n instruction.

Figure 6-4. Stack Switch on a Call to a Different Privilege Level

6-10 Vol. 2
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

3. Loads the segment selector and stack pointer for the new stack (that is, the stack
for the privilege level being called) from the TSS into the SS and ESP registers
and switches to the new stack.
4. Pushes the temporarily saved SS and ESP values for the calling procedure’s stack
onto the new stack.
5. Copies the parameters from the calling procedure’s stack to the new stack. A
value in the call gate descriptor determines how many parameters to copy to the
new stack.
6. Pushes the temporarily saved CS and EIP values for the calling procedure to the
new stack.
7. Loads the segment selector for the new code segment and the new instruction
pointer from the call gate into the CS and EIP registers, respectively.
8. Begins execution of the called procedure at the new privilege level.
When executing a return from the privileged procedure, the processor performs
these actions:
1. Performs a privilege check.
2. Restores the CS and EIP registers to their values prior to the call.
3. If the RET instruction has an optional n argument, increments the stack pointer
by the number of bytes specified with the n operand to release parameters from
the stack. If the call gate descriptor specifies that one or more parameters be
copied from one stack to the other, a RET n instruction must be used to release
the parameters from both stacks. Here, the n operand specifies the number of
bytes occupied on each stack by the parameters. On a return, the processor
increments ESP by n for each stack to step over (effectively remove) these
parameters from the stacks.
4. Restores the SS and ESP registers to their values prior to the call, which causes a
switch back to the stack of the calling procedure.
5. If the RET instruction has an optional n argument, increments the stack pointer
by the number of bytes specified with the n operand to release parameters from
the stack (see explanation in step 3).
6. Resumes execution of the calling procedure.
See Chapter 4, “Protection,” in the Intel® 64 and IA-32 Architectures Software
Developer’s Manual, Volume 3A, for detailed information on calls to privileged levels
and the call gate descriptor.

6.3.7 Branch Functions in 64-Bit Mode

The 64-bit extensions expand branching mechanisms to accommodate branches in
64-bit linear-address space. These are:
• Near-branch semantics are redefined in 64-bit mode

Vol. 2 6-11
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

• In 64-bit mode and compatibility mode, 64-bit call-gate descriptors for far calls
are available
In 64-bit mode, the operand size for all near branches (CALL, RET, JCC, JCXZ, JMP,
and LOOP) is forced to 64 bits. These instructions update the 64-bit RIP without the
need for a REX operand-size prefix.
The following aspects of near branches are controlled by the effective operand size:
• Truncation of the size of the instruction pointer
• Size of a stack pop or push, due to a CALL or RET
• Size of a stack-pointer increment or decrement, due to a CALL or RET
• Indirect-branch operand size
In 64-bit mode, all of the above actions are forced to 64 bits regardless of operand
size prefixes (operand size prefixes are silently ignored). However, the displacement
field for relative branches is still limited to 32 bits and the address size for near
branches is not forced in 64-bit mode.
Address sizes affect the size of RCX used for JCXZ and LOOP; they also impact the
address calculation for memory indirect branches. Such addresses are 64 bits by
default; but they can be overridden to 32 bits by an address size prefix.
Software typically uses far branches to change privilege levels. The legacy IA-32
architecture provides the call-gate mechanism to allow software to branch from one
privilege level to another, although call gates can also be used for branches that do
not change privilege levels. When call gates are used, the selector portion of the
direct or indirect pointer references a gate descriptor (the offset in the instruction is
ignored). The offset to the destination’s code segment is taken from the call-gate
descriptor.
64-bit mode redefines the type value of a 32-bit call-gate descriptor type to a 64-bit
call gate descriptor and expands the size of the 64-bit descriptor to hold a 64-bit
offset. The 64-bit mode call-gate descriptor allows far branches that reference any
location in the supported linear-address space. These call gates also hold the target
code selector (CS), allowing changes to privilege level and default size as a result of
the gate transition.
Because immediates are generally specified up to 32 bits, the only way to specify a
full 64-bit absolute RIP in 64-bit mode is with an indirect branch. For this reason,
direct far branches are eliminated from the instruction set in 64-bit mode.
64-bit mode also expands the semantics of the SYSENTER and SYSEXIT instructions
so that the instructions operate within a 64-bit memory space. The mode also intro-
duces two new instructions: SYSCALL and SYSRET (which are valid only in 64-bit
mode). For details, see “SYSENTER—Fast System Call” and “SYSEXIT—Fast Return
from Fast System Call” in Chapter 4, “Instruction Set Reference, N-Z,” of the Intel®
64 and IA-32 Architectures Software Developer’s Manual, Volume 2B.

6-12 Vol. 2
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

6.4 INTERRUPTS AND EXCEPTIONS

The processor provides two mechanisms for interrupting program execution, inter-
rupts and exceptions:
• An interrupt is an asynchronous event that is typically triggered by an I/O
device.
• An exception is a synchronous event that is generated when the processor
detects one or more predefined conditions while executing an instruction. The
IA-32 architecture specifies three classes of exceptions: faults, traps, and aborts.
The processor responds to interrupts and exceptions in essentially the same way.
When an interrupt or exception is signaled, the processor halts execution of the
current program or task and switches to a handler procedure that has been written
specifically to handle the interrupt or exception condition. The processor accesses
the handler procedure through an entry in the interrupt descriptor table (IDT). When
the handler has completed handling the interrupt or exception, program control is
returned to the interrupted program or task.
The operating system, executive, and/or device drivers normally handle interrupts
and exceptions independently from application programs or tasks. Application
programs can, however, access the interrupt and exception handlers incorporated in
an operating system or executive through assembly-language calls. The remainder
of this section gives a brief overview of the processor’s interrupt and exception
handling mechanism. See Chapter 5, “Interrupt and Exception Handling,” in the
Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volume 3B, for a
description of this mechanism.
The IA-32 Architecture defines 18 predefined interrupts and exceptions and 224 user
defined interrupts, which are associated with entries in the IDT. Each interrupt and
exception in the IDT is identified with a number, called a vector. Table 6-1 lists the
interrupts and exceptions with entries in the IDT and their respective vector
numbers. Vectors 0 through 8, 10 through 14, and 16 through 19 are the predefined
interrupts and exceptions, and vectors 32 through 255 are the user-defined inter-
rupts, called maskable interrupts.
Note that the processor defines several additional interrupts that do not point to
entries in the IDT; the most notable of these interrupts is the SMI interrupt. See
Chapter 5, “Interrupt and Exception Handling,” in the Intel® 64 and IA-32 Architec-
tures Software Developer’s Manual, Volume 3B, for more information about the
interrupts and exceptions.
When the processor detects an interrupt or exception, it does one of the following
things:
• Executes an implicit call to a handler procedure.
• Executes an implicit call to a handler task.

Vol. 2 6-13
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

6.4.1 Call and Return Operation for Interrupt or Exception

Handling Procedures
A call to an interrupt or exception handler procedure is similar to a procedure call to
another protection level (see Section 6.3.6, “CALL and RET Operation Between Privi-
lege Levels”). Here, the interrupt vector references one of two kinds of gates: an
interrupt gate or a trap gate. Interrupt and trap gates are similar to call gates in
that they provide the following information:
• Access rights information
• The segment selector for the code segment that contains the handler procedure
• An offset into the code segment to the first instruction of the handler procedure
The difference between an interrupt gate and a trap gate is as follows. If an interrupt
or exception handler is called through an interrupt gate, the processor clears the
interrupt enable (IF) flag in the EFLAGS register to prevent subsequent interrupts
from interfering with the execution of the handler. When a handler is called through
a trap gate, the state of the IF flag is not changed.

Table 6-1. Exceptions and Interrupts

Vector No. Mnemonic Description Source
0 #DE Divide Error DIV and IDIV instructions.
1 #DB Debug Any code or data reference.
2 NMI Interrupt Non-maskable external interrupt.
3 #BP Breakpoint INT 3 instruction.
4 #OF Overflow INTO instruction.
5 #BR BOUND Range Exceeded BOUND instruction.
6 #UD Invalid Opcode (UnDefined UD2 instruction or reserved opcode.1
Opcode)
7 #NM Device Not Available (No Math Floating-point or WAIT/FWAIT
Coprocessor) instruction.
8 #DF Double Fault Any instruction that can generate an
exception, an NMI, or an INTR.
9 #MF CoProcessor Segment Overrun Floating-point instruction.2
(reserved)
10 #TS Invalid TSS Task switch or TSS access.
11 #NP Segment Not Present Loading segment registers or accessing
system segments.
12 #SS Stack Segment Fault Stack operations and SS register loads.
13 #GP General Protection Any memory reference and other
protection checks.

6-14 Vol. 2
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

Table 6-1. Exceptions and Interrupts (Contd.)

Vector No. Mnemonic Description Source
14 #PF Page Fault Any memory reference.
15 Reserved
16 #MF Floating-Point Error (Math Floating-point or WAIT/FWAIT
Fault) instruction.
17 #AC Alignment Check Any data reference in memory.3
18 #MC Machine Check Error codes (if any) and source are model
dependent.4
19 #XM SIMD Floating-Point Exception SIMD Floating-Point Instruction5
20-31 Reserved
32-255 Maskable Interrupts External interrupt from INTR pin or INT n
instruction.
NOTES:
1. The UD2 instruction was introduced in the Pentium Pro processor.
2. IA-32 processors after the Intel386 processor do not generate this exception.
3. This exception was introduced in the Intel486 processor.
4. This exception was introduced in the Pentium processor and enhanced in the P6 family processors.
5. This exception was introduced in the Pentium III processor.

If the code segment for the handler procedure has the same privilege level as the
currently executing program or task, the handler procedure uses the current stack; if
the handler executes at a more privileged level, the processor switches to the stack
for the handler’s privilege level.
If no stack switch occurs, the processor does the following when calling an interrupt
or exception handler (see Figure 6-5):
1. Pushes the current contents of the EFLAGS, CS, and EIP registers (in that order)
on the stack.
2. Pushes an error code (if appropriate) on the stack.
3. Loads the segment selector for the new code segment and the new instruction
pointer (from the interrupt gate or trap gate) into the CS and EIP registers,
respectively.
4. If the call is through an interrupt gate, clears the IF flag in the EFLAGS register.
5. Begins execution of the handler procedure.

Vol. 2 6-15
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

Stack Usage with No

Privilege-Level Change
Interrupted Procedure’s
and Handler’s Stack

ESP Before
EFLAGS Transfer to Handler
CS
EIP
Error Code ESP After
Transfer to Handler

Stack Usage with

Privilege-Level Change
Interrupted Procedure’s Handler’s Stack
Stack

ESP Before
Transfer to Handler SS
ESP
EFLAGS
CS
EIP
ESP After Error Code
Transfer to Handler

Figure 6-5. Stack Usage on Transfers to Interrupt and Exception Handling Routines

If a stack switch does occur, the processor does the following:

1. Temporarily saves (internally) the current contents of the SS, ESP, EFLAGS, CS,
and EIP registers.
2. Loads the segment selector and stack pointer for the new stack (that is, the stack
for the privilege level being called) from the TSS into the SS and ESP registers
and switches to the new stack.
3. Pushes the temporarily saved SS, ESP, EFLAGS, CS, and EIP values for the
interrupted procedure’s stack onto the new stack.
4. Pushes an error code on the new stack (if appropriate).
5. Loads the segment selector for the new code segment and the new instruction
pointer (from the interrupt gate or trap gate) into the CS and EIP registers,
respectively.
6. If the call is through an interrupt gate, clears the IF flag in the EFLAGS register.
7. Begins execution of the handler procedure at the new privilege level.

6-16 Vol. 2
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

A return from an interrupt or exception handler is initiated with the IRET instruction.
The IRET instruction is similar to the far RET instruction, except that it also restores
the contents of the EFLAGS register for the interrupted procedure. When executing a
return from an interrupt or exception handler from the same privilege level as the
interrupted procedure, the processor performs these actions:
1. Restores the CS and EIP registers to their values prior to the interrupt or
exception.
2. Restores the EFLAGS register.
3. Increments the stack pointer appropriately.
4. Resumes execution of the interrupted procedure.
When executing a return from an interrupt or exception handler from a different priv-
ilege level than the interrupted procedure, the processor performs these actions:
1. Performs a privilege check.
2. Restores the CS and EIP registers to their values prior to the interrupt or
exception.
3. Restores the EFLAGS register.
4. Restores the SS and ESP registers to their values prior to the interrupt or
exception, resulting in a stack switch back to the stack of the interrupted
procedure.
5. Resumes execution of the interrupted procedure.

6.4.2 Calls to Interrupt or Exception Handler Tasks

Interrupt and exception handler routines can also be executed in a separate task.
Here, an interrupt or exception causes a task switch to a handler task. The handler
task is given its own address space and (optionally) can execute at a higher protec-
tion level than application programs or tasks.
The switch to the handler task is accomplished with an implicit task call that refer-
ences a task gate descriptor. The task gate provides access to the address space
for the handler task. As part of the task switch, the processor saves complete state
information for the interrupted program or task. Upon returning from the handler
task, the state of the interrupted program or task is restored and execution
continues. See Chapter 5, “Interrupt and Exception Handling,” in the Intel® 64 and
IA-32 Architectures Software Developer’s Manual, Volume 3B, for more information
on handling interrupts and exceptions through handler tasks.

6.4.3 Interrupt and Exception Handling in Real-Address Mode

When operating in real-address mode, the processor responds to an interrupt or
exception with an implicit far call to an interrupt or exception handler. The processor
uses the interrupt or exception vector number as an index into an interrupt table. The

Vol. 2 6-17
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

interrupt table contains instruction pointers to the interrupt and exception handler
procedures.
The processor saves the state of the EFLAGS register, the EIP register, the CS
register, and an optional error code on the stack before switching to the handler
procedure.
A return from the interrupt or exception handler is carried out with the IRET
instruction.
See Chapter 15, “8086 Emulation,” in the Intel® 64 and IA-32 Architectures Soft-
ware Developer’s Manual, Volume 3A, for more information on handling interrupts
and exceptions in real-address mode.

6.4.4 INT n, INTO, INT 3, and BOUND Instructions

The INT n, INTO, INT 3, and BOUND instructions allow a program or task to explicitly
call an interrupt or exception handler. The INT n instruction uses an interrupt vector
as an argument, which allows a program to call any interrupt handler.
The INTO instruction explicitly calls the overflow exception (#OF) handler if the over-
flow flag (OF) in the EFLAGS register is set. The OF flag indicates overflow on arith-
metic instructions, but it does not automatically raise an overflow exception. An
overflow exception can only be raised explicitly in either of the following ways:
• Execute the INTO instruction.
• Test the OF flag and execute the INT n instruction with an argument of 4 (the
vector number of the overflow exception) if the flag is set.
Both the methods of dealing with overflow conditions allow a program to test for
overflow at specific places in the instruction stream.
The INT 3 instruction explicitly calls the breakpoint exception (#BP) handler.
The BOUND instruction explicitly calls the BOUND-range exceeded exception (#BR)
handler if an operand is found to be not within predefined boundaries in memory.
This instruction is provided for checking references to arrays and other data struc-
tures. Like the overflow exception, the BOUND-range exceeded exception can only
be raised explicitly with the BOUND instruction or the INT n instruction with an argu-
ment of 5 (the vector number of the bounds-check exception). The processor does
not implicitly perform bounds checks and raise the BOUND-range exceeded excep-
tion.

6.4.5 Handling Floating-Point Exceptions

When operating on individual or packed floating-point values, the IA-32 architecture
supports a set of six floating-point exceptions. These exceptions can be generated
during operations performed by the x87 FPU instructions or by SSE/SSE2/SSE3
instructions. When an x87 FPU instruction (including the FISTTP instruction in SSE3)
generates one or more of these exceptions, it in turn generates floating-point error

6-18 Vol. 2
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

exception (#MF); when an SSE/SSE2/SSE3 instruction generates a floating-point

exception, it in turn generates SIMD floating-point exception (#XM).
See the following sections for further descriptions of the floating-point exceptions,
how they are generated, and how they are handled:
• Section 4.9.1, “Floating-Point Exception Conditions,” and Section 4.9.3, “Typical
Actions of a Floating-Point Exception Handler”
• Section 8.4, “x87 FPU Floating-Point Exception Handling,” and Section 8.5, “x87
FPU Floating-Point Exception Conditions”
• Section 11.5.1, “SIMD Floating-Point Exceptions”
• Interrupt Behavior

6.4.6 Interrupt and Exception Behavior in 64-Bit Mode

64-bit extensions expand the legacy IA-32 interrupt-processing and exception-
processing mechanism to allow support for 64-bit operating systems and applica-
tions. Changes include:
• All interrupt handlers pointed to by the IDT are 64-bit code (does not apply to the
SMI handler).
• The size of interrupt-stack pushes is fixed at 64 bits. The processor uses 8-byte,
zero extended stores.
• The stack pointer (SS:RSP) is pushed unconditionally on interrupts. In legacy
environments, this push is conditional and based on a change in current privilege
level (CPL).
• The new SS is set to NULL if there is a change in CPL.
• IRET behavior changes.
• There is a new interrupt stack-switch mechanism.
• The alignment of interrupt stack frame is different.

6.5 PROCEDURE CALLS FOR BLOCK-STRUCTURED

LANGUAGES
The IA-32 architecture supports an alternate method of performing procedure calls
with the ENTER (enter procedure) and LEAVE (leave procedure) instructions. These
instructions automatically create and release, respectively, stack frames for called
procedures. The stack frames have predefined spaces for local variables and the
necessary pointers to allow coherent returns from called procedures. They also allow
scope rules to be implemented so that procedures can access their own local vari-
ables and some number of other variables located in other stack frames.

Vol. 2 6-19
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

ENTER and LEAVE offer two benefits:

• They provide machine-language support for implementing block-structured
languages, such as C and Pascal.
• They simplify procedure entry and exit in compiler-generated code.

6.5.1 ENTER Instruction

The ENTER instruction creates a stack frame compatible with the scope rules typically
used in block-structured languages. In block-structured languages, the scope of a
procedure is the set of variables to which it has access. The rules for scope vary
among languages. They may be based on the nesting of procedures, the division of
the program into separately compiled files, or some other modularization scheme.
ENTER has two operands. The first specifies the number of bytes to be reserved on
the stack for dynamic storage for the procedure being called. Dynamic storage is the
memory allocated for variables created when the procedure is called, also known as
automatic variables. The second parameter is the lexical nesting level (from 0 to 31)
of the procedure. The nesting level is the depth of a procedure in a hierarchy of
procedure calls. The lexical level is unrelated to either the protection privilege level or
to the I/O privilege level of the currently running program or task.
ENTER, in the following example, allocates 2 Kbytes of dynamic storage on the stack
and sets up pointers to two previous stack frames in the stack frame for this proce-
dure:

ENTER 2048,3
The lexical nesting level determines the number of stack frame pointers to copy into
the new stack frame from the preceding frame. A stack frame pointer is a doubleword
used to access the variables of a procedure. The set of stack frame pointers used by
a procedure to access the variables of other procedures is called the display. The first
doubleword in the display is a pointer to the previous stack frame. This pointer is
used by a LEAVE instruction to undo the effect of an ENTER instruction by discarding
the current stack frame.
After the ENTER instruction creates the display for a procedure, it allocates the
dynamic local variables for the procedure by decrementing the contents of the ESP
register by the number of bytes specified in the first parameter. This new value in the
ESP register serves as the initial top-of-stack for all PUSH and POP operations within
the procedure.
To allow a procedure to address its display, the ENTER instruction leaves the EBP
register pointing to the first doubleword in the display. Because stacks grow down,
this is actually the doubleword with the highest address in the display. Data manipu-
lation instructions that specify the EBP register as a base register automatically
address locations within the stack segment instead of the data segment.
The ENTER instruction can be used in two ways: nested and non-nested. If the lexical
level is 0, the non-nested form is used. The non-nested form pushes the contents of

6-20 Vol. 2
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

the EBP register on the stack, copies the contents of the ESP register into the EBP
register, and subtracts the first operand from the contents of the ESP register to allo-
cate dynamic storage. The non-nested form differs from the nested form in that no
stack frame pointers are copied. The nested form of the ENTER instruction occurs
when the second parameter (lexical level) is not zero.
The following pseudo code shows the formal definition of the ENTER instruction.
STORAGE is the number of bytes of dynamic storage to allocate for local variables,
and LEVEL is the lexical nesting level.

PUSH EBP;
FRAME_PTR ← ESP;
IF LEVEL > 0
THEN
DO (LEVEL − 1) times
EBP ← EBP − 4;
PUSH Pointer(EBP); (* doubleword pointed to by EBP *)
OD;
PUSH FRAME_PTR;
FI;
EBP ← FRAME_PTR;
ESP ← ESP − STORAGE;
The main procedure (in which all other procedures are nested) operates at the
highest lexical level, level 1. The first procedure it calls operates at the next deeper
lexical level, level 2. A level 2 procedure can access the variables of the main
program, which are at fixed locations specified by the compiler. In the case of level 1,
the ENTER instruction allocates only the requested dynamic storage on the stack
because there is no previous display to copy.
A procedure that calls another procedure at a lower lexical level gives the called
procedure access to the variables of the caller. The ENTER instruction provides this
access by placing a pointer to the calling procedure's stack frame in the display.
A procedure that calls another procedure at the same lexical level should not give
access to its variables. In this case, the ENTER instruction copies only that part of the
display from the calling procedure which refers to previously nested procedures
operating at higher lexical levels. The new stack frame does not include the pointer
for addressing the calling procedure’s stack frame.
The ENTER instruction treats a re-entrant procedure as a call to a procedure at the
same lexical level. In this case, each succeeding iteration of the re-entrant procedure
can address only its own variables and the variables of the procedures within which it
is nested. A re-entrant procedure always can address its own variables; it does not
require pointers to the stack frames of previous iterations.
By copying only the stack frame pointers of procedures at higher lexical levels, the
ENTER instruction makes certain that procedures access only those variables of
higher lexical levels, not those at parallel lexical levels (see Figure 6-6).

Vol. 2 6-21
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

Main (Lexical Level 1)

Procedure A (Lexical Level 2)
Procedure B (Lexical Level 3)

Procedure C (Lexical Level 3)

Procedure D (Lexical Level 4)

Figure 6-6. Nested Procedures

Block-structured languages can use the lexical levels defined by ENTER to control
access to the variables of nested procedures. In Figure 6-6, for example, if procedure
A calls procedure B which, in turn, calls procedure C, then procedure C will have
access to the variables of the MAIN procedure and procedure A, but not those of
procedure B because they are at the same lexical level. The following definition
describes the access to variables for the nested procedures in Figure 6-6.
1. MAIN has variables at fixed locations.
2. Procedure A can access only the variables of MAIN.
3. Procedure B can access only the variables of procedure A and MAIN. Procedure B
cannot access the variables of procedure C or procedure D.
4. Procedure C can access only the variables of procedure A and MAIN. Procedure C
cannot access the variables of procedure B or procedure D.
5. Procedure D can access the variables of procedure C, procedure A, and MAIN.
Procedure D cannot access the variables of procedure B.
In Figure 6-7, an ENTER instruction at the beginning of the MAIN procedure creates
three doublewords of dynamic storage for MAIN, but copies no pointers from other
stack frames. The first doubleword in the display holds a copy of the last value in the
EBP register before the ENTER instruction was executed. The second doubleword
holds a copy of the contents of the EBP register following the ENTER instruction. After
the instruction is executed, the EBP register points to the first doubleword pushed on
the stack, and the ESP register points to the last doubleword in the stack frame.
When MAIN calls procedure A, the ENTER instruction creates a new display (see
Figure 6-8). The first doubleword is the last value held in MAIN's EBP register. The
second doubleword is a pointer to MAIN's stack frame which is copied from the
second doubleword in MAIN's display. This happens to be another copy of the last
value held in MAIN’s EBP register. Procedure A can access variables in MAIN because
MAIN is at level 1.

6-22 Vol. 2
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

Therefore the base address for the dynamic storage used in MAIN is the current
address in the EBP register, plus four bytes to account for the saved contents of
MAIN’s EBP register. All dynamic variables for MAIN are at fixed, positive offsets from
this value.

Old EBP EBP

Display
Main’s EBP

Dynamic
Storage
ESP

Figure 6-7. Stack Frame After Entering the MAIN Procedure

Old EBP
Main’s EBP

Main’s EBP EBP

Display
Main’s EBP
Procedure A’s EBP
Dynamic
Storage
ESP

Figure 6-8. Stack Frame After Entering Procedure A

When procedure A calls procedure B, the ENTER instruction creates a new display
(see Figure 6-9). The first doubleword holds a copy of the last value in procedure A’s
EBP register. The second and third doublewords are copies of the two stack frame
pointers in procedure A’s display. Procedure B can access variables in procedure A
and MAIN by using the stack frame pointers in its display.

Vol. 2 6-23
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

When procedure B calls procedure C, the ENTER instruction creates a new display for
procedure C (see Figure 6-10). The first doubleword holds a copy of the last value in
procedure B’s EBP register. This is used by the LEAVE instruction to restore procedure
B’s stack frame. The second and third doublewords are copies of the two stack frame
pointers in procedure A’s display. If procedure C were at the next deeper lexical level
from procedure B, a fourth doubleword would be copied, which would be the stack
frame pointer to procedure B’s local variables.
Note that procedure B and procedure C are at the same level, so procedure C is not
intended to access procedure B’s variables. This does not mean that procedure C is
completely isolated from procedure B; procedure C is called by procedure B, so the
pointer to the returning stack frame is a pointer to procedure B’s stack frame. In
addition, procedure B can pass parameters to procedure C either on the stack or
through variables global to both procedures (that is, variables in the scope of both
procedures).

Old EBP
Main’s EBP

Main’s EBP
Main’s EBP
Procedure A’s EBP

Procedure A’s EBP EBP

Main’s EBP
Display
Procedure A’s EBP
Procedure B’s EBP

Dynamic
Storage
ESP

Figure 6-9. Stack Frame After Entering Procedure B

6-24 Vol. 2
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

Old EBP
Main’s EBP

Main’s EBP
Main’s EBP
Procedure A’s EBP

Procedure A’s EBP

Main’s EBP
Procedure A’s EBP
Procedure B’s EBP

Procedure B’s EBP EBP

Main’s EBP
Display
Procedure A’s EBP
Procedure C’s EBP

Dynamic
Storage
ESP

Figure 6-10. Stack Frame After Entering Procedure C

6.5.2 LEAVE Instruction

The LEAVE instruction, which does not have any operands, reverses the action of the
previous ENTER instruction. The LEAVE instruction copies the contents of the EBP
register into the ESP register to release all stack space allocated to the procedure.
Then it restores the old value of the EBP register from the stack. This simultaneously
restores the ESP register to its original value. A subsequent RET instruction then can
remove any arguments and the return address pushed on the stack by the calling
program for use by the procedure.

Vol. 2 6-25
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

6-26 Vol. 2
 CHAPTER 7
PROGRAMMING WITH
GENERAL-PURPOSE INSTRUCTIONS

General-purpose (GP) instructions are a subset of the IA-32 instructions that repre-
sent the fundamental instruction set for the Intel IA-32 processors. These instruc-
tions were introduced into the IA-32 architecture with the first IA-32 processors (the
Intel 8086 and 8088). Additional instructions were added to the general-purpose
instruction set in subsequent families of IA-32 processors (the Intel 286, Intel386,
Intel486, Pentium, Pentium Pro, and Pentium II processors).
Intel 64 architecture further extends the capability of most general-purpose instruc-
tions so that they are able to handle 64-bit data in 64-bit mode. A small number of
general-purpose instructions (still supported in non-64-bit modes) are not supported
in 64-bit mode.
General-purpose instructions perform basic data movement, memory addressing,
arithmetic and logical, program flow control, input/output, and string operations on a
set of integer, pointer, and BCD data types. This chapter provides an overview of the
general-purpose instructions. See Intel® 64 and IA-32 Architectures Software
Developer’s Manual, Volumes 3A & 3B, for detailed descriptions of individual instruc-
tions.

7.1 PROGRAMMING ENVIRONMENT FOR GP

INSTRUCTIONS
The programming environment for the general-purpose instructions consists of the
set of registers and address space. The environment includes the following items:
• General-purpose registers — Eight 32-bit general-purpose registers (see
Section 3.4.1, “General-Purpose Registers”) are used in non-64-bit modes to
address operands in memory. These registers are referenced by the names EAX,
EBX, ECX, EDX, EBP, ESI EDI, and ESP.
• Segment registers — The six 16-bit segment registers contain segment
pointers for use in accessing memory (see Section 3.4.2, “Segment Registers”).
These registers are referenced by the names CS, DS, SS, ES, FS, and GS.
• EFLAGS register — This 32-bit register (see Section 3.4.3, “EFLAGS Register”)
is used to provide status and control for basic arithmetic, compare, and system
operations.
• EIP register — This 32-bit register contains the current instruction pointer (see
Section 3.4.3, “EFLAGS Register”).
General-purpose instructions operate on the following data types. The width of valid
data types is dependent on processor mode (see Chapter 4):

Vol. 1 7-1
PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

• Bytes, words, doublewords

• Signed and unsigned byte, word, doubleword integers
• Near and far pointers
• Bit fields
• BCD integers

7.2 PROGRAMMING ENVIRONMENT FOR GP

INSTRUCTIONS IN 64-BIT MODE
The programming environment for the general-purpose instructions in 64-bit mode is
similar to that described in Section 7.1.
• General-purpose registers — In 64-bit mode, sixteen general-purpose
registers available. These include the eight GPRs described in Section 7.1 and
eight new GPRs (R8D-R15D). R8D-R15D are available by using a REX prefix. All
sixteen GPRs can be promoted to 64 bits. The 64-bit registers are referenced as
RAX, RBX, RCX, RDX, RBP, RSI, RDI, RSP and R8-R15 (see Section 3.4.1.1,
“General-Purpose Registers in 64-Bit Mode”). Promotion to 64-bit operand
requires REX prefix encodings.
• Segment registers — In 64-bit mode, segmentation is available but it is set up
uniquely (see Section 3.4.2.1, “Segment Registers in 64-Bit Mode”).
• Flags and Status register — When the processor is running in 64-bit mode,
EFLAGS becomes the 64-bit RFLAGS register (see Section 3.4.3, “EFLAGS
Register”).
• Instruction Pointer register — In 64-bit mode, the EIP register becomes the
64-bit RIP register (see Section 3.5.1, “Instruction Pointer in 64-Bit Mode”).
General-purpose instructions operate on the following data types in 64-bit mode. The
width of valid data types is dependent on default operand size, address size, or a
prefix that overrides the default size:
• Bytes, words, doublewords, quadwords
• Signed and unsigned byte, word, doubleword, quadword integers
• Near and far pointers
• Bit fields
See also:
• Chapter 3, “Basic Execution Environment,” for more information about IA-32e
modes.
• Chapter 2, “Instruction Format,” in the Intel® 64 and IA-32 Architectures
Software Developer’s Manual, Volume 2A, for more detailed information about
REX prefixes.

7-2 Vol. 1
 PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

• Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volumes

2A & 2B for a complete listing of all instructions. This information documents the
behavior of individual instructions in the 64-bit mode context.

7.3 SUMMARY OF GP INSTRUCTIONS

General purpose instructions are divided into the following subgroups:
• Data transfer
• Binary arithmetic
• Decimal arithmetic
• Logical
• Shift and rotate
• Bit and byte
• Control transfer
• String
• I/O
• Enter and Leave
• Flag control
• Segment register
• Miscellaneous
Each sub-group of general-purpose instructions is discussed in the context of non-
64-bit mode operation first. Changes in 64-bit mode beyond those affected by the
use of the REX prefixes are discussed in separate sub-sections within each subgroup.
For a simple list of general-purpose instructions by subgroup, see Chapter 5.

7.3.1 Data Transfer Instructions

The data transfer instructions move bytes, words, doublewords, or quadwords both
between memory and the processor’s registers and between registers. For the
purpose of this discussion, these instructions are divided into subordinate subgroups
that provide for:
• General data movement
• Exchange
• Stack manipulation
• Type conversion

Vol. 1 7-3
PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

7.3.1.1 General Data Movement Instructions

Move instructions — The MOV (move) and CMOVcc (conditional move) instructions
transfer data between memory and registers or between registers.
The MOV instruction performs basic load data and store data operations between
memory and the processor’s registers and data movement operations between regis-
ters. It handles data transfers along the paths listed in Table 7-1. (See “MOV—Move
to/from Control Registers” and “MOV—Move to/from Debug Registers” in Chapter 3,
“Instruction Set Reference, A-M,” of the Intel® 64 and IA-32 Architectures Software
Developer’s Manual, Volume 2A, for information on moving data to and from the
control and debug registers.)
The MOV instruction cannot move data from one memory location to another or from
one segment register to another segment register. Memory-to-memory moves are
performed with the MOVS (string move) instruction (see Section 7.3.9, “String Oper-
ations”).
Conditional move instructions — The CMOVcc instructions are a group of instruc-
tions that check the state of the status flags in the EFLAGS register and perform a
move operation if the flags are in a specified state. These instructions can be used to
move a 16-bit or 32-bit value from memory to a general-purpose register or from
one general-purpose register to another. The flag state being tested is specified with
a condition code (cc) associated with the instruction. If the condition is not satisfied,
a move is not performed and execution continues with the instruction following the
CMOVcc instruction.

Table 7-1. Move Instruction Operations

Type of Data Movement Source → Destination
From memory to a register Memory location → General-purpose register
Memory location → Segment register
From a register to memory General-purpose register → Memory location
Segment register → Memory location
Between registers General-purpose register → General-purpose register
General-purpose register → Segment register
Segment register → General-purpose register
General-purpose register → Control register
Control register → General-purpose register
General-purpose register → Debug register
Debug register → General-purpose register
Immediate data to a register Immediate → General-purpose register
Immediate data to memory Immediate → Memory location

7-4 Vol. 1
 PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

Table 7-2 shows mnemonics for CMOVcc instructions and the conditions being tested
for each instruction. The condition code mnemonics are appended to the letters
“CMOV” to form the mnemonics for CMOVcc instructions. The instructions listed in
Table 7-2 as pairs (for example, CMOVA/CMOVNBE) are alternate names for the
same instruction. The assembler provides these alternate names to make it easier to
read program listings.
CMOVcc instructions are useful for optimizing small IF constructions. They also help
eliminate branching overhead for IF statements and the possibility of branch mispre-
dictions by the processor.
These conditional move instructions are supported in the P6 family, Pentium 4, and
Intel Xeon processors. Software can check if CMOVcc instructions are supported by
checking the processor’s feature information with the CPUID instruction.

7.3.1.2 Exchange Instructions

The exchange instructions swap the contents of one or more operands and, in some
cases, perform additional operations such as asserting the LOCK signal or modifying
flags in the EFLAGS register.
The XCHG (exchange) instruction swaps the contents of two operands. This instruc-
tion takes the place of three MOV instructions and does not require a temporary loca-
tion to save the contents of one operand location while the other is being loaded.
When a memory operand is used with the XCHG instruction, the processor’s LOCK
signal is automatically asserted. This instruction is thus useful for implementing
semaphores or similar data structures for process synchronization. See “Bus
Locking” in Chapter 7, “Multiple-Processor Management,” of the Intel® 64 and IA-32
Architectures Software Developer’s Manual, Volume 3A, for more information on bus
locking.
The BSWAP (byte swap) instruction reverses the byte order in a 32-bit register
operand. Bit positions 0 through 7 are exchanged with 24 through 31, and bit posi-
tions 8 through 15 are exchanged with 16 through 23. Executing this instruction
twice in a row leaves the register with the same value as before. The BSWAP instruc-
tion is useful for converting between “big-endian” and “little-endian” data formats.
This instruction also speeds execution of decimal arithmetic. (The XCHG instruction
can be used to swap the bytes in a word.)

Vol. 1 7-5
PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

Table 7-2. Conditional Move Instructions

Instruction Mnemonic Status Flag States Condition Description
Unsigned Conditional Moves
CMOVA/CMOVNBE (CF or ZF) = 0 Above/not below or equal
CMOVAE/CMOVNB CF = 0 Above or equal/not below
CMOVNC CF = 0 Not carry
CMOVB/CMOVNAE CF = 1 Below/not above or equal
CMOVC CF = 1 Carry
CMOVBE/CMOVNA (CF or ZF) = 1 Below or equal/not above
CMOVE/CMOVZ ZF = 1 Equal/zero
CMOVNE/CMOVNZ ZF = 0 Not equal/not zero
CMOVP/CMOVPE PF = 1 Parity/parity even
CMOVNP/CMOVPO PF = 0 Not parity/parity odd
Signed Conditional Moves
CMOVGE/CMOVNL (SF xor OF) = 0 Greater or equal/not less
CMOVL/CMOVNGE (SF xor OF) = 1 Less/not greater or equal
CMOVLE/CMOVNG ((SF xor OF) or ZF) = 1 Less or equal/not greater
CMOVO OF = 1 Overflow
CMOVNO OF = 0 Not overflow
CMOVS SF = 1 Sign (negative)
CMOVNS SF = 0 Not sign (non-negative)

The XADD (exchange and add) instruction swaps two operands and then stores the
sum of the two operands in the destination operand. The status flags in the EFLAGS
register indicate the result of the addition. This instruction can be combined with the
LOCK prefix (see “LOCK—Assert LOCK# Signal Prefix” in Chapter 3, “Instruction Set
Reference, A-M,” of the Intel® 64 and IA-32 Architectures Software Developer’s
Manual, Volume 2A) in a multiprocessing system to allow multiple processors to
execute one DO loop.
The CMPXCHG (compare and exchange) and CMPXCHG8B (compare and exchange
8 bytes) instructions are used to synchronize operations in systems that use
multiple processors. The CMPXCHG instruction requires three operands: a source
operand in a register, another source operand in the EAX register, and a destination
operand. If the values contained in the destination operand and the EAX register are
equal, the destination operand is replaced with the value of the other source
operand (the value not in the EAX register). Otherwise, the original value of the
destination operand is loaded in the EAX register. The status flags in the EFLAGS

7-6 Vol. 1
 PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

register reflect the result that would have been obtained by subtracting the destina-
tion operand from the value in the EAX register.
The CMPXCHG instruction is commonly used for testing and modifying semaphores.
It checks to see if a semaphore is free. If the semaphore is free, it is marked allo-
cated; otherwise it gets the ID of the current owner. This is all done in one uninter-
ruptible operation. In a single-processor system, the CMPXCHG instruction
eliminates the need to switch to protection level 0 (to disable interrupts) before
executing multiple instructions to test and modify a semaphore.
For multiple processor systems, CMPXCHG can be combined with the LOCK prefix to
perform the compare and exchange operation atomically. (See “Locked Atomic Oper-
ations” in Chapter 7, “Multiple-Processor Management,” of the Intel® 64 and IA-32
Architectures Software Developer’s Manual, Volume 3A, for more information on
atomic operations.)
The CMPXCHG8B instruction also requires three operands: a 64-bit value in
EDX:EAX, a 64-bit value in ECX:EBX, and a destination operand in memory. The
instruction compares the 64-bit value in the EDX:EAX registers with the destination
operand. If they are equal, the 64-bit value in the ECX:EBX register is stored in the
destination operand. If the EDX:EAX register and the destination are not equal, the
destination is loaded in the EDX:EAX register. The CMPXCHG8B instruction can be
combined with the LOCK prefix to perform the operation atomically.

7.3.1.3 Exchange Instructions in 64-Bit Mode

The CMPXCHG16B instruction is available in 64-bit mode only. It is an extension of
the functionality provided by CMPXCHG8B that operates on 128-bits of data.

7.3.1.4 Stack Manipulation Instructions

The PUSH, POP, PUSHA (push all registers), and POPA (pop all registers) instructions
move data to and from the stack. The PUSH instruction decrements the stack pointer
(contained in the ESP register), then copies the source operand to the top of stack
(see Figure 7-1). It operates on memory operands, immediate operands, and
register operands (including segment registers). The PUSH instruction is commonly
used to place parameters on the stack before calling a procedure. It can also be used
to reserve space on the stack for temporary variables.

Vol. 1 7-7
PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

Stack
Before Pushing Doubleword After Pushing Doubleword
Stack
Growth 31 0 31 0
n ESP
n−4 Doubleword Value ESP
n−8

Figure 7-1. Operation of the PUSH Instruction

The PUSHA instruction saves the contents of the eight general-purpose registers on
the stack (see Figure 7-2). This instruction simplifies procedure calls by reducing the
number of instructions required to save the contents of the general-purpose regis-
ters. The registers are pushed on the stack in the following order: EAX, ECX, EDX,
EBX, the initial value of ESP before EAX was pushed, EBP, ESI, and EDI.

Stack
Before Pushing Registers After Pushing Registers
Stack 31 0 31 0
Growth
n
n-4 ESP
n-8 EAX
n - 12 ECX
n - 16 EDX
n - 20 EBX
n - 24 Old ESP
n - 28 EBP
n - 32 ESI
n - 36 EDI ESP

Figure 7-2. Operation of the PUSHA Instruction

The POP instruction copies the word or doubleword at the current top of stack (indi-
cated by the ESP register) to the location specified with the destination operand. It
then increments the ESP register to point to the new top of stack (see Figure 7-3).
The destination operand may specify a general-purpose register, a segment register,
or a memory location.

7-8 Vol. 1
 PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

Stack
Before Popping Doubleword After Popping Doubleword
Stack
Growth 31 0 31 0

n
n-4 ESP
n-8 Doubleword Value ESP

Figure 7-3. Operation of the POP Instruction

The POPA instruction reverses the effect of the PUSHA instruction. It pops the top
eight words or doublewords from the top of the stack into the general-purpose regis-
ters, except for the ESP register (see Figure 7-4). If the operand-size attribute is 32,
the doublewords on the stack are transferred to the registers in the following order:
EDI, ESI, EBP, ignore doubleword, EBX, EDX, ECX, and EAX. The ESP register is
restored by the action of popping the stack. If the operand-size attribute is 16, the
words on the stack are transferred to the registers in the following order: DI, SI, BP,
ignore word, BX, DX, CX, and AX.

Stack
Before Popping Registers After Popping Registers
Stack 0 31 0 31
Growth
n
n-4 ESP
n-8 EAX
n - 12 ECX
n - 16 EDX
n - 20 EBX
n - 24 Ignored
n - 28 EBP
n - 32 ESI
n - 36 EDI ESP

Figure 7-4. Operation of the POPA Instruction

7.3.1.5 Stack Manipulation Instructions in 64-Bit Mode

In 64-bit mode, the stack pointer size is 64 bits and cannot be overridden by an
instruction prefix. In implicit stack references, address-size overrides are ignored.
Pushes and pops of 32-bit values on the stack are not possible in 64-bit mode. 16-bit

Vol. 1 7-9
PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

pushes and pops are supported by using the 66H operand-size prefix. PUSHA,
PUSHAD, POPA, and POPAD are not supported.

7.3.1.6 Type Conversion Instructions

The type conversion instructions convert bytes into words, words into doublewords,
and doublewords into quadwords. These instructions are especially useful for
converting integers to larger integer formats, because they perform sign extension
(see Figure 7-5).
Two kinds of type conversion instructions are provided: simple conversion and move
and convert.

7-10 Vol. 1
 PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

15 0
Before Sign
S N N N N N N N N N N N N N N N
Extension
31 15 0
After Sign
S S S S S S S S S S S S S S S S S N N N N N N N N N N N N N N N
Extension

Figure 7-5. Sign Extension

Simple conversion — The CBW (convert byte to word), CWDE (convert word to
doubleword extended), CWD (convert word to doubleword), and CDQ (convert
doubleword to quadword) instructions perform sign extension to double the size of
the source operand.
The CBW instruction copies the sign (bit 7) of the byte in the AL register into every bit
position of the upper byte of the AX register. The CWDE instruction copies the sign
(bit 15) of the word in the AX register into every bit position of the high word of the
EAX register.
The CWD instruction copies the sign (bit 15) of the word in the AX register into every
bit position in the DX register. The CDQ instruction copies the sign (bit 31) of the
doubleword in the EAX register into every bit position in the EDX register. The CWD
instruction can be used to produce a doubleword dividend from a word before a word
division, and the CDQ instruction can be used to produce a quadword dividend from
a doubleword before doubleword division.
Move with sign or zero extension — The MOVSX (move with sign extension) and
MOVZX (move with zero extension) instructions move the source operand into a
register then perform the sign extension.
The MOVSX instruction extends an 8-bit value to a 16-bit value or an 8-bit or 16-bit
value to a 32-bit value by sign extending the source operand, as shown in Figure 7-5.
The MOVZX instruction extends an 8-bit value to a 16-bit value or an 8-bit or 16-bit
value to a 32-bit value by zero extending the source operand.

7.3.1.7 Type Conversion Instructions in 64-Bit Mode

The MOVSXD instruction operates on 64-bit data. It sign-extends a 32-bit value to 64
bits. This instruction is not encodable in non-64-bit modes.

Vol. 1 7-11
PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

7.3.2 Binary Arithmetic Instructions

Binary arithmetic instructions operate on 8-, 16-, and 32-bit numeric data encoded
as signed or unsigned binary integers. The binary arithmetic instructions may also be
used in algorithms that operate on decimal (BCD) values.
For the purpose of this discussion, these instructions are divided subordinate
subgroups of instructions that:
• Add and subtract
• Increment and decrement
• Compare and change signs
• Multiply and divide

7.3.2.1 Addition and Subtraction Instructions

The ADD (add integers), ADC (add integers with carry), SUB (subtract integers), and
SBB (subtract integers with borrow) instructions perform addition and subtraction
operations on signed or unsigned integer operands.
The ADD instruction computes the sum of two integer operands.
The ADC instruction computes the sum of two integer operands, plus 1 if the CF flag
is set. This instruction is used to propagate a carry when adding numbers in stages.
The SUB instruction computes the difference of two integer operands.
The SBB instruction computes the difference of two integer operands, minus 1 if the
CF flag is set. This instruction is used to propagate a borrow when subtracting
numbers in stages.

7.3.2.2 Increment and Decrement Instructions

The INC (increment) and DEC (decrement) instructions add 1 to or subtract 1 from
an unsigned integer operand, respectively. A primary use of these instructions is for
implementing counters.

7.3.2.3 Increment and Decrement Instructions in 64-Bit Mode

The INC and DEC instructions are supported in 64-bit mode. However, some forms of
INC and DEC (the register operand being encoded using register extension field in
the MOD R/M byte) are not encodable in 64-bit mode because the opcodes are
treated as REX prefixes.

7.3.2.4 Comparison and Sign Change Instruction

The CMP (compare) instruction computes the difference between two integer oper-
ands and updates the OF, SF, ZF, AF, PF, and CF flags according to the result. The

7-12 Vol. 1
 PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

source operands are not modified, nor is the result saved. The CMP instruction is
commonly used in conjunction with a Jcc (jump) or SETcc (byte set on condition)
instruction, with the latter instructions performing an action based on the result of a
CMP instruction.
The NEG (negate) instruction subtracts a signed integer operand from zero. The
effect of the NEG instruction is to change the sign of a two's complement operand
while keeping its magnitude.

7.3.2.5 Multiplication and Divide Instructions

The processor provides two multiply instructions, MUL (unsigned multiply) and IMUL
signed multiply), and two divide instructions, DIV (unsigned divide) and IDIV (signed
divide).
The MUL instruction multiplies two unsigned integer operands. The result is
computed to twice the size of the source operands (for example, if word operands are
being multiplied, the result is a doubleword).
The IMUL instruction multiplies two signed integer operands. The result is computed
to twice the size of the source operands; however, in some cases the result is trun-
cated to the size of the source operands (see “IMUL—Signed Multiply” in Chapter 3,
“Instruction Set Reference, A-M,” of the Intel® 64 and IA-32 Architectures Software
Developer’s Manual, Volume 3A).
The DIV instruction divides one unsigned operand by another unsigned operand and
returns a quotient and a remainder.
The IDIV instruction is identical to the DIV instruction, except that IDIV performs a
signed division.

7.3.3 Decimal Arithmetic Instructions

Decimal arithmetic can be performed by combining the binary arithmetic instructions
ADD, SUB, MUL, and DIV (discussed in Section 7.3.2, “Binary Arithmetic Instruc-
tions”) with the decimal arithmetic instructions. The decimal arithmetic instructions
are provided to carry out the following operations:
• To adjust the results of a previous binary arithmetic operation to produce a valid
BCD result.
• To adjust the operands of a subsequent binary arithmetic operation so that the
operation will produce a valid BCD result.
These instructions operate on both packed and unpacked BCD values. For the
purpose of this discussion, the decimal arithmetic instructions are divided subordi-
nate subgroups of instructions that provide:
• Packed BCD adjustments
• Unpacked BCD adjustments

Vol. 1 7-13
PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

7.3.3.1 Packed BCD Adjustment Instructions

The DAA (decimal adjust after addition) and DAS (decimal adjust after subtraction)
instructions adjust the results of operations performed on packed BCD integers
(see Section 4.7, “BCD and Packed BCD Integers”). Adding two packed BCD values
requires two instructions: an ADD instruction followed by a DAA instruction. The ADD
instruction adds (binary addition) the two values and stores the result in the AL
register. The DAA instruction then adjusts the value in the AL register to obtain a
valid, 2-digit, packed BCD value and sets the CF flag if a decimal carry occurred as
the result of the addition.
Likewise, subtracting one packed BCD value from another requires a SUB instruction
followed by a DAS instruction. The SUB instruction subtracts (binary subtraction) one
BCD value from another and stores the result in the AL register. The DAS instruction
then adjusts the value in the AL register to obtain a valid, 2-digit, packed BCD value
and sets the CF flag if a decimal borrow occurred as the result of the subtraction.

7.3.3.2 Unpacked BCD Adjustment Instructions

The AAA (ASCII adjust after addition), AAS (ASCII adjust after subtraction), AAM
(ASCII adjust after multiplication), and AAD (ASCII adjust before division) instruc-
tions adjust the results of arithmetic operations performed in unpacked BCD
values (see Section 4.7, “BCD and Packed BCD Integers”). All these instructions
assume that the value to be adjusted is stored in the AL register or, in one instance,
the AL and AH registers.
The AAA instruction adjusts the contents of the AL register following the addition of
two unpacked BCD values. It converts the binary value in the AL register into a
decimal value and stores the result in the AL register in unpacked BCD format (the
decimal number is stored in the lower 4 bits of the register and the upper 4 bits are
cleared). If a decimal carry occurred as a result of the addition, the CF flag is set and
the contents of the AH register are incremented by 1.
The AAS instruction adjusts the contents of the AL register following the subtraction
of two unpacked BCD values. Here again, a binary value is converted into an
unpacked BCD value. If a borrow was required to complete the decimal subtract, the
CF flag is set and the contents of the AH register are decremented by 1.
The AAM instruction adjusts the contents of the AL register following a multiplication
of two unpacked BCD values. It converts the binary value in the AL register into a
decimal value and stores the least significant digit of the result in the AL register (in
unpacked BCD format) and the most significant digit, if there is one, in the AH
register (also in unpacked BCD format).
The AAD instruction adjusts a two-digit BCD value so that when the value is divided
with the DIV instruction, a valid unpacked BCD result is obtained. The instruction
converts the BCD value in registers AH (most significant digit) and AL (least signifi-
cant digit) into a binary value and stores the result in register AL. When the value in
AL is divided by an unpacked BCD value, the quotient and remainder will be automat-
ically encoded in unpacked BCD format.

7-14 Vol. 1
 PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

7.3.4 Decimal Arithmetic Instructions in 64-Bit Mode

Decimal arithmetic instructions are not supported in 64-bit mode, They are either
invalid or not encodable.

7.3.5 Logical Instructions

The logical instructions AND, OR, XOR (exclusive or), and NOT perform the standard
Boolean operations for which they are named. The AND, OR, and XOR instructions
require two operands; the NOT instruction operates on a single operand.

7.3.6 Shift and Rotate Instructions

The shift and rotate instructions rearrange the bits within an operand. For the
purpose of this discussion, these instructions are further divided subordinate
subgroups of instructions that:
• Shift bits
• Double-shift bits (move them between operands)
• Rotate bits

7.3.6.1 Shift Instructions

The SAL (shift arithmetic left), SHL (shift logical left), SAR (shift arithmetic right),
SHR (shift logical right) instructions perform an arithmetic or logical shift of the bits
in a byte, word, or doubleword.
The SAL and SHL instructions perform the same operation (see Figure 7-6). They
shift the source operand left by from 1 to 31 bit positions. Empty bit positions are
cleared. The CF flag is loaded with the last bit shifted out of the operand.

Vol. 1 7-15
PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

Initial State
CF Operand

X 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 1 1 1

After 1-bit SHL/SAL Instruction

0
1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 1 1 1 0

After 10-bit SHL/SAL Instruction

0
0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 1 1 1 0 0 0 0 0 0 0 0 0 0

Figure 7-6. SHL/SAL Instruction Operation

The SHR instruction shifts the source operand right by from 1 to 31 bit positions (see
Figure 7-7). As with the SHL/SAL instruction, the empty bit positions are cleared and
the CF flag is loaded with the last bit shifted out of the operand.

Initial State Operand CF

1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 1 1 1 X

After 1-bit SHR Instruction

0
0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 1 1 1

After 10-bit SHR Instruction

0
0 0 0 0 0 0 0 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0

Figure 7-7. SHR Instruction Operation

7-16 Vol. 1
 PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

The SAR instruction shifts the source operand right by from 1 to 31 bit positions (see
Figure 7-8). This instruction differs from the SHR instruction in that it preserves the
sign of the source operand by clearing empty bit positions if the operand is positive or
setting the empty bits if the operand is negative. Again, the CF flag is loaded with the
last bit shifted out of the operand.
The SAR and SHR instructions can also be used to perform division by powers of
2 (see “SAL/SAR/SHL/SHR—Shift Instructions” in Chapter 4, “Instruction Set Refer-
ence, N-Z,” of the Intel® 64 and IA-32 Architectures Software Developer’s Manual,
Volume 2B).

Initial State (Positive Operand) Operand CF

0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 1 1 X

After 1-bit SAR Instruction

0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 1 1

Initial State (Negative Operand)

CF
1 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 1 1 X

After 1-bit SAR Instruction

1 1 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 1 1

Figure 7-8. SAR Instruction Operation

7.3.6.2 Double-Shift Instructions

The SHLD (shift left double) and SHRD (shift right double) instructions shift a speci-
fied number of bits from one operand to another (see Figure 7-9). They are provided
to facilitate operations on unaligned bit strings. They can also be used to implement a
variety of bit string move operations.

Vol. 1 7-17
PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

SHLD Instruction
31 0
CF Destination (Memory or Register)

31 0
Source (Register)

SHRD Instruction
31 0
Source (Register)

31 0
Destination (Memory or Register) CF

Figure 7-9. SHLD and SHRD Instruction Operations

The SHLD instruction shifts the bits in the destination operand to the left and fills the
empty bit positions (in the destination operand) with bits shifted out of the source
operand. The destination and source operands must be the same length (either
words or doublewords). The shift count can range from 0 to 31 bits. The result of this
shift operation is stored in the destination operand, and the source operand is not
modified. The CF flag is loaded with the last bit shifted out of the destination operand.
The SHRD instruction operates the same as the SHLD instruction except bits are
shifted to the right in the destination operand, with the empty bit positions filled with
bits shifted out of the source operand.

7.3.6.3 Rotate Instructions

The ROL (rotate left), ROR (rotate right), RCL (rotate through carry left) and RCR
(rotate through carry right) instructions rotate the bits in the destination operand out
of one end and back through the other end (see Figure 7-10). Unlike a shift, no bits
are lost during a rotation. The rotate count can range from 0 to 31.

7-18 Vol. 1
 PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

ROL Instruction
31 0

CF Destination (Memory or Register)

31 ROR Instruction 0
Destination (Memory or Register) CF

RCL Instruction
31 0
CF Destination (Memory or Register)

RCR Instruction
31 0
Destination (Memory or Register) CF

Figure 7-10. ROL, ROR, RCL, and RCR Instruction Operations

The ROL instruction rotates the bits in the operand to the left (toward more signifi-
cant bit locations). The ROR instruction rotates the operand right (toward less signif-
icant bit locations).
The RCL instruction rotates the bits in the operand to the left, through the CF flag.
This instruction treats the CF flag as a one-bit extension on the upper end of the
operand. Each bit that exits from the most significant bit location of the operand
moves into the CF flag. At the same time, the bit in the CF flag enters the least signif-
icant bit location of the operand.
The RCR instruction rotates the bits in the operand to the right through the CF flag.
For all the rotate instructions, the CF flag always contains the value of the last bit
rotated out of the operand, even if the instruction does not use the CF flag as an
extension of the operand. The value of this flag can then be tested by a conditional
jump instruction (JC or JNC).

Vol. 1 7-19
PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

7.3.7 Bit and Byte Instructions

These instructions operate on bit or byte strings. For the purpose of this discussion,
they are further divided subordinate subgroups that:
• Test and modify a single bit
• Scan a bit string
• Set a byte given conditions
• Test operands and report results

7.3.7.1 Bit Test and Modify Instructions

The bit test and modify instructions (see Table 7-3) operate on a single bit, which can
be in an operand. The location of the bit is specified as an offset from the least signif-
icant bit of the operand. When the processor identifies the bit to be tested and modi-
fied, it first loads the CF flag with the current value of the bit. Then it assigns a new
value to the selected bit, as determined by the modify operation for the instruction.

Table 7-3. Bit Test and Modify Instructions

Instruction Effect on CF Flag Effect on Selected Bit
BT (Bit Test) CF flag ← Selected Bit No effect
BTS (Bit Test and Set) CF flag ← Selected Bit Selected Bit ← 1
BTR (Bit Test and Reset) CF flag ← Selected Bit Selected Bit ← 0
BTC (Bit Test and CF flag ← Selected Bit Selected Bit ← NOT (Selected Bit)
Complement)

7.3.7.2 Bit Scan Instructions

The BSF (bit scan forward) and BSR (bit scan reverse) instructions scan a bit string in
a source operand for a set bit and store the bit index of the first set bit found in a
destination register. The bit index is the offset from the least significant bit (bit 0) in
the bit string to the first set bit. The BSF instruction scans the source operand low-to-
high (from bit 0 of the source operand toward the most significant bit); the BSR
instruction scans high-to-low (from the most significant bit toward the least signifi-
cant bit).

7.3.7.3 Byte Set on Condition Instructions

The SETcc (set byte on condition) instructions set a destination-operand byte to 0 or
1, depending on the state of selected status flags (CF, OF, SF, ZF, and PF) in the
EFLAGS register. The suffix (cc) added to the SET mnemonic determines the condi-
tion being tested for.
For example, the SETO instruction tests for overflow. If the OF flag is set, the desti-
nation byte is set to 1; if OF is clear, the destination byte is cleared to 0. Appendix B,

7-20 Vol. 1
 PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

“EFLAGS Condition Codes,” lists the conditions it is possible to test for with this
instruction.

7.3.7.4 Test Instruction

The TEST instruction performs a logical AND of two operands and sets the SF, ZF, and
PF flags according to the results. The flags can then be tested by the conditional jump
or loop instructions or the SETcc instructions. The TEST instruction differs from the
AND instruction in that it does not alter either of the operands.

7.3.8 Control Transfer Instructions

The processor provides both conditional and unconditional control transfer instruc-
tions to direct the flow of program execution. Conditional transfers are taken only for
specified states of the status flags in the EFLAGS register. Unconditional control
transfers are always executed.
For the purpose of this discussion, these instructions are further divided subordinate
subgroups that process:
• Unconditional transfers
• Conditional transfers
• Software interrupts

7.3.8.1 Unconditional Transfer Instructions

The JMP, CALL, RET, INT, and IRET instructions transfer program control to another
location (destination address) in the instruction stream. The destination can be
within the same code segment (near transfer) or in a different code segment (far
transfer).
Jump instruction — The JMP (jump) instruction unconditionally transfers program
control to a destination instruction. The transfer is one-way; that is, a return address
is not saved. A destination operand specifies the address (the instruction pointer) of
the destination instruction. The address can be a relative address or an absolute
address.
A relative address is a displacement (offset) with respect to the address in the EIP
register. The destination address (a near pointer) is formed by adding the displace-
ment to the address in the EIP register. The displacement is specified with a signed
integer, allowing jumps either forward or backward in the instruction stream.
An absolute address is a offset from address 0 of a segment. It can be specified in
either of the following ways:
• An address in a general-purpose register — This address is treated as a near
pointer, which is copied into the EIP register. Program execution then continues at
the new address within the current code segment.

Vol. 1 7-21
PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

• An address specified using the standard addressing modes of the

processor — Here, the address can be a near pointer or a far pointer. If the
address is for a near pointer, the address is translated into an offset and copied
into the EIP register. If the address is for a far pointer, the address is translated
into a segment selector (which is copied into the CS register) and an offset
(which is copied into the EIP register).
In protected mode, the JMP instruction also allows jumps to a call gate, a task gate,
and a task-state segment.
Call and return instructions — The CALL (call procedure) and RET (return from
procedure) instructions allow a jump from one procedure (or subroutine) to another
and a subsequent jump back (return) to the calling procedure.
The CALL instruction transfers program control from the current (or calling proce-
dure) to another procedure (the called procedure). To allow a subsequent return to
the calling procedure, the CALL instruction saves the current contents of the EIP
register on the stack before jumping to the called procedure. The EIP register (prior
to transferring program control) contains the address of the instruction following the
CALL instruction. When this address is pushed on the stack, it is referred to as the
return instruction pointer or return address.
The address of the called procedure (the address of the first instruction in the proce-
dure being jumped to) is specified in a CALL instruction the same way as it is in a JMP
instruction (see “Jump instruction” on page 7-21). The address can be specified as a
relative address or an absolute address. If an absolute address is specified, it can be
either a near or a far pointer.
The RET instruction transfers program control from the procedure currently being
executed (the called procedure) back to the procedure that called it (the calling
procedure). Transfer of control is accomplished by copying the return instruction
pointer from the stack into the EIP register. Program execution then continues with
the instruction pointed to by the EIP register.
The RET instruction has an optional operand, the value of which is added to the
contents of the ESP register as part of the return operation. This operand allows the
stack pointer to be incremented to remove parameters from the stack that were
pushed on the stack by the calling procedure.
See Section 6.3, “Calling Procedures Using CALL and RET,” for more information on
the mechanics of making procedure calls with the CALL and RET instructions.
Return from interrupt instruction — When the processor services an interrupt, it
performs an implicit call to an interrupt-handling procedure. The IRET (return from
interrupt) instruction returns program control from an interrupt handler to the inter-
rupted procedure (that is, the procedure that was executing when the interrupt
occurred). The IRET instruction performs a similar operation to the RET instruction
(see “Call and return instructions” on page 7-22) except that it also restores the
EFLAGS register from the stack. The contents of the EFLAGS register are automati-
cally stored on the stack along with the return instruction pointer when the processor
services an interrupt.

7-22 Vol. 1
 PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

7.3.8.2 Conditional Transfer Instructions

The conditional transfer instructions execute jumps or loops that transfer program
control to another instruction in the instruction stream if specified conditions are
met. The conditions for control transfer are specified with a set of condition codes
that define various states of the status flags (CF, ZF, OF, PF, and SF) in the EFLAGS
register.
Conditional jump instructions — The Jcc (conditional) jump instructions transfer
program control to a destination instruction if the conditions specified with the condi-
tion code (cc) associated with the instruction are satisfied (see Table 7-4). If the
condition is not satisfied, execution continues with the instruction following the Jcc
instruction. As with the JMP instruction, the transfer is one-way; that is, a return
address is not saved.

Table 7-4. Conditional Jump Instructions

Instruction Mnemonic Condition (Flag States) Description
Unsigned Conditional Jumps
JA/JNBE (CF or ZF) = 0 Above/not below or equal
JAE/JNB CF = 0 Above or equal/not below
JB/JNAE CF = 1 Below/not above or equal
JBE/JNA (CF or ZF) = 1 Below or equal/not above
JC CF = 1 Carry
JE/JZ ZF = 1 Equal/zero
JNC CF = 0 Not carry
JNE/JNZ ZF = 0 Not equal/not zero
JNP/JPO PF = 0 Not parity/parity odd
JP/JPE PF = 1 Parity/parity even
JCXZ CX = 0 Register CX is zero
JECXZ ECX = 0 Register ECX is zero
Signed Conditional Jumps
JG/JNLE ((SF xor OF) or ZF) = 0 Greater/not less or equal
JGE/JNL (SF xor OF) = 0 Greater or equal/not less
JL/JNGE (SF xor OF) = 1 Less/not greater or equal
JLE/JNG ((SF xor OF) or ZF) = 1 Less or equal/not greater
JNO OF = 0 Not overflow
JNS SF = 0 Not sign (non-negative)
JO OF = 1 Overflow
JS SF = 1 Sign (negative)

Vol. 1 7-23
PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

The destination operand specifies a relative address (a signed offset with respect to
the address in the EIP register) that points to an instruction in the current code
segment. The Jcc instructions do not support far transfers; however, far transfers can
be accomplished with a combination of a Jcc and a JMP instruction (see “Jcc—Jump if
Condition Is Met” in Chapter 3, “Instruction Set Reference, A-M,” of the Intel® 64
and IA-32 Architectures Software Developer’s Manual, Volume 2A).
Table 7-4 shows the mnemonics for the Jcc instructions and the conditions being
tested for each instruction. The condition code mnemonics are appended to the letter
“J” to form the mnemonic for a Jcc instruction. The instructions are divided into two
groups: unsigned and signed conditional jumps. These groups correspond to the
results of operations performed on unsigned and signed integers respectively. Those
instructions listed as pairs (for example, JA/JNBE) are alternate names for the same
instruction. Assemblers provide alternate names to make it easier to read program
listings.
The JCXZ and JECXZ instructions test the CX and ECX registers, respectively, instead
of one or more status flags. See “Jump if zero instructions” on page 7-25 for more
information about these instructions.
Loop instructions — The LOOP, LOOPE (loop while equal), LOOPZ (loop while zero),
LOOPNE (loop while not equal), and LOOPNZ (loop while not zero) instructions are
conditional jump instructions that use the value of the ECX register as a count for the
number of times to execute a loop. All the loop instructions decrement the count in
the ECX register each time they are executed and terminate a loop when zero is
reached. The LOOPE, LOOPZ, LOOPNE, and LOOPNZ instructions also accept the ZF
flag as a condition for terminating the loop before the count reaches zero.
The LOOP instruction decrements the contents of the ECX register (or the CX register,
if the address-size attribute is 16), then tests the register for the loop-termination
condition. If the count in the ECX register is non-zero, program control is transferred
to the instruction address specified by the destination operand. The destination
operand is a relative address (that is, an offset relative to the contents of the EIP
register), and it generally points to the first instruction in the block of code that is to
be executed in the loop. When the count in the ECX register reaches zero, program
control is transferred to the instruction immediately following the LOOP instruc-
tion, which terminates the loop. If the count in the ECX register is zero when the
LOOP instruction is first executed, the register is pre-decremented to FFFFFFFFH,
causing the loop to be executed 232 times.
The LOOPE and LOOPZ instructions perform the same operation (they are
mnemonics for the same instruction). These instructions operate the same as the
LOOP instruction, except that they also test the ZF flag.
If the count in the ECX register is not zero and the ZF flag is set, program control is
transferred to the destination operand. When the count reaches zero or the ZF flag is
clear, the loop is terminated by transferring program control to the instruction imme-
diately following the LOOPE/LOOPZ instruction.

7-24 Vol. 1
 PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

The LOOPNE and LOOPNZ instructions (mnemonics for the same instruction) operate
the same as the LOOPE/LOOPPZ instructions, except that they terminate the loop if
the ZF flag is set.
Jump if zero instructions — The JECXZ (jump if ECX zero) instruction jumps to the
location specified in the destination operand if the ECX register contains the value
zero. This instruction can be used in combination with a loop instruction (LOOP,
LOOPE, LOOPZ, LOOPNE, or LOOPNZ) to test the ECX register prior to beginning a
loop. As described in “Loop instructions on page 7-24, the loop instructions decre-
ment the contents of the ECX register before testing for zero. If the value in the ECX
register is zero initially, it will be decremented to FFFFFFFFH on the first loop instruc-
tion, causing the loop to be executed 232 times. To prevent this problem, a JECXZ
instruction can be inserted at the beginning of the code block for the loop, causing a
jump out the loop if the EAX register count is initially zero. When used with repeated
string scan and compare instructions, the JECXZ instruction can determine whether
the loop terminated because the count reached zero or because the scan or compare
conditions were satisfied.
The JCXZ (jump if CX is zero) instruction operates the same as the JECXZ instruction
when the 16-bit address-size attribute is used. Here, the CX register is tested for
zero.

7.3.8.3 Control Transfer Instructions in 64-Bit Mode

In 64-bit mode, the operand size for all near branches (CALL, RET, JCC, JCXZ, JMP,
and LOOP) is forced to 64 bits. The listed instructions update the 64-bit RIP without
need for a REX operand-size prefix.
Near branches in the following operations are forced to 64-bits (regardless of
operand size prefixes):
• Truncation of the size of the instruction pointer
• Size of a stack pop or push, due to CALL or RET
• Size of a stack-pointer increment or decrement, due to CALL or RET
• Indirect-branch operand size
Note that the displacement field for relative branches is still limited to 32 bits and the
address size for near branches is not forced.
Address size determines the register size (CX/ECX/RCX) used for JCXZ and LOOP. It
also impacts the address calculation for memory indirect branches. Addresses size is
64 bits by default, although it can be over-ridden to 32 bits (using a prefix).

7.3.8.4 Software Interrupt Instructions

The INT n (software interrupt), INTO (interrupt on overflow), and BOUND (detect
value out of range) instructions allow a program to explicitly raise a specified inter-
rupt or exception, which in turn causes the handler routine for the interrupt or excep-
tion to be called.

Vol. 1 7-25
PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

The INT n instruction can raise any of the processor’s interrupts or exceptions by
encoding the vector number or the interrupt or exception in the instruction. This
instruction can be used to support software generated interrupts or to test the oper-
ation of interrupt and exception handlers.
The IRET (return from interrupt) instruction returns program control from an inter-
rupt handler to the interrupted procedure. The IRET instruction performs a similar
operation to the RET instruction.
The CALL (call procedure) and RET (return from procedure) instructions allow a jump
from one procedure to another and a subsequent return to the calling procedure.
EFLAGS register contents are automatically stored on the stack along with the return
instruction pointer when the processor services an interrupt.
The INTO instruction raises the overflow exception if the OF flag is set. If the flag is
clear, execution continues without raising the exception. This instruction allows soft-
ware to access the overflow exception handler explicitly to check for overflow condi-
tions.
The BOUND instruction compares a signed value against upper and lower bounds,
and raises the “BOUND range exceeded” exception if the value is less than the lower
bound or greater than the upper bound. This instruction is useful for operations such
as checking an array index to make sure it falls within the range defined for the array.

7.3.8.5 Software Interrupt Instructions in 64-bit Mode and Compatibility

Mode
In 64-bit mode, the stack size is 8 bytes wide. IRET must pop 8-byte items off the
stack. SS:RSP pops unconditionally. BOUND is not supported.
In compatibility mode, SS:RSP is popped only if the CPL changes.

7.3.9 String Operations

The MOVS (Move String), CMPS (Compare string), SCAS (Scan string), LODS (Load
string), and STOS (Store string) instructions permit large data structures, such as
alphanumeric character strings, to be moved and examined in memory. These
instructions operate on individual elements in a string, which can be a byte, word, or
doubleword. The string elements to be operated on are identified with the ESI
(source string element) and EDI (destination string element) registers. Both of these
registers contain absolute addresses (offsets into a segment) that point to a string
element.
By default, the ESI register addresses the segment identified with the DS segment
register. A segment-override prefix allows the ESI register to be associated with the
CS, SS, ES, FS, or GS segment register. The EDI register addresses the segment
identified with the ES segment register; no segment override is allowed for the EDI
register. The use of two different segment registers in the string instructions permits
operations to be performed on strings located in different segments. Or by associ-
ating the ESI register with the ES segment register, both the source and destination

7-26 Vol. 1
 PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

strings can be located in the same segment. (This latter condition can also be
achieved by loading the DS and ES segment registers with the same segment
selector and allowing the ESI register to default to the DS register.)
The MOVS instruction moves the string element addressed by the ESI register to the
location addressed by the EDI register. The assembler recognizes three “short forms”
of this instruction, which specify the size of the string to be moved: MOVSB (move
byte string), MOVSW (move word string), and MOVSD (move doubleword string).
The CMPS instruction subtracts the destination string element from the source string
element and updates the status flags (CF, ZF, OF, SF, PF, and AF) in the EFLAGS
register according to the results. Neither string element is written back to memory.
The assembler recognizes three “short forms” of the CMPS instruction: CMPSB
(compare byte strings), CMPSW (compare word strings), and CMPSD (compare
doubleword strings).
The SCAS instruction subtracts the destination string element from the contents of
the EAX, AX, or AL register (depending on operand length) and updates the status
flags according to the results. The string element and register contents are not modi-
fied. The following “short forms” of the SCAS instruction specify the operand length:
SCASB (scan byte string), SCASW (scan word string), and SCASD (scan doubleword
string).
The LODS instruction loads the source string element identified by the ESI register
into the EAX register (for a doubleword string), the AX register (for a word string), or
the AL register (for a byte string). The “short forms” for this instruction are LODSB
(load byte string), LODSW (load word string), and LODSD (load doubleword string).
This instruction is usually used in a loop, where other instructions process each
element of the string after they are loaded into the target register.
The STOS instruction stores the source string element from the EAX (doubleword
string), AX (word string), or AL (byte string) register into the memory location iden-
tified with the EDI register. The “short forms” for this instruction are STOSB (store
byte string), STOSW (store word string), and STOSD (store doubleword string). This
instruction is also normally used in a loop. Here a string is commonly loaded into
the register with a LODS instruction, operated on by other instructions, and then
stored again in memory with a STOS instruction.
The I/O instructions (see Section 7.3.11, “I/O Instructions”) also perform operations
on strings in memory.

7.3.9.1 Repeating String Operations

The string instructions described in Section 7.3.9, “String Operations”, perform one
iteration of a string operation. To operate strings longer than a doubleword, the
string instructions can be combined with a repeat prefix (REP) to create a repeating
instruction or be placed in a loop.
When used in string instructions, the ESI and EDI registers are automatically incre-
mented or decremented after each iteration of an instruction to point to the next
element (byte, word, or doubleword) in the string. String operations can thus begin

Vol. 1 7-27
PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

at higher addresses and work toward lower ones, or they can begin at lower
addresses and work toward higher ones. The DF flag in the EFLAGS register controls
whether the registers are incremented (DF = 0) or decremented (DF = 1). The STD
and CLD instructions set and clear this flag, respectively.
The following repeat prefixes can be used in conjunction with a count in the ECX
register to cause a string instruction to repeat:
• REP — Repeat while the ECX register not zero.
• REPE/REPZ — Repeat while the ECX register not zero and the ZF flag is set.
• REPNE/REPNZ — Repeat while the ECX register not zero and the ZF flag is clear.
When a string instruction has a repeat prefix, the operation executes until one of the
termination conditions specified by the prefix is satisfied. The REPE/REPZ and
REPNE/REPNZ prefixes are used only with the CMPS and SCAS instructions. Also,
note that a REP STOS instruction is the fastest way to initialize a large block of
memory.

7.3.10 String Operations in 64-Bit Mode

The behavior of MOVS (Move String), CMPS (Compare string), SCAS (Scan string),
LODS (Load string), and STOS (Store string) instructions in 64-bit mode is similar to
their behavior in non-64-bit modes, with the following differences:
• The source operand is specified by RSI or DS:ESI, depending on the address size
attribute of the operation.
• The destination operand is specified by RDI or DS:EDI, depending on the address
size attribute of the operation.
• Operation on 64-bit data is supported by using the REX.W prefix.

7.3.10.1 Repeating String Operations in 64-bit Mode

When using REP prefixes for string operations in 64-bit mode, the repeat count is
specified by RCX or ECX (depending on the address size attribute of the operation).
The default address size is 64 bits.

7.3.11 I/O Instructions

The IN (input from port to register), INS (input from port to string), OUT (output
from register to port), and OUTS (output string to port) instructions move data
between the processor’s I/O ports and either a register or memory.
The register I/O instructions (IN and OUT) move data between an I/O port and the
EAX register (32-bit I/O), the AX register (16-bit I/O), or the AL (8-bit I/O) register.
The I/O port being read or written to is specified with an immediate operand or an
address in the DX register.

7-28 Vol. 1
 PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

The block I/O instructions (INS and OUTS) instructions move blocks of data (strings)
between an I/O port and memory. These instructions operate similar to the string
instructions (see Section 7.3.9, “String Operations”). The ESI and EDI registers are
used to specify string elements in memory and the repeat prefixes (REP) are used to
repeat the instructions to implement block moves. The assembler recognizes the
following alternate mnemonics for these instructions: INSB (input byte), INSW (input
word), and INSD (input doubleword), and OUTB (output byte), OUTW (output word),
and OUTD (output doubleword).
The INS and OUTS instructions use an address in the DX register to specify the I/O
port to be read or written to.

7.3.12 I/O Instructions in 64-Bit Mode

For I/O instructions to and from memory, the differences in 64-bit mode are:
• The source operand is specified by RSI or DS:ESI, depending on the address size
attribute of the operation.
• The destination operand is specified by RDI or DS:EDI, depending on the address
size attribute of the operation.
• Operation on 64-bit data is not encodable and REX prefixes are silently ignored.

7.3.13 Enter and Leave Instructions

The ENTER and LEAVE instructions provide machine-language support for procedure
calls in block-structured languages, such as C and Pascal. These instructions and the
call and return mechanism that they support are described in detail in Section 6.5,
“Procedure Calls for Block-Structured Languages”.

7.3.14 Flag Control (EFLAG) Instructions

The Flag Control (EFLAG) instructions allow the state of selected flags in the EFLAGS
register to be read or modified. For the purpose of this discussion, these instructions
are further divided subordinate subgroups of instructions that manipulate:
• Carry and direction flags
• The EFLAGS register
• Interrupt flags

7.3.14.1 Carry and Direction Flag Instructions

The STC (set carry flag), CLC (clear carry flag), and CMC (complement carry flag)
instructions allow the CF flags in the EFLAGS register to be modified directly. They
are typically used to initialize the CF flag to a known state before an instruction that

Vol. 1 7-29
PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

uses the flag in an operation is executed. They are also used in conjunction with the
rotate-with-carry instructions (RCL and RCR).
The STD (set direction flag) and CLD (clear direction flag) instructions allow the DF
flag in the EFLAGS register to be modified directly. The DF flag determines the direc-
tion in which index registers ESI and EDI are stepped when executing string
processing instructions. If the DF flag is clear, the index registers are incremented
after each iteration of a string instruction; if the DF flag is set, the registers are
decremented.

7.3.14.2 EFLAGS Transfer Instructions

The EFLAGS transfer instructions allow groups of flags in the EFLAGS register to be
copied to a register or memory or be loaded from a register or memory.
The LAHF (load AH from flags) and SAHF (store AH into flags) instructions operate on
five of the EFLAGS status flags (SF, ZF, AF, PF, and CF). The LAHF instruction copies
the status flags to bits 7, 6, 4, 2, and 0 of the AH register, respectively. The contents
of the remaining bits in the register (bits 5, 3, and 1) are undefined, and the contents
of the EFLAGS register remain unchanged. The SAHF instruction copies bits 7, 6, 4,
2, and 0 from the AH register into the SF, ZF, AF, PF, and CF flags, respectively in the
EFLAGS register.
The PUSHF (push flags), PUSHFD (push flags double), POPF (pop flags), and POPFD
(pop flags double) instructions copy the flags in the EFLAGS register to and from the
stack. The PUSHF instruction pushes the lower word of the EFLAGS register onto the
stack (see Figure 7-11). The PUSHFD instruction pushes the entire EFLAGS register
onto the stack (with the RF and VM flags read as clear).

PUSHFD/POPFD

PUSHF/POPF

31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0
I
V V
A V R 0 N O O D I T S Z P C
0 0 0 0 0 0 0 0 0 0 I I I A
D C M F T P F F F F F F 0 F 0 F 1 F
P F
L

Figure 7-11. Flags Affected by the PUSHF, POPF, PUSHFD, and POPFD Instructions

The POPF instruction pops a word from the stack into the EFLAGS register. Only bits
11, 10, 8, 7, 6, 4, 2, and 0 of the EFLAGS register are affected with all uses of this
instruction. If the current privilege level (CPL) of the current code segment is 0 (most
privileged), the IOPL bits (bits 13 and 12) also are affected. If the I/O privilege level
(IOPL) is greater than or equal to the CPL, numerically, the IF flag (bit 9) also is
affected.

7-30 Vol. 1
 PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

The POPFD instruction pops a doubleword into the EFLAGS register. This instruction
can change the state of the AC bit (bit 18) and the ID bit (bit 21), as well as the bits
affected by a POPF instruction. The restrictions for changing the IOPL bits and the IF
flag that were given for the POPF instruction also apply to the POPFD instruction.

7.3.14.3 Interrupt Flag Instructions

The STI (set interrupt flag) and CTI (clear interrupt flag) instructions allow the inter-
rupt IF flag in the EFLAGS register to be modified directly. The IF flag controls the
servicing of hardware-generated interrupts (those received at the processor’s INTR
pin). If the IF flag is set, the processor services hardware interrupts; if the IF flag is
clear, hardware interrupts are masked.
The ability to execute these instructions depends on the operating mode of the
processor and the current privilege level (CPL) of the program or task attempting to
execute these instructions.

7.3.15 Flag Control (RFLAG) Instructions in 64-Bit Mode

In 64-bit mode, the LAHF and SAHF instructions are supported if
CPUID.80000001H:ECX.LAHF-SAHF[bit 0] = 1.
PUSHF and POPF behave the same in 64-bit mode as in non-64-bit mode. PUSHFD
always pushes 64-bit RFLAGS onto the stack (with the RF and VM flags read as clear).
POPFD always pops a 64-bit value from the top of the stack and loads the lower 32
bits into RFLAGS. It then zero extends the upper bits of RFLAGS.

7.3.16 Segment Register Instructions

The processor provides a variety of instructions that address the segment registers
of the processor directly. These instructions are only used when an operating system
or executive is using the segmented or the real-address mode memory model.
For the purpose of this discussion, these instructions are divided subordinate
subgroups of instructions that allow:
• Segment-register load and store
• Far control transfers
• Software interrupt calls
• Handling of far pointers

7.3.16.1 Segment-Register Load and Store Instructions

The MOV instruction (introduced in Section 7.3.1.1, “General Data Movement
Instructions”) and the PUSH and POP instructions (introduced in Section 7.3.1.4,
“Stack Manipulation Instructions”) can transfer 16-bit segment selectors to and from

Vol. 1 7-31
PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

segment registers (DS, ES, FS, GS, and SS). The transfers are always made to or
from a segment register and a general-purpose register or memory. Transfers
between segment registers are not supported.
The POP and MOV instructions cannot place a value in the CS register. Only the far
control-transfer versions of the JMP, CALL, and RET instructions (see Section
7.3.16.2, “Far Control Transfer Instructions”) affect the CS register directly.

7.3.16.2 Far Control Transfer Instructions

The JMP and CALL instructions (see Section 7.3.8, “Control Transfer Instructions”)
both accept a far pointer as a source operand to transfer program control to a
segment other than the segment currently being pointed to by the CS register. When
a far call is made with the CALL instruction, the current values of the EIP and CS
registers are both pushed on the stack.
The RET instruction (see “Call and return instructions” on page 7-22) can be used to
execute a far return. Here, program control is transferred from a code segment that
contains a called procedure back to the code segment that contained the calling
procedure. The RET instruction restores the values of the CS and EIP registers for the
calling procedure from the stack.

7.3.16.3 Software Interrupt Instructions

The software interrupt instructions INT, INTO, BOUND, and IRET (see Section
7.3.8.4, “Software Interrupt Instructions”) can also call and return from interrupt
and exception handler procedures that are located in a code segment other than the
current code segment. With these instructions, however, the switching of code
segments is handled transparently from the application program.

7.3.16.4 Load Far Pointer Instructions

The load far pointer instructions LDS (load far pointer using DS), LES (load far
pointer using ES), LFS (load far pointer using FS), LGS (load far pointer using GS),
and LSS (load far pointer using SS) load a far pointer from memory into a segment
register and a general-purpose general register. The segment selector part of the far
pointer is loaded into the selected segment register and the offset is loaded into the
selected general-purpose register.

7.3.17 Miscellaneous Instructions

The following instructions perform operations that are of interest to applications
programmers. For the purpose of this discussion, these instructions are further
divided into subordinate subgroups of instructions that provide for:
• Address computations
• Table lookup

7-32 Vol. 1
 PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

• Processor identification
• NOP and undefined instruction entry

7.3.17.1 Address Computation Instruction

The LEA (load effective address) instruction computes the effective address in
memory (offset within a segment) of a source operand and places it in a general-
purpose register. This instruction can interpret any of the processor’s addressing
modes and can perform any indexing or scaling that may be needed. It is especially
useful for initializing the ESI or EDI registers before the execution of string instruc-
tions or for initializing the EBX register before an XLAT instruction.

7.3.17.2 Table Lookup Instructions

The XLAT and XLATB (table lookup) instructions replace the contents of the AL
register with a byte read from a translation table in memory. The initial value in the
AL register is interpreted as an unsigned index into the translation table. This index
is added to the contents of the EBX register (which contains the base address of the
table) to calculate the address of the table entry. These instructions are used for
applications such as converting character codes from one alphabet into another (for
example, an ASCII code could be used to look up its EBCDIC equivalent in a table).

7.3.17.3 Processor Identification Instruction

The CPUID (processor identification) instruction returns information about the
processor on which the instruction is executed.

7.3.17.4 No-Operation and Undefined Instructions

The NOP (no operation) instruction increments the EIP register to point at the next
instruction, but affects nothing else.
The UD2 (undefined) instruction generates an invalid opcode exception. Intel
reserves the opcode for this instruction for this function. The instruction is provided
to allow software to test an invalid opcode exception handler.

Vol. 1 7-33
PROGRAMMING WITH GENERAL-PURPOSE INSTRUCTIONS

7-34 Vol. 1
 CHAPTER 8
PROGRAMMING WITH THE X87 FPU

The x87 Floating-Point Unit (FPU) provides high-performance floating-point

processing capabilities for use in graphics processing, scientific, engineering, and
business applications. It supports the floating-point, integer, and packed BCD integer
data types and the floating-point processing algorithms and exception handling
architecture defined in the IEEE Standard 754 for Binary Floating-Point Arithmetic.
This chapter describes the x87 FPU’s execution environment and instruction set. It
also provides exception handling information that is specific to the x87 FPU. Refer to
the following chapters or sections of chapters for additional information about x87
FPU instructions and floating-point operations:
• Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volumes
2A & 2B, provide detailed descriptions of x87 FPU instructions.
• Section 4.2.2, “Floating-Point Data Types,” Section 4.2.1.2, “Signed Integers,”
and Section 4.7, “BCD and Packed BCD Integers,” describe the floating-point,
integer, and BCD data types.
• Section 4.9, “Overview of Floating-Point Exceptions,” Section 4.9.1, “Floating-
Point Exception Conditions,” and Section 4.9.2, “Floating-Point Exception
Priority,” give an overview of the floating-point exceptions that the x87 FPU can
detect and report.

8.1 X87 FPU EXECUTION ENVIRONMENT

The x87 FPU represents a separate execution environment within the IA-32 architec-
ture (see Figure 8-1). This execution environment consists of eight data registers
(called the x87 FPU data registers) and the following special-purpose registers:
• Status register
• Control register
• Tag word register
• Last instruction pointer register
• Last data (operand) pointer register
• Opcode register
These registers are described in the following sections.
The x87 FPU executes instructions from the processor’s normal instruction stream.
The state of the x87 FPU is independent from the state of the basic execution envi-
ronment and from the state of SSE/SSE2/SSE3 extensions.
However, the x87 FPU and Intel MMX technology share state because the MMX regis-
ters are aliased to the x87 FPU data registers. Therefore, when writing code that uses

Vol. 1 8-1
PROGRAMMING WITH THE X87 FPU

x87 FPU and MMX instructions, the programmer must explicitly manage the x87 FPU
and MMX state (see Section 9.5, “Compatibility with x87 FPU Architecture”).

8.1.1 x87 FPU in 64-Bit Mode and Compatibility Mode

In compatibility mode and 64-bit mode, x87 FPU instructions function like they do in
protected mode. Memory operands are specified using the ModR/M, SIB encoding
that is described in Section 3.7.5, “Specifying an Offset.”

8.1.2 x87 FPU Data Registers

The x87 FPU data registers (shown in Figure 8-1) consist of eight 80-bit registers.
Values are stored in these registers in the double extended-precision floating-point
format shown in Figure 4-3. When floating-point, integer, or packed BCD integer
values are loaded from memory into any of the x87 FPU data registers, the values are
automatically converted into double extended-precision floating-point format (if they
are not already in that format). When computation results are subsequently trans-
ferred back into memory from any of the x87 FPU registers, the results can be left in
the double extended-precision floating-point format or converted back into a shorter
floating-point format, an integer format, or the packed BCD integer format. (See
Section 8.2, “x87 FPU Data Types,” for a description of the data types operated on by
the x87 FPU.)

8-2 Vol. 1
 PROGRAMMING WITH THE X87 FPU

Data Registers
Sign 79 78 64 63 0
R7 Exponent Significand
R6
R5
R4
R3
R2
R1
R0

15 0 47 0
Control Last Instruction Pointer
Register

Status Last Data (Operand) Pointer

Register

Tag 10 0
Register
Opcode

Figure 8-1. x87 FPU Execution Environment

The x87 FPU instructions treat the eight x87 FPU data registers as a register stack (see
Figure 8-2). All addressing of the data registers is relative to the register on the top of
the stack. The register number of the current top-of-stack register is stored in the
TOP (stack TOP) field in the x87 FPU status word. Load operations decrement TOP by
one and load a value into the new top-of-stack register, and store operations store
the value from the current TOP register in memory and then increment TOP by one.
(For the x87 FPU, a load operation is equivalent to a push and a store operation is
equivalent to a pop.) Note that load and store operations are also available that do
not push and pop the stack.

Vol. 1 8-3
PROGRAMMING WITH THE X87 FPU

FPU Data Register Stack

7
6
Growth
Stack 5 ST(2)
4 ST(1) Top
3 ST(0) 011B
2
1
0

Figure 8-2. x87 FPU Data Register Stack

If a load operation is performed when TOP is at 0, register wraparound occurs and

the new value of TOP is set to 7. The floating-point stack-overflow exception indicates
when wraparound might cause an unsaved value to be overwritten (see Section
8.5.1.1, “Stack Overflow or Underflow Exception (#IS)”).
Many floating-point instructions have several addressing modes that permit the
programmer to implicitly operate on the top of the stack, or to explicitly operate on
specific registers relative to the TOP. Assemblers support these register addressing
modes, using the expression ST(0), or simply ST, to represent the current stack top
and ST(i) to specify the ith register from TOP in the stack (0 ≤ i ≤ 7). For example, if
TOP contains 011B (register 3 is the top of the stack), the following instruction would
add the contents of two registers in the stack (registers 3 and 5):
FADD ST, ST(2);
Figure 8-3 shows an example of how the stack structure of the x87 FPU registers and
instructions are typically used to perform a series of computations. Here, a two-
dimensional dot product is computed, as follows:
1. The first instruction (FLD value1) decrements the stack register pointer (TOP)
and loads the value 5.6 from memory into ST(0). The result of this operation is
shown in snap-shot (a).
2. The second instruction multiplies the value in ST(0) by the value 2.4 from
memory and stores the result in ST(0), shown in snap-shot (b).
3. The third instruction decrements TOP and loads the value 3.8 in ST(0).
4. The fourth instruction multiplies the value in ST(0) by the value 10.3 from
memory and stores the result in ST(0), shown in snap-shot (c).
5. The fifth instruction adds the value and the value in ST(1) and stores the result in
ST(0), shown in snap-shot (d).

8-4 Vol. 1
 PROGRAMMING WITH THE X87 FPU

Computation
Dot Product = (5.6 x 2.4) + (3.8 x 10.3)

Code:
FLD value1 ;(a) value1 = 5.6
FMUL value2 ;(b) value2 = 2.4
FLD value3 ; value3 = 3.8
FMUL value4 ;(c)value4 = 10.3
FADD ST(1) ;(d)

(a) (b) (c) (d)

R7 R7 R7 R7
R6 R6 R6 R6
R5 R5 R5 R5
R4 5.6 ST(0) R4 13.44 ST(0) R4 13.44 ST(1) R4 13.44 ST(
R3 R3 R3 39.14 ST(0) R3 52.58 ST
R2 R2 R2 R2
R1 R1 R1 R1
R0 R0 R0 R0

Figure 8-3. Example x87 FPU Dot Product Computation

The style of programming demonstrated in this example is supported by the floating-

point instruction set. In cases where the stack structure causes computation bottle-
necks, the FXCH (exchange x87 FPU register contents) instruction can be used to
streamline a computation.

8.1.2.1 Parameter Passing With the x87 FPU Register Stack

Like the general-purpose registers, the contents of the x87 FPU data registers are
unaffected by procedure calls, or in other words, the values are maintained across
procedure boundaries. A calling procedure can thus use the x87 FPU data registers
(as well as the procedure stack) for passing parameter between procedures. The
called procedure can reference parameters passed through the register stack using
the current stack register pointer (TOP) and the ST(0) and ST(i) nomenclature. It is
also common practice for a called procedure to leave a return value or result in
register ST(0) when returning execution to the calling procedure or program.
When mixing MMX and x87 FPU instructions in the procedures or code sequences,
the programmer is responsible for maintaining the integrity of parameters being
passed in the x87 FPU data registers. If an MMX instruction is executed before the
parameters in the x87 FPU data registers have been passed to another procedure,
the parameters may be lost (see Section 9.5, “Compatibility with x87 FPU Architec-
ture”).

Vol. 1 8-5
PROGRAMMING WITH THE X87 FPU

8.1.3 x87 FPU Status Register

The 16-bit x87 FPU status register (see Figure 8-4) indicates the current state of the
x87 FPU. The flags in the x87 FPU status register include the FPU busy flag, top-of-
stack (TOP) pointer, condition code flags, error summary status flag, stack fault flag,
and exception flags. The x87 FPU sets the flags in this register to show the results of
operations.

FPU Busy
Top of Stack Pointer

15 14 13 11 10 9 8 7 6 5 4 3 2 1 0

C C C C E S P U O Z D I
B TOP
3 2 1 0 S F E E E E E E

Condition
Code
Error Summary Status
Stack Fault
Exception Flags
Precision
Underflow
Overflow
Zero Divide
Denormalized Operand
Invalid Operation

Figure 8-4. x87 FPU Status Word

The contents of the x87 FPU status register (referred to as the x87 FPU status word)
can be stored in memory using the FSTSW/FNSTSW, FSTENV/FNSTENV,
FSAVE/FNSAVE, and FXSAVE instructions. It can also be stored in the AX register of
the integer unit, using the FSTSW/FNSTSW instructions.

8.1.3.1 Top of Stack (TOP) Pointer

A pointer to the x87 FPU data register that is currently at the top of the x87 FPU
register stack is contained in bits 11 through 13 of the x87 FPU status word. This
pointer, which is commonly referred to as TOP (for top-of-stack), is a binary value
from 0 to 7. See Section 8.1.2, “x87 FPU Data Registers,” for more information
about the TOP pointer.

8.1.3.2 Condition Code Flags

The four condition code flags (C0 through C3) indicate the results of floating-point
comparison and arithmetic operations. Table 8-1 summarizes the manner in which
the floating-point instructions set the condition code flags. These condition code bits

8-6 Vol. 1
 PROGRAMMING WITH THE X87 FPU

are used principally for conditional branching and for storage of information used in
exception handling (see Section 8.1.4, “Branching and Conditional Moves on Condi-
tion Codes”).
As shown in Table 8-1, the C1 condition code flag is used for a variety of functions.
When both the IE and SF flags in the x87 FPU status word are set, indicating a stack
overflow or underflow exception (#IS), the C1 flag distinguishes between overflow
(C1 = 1) and underflow (C1 = 0). When the PE flag in the status word is set, indi-
cating an inexact (rounded) result, the C1 flag is set to 1 if the last rounding by the
instruction was upward. The FXAM instruction sets C1 to the sign of the value being
examined.
The C2 condition code flag is used by the FPREM and FPREM1 instructions to indicate
an incomplete reduction (or partial remainder). When a successful reduction has
been completed, the C0, C3, and C1 condition code flags are set to the three least-
significant bits of the quotient (Q2, Q1, and Q0, respectively). See “FPREM1—Partial
Remainder” in Chapter 3, “Instruction Set Reference, A-M,” of the Intel® 64 and
IA-32 Architectures Software Developer’s Manual, Volume 2A, for more information
on how these instructions use the condition code flags.
The FPTAN, FSIN, FCOS, and FSINCOS instructions set the C2 flag to 1 to indicate
that the source operand is beyond the allowable range of ±263 and clear the C2 flag
if the source operand is within the allowable range.
Where the state of the condition code flags are listed as undefined in Table 8-1, do
not rely on any specific value in these flags.

8.1.3.3 x87 FPU Floating-Point Exception Flags

The six x87 FPU floating-point exception flags (bits 0 through 5) of the x87 FPU
status word indicate that one or more floating-point exceptions have been detected
since the bits were last cleared. The individual exception flags (IE, DE, ZE, OE, UE,
and PE) are described in detail in Section 8.4, “x87 FPU Floating-Point Exception
Handling.” Each of the exception flags can be masked by an exception mask bit in the
x87 FPU control word (see Section 8.1.5, “x87 FPU Control Word”). The exception
summary status flag (ES, bit 7) is set when any of the unmasked exception flags are
set. When the ES flag is set, the x87 FPU exception handler is invoked, using one of
the techniques described in Section 8.7, “Handling x87 FPU Exceptions in Software.”
(Note that if an exception flag is masked, the x87 FPU will still set the appropriate
flag if the associated exception occurs, but it will not set the ES flag.)
The exception flags are “sticky” bits (once set, they remain set until explicitly
cleared). They can be cleared by executing the FCLEX/FNCLEX (clear exceptions)
instructions, by reinitializing the x87 FPU with the FINIT/FNINIT or FSAVE/FNSAVE
instructions, or by overwriting the flags with an FRSTOR or FLDENV instruction.
The B-bit (bit 15) is included for 8087 compatibility only. It reflects the contents of
the ES flag.

Vol. 1 8-7
PROGRAMMING WITH THE X87 FPU

Table 8-1. Condition Code Interpretation

Instruction C0 C3 C2 C1
FCOM, FCOMP, FCOMPP, Result of Comparison Operands 0 or #IS
FICOM, FICOMP, FTST, are not
FUCOM, FUCOMP, FUCOMPP Comparable
FCOMI, FCOMIP, FUCOMI, Undefined. (These instructions set the #IS
FUCOMIP status flags in the EFLAGS register.)
FXAM Operand class Sign
FPREM, FPREM1 Q2 Q1 0 = reduction Q0 or #IS
complete
1 = reduction
incomplete
F2XM1, FADD, FADDP, Undefined Roundup or #IS
FBSTP, FCMOVcc, FIADD,
FDIV, FDIVP, FDIVR, FDIVRP,
FIDIV, FIDIVR, FIMUL, FIST,
FISTP, FISUB, FISUBR,FMUL,
FMULP, FPATAN, FRNDINT,
FSCALE, FST, FSTP, FSUB,
FSUBP, FSUBR,
FSUBRP,FSQRT, FYL2X,
FYL2XP1
FCOS, FSIN, FSINCOS, Undefined 0 = source Roundup or #IS
FPTAN operand (Undefined if
within range C2 = 1)
1 = source
operand out
of range
FABS, FBLD, FCHS, Undefined 0 or #IS
FDECSTP, FILD, FINCSTP,
FLD, Load Constants, FSTP
(ext. prec.), FXCH, FXTRACT
FLDENV, FRSTOR Each bit loaded from memory
FFREE, FLDCW,
FCLEX/FNCLEX, FNOP, Undefined
FSTCW/FNSTCW,
FSTENV/FNSTENV,
FSTSW/FNSTSW,
FINIT/FNINIT, 0 0 0 0
FSAVE/FNSAVE

8-8 Vol. 1
 PROGRAMMING WITH THE X87 FPU

8.1.3.4 Stack Fault Flag

The stack fault flag (bit 6 of the x87 FPU status word) indicates that stack overflow or
stack underflow has occurred with data in the x87 FPU data register stack. The x87
FPU explicitly sets the SF flag when it detects a stack overflow or underflow condi-
tion, but it does not explicitly clear the flag when it detects an invalid-arithmetic-
operand condition.
When this flag is set, the condition code flag C1 indicates the nature of the fault:
overflow (C1 = 1) and underflow (C1 = 0). The SF flag is a “sticky” flag, meaning
that after it is set, the processor does not clear it until it is explicitly instructed to do
so (for example, by an FINIT/FNINIT, FCLEX/FNCLEX, or FSAVE/FNSAVE instruction).
See Section 8.1.7, “x87 FPU Tag Word,” for more information on x87 FPU stack faults.

8.1.4 Branching and Conditional Moves on Condition Codes

The x87 FPU (beginning with the P6 family processors) supports two mechanisms for
branching and performing conditional moves according to comparisons of two
floating-point values. These mechanism are referred to here as the “old mechanism”
and the “new mechanism.”
The old mechanism is available in x87 FPU’s prior to the P6 family processors and in
P6 family processors. This mechanism uses the floating-point compare instructions
(FCOM, FCOMP, FCOMPP, FTST, FUCOMPP, FICOM, and FICOMP) to compare two
floating-point values and set the condition code flags (C0 through C3) according to
the results. The contents of the condition code flags are then copied into the status
flags of the EFLAGS register using a two step process (see Figure 8-5):
1. The FSTSW AX instruction moves the x87 FPU status word into the AX register.
2. The SAHF instruction copies the upper 8 bits of the AX register, which includes the
condition code flags, into the lower 8 bits of the EFLAGS register.
When the condition code flags have been loaded into the EFLAGS register, conditional
jumps or conditional moves can be performed based on the new settings of the
status flags in the EFLAGS register.

Vol. 1 8-9
PROGRAMMING WITH THE X87 FPU

15 x87 FPU Status Word 0

Condition Status
Flag C C C C
Code 3 2 1 0
C0 CF
C1 (none)
C2 PF FSTSW AX Instruction
C3 ZF 15 AX Register 0
C C C C
3 2 1 0

SAHF Instruction

31 EFLAGS Register 7 0
Z P C
F F 1 F

Figure 8-5. Moving the Condition Codes to the EFLAGS Register

The new mechanism is available beginning with the P6 family processors. Using this
mechanism, the new floating-point compare and set EFLAGS instructions (FCOMI,
FCOMIP, FUCOMI, and FUCOMIP) compare two floating-point values and set the ZF,
PF, and CF flags in the EFLAGS register directly. A single instruction thus replaces the
three instructions required by the old mechanism.
Note also that the FCMOVcc instructions (also new in the P6 family processors) allow
conditional moves of floating-point values (values in the x87 FPU data registers)
based on the setting of the status flags (ZF, PF, and CF) in the EFLAGS register. These
instructions eliminate the need for an IF statement to perform conditional moves of
floating-point values.

8.1.5 x87 FPU Control Word

The 16-bit x87 FPU control word (see Figure 8-6) controls the precision of the x87
FPU and rounding method used. It also contains the x87 FPU floating-point exception
mask bits. The control word is cached in the x87 FPU control register. The contents of
this register can be loaded with the FLDCW instruction and stored in memory with the
FSTCW/FNSTCW instructions.

8-10 Vol. 1
 PROGRAMMING WITH THE X87 FPU

Infinity Control
Rounding Control
Precision Control

15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0

P U O Z D I
X RC PC M M M M M M

Exception Masks
Precision
Underflow
Overflow
Zero Divide
Denormal Operand
Invalid Operation

Reserved

Figure 8-6. x87 FPU Control Word

When the x87 FPU is initialized with either an FINIT/FNINIT or FSAVE/FNSAVE

instruction, the x87 FPU control word is set to 037FH, which masks all floating-point
exceptions, sets rounding to nearest, and sets the x87 FPU precision to 64 bits.

8.1.5.1 x87 FPU Floating-Point Exception Mask Bits

The exception-flag mask bits (bits 0 through 5 of the x87 FPU control word) mask the
6 floating-point exception flags in the x87 FPU status word. When one of these mask
bits is set, its corresponding x87 FPU floating-point exception is blocked from being
generated.

8.1.5.2 Precision Control Field

The precision-control (PC) field (bits 8 and 9 of the x87 FPU control word) determines
the precision (64, 53, or 24 bits) of floating-point calculations made by the x87 FPU
(see Table 8-2). The default precision is double extended precision, which uses the
full 64-bit significand available with the double extended-precision floating-point
format of the x87 FPU data registers. This setting is best suited for most applications,
because it allows applications to take full advantage of the maximum precision avail-
able with the x87 FPU data registers.

Vol. 1 8-11
PROGRAMMING WITH THE X87 FPU

Table 8-2. Precision Control Field (PC)

Precision PC Field
Single Precision (24 bits) 00B
Reserved 01B
Double Precision (53 bits) 10B
Double Extended Precision (64 bits) 11B

The double precision and single precision settings reduce the size of the significand to
53 bits and 24 bits, respectively. These settings are provided to support IEEE Stan-
dard 754 and to provide compatibility with the specifications of certain existing
programming languages. Using these settings nullifies the advantages of the double
extended-precision floating-point format's 64-bit significand length. When reduced
precision is specified, the rounding of the significand value clears the unused bits on
the right to zeros.
The precision-control bits only affect the results of the following floating-point
instructions: FADD, FADDP, FIADD, FSUB, FSUBP, FISUB, FSUBR, FSUBRP, FISUBR,
FMUL, FMULP, FIMUL, FDIV, FDIVP, FIDIV, FDIVR, FDIVRP, FIDIVR, and FSQRT.

8.1.5.3 Rounding Control Field

The rounding-control (RC) field of the x87 FPU control register (bits 10 and 11)
controls how the results of x87 FPU floating-point instructions are rounded. See
Section 4.8.4, “Rounding,” for a discussion of rounding of floating-point values; See
Section 4.8.4.1, “Rounding Control (RC) Fields”, for the encodings of the RC field.

8.1.6 Infinity Control Flag

The infinity control flag (bit 12 of the x87 FPU control word) is provided for compati-
bility with the Intel 287 Math Coprocessor; it is not meaningful for later version x87
FPU coprocessors or IA-32 processors. See Section 4.8.3.3, “Signed Infinities,” for
information on how the x87 FPUs handle infinity values.

8.1.7 x87 FPU Tag Word

The 16-bit tag word (see Figure 8-7) indicates the contents of each the 8 registers in
the x87 FPU data-register stack (one 2-bit tag per register). The tag codes indicate
whether a register contains a valid number, zero, or a special floating-point number
(NaN, infinity, denormal, or unsupported format), or whether it is empty. The x87
FPU tag word is cached in the x87 FPU in the x87 FPU tag word register. When the x87
FPU is initialized with either an FINIT/FNINIT or FSAVE/FNSAVE instruction, the x87
FPU tag word is set to FFFFH, which marks all the x87 FPU data registers as empty.

8-12 Vol. 1
 PROGRAMMING WITH THE X87 FPU

15 0

TAG(7) TAG(6) TAG(5) TAG(4) TAG(3) TAG(2) TAG(1) TAG(0)

TAG Values
00 — Valid
01 — Zero
10 — Special: invalid (NaN, unsupported), infinity, or denormal
11 — Empty

Figure 8-7. x87 FPU Tag Word

Each tag in the x87 FPU tag word corresponds to a physical register (numbers 0
through 7). The current top-of-stack (TOP) pointer stored in the x87 FPU status word
can be used to associate tags with registers relative to ST(0).
The x87 FPU uses the tag values to detect stack overflow and underflow conditions
(see Section 8.5.1.1, “Stack Overflow or Underflow Exception (#IS)”).
Application programs and exception handlers can use this tag information to check
the contents of an x87 FPU data register without performing complex decoding of the
actual data in the register. To read the tag register, it must be stored in memory using
either the FSTENV/FNSTENV or FSAVE/FNSAVE instructions. The location of the tag
word in memory after being saved with one of these instructions is shown in Figures
8-9 through 8-12.
Software cannot directly load or modify the tags in the tag register. The FLDENV and
FRSTOR instructions load an image of the tag register into the x87 FPU; however, the
x87 FPU uses those tag values only to determine if the data registers are empty
(11B) or non-empty (00B, 01B, or 10B).
If the tag register image indicates that a data register is empty, the tag in the tag
register for that data register is marked empty (11B); if the tag register image indi-
cates that the data register is non-empty, the x87 FPU reads the actual value in the
data register and sets the tag for the register accordingly. This action prevents a
program from setting the values in the tag register to incorrectly represent the actual
contents of non-empty data registers.

8.1.8 x87 FPU Instruction and Data (Operand) Pointers

The x87 FPU stores pointers to the instruction and data (operand) for the last non-
control instruction executed. These pointers are stored in two 48-bit registers: the
x87 FPU instruction pointer and x87 FPU operand (data) pointer registers (see
Figure 8-1). (These pointers are saved to provide state information for exception
handlers.)

Vol. 1 8-13
PROGRAMMING WITH THE X87 FPU

Note that the value in the x87 FPU data pointer register is always a pointer to a
memory operand, If the last non-control instruction that was executed did not have
a memory operand, the value in the data pointer register is undefined (reserved).
The contents of the x87 FPU instruction and data pointer registers remain unchanged
when any of the control instructions (FINIT/FNINIT, FCLEX/FNCLEX, FLDCW,
FSTCW/FNSTCW, FSTSW/FNSTSW, FSTENV/FNSTENV, FLDENV, FSAVE/FNSAVE,
FRSTOR, and WAIT/FWAIT) are executed.
The pointers stored in the x87 FPU instruction and data pointer registers consist of an
offset (stored in bits 0 through 31) and a segment selector (stored in bits 32
through 47).
These registers can be accessed by the FSTENV/FNSTENV, FLDENV, FINIT/FNINIT,
FSAVE/FNSAVE, FRSTOR, FXSAVE, and FXRSTOR instructions. The FINIT/FNINIT and
FSAVE/FNSAVE instructions clear these registers.
For all the x87 FPUs and NPXs except the 8087, the x87 FPU instruction pointer points
to any prefixes that preceded the instruction. For the 8087, the x87 FPU instruction
pointer points only to the actual opcode.

8.1.9 Last Instruction Opcode

The x87 FPU stores the opcode of the last non-control instruction executed in an
11-bit x87 FPU opcode register. (This information provides state information for
exception handlers.) Only the first and second opcode bytes (after all prefixes) are
stored in the x87 FPU opcode register. Figure 8-8 shows the encoding of these two
bytes. Since the upper 5 bits of the first opcode byte are the same for all floating-
point opcodes (11011B), only the lower 3 bits of this byte are stored in the opcode
register.

8.1.9.1 Fopcode Compatibility Sub-mode

Beginning with the Pentium 4 and Intel Xeon processors, the IA-32 architecture
provides program control over the storing of the last instruction opcode (sometimes
referred to as the fopcode). Here, bit 2 of the IA32_MISC_ENABLE MSR enables (set)
or disables (clear) the fopcode compatibility mode.
If FOP code compatibility mode is enabled, the FOP is defined as it has always been
in previous IA32 implementations (always defined as the FOP of the last non-trans-
parent FP instruction executed before a FSAVE/FSTENV/FXSAVE). If FOP code
compatibility mode is disabled (default), FOP is only valid if the last non-transparent
FP instruction executed before a FSAVE/FSTENV/FXSAVE had an unmasked exception.

8-14 Vol. 1
 PROGRAMMING WITH THE X87 FPU

1st Instruction Byte 2nd Instruction Byte

7 2 0 7 0

10 8 7 0

x87 FPU Opcode Register

Figure 8-8. Contents of x87 FPU Opcode Registers

The fopcode compatibility mode should be enabled only when x87 FPU floating-point
exception handlers are designed to use the fopcode to analyze program performance
or restart a program after an exception has been handled.

8.1.10 Saving the x87 FPU’s State with FSTENV/FNSTENV and

FSAVE/FNSAVE
The FSTENV/FNSTENV and FSAVE/FNSAVE instructions store x87 FPU state informa-
tion in memory for use by exception handlers and other system and application soft-
ware. The FSTENV/FNSTENV instruction saves the contents of the status, control,
tag, x87 FPU instruction pointer, x87 FPU operand pointer, and opcode registers. The
FSAVE/FNSAVE instruction stores that information plus the contents of the x87 FPU
data registers. Note that the FSAVE/FNSAVE instruction also initializes the x87 FPU to
default values (just as the FINIT/FNINIT instruction does) after it has saved the orig-
inal state of the x87 FPU.
The manner in which this information is stored in memory depends on the operating
mode of the processor (protected mode or real-address mode) and on the operand-
size attribute in effect (32-bit or 16-bit). See Figures 8-9 through 8-12. In virtual-
8086 mode or SMM, the real-address mode formats shown in Figure 8-12 is used.
See Chapter 24, “System Management,” of the Intel® 64 and IA-32 Architectures
Software Developer’s Manual, Volume 3B, for information on using the x87 FPU while
in SMM.
The FLDENV and FRSTOR instructions allow x87 FPU state information to be loaded
from memory into the x87 FPU. Here, the FLDENV instruction loads only the status,
control, tag, x87 FPU instruction pointer, x87 FPU operand pointer, and opcode regis-
ters, and the FRSTOR instruction loads all the x87 FPU registers, including the x87
FPU stack registers.

Vol. 1 8-15
PROGRAMMING WITH THE X87 FPU

32-Bit Protected Mode Format

31 16 15 0
Control Word 0
Status Word 4
Tag Word 8
FPU Instruction Pointer Offset 12
00000 Opcode 10...00 FPU Instruction Pointer Selector 16
FPU Operand Pointer Offset 20
FPU Operand Pointer Selector 24

For instructions that also store x87 FPU data registers, the eight
80-bit registers (R0-R7) follow the above structure in sequence.

Figure 8-9. Protected Mode x87 FPU State Image in Memory, 32-Bit Format

32-Bit Real-Address Mode Format

31 16 15 0
Control Word 0
Status Word 4
Tag Word 8
FPU Instruction Pointer 15...00 12
0000 FPU Instruction Pointer 31...16 0 Opcode 10...00 16
FPU Operand Pointer 15...00 20
0000 FPU Operand Pointer 31...16 000000000000 24

For instructions that also store x87 FPU data registers, the eight
80-bit registers (R0-R7) follow the above structure in sequence.

Figure 8-10. Real Mode x87 FPU State Image in Memory, 32-Bit Format

8-16 Vol. 1
 PROGRAMMING WITH THE X87 FPU

16-Bit Protected Mode Format

15 0
Control Word 0
Status Word 2
Tag Word 4
FPU Instruction Pointer Offset 6
FPU Instruction Pointer Selector 8
FPU Operand Pointer Offset 10
FPU Operand Pointer Selector 12

Figure 8-11. Protected Mode x87 FPU State Image in Memory, 16-Bit Format

16-Bit Real-Address Mode and

Virtual-8086 Mode Format
15 0
Control Word 0
Status Word 2
Tag Word 4
FPU Instruction Pointer 15...00 6
IP 19..16 0 Opcode 10...00 8
FPU Operand Pointer 15...00 10
OP 19..16 0 0 0 0 0 0 0 0 0 0 0 0 12

Figure 8-12. Real Mode x87 FPU State Image in Memory, 16-Bit Format

8.1.11 Saving the x87 FPU’s State with FXSAVE

The FXSAVE and FXRSTOR instructions save and restore, respectively, the x87 FPU
state along with the state of the XMM registers and the MXCSR register. Using the
FXSAVE instruction to save the x87 FPU state has two benefits: (1) FXSAVE executes
faster than FSAVE, and (2) FXSAVE saves the entire x87 FPU, MMX, and XMM state in
one operation. See Section 10.5, “FXSAVE and FXRSTOR Instructions,” for additional
information about these instructions.

8.2 X87 FPU DATA TYPES

The x87 FPU recognizes and operates on the following seven data types (see Figures
8-13): single-precision floating point, double-precision floating point, double

Vol. 1 8-17
PROGRAMMING WITH THE X87 FPU

extended-precision floating point, signed word integer, signed doubleword integer,

signed quadword integer, and packed BCD decimal integers.
For detailed information about these data types, see Section 4.2.2, “Floating-Point
Data Types,” Section 4.2.1.2, “Signed Integers,” and Section 4.7, “BCD and Packed
BCD Integers.”
With the exception of the 80-bit double extended-precision floating-point format, all
of these data types exist in memory only. When they are loaded into x87 FPU data
registers, they are converted into double extended-precision floating-point format
and operated on in that format.
Denormal values are also supported in each of the floating-point types, as required
by IEEE Standard 754. When a denormal number in single-precision or double-preci-
sion floating-point format is used as a source operand and the denormal exception is
masked, the x87 FPU automatically normalizes the number when it is converted to
double extended-precision format.
When stored in memory, the least significant byte of an x87 FPU data-type value is
stored at the initial address specified for the value. Successive bytes from the value
are then stored in successively higher addresses in memory. The floating-point
instructions load and store memory operands using only the initial address of the
operand.

8-18 Vol. 1
 PROGRAMMING WITH THE X87 FPU

Single-Precision Floating-Point
Sign Exp. Fraction
3130 23 22 Implied Integer 0

Double-Precision Floating-Point
Sign Exponent Fraction
63 62 52 51 Implied Integer 0
Sign
Double Extended-Precision Floating-Point
Exponent Fraction
79 78 6463 62 Integer 0
Word Integer
Sign
15 14 0
Doubleword Integer
Sign
31 30 0
Quadword Integer
Sign
Sign 63 62 0
Packed BCD Integers
X D17 D16 D15 D14 D13 D12 D11 D10 D9 D8 D7 D6 D5 D4 D3 D2 D1 D0
79 78 72 71 4 Bits = 1 BCD Digit 0

Figure 8-13. x87 FPU Data Type Formats

As a general rule, values should be stored in memory in double-precision format. This

format provides sufficient range and precision to return correct results with a
minimum of programmer attention. The single-precision format is useful for debug-
ging algorithms, because rounding problems will manifest themselves more quickly
in this format. The double extended-precision format is normally reserved for holding
intermediate results in the x87 FPU registers and constants. Its extra length is
designed to shield final results from the effects of rounding and overflow/underflow
in intermediate calculations. However, when an application requires the maximum
range and precision of the x87 FPU (for data storage, computations, and results),
values can be stored in memory in double extended-precision format.

8.2.1 Indefinites
For each x87 FPU data type, one unique encoding is reserved for representing the
special value indefinite. The x87 FPU produces indefinite values as responses to
some masked floating-point invalid-operation exceptions. See Tables 4-1, 4-3, and

Vol. 1 8-19
PROGRAMMING WITH THE X87 FPU

4-4 for the encoding of the integer indefinite, QNaN floating-point indefinite, and
packed BCD integer indefinite, respectively.
The binary integer encoding 100..00B represents either of two things, depending on
the circumstances of its use:
• The largest negative number supported by the format (–215, –231, or –263)
• The integer indefinite value
If this encoding is used as a source operand (as in an integer load or integer arith-
metic instruction), the x87 FPU interprets it as the largest negative number repre-
sentable in the format being used. If the x87 FPU detects an invalid operation when
storing an integer value in memory with an FIST/FISTP instruction and the invalid-
operation exception is masked, the x87 FPU stores the integer indefinite encoding in
the destination operand as a masked response to the exception. In situations where
the origin of a value with this encoding may be ambiguous, the invalid-operation
exception flag can be examined to see if the value was produced as a response to an
exception.

8.2.2 Unsupported Double Extended-Precision

Floating-Point Encodings and Pseudo-Denormals
The double extended-precision floating-point format permits many encodings that do
not fall into any of the categories shown in Table 4-3. Table 8-3 shows these unsup-
ported encodings. Some of these encodings were supported by the Intel 287 math
coprocessor; however, most of them are not supported by the Intel 387 math copro-
cessor and later IA-32 processors. These encodings are no longer supported due to
changes made in the final version of IEEE Standard 754 that eliminated these encod-
ings.
Specifically, the categories of encodings formerly known as pseudo-NaNs, pseudo-
infinities, and un-normal numbers are not supported and should not be used as
operand values. The Intel 387 math coprocessor and later IA-32 processors generate
an invalid-operation exception when these encodings are encountered as operands.
Beginning with the Intel 387 math coprocessor, the encodings formerly known as
pseudo-denormal numbers are not generated by IA-32 processors. When encoun-
tered as operands, however, they are handled correctly; that is, they are treated as
denormals and a denormal exception is generated. Pseudo-denormal numbers
should not be used as operand values. They are supported by current IA-32 proces-
sors (as described here) to support legacy code.

8-20 Vol. 1
 PROGRAMMING WITH THE X87 FPU

Table 8-3. Unsupported Double Extended-Precision Floating-Point Encodings and

Pseudo-Denormals
Significand
Class Sign Biased Exponent Integer Fraction
Positive 0 11..11 0 11..11
Pseudo-NaNs Quiet . . .
0 11..11 10..00
0 11..11 0 01..11
Signaling . . .
0 11..11 00..01
Positive Floating Pseudo-infinity 0 11..11 0 00..00
Point 0 11..10 0 11..11
Unnormals . . .
0 00..01 00..00
Pseudo-denormals 0 00..00 1 11..11
. . .
0 00..00 00..00
Negative Pseudo-denormals 1 00..00 1 11..11
Floating Point . . .
1 00..00 00..00
1 11..10 0 11..01
Unnormals . . .
1 00..01 00..00
Pseudo-infinity 1 11..11 0 00..00
Negative 1 11..11 0 01..11
Pseudo-NaNs Signaling . . .
1 11..11 00..01
1 11..11 0 11..11
Quiet . . .
1 11..11 10..00
← 15 bits → ← 63 bits →

8.3 X86 FPU INSTRUCTION SET

The floating-point instructions that the x87 FPU supports can be grouped into six
functional categories:
• Data transfer instructions
• Basic arithmetic instructions
• Comparison instructions
• Transcendental instructions

Vol. 1 8-21
PROGRAMMING WITH THE X87 FPU

• Load constant instructions

• x87 FPU control instructions
See Section 5.2, “x87 FPU Instructions,” for a list of the floating-point instructions by
category.
The following section briefly describes the instructions in each category. Detailed
descriptions of the floating-point instructions are given in the Intel® 64 and IA-32
Architectures Software Developer’s Manual, Volumes 3A & 3B.

8.3.1 Escape (ESC) Instructions

All of the instructions in the x87 FPU instruction set fall into a class of instructions
known as escape (ESC) instructions. All of these instructions have a common opcode
format, where the first byte of the opcode is one of the numbers from D8H through
DFH.

8.3.2 x87 FPU Instruction Operands

Most floating-point instructions require one or two operands, located on the x87 FPU
data-register stack or in memory. (None of the floating-point instructions accept
immediate operands.)
When an operand is located in a data register, it is referenced relative to the ST(0)
register (the register at the top of the register stack), rather than by a physical
register number. Often the ST(0) register is an implied operand.
Operands in memory can be referenced using the same operand addressing methods
described in Section 3.7, “Operand Addressing.”

8.3.3 Data Transfer Instructions

The data transfer instructions (see Table 8-4) perform the following operations:
• Load a floating-point, integer, or packed BCD operand from memory into the
ST(0) register.
• Store the value in an ST(0) register to memory in floating-point, integer, or
packed BCD format.
• Move values between registers in the x87 FPU register stack.
The FLD (load floating point) instruction pushes a floating-point operand from
memory onto the top of the x87 FPU data-register stack. If the operand is in single-
precision or double-precision floating-point format, it is automatically converted to
double extended-precision floating-point format. This instruction can also be used to
push the value in a selected x87 FPU data register onto the top of the register stack.

8-22 Vol. 1
 PROGRAMMING WITH THE X87 FPU

The FILD (load integer) instruction converts an integer operand in memory into
double extended-precision floating-point format and pushes the value onto the top of
the register stack. The FBLD (load packed decimal) instruction performs the same
load operation for a packed BCD operand in memory.

Table 8-4. Data Transfer Instructions

Floating Point Integer Packed Decimal
FLD Load Floating FILD Load Integer FBLD Load Packed
Point Decimal
FST Store Floating FIST Store Integer
Point
FSTP Store Floating FISTP Store Integer FBSTP Store Packed
Point and Pop and Pop Decimal and Pop
FXCH Exchange Register
Contents
FCMOVcc Conditional Move

The FST (store floating point) and FIST (store integer) instructions store the value in
register ST(0) in memory in the destination format (floating point or integer, respec-
tively). Again, the format conversion is carried out automatically.
The FSTP (store floating point and pop), FISTP (store integer and pop), and FBSTP
(store packed decimal and pop) instructions store the value in the ST(0) registers
into memory in the destination format (floating point, integer, or packed BCD), then
performs a pop operation on the register stack. A pop operation causes the ST(0)
register to be marked empty and the stack pointer (TOP) in the x87 FPU control work
to be incremented by 1. The FSTP instruction can also be used to copy the value in
the ST(0) register to another x87 FPU register [ST(i)].
The FXCH (exchange register contents) instruction exchanges the value in a selected
register in the stack [ST(i)] with the value in ST(0).
The FCMOVcc (conditional move) instructions move the value in a selected register in
the stack [ST(i)] to register ST(0) if a condition specified with a condition code (cc) is
satisfied (see Table 8-5). The condition being tested for is represented by the status
flags in the EFLAGS register. The condition code mnemonics are appended to the
letters “FCMOV” to form the mnemonic for a FCMOVcc instruction.

Table 8-5. Floating-Point Conditional Move Instructions

Instruction Mnemonic Status Flag States Condition Description
FCMOVB CF=1 Below
FCMOVNB CF=0 Not below
FCMOVE ZF=1 Equal
FCMOVNE ZF=0 Not equal

Vol. 1 8-23
PROGRAMMING WITH THE X87 FPU

Table 8-5. Floating-Point Conditional Move Instructions (Contd.)

Instruction Mnemonic Status Flag States Condition Description
FCMOVBE CF=1 or ZF=1 Below or equal
FCMOVNBE CF=0 or ZF=0 Not below nor equal
FCMOVU PF=1 Unordered
FCMOVNU PF=0 Not unordered

Like the CMOVcc instructions, the FCMOVcc instructions are useful for optimizing
small IF constructions. They also help eliminate branching overhead for IF operations
and the possibility of branch mispredictions by the processor.
Software can check if the FCMOVcc instructions are supported by checking the
processor’s feature information with the CPUID instruction.

8.3.4 Load Constant Instructions

The following instructions push commonly used constants onto the top [ST(0)] of the
x87 FPU register stack:

FLDZ Load +0.0

FLD1 Load +1.0
FLDPI Load π
FLDL2T Load log2 10
FLDL2E Load log2e
FLDLG2 Load log102
FLDLN2 Load loge2

The constant values have full double extended-precision floating-point precision

(64 bits) and are accurate to approximately 19 decimal digits. They are stored
internally in a format more precise than double extended-precision floating point.
When loading the constant, the x87 FPU rounds the more precise internal constant
according to the RC (rounding control) field of the x87 FPU control word. The
inexact-result exception (#P) is not generated as a result of this rounding, nor is
the C1 flag set in the x87 FPU status word if the value is rounded up. See Section
8.3.8, “Pi,” for information on the π constant.

8.3.5 Basic Arithmetic Instructions

The following floating-point instructions perform basic arithmetic operations on
floating-point numbers. Where applicable, these instructions match IEEE Standard
754:
FADD/FADDP Add floating point

8-24 Vol. 1
 PROGRAMMING WITH THE X87 FPU

FIADD Add integer to floating point

FSUB/FSUBP Subtract floating point
FISUB Subtract integer from floating point
FSUBR/FSUBRP Reverse subtract floating point
FISUBR Reverse subtract floating point from integer
FMUL/FMULP Multiply floating point
FIMUL Multiply integer by floating point
FDIV/FDIVP Divide floating point
FIDIV Divide floating point by integer
FDIVR/FDIVRP Reverse divide
FIDIVR Reverse divide integer by floating point
FABS Absolute value
FCHS Change sign
FSQRT Square root
FPREM Partial remainder
FPREM1 IEEE partial remainder
FRNDINT Round to integral value
FXTRACT Extract exponent and significand

The add, subtract, multiply and divide instructions operate on the following types of
operands:
• Two x87 FPU data registers
• An x87 FPU data register and a floating-point or integer value in memory
See Section 8.1.2, “x87 FPU Data Registers,” for a description of how operands are
referenced on the data register stack.
Operands in memory can be in single-precision floating-point, double-precision
floating-point, word-integer, or doubleword-integer format. They are converted to
double extended-precision floating-point format automatically.
Reverse versions of the subtract (FSUBR) and divide (FDIVR) instructions enable effi-
cient coding. For example, the following options are available with the FSUB and
FSUBR instructions for operating on values in a specified x87 FPU data register ST(i)
and the ST(0) register:
FSUB:
ST(0) ← ST(0) − ST(i)
ST(i) ← ST(i) − ST(0)
FSUBR:
ST(0) ← ST(i) − ST(0)
ST(i) ← ST(0) − ST(i)
These instructions eliminate the need to exchange values between the ST(0) register
and another x87 FPU register to perform a subtraction or division.

Vol. 1 8-25
PROGRAMMING WITH THE X87 FPU

The pop versions of the add, subtract, multiply, and divide instructions offer the
option of popping the x87 FPU register stack following the arithmetic operation.
These instructions operate on values in the ST(i) and ST(0) registers, store the result
in the ST(i) register, and pop the ST(0) register.
The FPREM instruction computes the remainder from the division of two operands in
the manner used by the Intel 8087 and Intel 287 math coprocessors; the FPREM1
instruction computes the remainder in the manner specified in IEEE Standard 754.
The FSQRT instruction computes the square root of the source operand.
The FRNDINT instruction returns a floating-point value that is the integral value
closest to the source value in the direction of the rounding mode specified in the RC
field of the x87 FPU control word.
The FABS, FCHS, and FXTRACT instructions perform convenient arithmetic opera-
tions. The FABS instruction produces the absolute value of the source operand. The
FCHS instruction changes the sign of the source operand. The FXTRACT instruction
separates the source operand into its exponent and fraction and stores each value in
a register in floating-point format.

8.3.6 Comparison and Classification Instructions

The following instructions compare or classify floating-point values:
FCOM/FCOMP/FCOMPPCompare floating point and set x87 FPU
condition code flags.
FUCOM/FUCOMP/FUCOMPPUnordered compare floating point and set
x87 FPU condition code flags.
FICOM/FICOMPCompare integer and set x87 FPU
condition code flags.
FCOMI/FCOMIPCompare floating point and set EFLAGS
status flags.
FUCOMI/FUCOMIPUnordered compare floating point and
set EFLAGS status flags.
FTST Test (compare floating point with 0.0).
FXAMExamine.
Comparison of floating-point values differ from comparison of integers because
floating-point values have four (rather than three) mutually exclusive relationships:
less than, equal, greater than, and unordered.
The unordered relationship is true when at least one of the two values being
compared is a NaN or in an unsupported format. This additional relationship is
required because, by definition, NaNs are not numbers, so they cannot have less
than, equal, or greater than relationships with other floating-point values.

8-26 Vol. 1
 PROGRAMMING WITH THE X87 FPU

The FCOM, FCOMP, and FCOMPP instructions compare the value in register ST(0) with
a floating-point source operand and set the condition code flags (C0, C2, and C3) in
the x87 FPU status word according to the results (see Table 8-6).
If an unordered condition is detected (one or both of the values are NaNs or in an
undefined format), a floating-point invalid-operation exception is generated.
The pop versions of the instruction pop the x87 FPU register stack once or twice after
the comparison operation is complete.
The FUCOM, FUCOMP, and FUCOMPP instructions operate the same as the FCOM,
FCOMP, and FCOMPP instructions. The only difference is that with the FUCOM,
FUCOMP, and FUCOMPP instructions, if an unordered condition is detected because
one or both of the operands are QNaNs, the floating-point invalid-operation excep-
tion is not generated.

Table 8-6. Setting of x87 FPU Condition Code Flags for Floating-Point Number
Comparisons
Condition C3 C2 C0
ST(0) > Source Operand 0 0 0
ST(0) < Source Operand 0 0 1
ST(0) = Source Operand 1 0 0
Unordered 1 1 1

The FICOM and FICOMP instructions also operate the same as the FCOM and FCOMP
instructions, except that the source operand is an integer value in memory. The
integer value is automatically converted into an double extended-precision floating-
point value prior to making the comparison. The FICOMP instruction pops the x87
FPU register stack following the comparison operation.
The FTST instruction performs the same operation as the FCOM instruction, except
that the value in register ST(0) is always compared with the value 0.0.
The FCOMI and FCOMIP instructions were introduced into the IA-32 architecture in
the P6 family processors. They perform the same comparison as the FCOM and
FCOMP instructions, except that they set the status flags (ZF, PF, and CF) in the
EFLAGS register to indicate the results of the comparison (see Table 8-7) instead of
the x87 FPU condition code flags. The FCOMI and FCOMIP instructions allow condition
branch instructions (Jcc) to be executed directly from the results of their comparison.

Vol. 1 8-27
PROGRAMMING WITH THE X87 FPU

Table 8-7. Setting of EFLAGS Status Flags for Floating-Point Number Comparisons
Comparison Results ZF PF CF
ST0 > ST(i) 0 0 0
ST0 < ST(i) 0 0 1
ST0 = ST(i) 1 0 0
Unordered 1 1 1

Software can check if the FCOMI and FCOMIP instructions are supported by checking
the processor’s feature information with the CPUID instruction.
The FUCOMI and FUCOMIP instructions operate the same as the FCOMI and FCOMIP
instructions, except that they do not generate a floating-point invalid-operation
exception if the unordered condition is the result of one or both of the operands being
a QNaN. The FCOMIP and FUCOMIP instructions pop the x87 FPU register stack
following the comparison operation.
The FXAM instruction determines the classification of the floating-point value in the
ST(0) register (that is, whether the value is zero, a denormal number, a normal finite
number, ∞, a NaN, or an unsupported format) or that the register is empty. It sets the
x87 FPU condition code flags to indicate the classification (see “FXAM—Examine” in
Chapter 3, “Instruction Set Reference, A-M,” of the Intel® 64 and IA-32 Architec-
tures Software Developer’s Manual, Volume 2A). It also sets the C1 flag to indicate
the sign of the value.

8.3.6.1 Branching on the x87 FPU Condition Codes

The processor does not offer any control-flow instructions that branch on the setting
of the condition code flags (C0, C2, and C3) in the x87 FPU status word. To branch on
the state of these flags, the x87 FPU status word must first be moved to the AX
register in the integer unit. The FSTSW AX (store status word) instruction can be
used for this purpose. When these flags are in the AX register, the TEST instruction
can be used to control conditional branching as follows:
1. Check for an unordered result. Use the TEST instruction to compare the contents
of the AX register with the constant 0400H (see Table 8-8). This operation will
clear the ZF flag in the EFLAGS register if the condition code flags indicate an
unordered result; otherwise, the ZF flag will be set. The JNZ instruction can then
be used to transfer control (if necessary) to a procedure for handling unordered
operands.

8-28 Vol. 1
 PROGRAMMING WITH THE X87 FPU

Table 8-8. TEST Instruction Constants for Conditional Branching

Order Constant Branch
ST(0) > Source Operand 4500H JZ
ST(0) < Source Operand 0100H JNZ
ST(0) = Source Operand 4000H JNZ
Unordered 0400H JNZ

2. Check ordered comparison result. Use the constants given in Table 8-8 in the
TEST instruction to test for a less than, equal to, or greater than result, then use
the corresponding conditional branch instruction to transfer program control to
the appropriate procedure or section of code.
If a program or procedure has been thoroughly tested and it incorporates periodic
checks for QNaN results, then it is not necessary to check for the unordered result
every time a comparison is made.
See Section 8.1.4, “Branching and Conditional Moves on Condition Codes,” for
another technique for branching on x87 FPU condition codes.
Some non-comparison x87 FPU instructions update the condition code flags in the
x87 FPU status word. To ensure that the status word is not altered inadvertently,
store it immediately following a comparison operation.

8.3.7 Trigonometric Instructions

The following instructions perform four common trigonometric functions:

FSIN Sine
FCOS Cosine
FSINCOS Sine and cosine
FPTAN Tangent
FPATAN Arctangent
These instructions operate on the top one or two registers of the x87 FPU register
stack and they return their results to the stack. The source operands for the FSIN,
FCOS, FSINCOS, and FPTAN instructions must be given in radians; the source
operand for the FPATAN instruction is given in rectangular coordinate units.
The FSINCOS instruction returns both the sine and the cosine of a source operand
value. It operates faster than executing the FSIN and FCOS instructions in succes-
sion.
The FPATAN instruction computes the arctangent of ST(1) divided by ST(0),
returning a result in radians. It is useful for converting rectangular coordinates to
polar coordinates.

Vol. 1 8-29
PROGRAMMING WITH THE X87 FPU

8.3.8 Pi
When the argument (source operand) of a trigonometric function is within the range
of the function, the argument is automatically reduced by the appropriate multiple of
2π through the same reduction mechanism used by the FPREM and FPREM1 instruc-
tions. The internal value of π that the x87 FPU uses for argument reduction and other
computations is as follows:
π = 0.f ∗ 22
where:
f = C90FDAA2 2168C234 C
(The spaces in the fraction above indicate 32-bit boundaries.)
This internal π value has a 66-bit mantissa, which is 2 bits more than is allowed in the
significand of an double extended-precision floating-point value. (Since 66 bits is not
an even number of hexadecimal digits, two additional zeros have been added to the
value so that it can be represented in hexadecimal format. The least-significant
hexadecimal digit (C) is thus 1100B, where the two least-significant bits represent
bits 67 and 68 of the mantissa.)
This value of π has been chosen to guarantee no loss of significance in a source
operand, provided the operand is within the specified range for the instruction.
If the results of computations that explicitly use π are to be used in the FSIN, FCOS,
FSINCOS, or FPTAN instructions, the full 66-bit fraction of π should be used. This
insures that the results are consistent with the argument-reduction algorithms that
these instructions use. Using a rounded version of π can cause inaccuracies in result
values, which if propagated through several calculations, might result in meaningless
results.
A common method of representing the full 66-bit fraction of π is to separate the value
into two numbers (highπ and lowπ) that when added together give the value for π
shown earlier in this section with the full 66-bit fraction:
π = highπ + lowπ
For example, the following two values (given in scientific notation with the fraction in
hexadecimal and the exponent in decimal) represent the 33 most-significant and the
33 least-significant bits of the fraction:
highπ (unnormalized) = 0.C90FDAA20 * 2+2
lowπ (unnormalized) = 0.42D184698 * 2− 31
These values encoded in the IEEE double-precision floating-point format are as
follows:
highπ = 400921FB 54400000
lowπ = 3DE0B461 1A600000
(Note that in the IEEE double-precision floating-point format, the exponents are
biased (by 1023) and the fractions are normalized.)
Similar versions of π can also be written in double extended-precision floating-point
format.

8-30 Vol. 1
 PROGRAMMING WITH THE X87 FPU

When using this two-part π value in an algorithm, parallel computations should be

performed on each part, with the results kept separate. When all the computations
are complete, the two results can be added together to form the final result.
The complications of maintaining a consistent value of π for argument reduction can
be avoided, either by applying the trigonometric functions only to arguments within
the range of the automatic reduction mechanism, or by performing all argument
reductions (down to a magnitude less than π/4) explicitly in software.

8.3.9 Logarithmic, Exponential, and Scale

The following instructions provide two different logarithmic functions, an exponential
function and a scale function:

FYL2X Logarithm
FYL2XP1 Logarithm epsilon
F2XM1 Exponential
FSCALE Scale

The FYL2X and FYL2XP1 instructions perform two different base 2 logarithmic opera-
tions. The FYL2X instruction computes (y ∗ log2x). This operation permits the calcu-
lation of the log of any base using the following equation:
logb x = (1/log2 b) ∗ log2 x
The FYL2XP1 instruction computes (y ∗ log2(x + 1)). This operation provides
optimum accuracy for values of x that are close to 0.
The F2XM1 instruction computes (2x − 1). This instruction only operates on source
values in the range −1.0 to +1.0.
The FSCALE instruction multiplies the source operand by a power of 2.

8.3.10 Transcendental Instruction Accuracy

New transcendental instruction algorithms were incorporated into the IA-32 architec-
ture beginning with the Pentium processors. These new algorithms (used in tran-
scendental instructions FSIN, FCOS, FSINCOS, FPTAN, FPATAN, F2XM1, FYL2X, and
FYL2XP1) allow a higher level of accuracy than was possible in earlier IA-32 proces-
sors and x87 math coprocessors. The accuracy of these instructions is measured in
terms of units in the last place (ulp). For a given argument x, let f(x) and F(x) be
the correct and computed (approximate) function values, respectively. The error in
ulps is defined to be:

( x ) – F ( x )-
error = f--------------------------
k – 63
2

Vol. 1 8-31
PROGRAMMING WITH THE X87 FPU

where k is an integer such that:

–k
1≤2 f ( x ) < 2.

With the Pentium processor and later IA-32 processors, the worst case error on
transcendental functions is less than 1 ulp when rounding to the nearest (even) and
less than 1.5 ulps when rounding in other modes. The functions are guaranteed to be
monotonic, with respect to the input operands, throughout the domain supported by
the instruction.
The instructions FYL2X and FYL2XP1 are two operand instructions and are guaran-
teed to be within 1 ulp only when y equals 1. When y is not equal to 1, the maximum
ulp error is always within 1.35 ulps in round to nearest mode. (For the two operand
functions, monotonicity was proved by holding one of the operands constant.)

8.3.11 x87 FPU Control Instructions

The following instructions control the state and modes of operation of the x87 FPU.
They also allow the status of the x87 FPU to be examined:
FINIT/FNINIT Initialize x87 FPU
FLDCW Load x87 FPU control word
FSTCW/FNSTCWStore x87 FPU control word
FSTSW/FNSTSWStore x87 FPU status word
FCLEX/FNCLEXClear x87 FPU exception flags
FLDENV Load x87 FPU environment
FSTENV/FNSTENVStore x87 FPU environment
FRSTOR Restore x87 FPU state
FSAVE/FNSAVESave x87 FPU state
FINCSTP Increment x87 FPU register stack pointer
FDECSTP Decrement x87 FPU register stack pointer
FFREE Free x87 FPU register
FNOP No operation
WAIT/FWAIT Check for and handle pending unmasked
x87 FPU exceptions
The FINIT/FNINIT instructions initialize the x87 FPU and its internal registers to
default values.
The FLDCW instructions loads the x87 FPU control word register with a value from
memory. The FSTCW/FNSTCW and FSTSW/FNSTSW instructions store the x87 FPU

8-32 Vol. 1
 PROGRAMMING WITH THE X87 FPU

control and status words, respectively, in memory (or for an FSTSW/FNSTSW

instruction in a general-purpose register).
The FSTENV/FNSTENV and FSAVE/FNSAVE instructions save the x87 FPU environ-
ment and state, respectively, in memory. The x87 FPU environment includes all the
x87 FPU’s control and status registers; the x87 FPU state includes the x87 FPU envi-
ronment and the data registers in the x87 FPU register stack. (The FSAVE/FNSAVE
instruction also initializes the x87 FPU to default values, like the FINIT/FNINIT
instruction, after it saves the original state of the x87 FPU.)
The FLDENV and FRSTOR instructions load the x87 FPU environment and state,
respectively, from memory into the x87 FPU. These instructions are commonly used
when switching tasks or contexts.
The WAIT/FWAIT instructions are synchronization instructions. (They are actually
mnemonics for the same opcode.) These instructions check the x87 FPU status word
for pending unmasked x87 FPU exceptions. If any pending unmasked x87 FPU excep-
tions are found, they are handled before the processor resumes execution of the
instructions (integer, floating-point, or system instruction) in the instruction stream.
The WAIT/FWAIT instructions are provided to allow synchronization of instruction
execution between the x87 FPU and the processor’s integer unit. See Section 8.6,
“x87 FPU Exception Synchronization,” for more information on the use of the
WAIT/FWAIT instructions.

8.3.12 Waiting vs. Non-waiting Instructions

All of the x87 FPU instructions except a few special control instructions perform a wait
operation (similar to the WAIT/FWAIT instructions), to check for and handle pending
unmasked x87 FPU floating-point exceptions, before they perform their primary
operation (such as adding two floating-point numbers). These instructions are called
waiting instructions. Some of the x87 FPU control instructions, such as
FSTSW/FNSTSW, have both a waiting and a non-waiting version. The waiting version
(with the “F” prefix) executes a wait operation before it performs its primary opera-
tion; whereas, the non-waiting version (with the “FN” prefix) ignores pending
unmasked exceptions.
Non-waiting instructions allow software to save the current x87 FPU state without
first handling pending exceptions or to reset or reinitialize the x87 FPU without
regard for pending exceptions.

NOTES
When operating a Pentium or Intel486 processor in MS-DOS compat-
ibility mode, it is possible (under unusual circumstances) for a non-
waiting instruction to be interrupted prior to being executed to
handle a pending x87 FPU exception. The circumstances where this
can happen and the resulting action of the processor are described in

Vol. 1 8-33
PROGRAMMING WITH THE X87 FPU

Section D.2.1.3, “No-Wait x87 FPU Instructions Can Get x87 FPU
Interrupt in Window.”
When operating a P6 family, Pentium 4, or Intel Xeon processor in
MS-DOS compatibility mode, non-waiting instructions can not be
interrupted in this way (see Section D.2.2, “MS-DOS* Compatibility
Sub-mode in the P6 Family and Pentium® 4 Processors”).

8.3.13 Unsupported x87 FPU Instructions

The Intel 8087 instructions FENI and FDISI and the Intel 287 math coprocessor
instruction FSETPM perform no function in the Intel 387 math coprocessor and later
IA-32 processors. If these opcodes are detected in the instruction stream, the x87
FPU performs no specific operation and no internal x87 FPU states are affected.

8.4 X87 FPU FLOATING-POINT EXCEPTION HANDLING

The x87 FPU detects the six classes of exception conditions described in Section 4.9,
“Overview of Floating-Point Exceptions”:
• Invalid operation (#I), with two subclasses:
— Stack overflow or underflow (#IS)
— Invalid arithmetic operation (#IA)
• Denormalized operand (#D)
• Divide-by-zero (#Z)
• Numeric overflow (#O)
• Numeric underflow (#U)
• Inexact result (precision) (#P)
Each of the six exception classes has a corresponding flag bit in the x87 FPU status
word and a mask bit in the x87 FPU control word (see Section 8.1.3, “x87 FPU Status
Register,” and Section 8.1.5, “x87 FPU Control Word,” respectively). In addition, the
exception summary (ES) flag in the status word indicates when one or more
unmasked exceptions has been detected. The stack fault (SF) flag (also in the status
word) distinguishes between the two types of invalid-operation exceptions.
The mask bits can be set with FLDCW, FRSTOR, or FXRSTOR; they can be read with
either FSTCW/FNSTCW, FSAVE/FNSAVE, or FXSAVE. The flag bits can be read with
the FSTSW/FNSTSW, FSAVE/FNSAVE, or FXSAVE instruction.

NOTE
Section 4.9.1, “Floating-Point Exception Conditions,” provides a
general overview of how the IA-32 processor detects and handles the

8-34 Vol. 1
 PROGRAMMING WITH THE X87 FPU

various classes of floating-point exceptions. This information pertains

to x87 FPU as well as SSE/SSE2/SSE3 extensions.
The following sections give specific information about how the x87 FPU handles
floating-point exceptions that are unique to the x87 FPU.

8.4.1 Arithmetic vs. Non-arithmetic Instructions

When dealing with floating-point exceptions, it is useful to distinguish between
arithmetic instructions and non-arithmetic instructions. Non-arithmetic
instructions have no operands or do not make substantial changes to their operands.
Arithmetic instructions do make significant changes to their operands; in particular,
they make changes that could result in floating-point exceptions being signaled.
Table 8-9 lists the non-arithmetic and arithmetic instructions. It should be noted that
some non-arithmetic instructions can signal a floating-point stack (fault) exception,
but this exception is not the result of an operation on an operand.

Table 8-9. Arithmetic and Non-arithmetic Instructions

Non-arithmetic Instructions Arithmetic Instructions
FABS F2XM1
FCHS FADD/FADDP
FCLEX FBLD
FDECSTP FBSTP
FFREE FCOM/FCOMP/FCOMPP
FINCSTP FCOS
FINIT/FNINIT FDIV/FDIVP/FDIVR/FDIVRP
FLD (register-to-register) FIADD
FLD (extended format from memory) FICOM/FICOMP
FLD constant FIDIV/FIDIVR
FLDCW FILD
FLDENV FIMUL
FNOP FIST/FISTP1
FRSTOR FISUB/FISUBR
FSAVE/FNSAVE FLD (single and double)
FST/FSTP (register-to-register) FMUL/FMULP
FSTP (extended format to memory) FPATAN
FSTCW/FNSTCW FPREM/FPREM1
FSTENV/FNSTENV FPTAN

Vol. 1 8-35
PROGRAMMING WITH THE X87 FPU

Table 8-9. Arithmetic and Non-arithmetic Instructions (Contd.)

Non-arithmetic Instructions Arithmetic Instructions
FSTSW/FNSTSW FRNDINT
WAIT/FWAIT FSCALE
FXAM FSIN
FXCH FSINCOS
FSQRT
FST/FSTP (single and double)
FSUB/FSUBP/FSUBR/FSUBRP
FTST
FUCOM/FUCOMP/FUCOMPP
FXTRACT
FYL2X/FYL2XP1
NOTE:
1. The FISTTP instruction in SSE3 is an arithmetic x87 FPU instruction.

8.5 X87 FPU FLOATING-POINT EXCEPTION CONDITIONS

The following sections describe the various conditions that cause a floating-point
exception to be generated by the x87 FPU and the masked response of the x87 FPU
when these conditions are detected. Intel® 64 and IA-32 Architectures Software
Developer’s Manual, Volumes 2A & 2B, list the floating-point exceptions that can be
signaled for each floating-point instruction.
See Section 4.9.2, “Floating-Point Exception Priority,” for a description of the rules for
exception precedence when more than one floating-point exception condition is
detected for an instruction.

8.5.1 Invalid Operation Exception

The floating-point invalid-operation exception occurs in response to two sub-classes
of operations:
• Stack overflow or underflow (#IS)
• Invalid arithmetic operand (#IA)
The flag for this exception (IE) is bit 0 of the x87 FPU status word, and the mask bit
(IM) is bit 0 of the x87 FPU control word. The stack fault flag (SF) of the x87 FPU
status word indicates the type of operation that caused the exception. When the SF
flag is set to 1, a stack operation has resulted in stack overflow or underflow; when

8-36 Vol. 1
 PROGRAMMING WITH THE X87 FPU

the flag is cleared to 0, an arithmetic instruction has encountered an invalid operand.

Note that the x87 FPU explicitly sets the SF flag when it detects a stack overflow or
underflow condition, but it does not explicitly clear the flag when it detects an invalid-
arithmetic-operand condition. As a result, the state of the SF flag can be 1 following
an invalid-arithmetic-operation exception, if it was not cleared from the last time a
stack overflow or underflow condition occurred. See Section 8.1.3.4, “Stack Fault
Flag,” for more information about the SF flag.

8.5.1.1 Stack Overflow or Underflow Exception (#IS)

The x87 FPU tag word keeps track of the contents of the registers in the x87 FPU
register stack (see Section 8.1.7, “x87 FPU Tag Word”). It then uses this information
to detect two different types of stack faults:
• Stack overflow — An instruction attempts to load a non-empty x87 FPU register
from memory. A non-empty register is defined as a register containing a zero
(tag value of 01), a valid value (tag value of 00), or a special value (tag value of
10).
• Stack underflow — An instruction references an empty x87 FPU register as a
source operand, including attempting to write the contents of an empty register
to memory. An empty register has a tag value of 11.

NOTES
The term stack overflow originates from the situation where the
program has loaded (pushed) eight values from memory onto the
x87 FPU register stack and the next value pushed on the stack causes
a stack wraparound to a register that already contains a value.
The term stack underflow originates from the opposite situation.
Here, a program has stored (popped) eight values from the x87 FPU
register stack to memory and the next value popped from the stack
causes stack wraparound to an empty register.
When the x87 FPU detects stack overflow or underflow, it sets the IE flag (bit 0) and
the SF flag (bit 6) in the x87 FPU status word to 1. It then sets condition-code flag C1
(bit 9) in the x87 FPU status word to 1 if stack overflow occurred or to 0 if stack
underflow occurred.
If the invalid-operation exception is masked, the x87 FPU returns the floating point,
integer, or packed decimal integer indefinite value to the destination operand,
depending on the instruction being executed. This value overwrites the destination
register or memory location specified by the instruction.
If the invalid-operation exception is not masked, a software exception handler is
invoked (see Section 8.7, “Handling x87 FPU Exceptions in Software”) and the top-
of-stack pointer (TOP) and source operands remain unchanged.

Vol. 1 8-37
PROGRAMMING WITH THE X87 FPU

8.5.1.2 Invalid Arithmetic Operand Exception (#IA)

The x87 FPU is able to detect a variety of invalid arithmetic operations that can be
coded in a program. These operations are listed in Table 8-10. (This list includes the
invalid operations defined in IEEE Standard 754.)
When the x87 FPU detects an invalid arithmetic operand, it sets the IE flag (bit 0) in
the x87 FPU status word to 1. If the invalid-operation exception is masked, the x87
FPU then returns an indefinite value or QNaN to the destination operand and/or sets
the floating-point condition codes as shown in Table 8-10. If the invalid-operation
exception is not masked, a software exception handler is invoked (see Section 8.7,
“Handling x87 FPU Exceptions in Software”) and the top-of-stack pointer (TOP) and
source operands remain unchanged.

Table 8-10. Invalid Arithmetic Operations and the

Masked Responses to Them
Condition Masked Response
Any arithmetic operation on an operand that is in Return the QNaN floating-point indefinite
an unsupported format. value to the destination operand.
Any arithmetic operation on a SNaN. Return a QNaN to the destination operand (see
Table 4-7).
Ordered compare and test operations: one or both Set the condition code flags (C0, C2, and C3) in
operands are NaNs. the x87 FPU status word or the CF, PF, and ZF
flags in the EFLAGS register to 111B (not
comparable).
Addition: operands are opposite-signed infinities. Return the QNaN floating-point indefinite
Subtraction: operands are like-signed infinities. value to the destination operand.
Multiplication: ∞ by 0; 0 by ∞ . Return the QNaN floating-point indefinite
value to the destination operand.
Division: ∞ by ∞ ; 0 by 0. Return the QNaN floating-point indefinite
value to the destination operand.
Remainder instructions FPREM, FPREM1: modulus Return the QNaN floating-point indefinite;
(divisor) is 0 or dividend is ∞ . clear condition code flag C2 to 0.
Trigonometric instructions FCOS, FPTAN, FSIN, Return the QNaN floating-point indefinite;
FSINCOS: source operand is ∞ . clear condition code flag C2 to 0.
FSQRT: negative operand (except FSQRT (–0) = – Return the QNaN floating-point indefinite
0); FYL2X: negative operand (except FYL2X (–0) = value to the destination operand.
–∞); FYL2XP1: operand more negative than –1.
FBSTP: Converted value cannot be represented in Store packed BCD integer indefinite value in
18 decimal digits, or source value is an SNaN, the destination operand.
QNaN, ± ∞ , or in an unsupported format.

8-38 Vol. 1
 PROGRAMMING WITH THE X87 FPU

Table 8-10. Invalid Arithmetic Operations and the

Masked Responses to Them (Contd.)
FIST/FISTP: Converted value exceeds Store integer indefinite value in the
representable integer range of the destination destination operand.
operand, or source value is an SNaN, QNaN, ±∞, or
in an unsupported format.
FXCH: one or both registers are tagged empty. Load empty registers with the QNaN floating-
point indefinite value, then perform the
exchange.

Normally, when one or both of the source operands is a QNaN (and neither is an
SNaN or in an unsupported format), an invalid-operand exception is not generated.
An exception to this rule is most of the compare instructions (such as the FCOM and
FCOMI instructions) and the floating-point to integer conversion instructions
(FIST/FISTP and FBSTP). With these instructions, a QNaN source operand will
generate an invalid-operand exception.

8.5.2 Denormal Operand Exception (#D)

The x87 FPU signals the denormal-operand exception under the following conditions:
• If an arithmetic instruction attempts to operate on a denormal operand (see
Section 4.8.3.2, “Normalized and Denormalized Finite Numbers”).
• If an attempt is made to load a denormal single-precision or double-precision
floating-point value into an x87 FPU register. (If the denormal value being loaded
is a double extended-precision floating-point value, the denormal-operand
exception is not reported.)
The flag (DE) for this exception is bit 1 of the x87 FPU status word, and the mask bit
(DM) is bit 1 of the x87 FPU control word.
When a denormal-operand exception occurs and the exception is masked, the x87
FPU sets the DE flag, then proceeds with the instruction. The denormal operand in
single- or double-precision floating-point format is automatically normalized when
converted to the double extended-precision floating-point format. Subsequent oper-
ations will benefit from the additional precision of the internal double extended-preci-
sion floating-point format.
When a denormal-operand exception occurs and the exception is not masked, the DE
flag is set and a software exception handler is invoked (see Section 8.7, “Handling
x87 FPU Exceptions in Software”). The top-of-stack pointer (TOP) and source oper-
ands remain unchanged.
For additional information about the denormal-operation exception, see Section
4.9.1.2, “Denormal Operand Exception (#D).”

Vol. 1 8-39
PROGRAMMING WITH THE X87 FPU

8.5.3 Divide-By-Zero Exception (#Z)

The x87 FPU reports a floating-point divide-by-zero exception whenever an instruc-
tion attempts to divide a finite non-zero operand by 0. The flag (ZE) for this exception
is bit 2 of the x87 FPU status word, and the mask bit (ZM) is bit 2 of the x87 FPU
control word. The FDIV, FDIVP, FDIVR, FDIVRP, FIDIV, and FIDIVR instructions and
the other instructions that perform division internally (FYL2X and FXTRACT) can
report the divide-by-zero exception.
When a divide-by-zero exception occurs and the exception is masked, the x87 FPU
sets the ZE flag and returns the values shown in Table 8-10. If the divide-by-zero
exception is not masked, the ZE flag is set, a software exception handler is invoked
(see Section 8.7, “Handling x87 FPU Exceptions in Software”), and the top-of-stack
pointer (TOP) and source operands remain unchanged.

Table 8-11. Divide-By-Zero Conditions and the Masked Responses to Them

Condition Masked Response
Divide or reverse divide operation Returns an ∞ signed with the exclusive OR of the sign of the
with a 0 divisor. two operands to the destination operand.
FYL2X instruction. Returns an ∞ signed with the opposite sign of the non-zero
operand to the destination operand.
FXTRACT instruction. ST(1) is set to –∞; ST(0) is set to 0 with the same sign as the
source operand.

8.5.4 Numeric Overflow Exception (#O)

The x87 FPU reports a floating-point numeric overflow exception (#O) whenever the
rounded result of an arithmetic instruction exceeds the largest allowable finite value
that will fit into the floating-point format of the destination operand. (See Section
4.9.1.4, “Numeric Overflow Exception (#O),” for additional information about the
numeric overflow exception.)
When using the x87 FPU, numeric overflow can occur on arithmetic operations where
the result is stored in an x87 FPU data register. It can also occur on store floating-
point operations (using the FST and FSTP instructions), where a within-range value
in a data register is stored in memory in a single-precision or double-precision
floating-point format. The numeric overflow exception cannot occur when storing
values in an integer or BCD integer format. Instead, the invalid-arithmetic-operand
exception is signaled.
The flag (OE) for the numeric-overflow exception is bit 3 of the x87 FPU status word,
and the mask bit (OM) is bit 3 of the x87 FPU control word.
When a numeric-overflow exception occurs and the exception is masked, the x87
FPU sets the OE flag and returns one of the values shown in Table 4-10. The value
returned depends on the current rounding mode of the x87 FPU (see Section 8.1.5.3,
“Rounding Control Field”).

8-40 Vol. 1
 PROGRAMMING WITH THE X87 FPU

The action that the x87 FPU takes when numeric overflow occurs and the numeric-
overflow exception is not masked, depends on whether the instruction is supposed to
store the result in memory or on the register stack.
• Destination is a memory location — The OE flag is set and a software
exception handler is invoked (see Section 8.7, “Handling x87 FPU Exceptions in
Software”). The top-of-stack pointer (TOP) and source and destination operands
remain unchanged. Because the data in the stack is in double extended-precision
format, the exception handler has the option either of re-executing the store
instruction after proper adjustment of the operand or of rounding the significand
on the stack to the destination's precision as the standard requires. The
exception handler should ultimately store a value into the destination location in
memory if the program is to continue.
• Destination is the register stack — The significand of the result is rounded
according to current settings of the precision and rounding control bits in the x87
FPU control word and the exponent of the result is adjusted by dividing it by
224576. (For instructions not affected by the precision field, the significand is
rounded to double-extended precision.) The resulting value is stored in the
destination operand. Condition code bit C1 in the x87 FPU status word (called in
this situation the “round-up bit”) is set if the significand was rounded upward and
cleared if the result was rounded toward 0. After the result is stored, the OE flag
is set and a software exception handler is invoked. The scaling bias value 24,576
is equal to 3 ∗ 213. Biasing the exponent by 24,576 normally translates the
number as nearly as possible to the middle of the double extended-precision
floating-point exponent range so that, if desired, it can be used in subsequent
scaled operations with less risk of causing further exceptions.
When using the FSCALE instruction, massive overflow can occur, where the result
is too large to be represented, even with a bias-adjusted exponent. Here, if
overflow occurs again, after the result has been biased, a properly signed ∞ is
stored in the destination operand.

8.5.5 Numeric Underflow Exception (#U)

The x87 FPU detects a floating-point numeric underflow condition whenever the
rounded result of an arithmetic instruction is tiny; that is, less than the smallest
possible normalized, finite value that will fit into the floating-point format of the
destination operand. (See Section 4.9.1.5, “Numeric Underflow Exception (#U),” for
additional information about the numeric underflow exception.)
Like numeric overflow, numeric underflow can occur on arithmetic operations where
the result is stored in an x87 FPU data register. It can also occur on store floating-
point operations (with the FST and FSTP instructions), where a within-range value in
a data register is stored in memory in the smaller single-precision or double-preci-
sion floating-point formats. A numeric underflow exception cannot occur when
storing values in an integer or BCD integer format, because a tiny value is always
rounded to an integral value of 0 or 1, depending on the rounding mode in effect.

Vol. 1 8-41
PROGRAMMING WITH THE X87 FPU

The flag (UE) for the numeric-underflow exception is bit 4 of the x87 FPU status
word, and the mask bit (UM) is bit 4 of the x87 FPU control word.
When a numeric-underflow condition occurs and the exception is masked, the x87
FPU performs the operation described in Section 4.9.1.5, “Numeric Underflow Excep-
tion (#U).”
When the exception is not masked, the action of the x87 FPU depends on whether the
instruction is supposed to store the result in a memory location or on the x87 FPU
resister stack.
• Destination is a memory location — (Can occur only with a store instruction.)
The UE flag is set and a software exception handler is invoked (see Section 8.7,
“Handling x87 FPU Exceptions in Software”). The top-of-stack pointer (TOP) and
source and destination operands remain unchanged, and no result is stored in
memory.
Because the data in the stack is in double extended-precision format, the
exception handler has the option either of re-exchanges the store instruction
after proper adjustment of the operand or of rounding the significand on the
stack to the destination's precision as the standard requires. The exception
handler should ultimately store a value into the destination location in memory if
the program is to continue.
• Destination is the register stack — The significand of the result is rounded
according to current settings of the precision and rounding control bits in the x87
FPU control word and the exponent of the result is adjusted by multiplying it by
224576. (For instructions not affected by the precision field, the significand is
rounded to double extended precision.) The resulting value is stored in the
destination operand. Condition code bit C1 in the x87 FPU status register (acting
here as a “round-up bit”) is set if the significand was rounded upward and cleared
if the result was rounded toward 0. After the result is stored, the UE flag is set
and a software exception handler is invoked. The scaling bias value 24,576 is the
same as is used for the overflow exception and has the same effect, which is to
translate the result as nearly as possible to the middle of the double extended-
precision floating-point exponent range.
When using the FSCALE instruction, massive underflow can occur, where the
result is too tiny to be represented, even with a bias-adjusted exponent. Here, if
underflow occurs again after the result has been biased, a properly signed 0 is
stored in the destination operand.

8.5.6 Inexact-Result (Precision) Exception (#P)

The inexact-result exception (also called the precision exception) occurs if the result
of an operation is not exactly representable in the destination format. (See Section
4.9.1.6, “Inexact-Result (Precision) Exception (#P),” for additional information about
the numeric overflow exception.) Note that the transcendental instructions (FSIN,
FCOS, FSINCOS, FPTAN, FPATAN, F2XM1, FYL2X, and FYL2XP1) by nature produce
inexact results.

8-42 Vol. 1
 PROGRAMMING WITH THE X87 FPU

The inexact-result exception flag (PE) is bit 5 of the x87 FPU status word, and the
mask bit (PM) is bit 5 of the x87 FPU control word.
If the inexact-result exception is masked when an inexact-result condition occurs and
a numeric overflow or underflow condition has not occurred, the x87 FPU handles the
exception as describe in Section 4.9.1.6, “Inexact-Result (Precision) Exception (#P),”
with one additional action. The C1 (round-up) bit in the x87 FPU status word is set to
indicate whether the inexact result was rounded up (C1 is set) or “not rounded up”
(C1 is cleared). In the “not rounded up” case, the least-significant bits of the inexact
result are truncated so that the result fits in the destination format.
If the inexact-result exception is not masked when an inexact result occurs and
numeric overflow or underflow has not occurred, the x87 FPU handles the exception
as described in the previous paragraph and, in addition, invokes a software exception
handler.
If an inexact result occurs in conjunction with numeric overflow or underflow, the x87
FPU carries out one of the following operations:
• If an inexact result occurs in conjunction with masked overflow or underflow, the
OE or UE flag and the PE flag are set and the result is stored as described for the
overflow or underflow exceptions (see Section 8.5.4, “Numeric Overflow
Exception (#O),” or Section 8.5.5, “Numeric Underflow Exception (#U)”). If the
inexact result exception is unmasked, the x87 FPU also invokes a software
exception handler.
• If an inexact result occurs in conjunction with unmasked overflow or underflow
and the destination operand is a register, the OE or UE flag and the PE flag are
set, the result is stored as described for the overflow or underflow exceptions
(see Section 8.5.4, “Numeric Overflow Exception (#O),” or Section 8.5.5,
“Numeric Underflow Exception (#U)”) and a software exception handler is
invoked.
If an unmasked numeric overflow or underflow exception occurs and the destination
operand is a memory location (which can happen only for a floating-point store), the
inexact-result condition is not reported and the C1 flag is cleared.

8.6 X87 FPU EXCEPTION SYNCHRONIZATION

Because the integer unit and x87 FPU are separate execution units, it is possible for
the processor to execute floating-point, integer, and system instructions concur-
rently. No special programming techniques are required to gain the advantages of
concurrent execution. (Floating-point instructions are placed in the instruction
stream along with the integer and system instructions.) However, concurrent execu-
tion can cause problems for floating-point exception handlers.
This problem is related to the way the x87 FPU signals the existence of unmasked
floating-point exceptions. (Special exception synchronization is not required for
masked floating-point exceptions, because the x87 FPU always returns a masked
result to the destination operand.)

Vol. 1 8-43
PROGRAMMING WITH THE X87 FPU

When a floating-point exception is unmasked and the exception condition occurs, the
x87 FPU stops further execution of the floating-point instruction and signals the
exception event. On the next occurrence of a floating-point instruction or a
WAIT/FWAIT instruction in the instruction stream, the processor checks the ES flag in
the x87 FPU status word for pending floating-point exceptions. If floating-point
exceptions are pending, the x87 FPU makes an implicit call (traps) to the floating-
point software exception handler. The exception handler can then execute recovery
procedures for selected or all floating-point exceptions.
Synchronization problems occur in the time between the moment when the excep-
tion is signaled and when it is actually handled. Because of concurrent execution,
integer or system instructions can be executed during this time. It is thus possible for
the source or destination operands for a floating-point instruction that faulted to be
overwritten in memory, making it impossible for the exception handler to analyze or
recover from the exception.
To solve this problem, an exception synchronizing instruction (either a floating-point
instruction or a WAIT/FWAIT instruction) can be placed immediately after any
floating-point instruction that might present a situation where state information
pertaining to a floating-point exception might be lost or corrupted. Floating-point
instructions that store data in memory are prime candidates for synchronization. For
example, the following three lines of code have the potential for exception synchro-
nization problems:
FILD COUNT ;Floating-point instruction
INC COUNT ;Integer instruction
FSQRT ;Subsequent floating-point instruction
In this example, the INC instruction modifies the source operand of the floating-point
instruction, FILD. If an exception is signaled during the execution of the FILD instruc-
tion, the INC instruction would be allowed to overwrite the value stored in the COUNT
memory location before the floating-point exception handler is called. With the
COUNT variable modified, the floating-point exception handler would not be able to
recover from the error.
Rearranging the instructions, as follows, so that the FSQRT instruction follows the
FILD instruction, synchronizes floating-point exception handling and eliminates the
possibility of the COUNT variable being overwritten before the floating-point excep-
tion handler is invoked.
FILD COUNT ;Floating-point instruction
FSQRT ;Subsequent floating-point instruction synchronizes
;any exceptions generated by the FILD instruction.
INC COUNT ;Integer instruction
The FSQRT instruction does not require any synchronization, because the results of
this instruction are stored in the x87 FPU data registers and will remain there, undis-
turbed, until the next floating-point or WAIT/FWAIT instruction is executed. To abso-
lutely insure that any exceptions emanating from the FSQRT instruction are handled
(for example, prior to a procedure call), a WAIT instruction can be placed directly
after the FSQRT instruction.

8-44 Vol. 1
 PROGRAMMING WITH THE X87 FPU

Note that some floating-point instructions (non-waiting instructions) do not check for
pending unmasked exceptions (see Section 8.3.11, “x87 FPU Control Instructions”).
They include the FNINIT, FNSTENV, FNSAVE, FNSTSW, FNSTCW, and FNCLEX instruc-
tions. When an FNINIT, FNSTENV, FNSAVE, or FNCLEX instruction is executed, all
pending exceptions are essentially lost (either the x87 FPU status register is cleared
or all exceptions are masked). The FNSTSW and FNSTCW instructions do not check
for pending interrupts, but they do not modify the x87 FPU status and control regis-
ters. A subsequent “waiting” floating-point instruction can then handle any pending
exceptions.

8.7 HANDLING X87 FPU EXCEPTIONS IN SOFTWARE

The x87 FPU in Pentium and later IA-32 processors provides two different modes of
operation for invoking a software exception handler for floating-point exceptions:
native mode and MS-DOS compatibility mode. The mode of operation is selected by
CR0.NE[bit 5]. (See Chapter 2, “System Architecture Overview,” in the Intel® 64 and
IA-32 Architectures Software Developer’s Manual, Volume 3A, for more information
about the NE flag.)

8.7.1 Native Mode

The native mode for handling floating-point exceptions is selected by setting
CR0.NE[bit 5] to 1. In this mode, if the x87 FPU detects an exception condition while
executing a floating-point instruction and the exception is unmasked (the mask bit
for the exception is cleared), the x87 FPU sets the flag for the exception and the ES
flag in the x87 FPU status word. It then invokes the software exception handler
through the floating-point-error exception (#MF, vector 16), immediately before
execution of any of the following instructions in the processor’s instruction stream:
• The next floating-point instruction, unless it is one of the non-waiting instructions
(FNINIT, FNCLEX, FNSTSW, FNSTCW, FNSTENV, and FNSAVE).
• The next WAIT/FWAIT instruction.
• The next MMX instruction.
If the next floating-point instruction in the instruction stream is a non-waiting
instruction, the x87 FPU executes the instruction without invoking the software
exception handler.

8.7.2 MS-DOS* Compatibility Sub-mode

If CR0.NE[bit 5] is 0, the MS-DOS compatibility mode for handling floating-point
exceptions is selected. In this mode, the software exception handler for floating-
point exceptions is invoked externally using the processor’s FERR#, INTR, and
IGNNE# pins. This method of reporting floating-point errors and invoking an excep-

Vol. 1 8-45
PROGRAMMING WITH THE X87 FPU

tion handler is provided to support the floating-point exception handling mechanism

used in PC systems that are running the MS-DOS or Windows* 95 operating system.
Using FERR# and IGNNE# to handle floating-point exception is deprecated by
modern operating systems, this approach also limits newer processors to operate
with one logical processor active.
The MS-DOS compatibility mode is typically used as follows to invoke the floating-
point exception handler:
1. If the x87 FPU detects an unmasked floating-point exception, it sets the flag for
the exception and the ES flag in the x87 FPU status word.
2. If the IGNNE# pin is deasserted, the x87 FPU then asserts the FERR# pin either
immediately, or else delayed (deferred) until just before the execution of the next
waiting floating-point instruction or MMX instruction. Whether the FERR# pin is
asserted immediately or delayed depends on the type of processor, the
instruction, and the type of exception.
3. If a preceding floating-point instruction has set the exception flag for an
unmasked x87 FPU exception, the processor freezes just before executing the
next WAIT instruction, waiting floating-point instruction, or MMX instruction.
Whether the FERR# pin was asserted at the preceding floating-point instruction
or is just now being asserted, the freezing of the processor assures that the x87
FPU exception handler will be invoked before the new floating-point (or MMX)
instruction gets executed.
4. The FERR# pin is connected through external hardware to IRQ13 of a cascaded,
programmable interrupt controller (PIC). When the FERR# pin is asserted, the
PIC is programmed to generate an interrupt 75H.
5. The PIC asserts the INTR pin on the processor to signal the interrupt 75H.
6. The BIOS for the PC system handles the interrupt 75H by branching to the
interrupt 02H (NMI) interrupt handler.
7. The interrupt 02H handler determines if the interrupt is the result of an NMI
interrupt or a floating-point exception.
8. If a floating-point exception is detected, the interrupt 02H handler branches to
the floating-point exception handler.
If the IGNNE# pin is asserted, the processor ignores floating-point error conditions.
This pin is provided to inhibit floating-point exceptions from being generated while
the floating-point exception handler is servicing a previously signaled floating-point
exception.
Appendix D, “Guidelines for Writing x87 FPU Exception Handlers,” describes the
MS-DOS compatibility mode in much greater detail. This mode is somewhat more
complicated in the Intel486 and Pentium processor implementations, as described in
Appendix D.

8-46 Vol. 1
 PROGRAMMING WITH THE X87 FPU

8.7.3 Handling x87 FPU Exceptions in Software

Section 4.9.3, “Typical Actions of a Floating-Point Exception Handler,” shows actions
that may be carried out by a floating-point exception handler. The state of the x87
FPU can be saved with the FSTENV/FNSTENV or FSAVE/FNSAVE instructions (see
Section 8.1.10, “Saving the x87 FPU’s State with FSTENV/FNSTENV and
FSAVE/FNSAVE”).
If the faulting floating-point instruction is followed by one or more non-floating-point
instructions, it may not be useful to re-execute the faulting instruction. See Section
8.6, “x87 FPU Exception Synchronization,” for more information on synchronizing
floating-point exceptions.
In cases where the handler needs to restart program execution with the faulting
instruction, the IRET instruction cannot be used directly. The reason for this is that
because the exception is not generated until the next floating-point or WAIT/FWAIT
instruction following the faulting floating-point instruction, the return instruction
pointer on the stack may not point to the faulting instruction. To restart program
execution at the faulting instruction, the exception handler must obtain a pointer to
the instruction from the saved x87 FPU state information, load it into the return
instruction pointer location on the stack, and then execute the IRET instruction.
See Section D.3.4, “x87 FPU Exception Handling Examples,” for general examples of
floating-point exception handlers and for specific examples of how to write a floating-
point exception handler when using the MS-DOS compatibility mode.

Vol. 1 8-47
PROGRAMMING WITH THE X87 FPU

8-48 Vol. 1
 CHAPTER 9
PROGRAMMING WITH INTEL® MMX™ TECHNOLOGY

The Intel MMX technology was introduced into the IA-32 architecture in the
Pentium II processor family and Pentium processor with MMX technology. The exten-
sions introduced in MMX technology support a single-instruction, multiple-data
(SIMD) execution model that is designed to accelerate the performance of advanced
media and communications applications.
This chapter describes MMX technology.

9.1 OVERVIEW OF MMX TECHNOLOGY

MMX technology defines a simple and flexible SIMD execution model to handle 64-bit
packed integer data. This model adds the following features to the IA-32 architec-
ture, while maintaining backwards compatibility with all IA-32 applications and
operating-system code:
• Eight new 64-bit data registers, called MMX registers
• Three new packed data types:
— 64-bit packed byte integers (signed and unsigned)
— 64-bit packed word integers (signed and unsigned)
— 64-bit packed doubleword integers (signed and unsigned)
• Instructions that support the new data types and to handle MMX state
management
• Extensions to the CPUID instruction
MMX technology is accessible from all the IA32-architecture execution modes
(protected mode, real address mode, and virtual 8086 mode). It does not add any
new modes to the architecture.
The following sections of this chapter describe MMX technology’s programming envi-
ronment, including MMX register set, data types, and instruction set. Additional
instructions that operate on MMX registers have been added to the IA-32 architec-
ture by the SSE/SSE2 extensions.
For more information, see:
• Section 10.4.4, “SSE 64-Bit SIMD Integer Instructions,” describes MMX instruc-
tions added to the IA-32 architecture with the SSE extensions.
• Section 11.4.2, “SSE2 64-Bit and 128-Bit SIMD Integer Instructions,” describes
MMX instructions added to the IA-32 architecture with SSE2 extensions.
• Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volumes
2A & 2B, give detailed descriptions of MMX instructions.

Vol. 1 9-1
PROGRAMMING WITH INTEL® MMX™ TECHNOLOGY

• Chapter 11, “Intel® MMX™ Technology System Programming,” in the Intel® 64

and IA-32 Architectures Software Developer’s Manual, Volume 3B, describes the
manner in which MMX technology is integrated into the IA-32 system
programming model.

9.2 THE MMX TECHNOLOGY PROGRAMMING

ENVIRONMENT
Figure 9-1 shows the execution environment for MMX technology. All MMX instruc-
tions operate on MMX registers, the general-purpose registers, and/or memory as
follows:
• MMX registers — These eight registers (see Figure 9-1) are used to perform
operations on 64-bit packed integer data. They are named MM0 through MM7.

Address Space
232 -1
MMX Registers
Eight 64-Bit

General-Purpose
Registers
Eight 32-Bit

Figure 9-1. MMX Technology Execution Environment

• General-purpose registers — The eight general-purpose registers (see

Figure 3-5) are used with existing IA-32 addressing modes to address operands
in memory. (MMX registers cannot be used to address memory). General-
purpose registers are also used to hold operands for some MMX technology
operations. They are EAX, EBX, ECX, EDX, EBP, ESI, EDI, and ESP.

9.2.1 MMX Technology in 64-Bit Mode and Compatibility Mode

In compatibility mode and 64-bit mode, MMX instructions function like they do in
protected mode. Memory operands are specified using the ModR/M, SIB encoding
described in Section 3.7.5.

9-2 Vol. 1
 PROGRAMMING WITH INTEL® MMX™ TECHNOLOGY

9.2.2 MMX Registers

The MMX register set consists of eight 64-bit registers (see Figure 9-2), that are used
to perform calculations on the MMX packed integer data types. Values in MMX regis-
ters have the same format as a 64-bit quantity in memory.
The MMX registers have two data access modes: 64-bit access mode and 32-bit
access mode. The 64-bit access mode is used for:
• 64-bit memory accesses
• 64-bit transfers between MMX registers
• All pack, logical, and arithmetic instructions
• Some unpack instructions
The 32-bit access mode is used for:
• 32-bit memory accesses
• 32-bit transfer between general-purpose registers and MMX registers
• Some unpack instructions

63 0
MM7

MM6

MM5

MM4

MM3

MM2

MM1

MM0

Figure 9-2. MMX Register Set

Although MMX registers are defined in the IA-32 architecture as separate registers,
they are aliased to the registers in the FPU data register stack (R0 through R7).
See also Section 9.5, “Compatibility with x87 FPU Architecture.”

9.2.3 MMX Data Types

MMX technology introduced the following 64-bit data types to the IA-32 architecture
(see Figure 9-3):
• 64-bit packed byte integers — eight packed bytes

Vol. 1 9-3
PROGRAMMING WITH INTEL® MMX™ TECHNOLOGY

• 64-bit packed word integers — four packed words

• 64-bit packed doubleword integers — two packed doublewords
MMX instructions move 64-bit packed data types (packed bytes, packed words, or
packed doublewords) and the quadword data type between MMX registers and
memory or between MMX registers in 64-bit blocks. However, when performing arith-
metic or logical operations on the packed data types, MMX instructions operate in
parallel on the individual bytes, words, or doublewords contained in MMX registers
(see Section 9.2.5, “Single Instruction, Multiple Data (SIMD) Execution Model”).

Packed Byte Integers

63 0

Packed Word Integers

63 0

Packed Doubleword Integers

63 0

Figure 9-3. Data Types Introduced with the MMX Technology

9.2.4 Memory Data Formats

When stored in memory: bytes, words and doublewords in the packed data types are
stored in consecutive addresses. The least significant byte, word, or doubleword is
stored at the lowest address and the most significant byte, word, or doubleword is
stored at the high address. The ordering of bytes, words, or doublewords in memory
is always little endian. That is, the bytes with the low addresses are less significant
than the bytes with high addresses.

9.2.5 Single Instruction, Multiple Data (SIMD) Execution Model

MMX technology uses the single instruction, multiple data (SIMD) technique for
performing arithmetic and logical operations on bytes, words, or doublewords packed
into MMX registers (see Figure 9-4). For example, the PADDSW instruction adds 4
signed word integers from one source operand to 4 signed word integers in a second
source operand and stores 4 word integer results in a destination operand. This SIMD
technique speeds up software performance by allowing the same operation to be
carried out on multiple data elements in parallel. MMX technology supports parallel
operations on byte, word, and doubleword data elements when contained in MMX
registers.

9-4 Vol. 1
 PROGRAMMING WITH INTEL® MMX™ TECHNOLOGY

The SIMD execution model supported in the MMX technology directly addresses the
needs of modern media, communications, and graphics applications, which often use
sophisticated algorithms that perform the same operations on a large number of
small data types (bytes, words, and doublewords). For example, most audio data is
represented in 16-bit (word) quantities. The MMX instructions can operate on 4
words simultaneously with one instruction. Video and graphics information is
commonly represented as palletized 8-bit (byte) quantities. In Figure 9-4, one MMX
instruction operates on 8 bytes simultaneously.

Source 1 X3 X2 X1 X0

Source 2 Y3 Y2 Y1 Y0

OP OP OP OP

Destination X3 OP Y3 X2 OP Y2 X1 OP Y1 X0 OP Y0

Figure 9-4. SIMD Execution Model

9.3 SATURATION AND WRAPAROUND MODES

When performing integer arithmetic, an operation may result in an out-of-range
condition, where the true result cannot be represented in the destination format. For
example, when performing arithmetic on signed word integers, positive overflow can
occur when the true signed result is larger than 16 bits.
The MMX technology provides three ways of handling out-of-range conditions:
• Wraparound arithmetic — With wraparound arithmetic, a true out-of-range
result is truncated (that is, the carry or overflow bit is ignored and only the least
significant bits of the result are returned to the destination). Wraparound
arithmetic is suitable for applications that control the range of operands to
prevent out-of-range results. If the range of operands is not controlled, however,
wraparound arithmetic can lead to large errors. For example, adding two large
signed numbers can cause positive overflow and produce a negative result.
• Signed saturation arithmetic — With signed saturation arithmetic, out-of-
range results are limited to the representable range of signed integers for the
integer size being operated on (see Table 9-1). For example, if positive overflow
occurs when operating on signed word integers, the result is “saturated” to
7FFFH, which is the largest positive integer that can be represented in 16 bits; if
negative overflow occurs, the result is saturated to 8000H.

Vol. 1 9-5
PROGRAMMING WITH INTEL® MMX™ TECHNOLOGY

• Unsigned saturation arithmetic — With unsigned saturation arithmetic, out-

of-range results are limited to the representable range of unsigned integers for
the integer size. So, positive overflow when operating on unsigned byte integers
results in FFH being returned and negative overflow results in 00H being
returned.
.

Table 9-1. Data Range Limits for Saturation

Data Type Lower Limit Upper Limit
Hexadecimal Decimal Hexadecimal Decimal
Signed Byte 80H -128 7FH 127
Signed Word 8000H -32,768 7FFFH 32,767
Unsigned Byte 00H 0 FFH 255
Unsigned Word 0000H 0 FFFFH 65,535

Saturation arithmetic provides an answer for many overflow situations. For example,
in color calculations, saturation causes a color to remain pure black or pure white
without allowing inversion. It also prevents wraparound artifacts from entering into
computations when range checking of source operands it not used.
MMX instructions do not indicate overflow or underflow occurrence by generating
exceptions or setting flags in the EFLAGS register.

9.4 MMX INSTRUCTIONS

The MMX instruction set consists of 47 instructions, grouped into the following cate-
gories:
• Data transfer
• Arithmetic
• Comparison
• Conversion
• Unpacking
• Logical
• Shift
• Empty MMX state instruction (EMMS)
Table 9-2 gives a summary of the instructions in the MMX instruction set. The
following sections give a brief overview of the instructions within each group.

9-6 Vol. 1
 PROGRAMMING WITH INTEL® MMX™ TECHNOLOGY

NOTES
The MMX instructions described in this chapter are those instructions
that are available in an IA-32 processor when
CPUID.01H:EDX.MMX[bit 23] = 0.
Section 10.4.4, “SSE 64-Bit SIMD Integer Instructions,” and Section
11.4.2, “SSE2 64-Bit and 128-Bit SIMD Integer Instructions,” list
additional instructions included with SSE/SSE2 extensions that
operate on the MMX registers but are not considered part of the MMX
instruction set.

Table 9-2. MMX Instruction Set Summary

Category Wraparound Signed Unsigned Saturation
Saturation
Arithmetic Addition PADDB, PADDW, PADDSB, PADDSW PADDUSB, PADDUSW
PADDD PSUBSB, PSUBSW PSUBUSB, PSUBUSW
Subtraction PSUBB, PSUBW,
PSUBD
Multiplication PMULL, PMULH
Multiply and Add PMADD
Comparison Compare for Equal PCMPEQB,
PCMPEQW,
PCMPEQD
Compare for PCMPGTPB,
Greater Than PCMPGTPW,
PCMPGTPD
Conversion Pack PACKSSWB, PACKUSWB
PACKSSDW
Unpack Unpack High PUNPCKHBW,
PUNPCKHWD,
PUNPCKHDQ
Unpack Low PUNPCKLBW,
PUNPCKLWD,
PUNPCKLDQ
Packed Full Quadword
Logical And PAND
And Not PANDN
Or POR
Exclusive OR PXOR

Vol. 1 9-7
PROGRAMMING WITH INTEL® MMX™ TECHNOLOGY

Table 9-2. MMX Instruction Set Summary (Contd.)

Category Wraparound Signed Unsigned Saturation
Saturation
Shift Shift Left Logical PSLLW, PSLLD PSLLQ
Shift Right Logical PSRLW, PSRLD PSRLQ
Shift Right PSRAW, PSRAD
Arithmetic
Doubleword Transfers Quadword Transfers
Data Register to MOVD MOVQ
Transfer Register MOVD MOVQ
Load from MOVD MOVQ
Memory
Store to Memory
Empty MMX EMMS
State

9.4.1 Data Transfer Instructions

The MOVD (Move 32 Bits) instruction transfers 32 bits of packed data from memory
to an MMX register and vice versa; or from a general-purpose register to an MMX
register and vice versa.
The MOVQ (Move 64 Bits) instruction transfers 64 bits of packed data from memory
to an MMX register and vice versa; or transfers data between MMX registers.

9.4.2 Arithmetic Instructions

The arithmetic instructions perform addition, subtraction, multiplication, and
multiply/add operations on packed data types.
The PADDB/PADDW/PADDD (add packed integers) instructions and the
PSUBB/PSUBW/ PSUBD (subtract packed integers) instructions add or subtract the
corresponding signed or unsigned data elements of the source and destination oper-
ands in wraparound mode. These instructions operate on packed byte, word, and
doubleword data types.
The PADDSB/PADDSW (add packed signed integers with signed saturation) instruc-
tions and the PSUBSB/PSUBSW (subtract packed signed integers with signed satura-
tion) instructions add or subtract the corresponding signed data elements of the
source and destination operands and saturate the result to the limits of the signed
data-type range. These instructions operate on packed byte and word data types.
The PADDUSB/PADDUSW (add packed unsigned integers with unsigned saturation)
instructions and the PSUBUSB/PSUBUSW (subtract packed unsigned integers with

9-8 Vol. 1
 PROGRAMMING WITH INTEL® MMX™ TECHNOLOGY

unsigned saturation) instructions add or subtract the corresponding unsigned data

elements of the source and destination operands and saturate the result to the limits
of the unsigned data-type range. These instructions operate on packed byte and
word data types.
The PMULHW (multiply packed signed integers and store high result) and PMULLW
(multiply packed signed integers and store low result) instructions perform a signed
multiply of the corresponding words of the source and destination operands and write
the high-order or low-order 16 bits of each of the results, respectively, to the desti-
nation operand.
The PMADDWD (multiply and add packed integers) instruction computes the prod-
ucts of the corresponding signed words of the source and destination operands. The
four intermediate 32-bit doubleword products are summed in pairs (high-order pair
and low-order pair) to produce two 32-bit doubleword results.

9.4.3 Comparison Instructions

The PCMPEQB/PCMPEQW/PCMPEQD (compare packed data for equal) instructions
and the PCMPGTB/PCMPGTW/PCMPGTD (compare packed signed integers for greater
than) instructions compare the corresponding signed data elements (bytes, words,
or doublewords) in the source and destination operands for equal to or greater than,
respectively.
These instructions generate a mask of ones or zeros which are written to the destina-
tion operand. Logical operations can use the mask to select packed elements. This
can be used to implement a packed conditional move operation without a branch or a
set of branch instructions. No flags in the EFLAGS register are affected.

9.4.4 Conversion Instructions

The PACKSSWB (pack words into bytes with signed saturation) and PACKSSDW (pack
doublewords into words with signed saturation) instructions convert signed words
into signed bytes and signed doublewords into signed words, respectively, using
signed saturation.
PACKUSWB (pack words into bytes with unsigned saturation) converts signed words
into unsigned bytes, using unsigned saturation.

9.4.5 Unpack Instructions

The PUNPCKHBW/PUNPCKHWD/PUNPCKHDQ (unpack high-order data elements)
instructions and the PUNPCKLBW/PUNPCKLWD/PUNPCKLDQ (unpack low-order data
elements) instructions unpack bytes, words, or doublewords from the high- or low-
order data elements of the source and destination operands and interleave them in
the destination operand. By placing all 0s in the source operand, these instructions

Vol. 1 9-9
PROGRAMMING WITH INTEL® MMX™ TECHNOLOGY

can be used to convert byte integers to word integers, word integers to doubleword
integers, or doubleword integers to quadword integers.

9.4.6 Logical Instructions

PAND (bitwise logical AND), PANDN (bitwise logical AND NOT), POR (bitwise logical
OR), and PXOR (bitwise logical exclusive OR) perform bitwise logical operations on
the quadword source and destination operands.

9.4.7 Shift Instructions

The logical shift left, logical shift right and arithmetic shift right instructions shift each
element by a specified number of bit positions.
The PSLLW/PSLLD/PSLLQ (shift packed data left logical) instructions and the
PSRLW/PSRLD/PSRLQ (shift packed data right logical) instructions perform a logical
left or right shift of the data elements and fill the empty high or low order bit posi-
tions with zeros. These instructions operate on packed words, doublewords, and
quadwords.
The PSRAW/PSRAD (shift packed data right arithmetic) instructions perform an arith-
metic right shift, copying the sign bit for each data element into empty bit positions
on the upper end of each data element. This instruction operates on packed words
and doublewords.

9.4.8 EMMS Instruction

The EMMS instruction empties the MMX state by setting the tags in x87 FPU tag word
to 11B, indicating empty registers. This instruction must be executed at the end of an
MMX routine before calling other routines that can execute floating-point instruc-
tions. See Section 9.6.3, “Using the EMMS Instruction,” for more information on the
use of this instruction.

9.5 COMPATIBILITY WITH X87 FPU ARCHITECTURE

The MMX state is aliased to the x87 FPU state. No new states or modes have been
added to IA-32 architecture to support the MMX technology. The same floating-point
instructions that save and restore the x87 FPU state also handle the MMX state (for
example, during context switching).
MMX technology uses the same interface techniques between the x87 FPU and the
operating system (primarily for task switching purposes). For more details, see
Chapter 11, “Intel® MMX™ Technology System Programming,” in the Intel® 64 and
IA-32 Architectures Software Developer’s Manual, Volume 3A.

9-10 Vol. 1
 PROGRAMMING WITH INTEL® MMX™ TECHNOLOGY

9.5.1 MMX Instructions and the x87 FPU Tag Word

After each MMX instruction, the entire x87 FPU tag word is set to valid (00B). The
EMMS instruction (empty MMX state) sets the entire x87 FPU tag word to empty
(11B).
Chapter 11, “Intel® MMX™ Technology System Programming,” in the Intel® 64 and
IA-32 Architectures Software Developer’s Manual, Volume 3A, provides additional
information about the effects of x87 FPU and MMX instructions on the x87 FPU tag
word. For a description of the tag word, see Section 8.1.7, “x87 FPU Tag Word.”

9.6 WRITING APPLICATIONS WITH MMX CODE

The following sections give guidelines for writing application code that uses MMX
technology.

9.6.1 Checking for MMX Technology Support

Before an application attempts to use the MMX technology, it should check that it is
present on the processor. Check by following these steps:
1. Check that the processor supports the CPUID instruction by attempting to
execute the CPUID instruction. If the processor does not support the CPUID
instruction, this will generate an invalid-opcode exception (#UD).
2. Check that the processor supports the MMX technology
(if CPUID.01H:EDX.MMX[bit 23] = 1).
3. Check that emulation of the x87 FPU is disabled (if CR0.EM[bit 2] = 0).
If the processor attempts to execute an unsupported MMX instruction or attempts to
execute an MMX instruction with CR0.EM[bit 2] set, this generates an invalid-opcode
exception (#UD).
Example 9-1 illustrates how to use the CPUID instruction to detect the MMX tech-
nology. This example does not represent the entire CPUID sequence, but shows the
portion used for detection of MMX technology.

Example 9-1. Partial Routine for Detecting MMX Technology with the CPUID Instruction
... ; identify existence of CPUID instruction
... ; identify Intel processor
mov EAX, 1 ; request for feature flags
CPUID ; 0FH, 0A2H CPUID instruction
test EDX, 00800000H ; Is IA MMX technology bit (Bit 23 of EDX) set?
jnz ; MMX_Technology_Found

Vol. 1 9-11
PROGRAMMING WITH INTEL® MMX™ TECHNOLOGY

9.6.2 Transitions Between x87 FPU and MMX Code

Applications can contain both x87 FPU floating-point and MMX instructions. However,
because the MMX registers are aliased to the x87 FPU register stack, care must be
taken when making transitions between x87 FPU instructions and MMX instructions
to prevent incoherent or unexpected results.
When an MMX instruction (other than the EMMS instruction) is executed, the
processor changes the x87 FPU state as follows:
• The TOS (top of stack) value of the x87 FPU status word is set to 0.
• The entire x87 FPU tag word is set to the valid state (00B in all tag fields).
• When an MMX instruction writes to an MMX register, it writes ones (11B) to the
exponent part of the corresponding floating-point register (bits 64 through 79).
The net result of these actions is that any x87 FPU state prior to the execution of the
MMX instruction is essentially lost.
When an x87 FPU instruction is executed, the processor assumes that the current
state of the x87 FPU register stack and control registers is valid and executes the
instruction without any preparatory modifications to the x87 FPU state.
If the application contains both x87 FPU floating-point and MMX instructions, the
following guidelines are recommended:
• When transitioning between x87 FPU and MMX code, save the state of any x87
FPU data or control registers that need to be preserved for future use. The FSAVE
and FXSAVE instructions save the entire x87 FPU state.
• When transitioning between MMX and x87 FPU code, do the following:
— Save any data in the MMX registers that needs to be preserved for future use.
FSAVE and FXSAVE also save the state of MMX registers.
— Execute the EMMS instruction to clear the MMX state from the x87 data and
control registers.
The following sections describe the use of the EMMS instruction and give additional
guidelines for mixing x87 FPU and MMX code.

9.6.3 Using the EMMS Instruction

As described in Section 9.6.2, “Transitions Between x87 FPU and MMX Code,” when
an MMX instruction executes, the x87 FPU tag word is marked valid (00B). In this
state, the execution of subsequent x87 FPU instructions may produce unexpected
x87 FPU floating-point exceptions and/or incorrect results because the x87 FPU
register stack appears to contain valid data. The EMMS instruction is provided to
prevent this problem by marking the x87 FPU tag word as empty.
The EMMS instruction should be used in each of the following cases:
• When an application using the x87 FPU instructions calls an MMX technology
library/DLL (use the EMMS instruction at the end of the MMX code).

9-12 Vol. 1
 PROGRAMMING WITH INTEL® MMX™ TECHNOLOGY

• When an application using MMX instructions calls a x87 FPU floating-point

library/DLL (use the EMMS instruction before calling the x87 FPU code).
• When a switch is made between MMX code in a task or thread and other tasks or
threads in cooperative operating systems, unless it is certain that more MMX
instructions will be executed before any x87 FPU code.
EMMS is not required when mixing MMX technology instructions with
SSE/SSE2/SSE3 instructions (see Section 11.6.7, “Interaction of SSE/SSE2 Instruc-
tions with x87 FPU and MMX Instructions”).

9.6.4 Mixing MMX and x87 FPU Instructions

An application can contain both x87 FPU floating-point and MMX instructions.
However, frequent transitions between MMX and x87 FPU instructions are not recom-
mended, because they can degrade performance in some processor implementa-
tions. When mixing MMX code with x87 FPU code, follow these guidelines:
• Keep the code in separate modules, procedures, or routines.
• Do not rely on register contents across transitions between x87 FPU and MMX
code modules.
• When transitioning between MMX code and x87 FPU code, save the MMX register
state (if it will be needed in the future) and execute an EMMS instruction to empty
the MMX state.
• When transitioning between x87 FPU code and MMX code, save the x87 FPU state
if it will be needed in the future.

9.6.5 Interfacing with MMX Code

MMX technology enables direct access to all the MMX registers. This means that all
existing interface conventions that apply to the use of the processor’s general-
purpose registers (EAX, EBX, etc.) also apply to the use of MMX registers.
An efficient interface to MMX routines might pass parameters and return values
through the MMX registers or through a combination of memory locations (via the
stack) and MMX registers. Do not use the EMMS instruction or mix MMX and x87 FPU
code when using to the MMX registers to pass parameters.
If a high-level language that does not support the MMX data types directly is used,
the MMX data types can be defined as a 64-bit structure containing packed data
types.
When implementing MMX instructions in high-level languages, other approaches can
be taken, such as:
• Passing parameters to an MMX routine by passing a pointer to a structure via the
stack.
• Returning a value from a function by returning a pointer to a structure.

Vol. 1 9-13
PROGRAMMING WITH INTEL® MMX™ TECHNOLOGY

9.6.6 Using MMX Code in a Multitasking Operating System

Environment
An application needs to identify the nature of the multitasking operating system on
which it runs. Each task retains its own state which must be saved when a task switch
occurs. The processor state (context) consists of the general-purpose registers and
the floating-point and MMX registers.
Operating systems can be classified into two types:
• Cooperative multitasking operating system
• Preemptive multitasking operating system
Cooperative multitasking operating systems do not save the FPU or MMX state when
performing a context switch. Therefore, the application needs to save the relevant
state before relinquishing direct or indirect control to the operating system.
Preemptive multitasking operating systems are responsible for saving and restoring
the FPU and MMX state when performing a context switch. Therefore, the application
does not have to save or restore the FPU and MMX state.

9.6.7 Exception Handling in MMX Code

MMX instructions generate the same type of memory-access exceptions as other
IA-32 instructions (page fault, segment not present, and limit violations). Existing
exception handlers do not have to be modified to handle these types of exceptions for
MMX code.
Unless there is a pending floating-point exception, MMX instructions do not generate
numeric exceptions. Therefore, there is no need to modify existing exception
handlers or add new ones to handle numeric exceptions.
If a floating-point exception is pending, the subsequent MMX instruction generates a
numeric error exception (interrupt 16 and/or assertion of the FERR# pin). The MMX
instruction resumes execution upon return from the exception handler.

9.6.8 Register Mapping

MMX registers and their tags are mapped to physical locations of the floating-point
registers and their tags. Register aliasing and mapping is described in more detail in
Chapter 11, “Intel® MMX™ Technology System Programming,” in the Intel® 64 and
IA-32 Architectures Software Developer’s Manual, Volume 3A.

9.6.9 Effect of Instruction Prefixes on MMX Instructions

Table 9-3 describes the effect of instruction prefixes on MMX instructions. Unpredict-
able behavior can range from being treated as a reserved operation on one genera-

9-14 Vol. 1
 PROGRAMMING WITH INTEL® MMX™ TECHNOLOGY

tion of IA-32 processors to generating an invalid opcode exception on another

generation of processors.

Table 9-3. Effect of Prefixes on MMX Instructions

Prefix Type Effect on MMX Instructions
Address Size Prefix (67H) Affects instructions with a memory operand.
Reserved for instructions without a memory operand and
may result in unpredictable behavior.
Operand Size (66H) Reserved and may result in unpredictable behavior.
Segment Override (2EH, 36H, Affects instructions with a memory operand.
3EH, 26H, 64H, 65H)
Reserved for instructions without a memory operand and
may result in unpredictable behavior.
Repeat Prefix (F3H) Reserved and may result in unpredictable behavior.
Repeat NE Prefix(F2H) Reserved and may result in unpredictable behavior.
Lock Prefix (F0H) Reserved; generates invalid opcode exception (#UD).
Branch Hint Prefixes (2EH and Reserved and may result in unpredictable behavior.
3EH)

See “Instruction Prefixes” in Chapter 2, “Instruction Format,” of the Intel® 64 and

IA-32 Architectures Software Developer’s Manual, Volume 2A, for a description of the
instruction prefixes.

Vol. 1 9-15
PROGRAMMING WITH INTEL® MMX™ TECHNOLOGY

9-16 Vol. 1
 CHAPTER 10
PROGRAMMING WITH
STREAMING SIMD EXTENSIONS (SSE)

The streaming SIMD extensions (SSE) were introduced into the IA-32 architecture in
the Pentium III processor family. These extensions enhance the performance of IA-32
processors for advanced 2-D and 3-D graphics, motion video, image processing,
speech recognition, audio synthesis, telephony, and video conferencing.
This chapter describes SSE. Chapter 11, “Programming with Streaming SIMD Exten-
sions 2 (SSE2),” provides information to assist in writing application programs that
use SSE2 extensions. Chapter 12, “Programming with SSE3, SSSE3, AND SSE4,”
provides this information for SSE3 extensions.

10.1 OVERVIEW OF SSE EXTENSIONS

Intel MMX technology introduced single-instruction multiple-data (SIMD) capability
into the IA-32 architecture, with the 64-bit MMX registers, 64-bit packed integer data
types, and instructions that allowed SIMD operations to be performed on packed
integers. SSE extensions expand the SIMD execution model by adding facilities for
handling packed and scalar single-precision floating-point values contained in
128-bit registers.
If CPUID.01H:EDX.SSE[bit 25] = 1, SSE extensions are present.
SSE extensions add the following features to the IA-32 architecture, while main-
taining backward compatibility with all existing IA-32 processors, applications and
operating systems.
• Eight 128-bit data registers (called XMM registers) in non-64-bit modes; sixteen
XMM registers are available in 64-bit mode.
• The 32-bit MXCSR register, which provides control and status bits for operations
performed on XMM registers.
• The 128-bit packed single-precision floating-point data type (four IEEE single-
precision floating-point values packed into a double quadword).
• Instructions that perform SIMD operations on single-precision floating-point
values and that extend SIMD operations that can be performed on integers:
— 128-bit Packed and scalar single-precision floating-point instructions that
operate on data located in MMX registers
— 64-bit SIMD integer instructions that support additional operations on packed
integer operands located in MMX registers
• Instructions that save and restore the state of the MXCSR register.

Vol. 1 10-1
PROGRAMMING WITH STREAMING SIMD EXTENSIONS (SSE)

• Instructions that support explicit prefetching of data, control of the cacheability

of data, and control the ordering of store operations.
• Extensions to the CPUID instruction.
These features extend the IA-32 architecture’s SIMD programming model in four
important ways:
• The ability to perform SIMD operations on four packed single-precision floating-
point values enhances the performance of IA-32 processors for advanced media
and communications applications that use computation-intensive algorithms to
perform repetitive operations on large arrays of simple, native data elements.
• The ability to perform SIMD single-precision floating-point operations in XMM
registers and SIMD integer operations in MMX registers provides greater
flexibility and throughput for executing applications that operate on large arrays
of floating-point and integer data.
• Cache control instructions provide the ability to stream data in and out of XMM
registers without polluting the caches and the ability to prefetch data to selected
cache levels before it is actually used. Applications that require regular access to
large amounts of data benefit from these prefetching and streaming store
capabilities.
• The SFENCE (store fence) instruction provides greater control over the ordering
of store operations when using weakly-ordered memory types.
SSE extensions are fully compatible with all software written for IA-32 processors. All
existing software continues to run correctly, without modification, on processors that
incorporate SSE extensions. Enhancements to CPUID permit detection of SSE exten-
sions. SSE extensions are accessible from all IA-32 execution modes: protected
mode, real address mode, and virtual-8086 mode.
The following sections of this chapter describe the programming environment for SSE
extensions, including: XMM registers, the packed single-precision floating-point data
type, and SSE instructions. For additional information, see:
• Section 11.6, “Writing Applications with SSE/SSE2 Extensions”.
• Section 11.5, “SSE, SSE2, and SSE3 Exceptions,” describes the exceptions that
can be generated with SSE/SSE2/SSE3 instructions.
• Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volumes
2A & 2B, provide a detailed description of these instructions.
• Chapter 12, “System Programming for Streaming SIMD Instruction Sets,” in the
Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volume 3A,
gives guidelines for integrating these extensions into an operating-system
environment.

10-2 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS (SSE)

10.2 SSE PROGRAMMING ENVIRONMENT

Figure 10-1 shows the execution environment for the SSE extensions. All SSE
instructions operate on the XMM registers, MMX registers, and/or memory as
follows:
• XMM registers — These eight registers (see Figure 10-2 and Section 10.2.2,
“XMM Registers”) are used to operate on packed or scalar single-precision
floating-point data. Scalar operations are operations performed on individual
(unpacked) single-precision floating-point values stored in the low doubleword of
an XMM register. XMM registers are referenced by the names XMM0 through
XMM7.

Address Space

XMM Registers 232 -1

Eight 128-Bit

MXCSR Register 32 Bits

MMX Registers
Eight 64-Bit

General-Purpose
Registers
Eight 32-Bit

0
EFLAGS Register 32 Bits

Figure 10-1. SSE Execution Environment

• MXCSR register — This 32-bit register (see Figure 10-3 and Section 10.2.3,
“MXCSR Control and Status Register”) provides status and control bits used in
SIMD floating-point operations.
• MMX registers — These eight registers (see Figure 9-2) are used to perform
operations on 64-bit packed integer data. They are also used to hold operands for
some operations performed between the MMX and XMM registers. MMX registers
are referenced by the names MM0 through MM7.
• General-purpose registers — The eight general-purpose registers (see
Figure 3-5) are used along with the existing IA-32 addressing modes to address
operands in memory. (MMX and XMM registers cannot be used to address
memory). The general-purpose registers are also used to hold operands for some

Vol. 1 10-3
PROGRAMMING WITH STREAMING SIMD EXTENSIONS (SSE)

SSE instructions and are referenced as EAX, EBX, ECX, EDX, EBP, ESI, EDI, and
ESP.
• EFLAGS register — This 32-bit register (see Figure 3-8) is used to record result
of some compare operations.

10.2.1 SSE in 64-Bit Mode and Compatibility Mode

In compatibility mode, SSE extensions function like they do in protected mode. In
64-bit mode, eight additional XMM registers are accessible. Registers XMM8-XMM15
are accessed by using REX prefixes. Memory operands are specified using the
ModR/M, SIB encoding described in Section 3.7.5.
Some SSE instructions may be used to operate on general-purpose registers. Use the
REX.W prefix to access 64-bit general-purpose registers. Note that if a REX prefix is
used when it has no meaning, the prefix is ignored.

10.2.2 XMM Registers

Eight 128-bit XMM data registers were introduced into the IA-32 architecture with
SSE extensions (see Figure 10-2). These registers can be accessed directly using the
names XMM0 to XMM7; and they can be accessed independently from the x87 FPU
and MMX registers and the general-purpose registers (that is, they are not aliased to
any other of the processor’s registers).

127 0
XMM7

XMM6

XMM5

XMM4

XMM3

XMM2

XMM1

XMM0

Figure 10-2. XMM Registers

SSE instructions use the XMM registers only to operate on packed single-precision
floating-point operands. SSE2 extensions expand the functions of the XMM registers
to operand on packed or scalar double-precision floating-point operands and packed

10-4 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS (SSE)

integer operands (see Section 11.2, “SSE2 Programming Environment,” and Section
12.1, “Programming Environment and Data types”).
XMM registers can only be used to perform calculations on data; they cannot be used
to address memory. Addressing memory is accomplished by using the general-
purpose registers.
Data can be loaded into XMM registers or written from the registers to memory in
32-bit, 64-bit, and 128-bit increments. When storing the entire contents of an XMM
register in memory (128-bit store), the data is stored in 16 consecutive bytes, with
the low-order byte of the register being stored in the first byte in memory.

10.2.3 MXCSR Control and Status Register

The 32-bit MXCSR register (see Figure 10-3) contains control and status information
for SSE, SSE2, and SSE3 SIMD floating-point operations. This register contains:
• flag and mask bits for SIMD floating-point exceptions
• rounding control field for SIMD floating-point operations
• flush-to-zero flag that provides a means of controlling underflow conditions on
SIMD floating-point operations
• denormals-are-zeros flag that controls how SIMD floating-point instructions
handle denormal source operands
The contents of this register can be loaded from memory with the LDMXCSR and
FXRSTOR instructions and stored in memory with STMXCSR and FXSAVE.
Bits 16 through 31 of the MXCSR register are reserved and are cleared on a power-
up or reset of the processor; attempting to write a non-zero value to these bits, using
either the FXRSTOR or LDMXCSR instructions, will result in a general-protection
exception (#GP) being generated.

Vol. 1 10-5
PROGRAMMING WITH STREAMING SIMD EXTENSIONS (SSE)

31 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0

F R P U O Z D I D P U O Z D I
Reserved A
Z C M M M M M M E E E E E E
Z

Flush to Zero
Rounding Control
Precision Mask
Underflow Mask
Overflow Mask
Divide-by-Zero Mask
Denormal Operation Mask
Invalid Operation Mask
Denormals Are Zeros*
Precision Flag
Underflow Flag
Overflow Flag
Divide-by-Zero Flag
Denormal Flag
Invalid Operation Flag
* The denormals-are-zeros flag was introduced in the Pentium 4 and Intel Xeon processor.

Figure 10-3. MXCSR Control/Status Register

10.2.3.1 SIMD Floating-Point Mask and Flag Bits

Bits 0 through 5 of the MXCSR register indicate whether a SIMD floating-point excep-
tion has been detected. They are “sticky” flags. That is, after a flag is set, it remains
set until explicitly cleared. To clear these flags, use the LDMXCSR or the FXRSTOR
instruction to write zeroes to them.
Bits 7 through 12 provide individual mask bits for the SIMD floating-point exceptions.
An exception type is masked if the corresponding mask bit is set, and it is unmasked
if the bit is clear. These mask bits are set upon a power-up or reset. This causes all
SIMD floating-point exceptions to be initially masked.
If LDMXCSR or FXRSTOR clears a mask bit and sets the corresponding exception flag
bit, a SIMD floating-point exception will not be generated as a result of this change.
The unmasked exception will be generated only upon the execution of the next
SSE/SSE2/SSE3 instruction that detects the unmasked exception condition.
For more information about the use of the SIMD floating-point exception mask and
flag bits, see Section 11.5, “SSE, SSE2, and SSE3 Exceptions,” and Section 12.8,
“SSE3/SSSE3 and SSE4 Exceptions.”

10-6 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS (SSE)

10.2.3.2 SIMD Floating-Point Rounding Control Field

Bits 13 and 14 of the MXCSR register (the rounding control [RC] field) control how the
results of SIMD floating-point instructions are rounded. See Section 4.8.4,
“Rounding,” for a description of the function and encoding of the rounding control bits.

10.2.3.3 Flush-To-Zero
Bit 15 (FZ) of the MXCSR register enables the flush-to-zero mode, which controls the
masked response to a SIMD floating-point underflow condition. When the underflow
exception is masked and the flush-to-zero mode is enabled, the processor performs
the following operations when it detects a floating-point underflow condition:
• Returns a zero result with the sign of the true result
• Sets the precision and underflow exception flags
If the underflow exception is not masked, the flush-to-zero bit is ignored.
The flush-to-zero mode is not compatible with IEEE Standard 754. The IEEE-
mandated masked response to underflow is to deliver the denormalized result (see
Section 4.8.3.2, “Normalized and Denormalized Finite Numbers”). The flush-to-zero
mode is provided primarily for performance reasons. At the cost of a slight precision
loss, faster execution can be achieved for applications where underflows are common
and rounding the underflow result to zero can be tolerated.
The flush-to-zero bit is cleared upon a power-up or reset of the processor, disabling
the flush-to-zero mode.

10.2.3.4 Denormals-Are-Zeros
Bit 6 (DAZ) of the MXCSR register enables the denormals-are-zeros mode, which
controls the processor’s response to a SIMD floating-point denormal operand condi-
tion. When the denormals-are-zeros flag is set, the processor converts all denormal
source operands to a zero with the sign of the original operand before performing any
computations on them. The processor does not set the denormal-operand exception
flag (DE), regardless of the setting of the denormal-operand exception mask bit
(DM); and it does not generate a denormal-operand exception if the exception is
unmasked.
The denormals-are-zeros mode is not compatible with IEEE Standard 754 (see
Section 4.8.3.2, “Normalized and Denormalized Finite Numbers”). The denormals-
are-zeros mode is provided to improve processor performance for applications such
as streaming media processing, where rounding a denormal operand to zero does
not appreciably affect the quality of the processed data.
The denormals-are-zeros flag is cleared upon a power-up or reset of the processor,
disabling the denormals-are-zeros mode.
The denormals-are-zeros mode was introduced in the Pentium 4 and Intel Xeon
processor with the SSE2 extensions; however, it is fully compatible with the SSE

Vol. 1 10-7
PROGRAMMING WITH STREAMING SIMD EXTENSIONS (SSE)

SIMD floating-point instructions (that is, the denormals-are-zeros flag affects the
operation of the SSE SIMD floating-point instructions). In earlier IA-32 processors
and in some models of the Pentium 4 processor, this flag (bit 6) is reserved. See
Section 11.6.3, “Checking for the DAZ Flag in the MXCSR Register,” for instructions
for detecting the availability of this feature.
Attempting to set bit 6 of the MXCSR register on processors that do not support the
DAZ flag will cause a general-protection exception (#GP). See Section 11.6.6,
“Guidelines for Writing to the MXCSR Register,” for instructions for preventing such
general-protection exceptions by using the MXCSR_MASK value returned by the
FXSAVE instruction.

10.2.4 Compatibility of SSE Extensions with SSE2/SSE3/MMX and

the x87 FPU
The state (XMM registers and MXCSR register) introduced into the IA-32 execution
environment with the SSE extensions is shared with SSE2 and SSE3 extensions.
SSE/SSE2/SSE3 instructions are fully compatible; they can be executed together in
the same instruction stream with no need to save state when switching between
instruction sets.
XMM registers are independent of the x87 FPU and MMX registers, so
SSE/SSE2/SSE3 operations performed on the XMM registers can be performed in
parallel with operations on the x87 FPU and MMX registers (see Section 11.6.7,
“Interaction of SSE/SSE2 Instructions with x87 FPU and MMX Instructions”).
The FXSAVE and FXRSTOR instructions save and restore the SSE/SSE2/SSE3 states
along with the x87 FPU and MMX state.

10.3 SSE DATA TYPES

SSE extensions introduced one data type, the 128-bit packed single-precision
floating-point data type, to the IA-32 architecture (see Figure 10-4). This data type
consists of four IEEE 32-bit single-precision floating-point values packed into a
double quadword. (See Figure 4-3 for the layout of a single-precision floating-point
value; refer to Section 4.2.2, “Floating-Point Data Types,” for a detailed description of
the single-precision floating-point format.)

Contains 4 Single-Precision
Floating-Point Values
127 96 95 64 63 32 31 0

Figure 10-4. 128-Bit Packed Single-Precision Floating-Point Data Type

10-8 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS (SSE)

This 128-bit packed single-precision floating-point data type is operated on in the

XMM registers or in memory. Conversion instructions are provided to convert two
packed single-precision floating-point values into two packed doubleword integers or
a scalar single-precision floating-point value into a doubleword integer (see
Figure 11-8).
SSE extensions provide conversion instructions between XMM registers and MMX
registers, and between XMM registers and general-purpose bit registers. See
Figure 11-8.
The address of a 128-bit packed memory operand must be aligned on a 16-byte
boundary, except in the following cases:
• The MOVUPS instruction supports unaligned accesses.
• Scalar instructions that use a 4-byte memory operand that is not subject to
alignment requirements.
Figure 4-2 shows the byte order of 128-bit (double quadword) data types in memory.

10.4 SSE INSTRUCTION SET

SSE instructions are divided into four functional groups
• Packed and scalar single-precision floating-point instructions
• 64-bit SIMD integer instructions
• State management instructions
• Cacheability control, prefetch, and memory ordering instructions
The following sections give an overview of each of the instructions in these groups.

10.4.1 SSE Packed and Scalar Floating-Point Instructions

The packed and scalar single-precision floating-point instructions are divided into the
following subgroups:
• Data movement instructions
• Arithmetic instructions
• Logical instructions
• Comparison instructions
• Shuffle instructions
• Conversion instructions
The packed single-precision floating-point instructions perform SIMD operations on
packed single-precision floating-point operands (see Figure 10-5). Each source
operand contains four single-precision floating-point values, and the destination

Vol. 1 10-9
PROGRAMMING WITH STREAMING SIMD EXTENSIONS (SSE)

operand contains the results of the operation (OP) performed in parallel on the corre-
sponding values (X0 and Y0, X1 and Y1, X2 and Y2, and X3 and Y3) in each operand.

X3 X2 X1 X0

Y3 Y2 Y1 Y0

OP OP OP OP

X3 OP Y3 X2 OP Y2 X1 OP Y1 X0 OP Y0

Figure 10-5. Packed Single-Precision Floating-Point Operation

The scalar single-precision floating-point instructions operate on the low (least

significant) doublewords of the two source operands (X0 and Y0); see Figure 10-6.
The three most significant doublewords (X1, X2, and X3) of the first source operand
are passed through to the destination. The scalar operations are similar to the
floating-point operations performed in the x87 FPU data registers with the precision
control field in the x87 FPU control word set for single precision (24-bit significand),
except that x87 stack operations use a 15-bit exponent range for the result, while
SSE operations use an 8-bit exponent range.

X3 X2 X1 X0

Y3 Y2 Y1 Y0

OP

X3 X2 X1 X0 OP Y0

Figure 10-6. Scalar Single-Precision Floating-Point Operation

10-10 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS (SSE)

10.4.1.1 SSE Data Movement Instructions

SSE data movement instructions move single-precision floating-point data between
XMM registers and between an XMM register and memory.
The MOVAPS (move aligned packed single-precision floating-point values) instruction
transfers a double quadword operand containing four packed single-precision
floating-point values from memory to an XMM register and vice versa, or between
XMM registers. The memory address must be aligned to a 16-byte boundary; other-
wise, a general-protection exception (#GP) is generated.
The MOVUPS (move unaligned packed single-precision, floating-point) instruction
performs the same operations as the MOVAPS instruction, except that 16-byte align-
ment of a memory address is not required.
The MOVSS (move scalar single-precision floating-point) instruction transfers a 32-
bit single-precision floating-point operand from memory to the low doubleword of an
XMM register and vice versa, or between XMM registers.
The MOVLPS (move low packed single-precision floating-point) instruction moves
two packed single-precision floating-point values from memory to the low quadword
of an XMM register and vice versa. The high quadword of the register is left
unchanged.
The MOVHPS (move high packed single-precision floating-point) instruction moves
two packed single-precision floating-point values from memory to the high quadword
of an XMM register and vice versa. The low quadword of the register is left
unchanged.
The MOVLHPS (move packed single-precision floating-point low to high) instruction
moves two packed single-precision floating-point values from the low quadword of
the source XMM register into the high quadword of the destination XMM register. The
low quadword of the destination register is left unchanged.
The MOVHLPS (move packed single-precision floating-point high to low) instruction
moves two packed single-precision floating-point values from the high quadword of
the source XMM register into the low quadword of the destination XMM register. The
high quadword of the destination register is left unchanged.
The MOVMSKPS (move packed single-precision floating-point mask) instruction
transfers the most significant bit of each of the four packed single-precision floating-
point numbers in an XMM register to a general-purpose register. This 4-bit value can
then be used as a condition to perform branching.

10.4.1.2 SSE Arithmetic Instructions

SSE arithmetic instructions perform addition, subtraction, multiply, divide, recip-
rocal, square root, reciprocal of square root, and maximum/minimum operations on
packed and scalar single-precision floating-point values.

Vol. 1 10-11
PROGRAMMING WITH STREAMING SIMD EXTENSIONS (SSE)

The ADDPS (add packed single-precision floating-point values) and SUBPS (subtract
packed single-precision floating-point values) instructions add and subtract, respec-
tively, two packed single-precision floating-point operands.
The ADDSS (add scalar single-precision floating-point values) and SUBSS (subtract
scalar single-precision floating-point values) instructions add and subtract, respec-
tively, the low single-precision floating-point values of two operands and store the
result in the low doubleword of the destination operand.
The MULPS (multiply packed single-precision floating-point values) instruction multi-
plies two packed single-precision floating-point operands.
The MULSS (multiply scalar single-precision floating-point values) instruction multi-
plies the low single-precision floating-point values of two operands and stores the
result in the low doubleword of the destination operand.
The DIVPS (divide packed, single-precision floating-point values) instruction divides
two packed single-precision floating-point operands.
The DIVSS (divide scalar single-precision floating-point values) instruction divides
the low single-precision floating-point values of two operands and stores the result in
the low doubleword of the destination operand.
The RCPPS (compute reciprocals of packed single-precision floating-point values)
instruction computes the approximate reciprocals of values in a packed single-preci-
sion floating-point operand.
The RCPSS (compute reciprocal of scalar single-precision floating-point values)
instruction computes the approximate reciprocal of the low single-precision floating-
point value in the source operand and stores the result in the low doubleword of the
destination operand.
The SQRTPS (compute square roots of packed single-precision floating-point values)
instruction computes the square roots of the values in a packed single-precision
floating-point operand.
The SQRTSS (compute square root of scalar single-precision floating-point values)
instruction computes the square root of the low single-precision floating-point value
in the source operand and stores the result in the low doubleword of the destination
operand.
The RSQRTPS (compute reciprocals of square roots of packed single-precision
floating-point values) instruction computes the approximate reciprocals of the
square roots of the values in a packed single-precision floating-point operand.
The RSQRTSS (reciprocal of square root of scalar single-precision floating-point
value) instruction computes the approximate reciprocal of the square root of the low
single-precision floating-point value in the source operand and stores the result in
the low doubleword of the destination operand.
The MAXPS (return maximum of packed single-precision floating-point values)
instruction compares the corresponding values from two packed single-precision
floating-point operands and returns the numerically greater value from each compar-
ison to the destination operand.

10-12 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS (SSE)

The MAXSS (return maximum of scalar single-precision floating-point values)

instruction compares the low values from two packed single-precision floating-point
operands and returns the numerically greater value from the comparison to the low
doubleword of the destination operand.
The MINPS (return minimum of packed single-precision floating-point values)
instruction compares the corresponding values from two packed single-precision
floating-point operands and returns the numerically lesser value from each compar-
ison to the destination operand.
The MINSS (return minimum of scalar single-precision floating-point values) instruc-
tion compares the low values from two packed single-precision floating-point oper-
ands and returns the numerically lesser value from the comparison to the low
doubleword of the destination operand.

10.4.2 SSE Logical Instructions

SSE logical instructions perform AND, AND NOT, OR, and XOR operations on packed
single-precision floating-point values.
The ANDPS (bitwise logical AND of packed single-precision floating-point values)
instruction returns the logical AND of two packed single-precision floating-point
operands.
The ANDNPS (bitwise logical AND NOT of packed single-precision, floating-point
values) instruction returns the logical AND NOT of two packed single-precision
floating-point operands.
The ORPS (bitwise logical OR of packed single-precision, floating-point values)
instruction returns the logical OR of two packed single-precision floating-point oper-
ands.
The XORPS (bitwise logical XOR of packed single-precision, floating-point values)
instruction returns the logical XOR of two packed single-precision floating-point oper-
ands.

10.4.2.1 SSE Comparison Instructions

The compare instructions compare packed and scalar single-precision floating-point
values and return the results of the comparison either to the destination operand or
to the EFLAGS register.
The CMPPS (compare packed single-precision floating-point values) instruction
compares the corresponding values from two packed single-precision floating-point
operands, using an immediate operand as a predicate, and returns a 32-bit mask
result of all 1s or all 0s for each comparison to the destination operand. The value of
the immediate operand allows the selection of any of 8 compare conditions: equal,
less than, less than equal, unordered, not equal, not less than, not less than or equal,
or ordered.

Vol. 1 10-13
PROGRAMMING WITH STREAMING SIMD EXTENSIONS (SSE)

The CMPSS (compare scalar single-precision, floating-point values) instruction

compares the low values from two packed single-precision floating-point operands,
using an immediate operand as a predicate, and returns a 32-bit mask result of all 1s
or all 0s for the comparison to the low doubleword of the destination operand. The
immediate operand selects the compare conditions as with the CMPPS instruction.
The COMISS (compare scalar single-precision floating-point values and set EFLAGS)
and UCOMISS (unordered compare scalar single-precision floating-point values and
set EFLAGS) instructions compare the low values of two packed single-precision
floating-point operands and set the ZF, PF, and CF flags in the EFLAGS register to
show the result (greater than, less than, equal, or unordered). These two instruc-
tions differ as follows: the COMISS instruction signals a floating-point invalid-opera-
tion (#I) exception when a source operand is either a QNaN or an SNaN; the
UCOMISS instruction only signals an invalid-operation exception when a source
operand is an SNaN.

10.4.2.2 SSE Shuffle and Unpack Instructions

SSE shuffle and unpack instructions shuffle or interleave the contents of two packed
single-precision floating-point values and store the results in the destination
operand.
The SHUFPS (shuffle packed single-precision floating-point values) instruction places
any two of the four packed single-precision floating-point values from the destination
operand into the two low-order doublewords of the destination operand, and places
any two of the four packed single-precision floating-point values from the source
operand in the two high-order doublewords of the destination operand (see
Figure 10-7). By using the same register for the source and destination operands,
the SHUFPS instruction can shuffle four single-precision floating-point values into
any order.

DEST X3 X2 X1 X0

SRC Y3 Y2 Y1 Y0

DEST Y3 ... Y0 Y3 ... Y0 X3 ... X0 X3 ... X0

Figure 10-7. SHUFPS Instruction, Packed Shuffle Operation

10-14 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS (SSE)

The UNPCKHPS (unpack and interleave high packed single-precision floating-point

values) instruction performs an interleaved unpack of the high-order single-precision
floating-point values from the source and destination operands and stores the result
in the destination operand (see Figure 10-8).

DEST X3 X2 X1 X0

SRC Y3 Y2 Y1 Y0

DEST Y3 X3 Y2 X2

Figure 10-8. UNPCKHPS Instruction, High Unpack and Interleave Operation

The UNPCKLPS (unpack and interleave low packed single-precision floating-point

values) instruction performs an interleaved unpack of the low-order single-precision
floating-point values from the source and destination operands and stores the result
in the destination operand (see Figure 10-9).

DEST X3 X2 X1 X0

SRC Y3 Y2 Y1 Y0

DEST Y1 X1 Y0 X0

Figure 10-9. UNPCKLPS Instruction, Low Unpack and Interleave Operation

10.4.3 SSE Conversion Instructions

SSE conversion instructions (see Figure 11-8) support packed and scalar conversions
between single-precision floating-point and doubleword integer formats.

Vol. 1 10-15
PROGRAMMING WITH STREAMING SIMD EXTENSIONS (SSE)

The CVTPI2PS (convert packed doubleword integers to packed single-precision

floating-point values) instruction converts two packed signed doubleword integers
into two packed single-precision floating-point values. When the conversion is
inexact, the result is rounded according to the rounding mode selected in the MXCSR
register.
The CVTSI2SS (convert doubleword integer to scalar single-precision floating-point
value) instruction converts a signed doubleword integer into a single-precision
floating-point value. When the conversion is inexact, the result is rounded according
to the rounding mode selected in the MXCSR register.
The CVTPS2PI (convert packed single-precision floating-point values to packed
doubleword integers) instruction converts two packed single-precision floating-point
values into two packed signed doubleword integers. When the conversion is inexact,
the result is rounded according to the rounding mode selected in the MXCSR register.
The CVTTPS2PI (convert with truncation packed single-precision floating-point
values to packed doubleword integers) instruction is similar to the CVTPS2PI instruc-
tion, except that truncation is used to round a source value to an integer value (see
Section 4.8.4.2, “Truncation with SSE and SSE2 Conversion Instructions”).
The CVTSS2SI (convert scalar single-precision floating-point value to doubleword
integer) instruction converts a single-precision floating-point value into a signed
doubleword integer. When the conversion is inexact, the result is rounded according
to the rounding mode selected in the MXCSR register. The CVTTSS2SI (convert with
truncation scalar single-precision floating-point value to doubleword integer) instruc-
tion is similar to the CVTSS2SI instruction, except that truncation is used to round
the source value to an integer value (see Section 4.8.4.2, “Truncation with SSE and
SSE2 Conversion Instructions”).

10.4.4 SSE 64-Bit SIMD Integer Instructions

SSE extensions add the following 64-bit packed integer instructions to the IA-32
architecture. These instructions operate on data in MMX registers and 64-bit memory
locations.

NOTE
When SSE2 extensions are present in an IA-32 processor, these
instructions are extended to operate on 128-bit operands in XMM
registers and 128-bit memory locations.

The PAVGB (compute average of packed unsigned byte integers) and PAVGW
(compute average of packed unsigned word integers) instructions compute a SIMD
average of two packed unsigned byte or word integer operands, respectively. For
each corresponding pair of data elements in the packed source operands, the
elements are added together, a 1 is added to the temporary sum, and that result is
shifted right one bit position.

10-16 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS (SSE)

The PEXTRW (extract word) instruction copies a selected word from an MMX register
into a general-purpose register.
The PINSRW (insert word) instruction copies a word from a general-purpose register
or from memory into a selected word location in an MMX register.
The PMAXUB (maximum of packed unsigned byte integers) instruction compares the
corresponding unsigned byte integers in two packed operands and returns the
greater of each comparison to the destination operand.
The PMINUB (minimum of packed unsigned byte integers) instruction compares the
corresponding unsigned byte integers in two packed operands and returns the lesser
of each comparison to the destination operand.
The PMAXSW (maximum of packed signed word integers) instruction compares the
corresponding signed word integers in two packed operands and returns the greater
of each comparison to the destination operand.
The PMINSW (minimum of packed signed word integers) instruction compares the
corresponding signed word integers in two packed operands and returns the lesser of
each comparison to the destination operand.
The PMOVMSKB (move byte mask) instruction creates an 8-bit mask from the packed
byte integers in an MMX register and stores the result in the low byte of a general-
purpose register. The mask contains the most significant bit of each byte in the MMX
register. (When operating on 128-bit operands, a 16-bit mask is created.)
The PMULHUW (multiply packed unsigned word integers and store high result)
instruction performs a SIMD unsigned multiply of the words in the two source oper-
ands and returns the high word of each result to an MMX register.
The PSADBW (compute sum of absolute differences) instruction computes the SIMD
absolute differences of the corresponding unsigned byte integers in two source oper-
ands, sums the differences, and stores the sum in the low word of the destination
operand.
The PSHUFW (shuffle packed word integers) instruction shuffles the words in the
source operand according to the order specified by an 8-bit immediate operand and
returns the result to the destination operand.

10.4.5 MXCSR State Management Instructions

The MXCSR state management instructions (LDMXCSR and STMXCSR) load and save
the state of the MXCSR register, respectively. The LDMXCSR instruction loads the
MXCSR register from memory, while the STMXCSR instruction stores the contents of
the register to memory.

Vol. 1 10-17
PROGRAMMING WITH STREAMING SIMD EXTENSIONS (SSE)

10.4.6 Cacheability Control, Prefetch, and Memory Ordering

Instructions
SSE extensions introduce several new instructions to give programs more control
over the caching of data. They also introduces the PREFETCHh instructions, which
provide the ability to prefetch data to a specified cache level, and the SFENCE
instruction, which enforces program ordering on stores. These instructions are
described in the following sections.

10.4.6.1 Cacheability Control Instructions

The following three instructions enable data from the MMX and XMM registers to be
stored to memory using a non-temporal hint. The non-temporal hint directs the
processor to when possible store the data to memory without writing the data into
the cache hierarchy. See Section 10.4.6.2, “Caching of Temporal vs. Non-Temporal
Data,” for information about non-temporal stores and hints.
The MOVNTQ (store quadword using non-temporal hint) instruction stores packed
integer data from an MMX register to memory, using a non-temporal hint.
The MOVNTPS (store packed single-precision floating-point values using non-
temporal hint) instruction stores packed floating-point data from an XMM register to
memory, using a non-temporal hint.
The MASKMOVQ (store selected bytes of quadword) instruction stores selected byte
integers from an MMX register to memory, using a byte mask to selectively write the
individual bytes. This instruction also uses a non-temporal hint.

10.4.6.2 Caching of Temporal vs. Non-Temporal Data

Data referenced by a program can be temporal (data will be used again) or non-
temporal (data will be referenced once and not reused in the immediate future). For
example, program code is generally temporal, whereas, multimedia data, such as the
display list in a 3-D graphics application, is often non-temporal. To make efficient use
of the processor’s caches, it is generally desirable to cache temporal data and not
cache non-temporal data. Overloading the processor’s caches with non-temporal
data is sometimes referred to as “polluting the caches.” The SSE and SSE2 cache-
ability control instructions enable a program to write non-temporal data to memory
in a manner that minimizes pollution of caches.
These SSE and SSE2 non-temporal store instructions minimize cache pollutions by
treating the memory being accessed as the write combining (WC) type. If a program
specifies a non-temporal store with one of these instructions and the destination
region is mapped as cacheable memory (write back [WB], write through [WT] or WC
memory type), the processor will do the following:
• If the memory location being written to is present in the cache hierarchy, the data
in the caches is evicted.
• The non-temporal data is written to memory with WC semantics.

10-18 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS (SSE)

See also: Chapter 10, “Memory Cache Control,” in the Intel® 64 and IA-32 Architec-
tures Software Developer’s Manual, Volume 3A.
Using the WC semantics, the store transaction will be weakly ordered, meaning that
the data may not be written to memory in program order, and the store will not write
allocate (that is, the processor will not fetch the corresponding cache line into the
cache hierarchy, prior to performing the store). Also, different processor implemen-
tations may choose to collapse and combine these stores.
The memory type of the region being written to can override the non-temporal hint,
if the memory address specified for the non-temporal store is in uncacheable
memory. Uncacheable as referred to here means that the region being written to has
been mapped with either an uncacheable (UC) or write protected (WP) memory type.
In general, WC semantics require software to ensure coherence, with respect to
other processors and other system agents (such as graphics cards). Appropriate use
of synchronization and fencing must be performed for producer-consumer usage
models. Fencing ensures that all system agents have global visibility of the stored
data; for instance, failure to fence may result in a written cache line staying within a
processor and not being visible to other agents.
For processors that implement non-temporal stores by updating data in-place that
already resides in the cache hierarchy, the destination region should also be mapped
as WC. If mapped as WB or WT, there is the potential for speculative processor reads
to bring the data into the caches; in this case, non-temporal stores would then
update in place, and data would not be flushed from the processor by a subsequent
fencing operation.
The memory type visible on the bus in the presence of memory type aliasing is imple-
mentation specific. As one possible example, the memory type written to the bus
may reflect the memory type for the first store to this line, as seen in program order;
other alternatives are possible. This behavior should be considered reserved, and
dependence on the behavior of any particular implementation risks future incompat-
ibility.

10.4.6.3 PREFETCHh Instructions

The PREFETCHh instructions permit programs to load data into the processor at a
suggested cache level, so that the data is closer to the processor’s load and store unit
when it is needed. These instructions fetch 32 aligned bytes (or more, depending on
the implementation) containing the addressed byte to a location in the cache hier-
archy specified by the temporal locality hint (see Table 10-1). In this table, the first-
level cache is closest to the processor and second-level cache is farther away from
the processor than the first-level cache. The hints specify a prefetch of either
temporal or non-temporal data (see Section 10.4.6.2, “Caching of Temporal vs. Non-
Temporal Data”). Subsequent accesses to temporal data are treated like normal
accesses, while those to non-temporal data will continue to minimize cache pollution.
If the data is already present at a level of the cache hierarchy that is closer to the
processor, the PREFETCHh instruction will not result in any data movement. The
PREFETCHh instructions do not affect functional behavior of the program.

Vol. 1 10-19
PROGRAMMING WITH STREAMING SIMD EXTENSIONS (SSE)

See Section 11.6.13, “Cacheability Hint Instructions,” for additional information

about the PREFETCHh instructions.

Table 10-1. PREFETCHh Instructions Caching Hints

PREFETCHh
Instruction Mnemonic Actions
PREFETCHT0 Temporal data—fetch data into all levels of cache hierarchy:
• Pentium III processor—1st-level cache or 2nd-level cache
• Pentium 4 and Intel Xeon processor—2nd-level cache
PREFETCHT1 Temporal data—fetch data into level 2 cache and higher
• Pentium III processor—2nd-level cache
• Pentium 4 and Intel Xeon processor—2nd-level cache
PREFETCHT2 Temporal data—fetch data into level 2 cache and higher
• Pentium III processor—2nd-level cache
• Pentium 4 and Intel Xeon processor—2nd-level cache
PREFETCHNTA Non-temporal data—fetch data into location close to the processor,
minimizing cache pollution
• Pentium III processor—1st-level cache
• Pentium 4 and Intel Xeon processor—2nd-level cache

10.4.6.4 SFENCE Instruction

The SFENCE (Store Fence) instruction controls write ordering by creating a fence for
memory store operations. This instruction guarantees that the result of every store
instruction that precedes the store fence in program order is globally visible before
any store instruction that follows the fence. The SFENCE instruction provides an effi-
cient way of ensuring ordering between procedures that produce weakly-ordered
data and procedures that consume that data.

10.5 FXSAVE AND FXRSTOR INSTRUCTIONS

The FXSAVE and FXRSTOR instructions were introduced into the IA-32 architecture in
the Pentium II processor family (prior to the introduction of the SSE extensions). The
original versions of these instructions performed a fast save and restore, respec-
tively, of the x87 FPU register state. (By saving the state of the x87 FPU data regis-
ters, the FXSAVE and FXRSTOR instructions implicitly save and restore the state of
the MMX registers.)
The SSE extensions expanded the scope of these instructions to save and restore the
states of the XMM registers and the MXCSR register, along with the x87 FPU and MMX
state.

10-20 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS (SSE)

The FXSAVE and FXRSTOR instructions can be used in place of the FSAVE/FNSAVE
and FRSTOR instructions; however, the operation of the FXSAVE and FXRSTOR
instructions are not identical to the operation of FSAVE/FNSAVE and FRSTOR.

NOTE
The FXSAVE and FXRSTOR instructions are not considered part
of the SSE instruction group. They have a separate CPUID
feature bit to indicate whether they are present (if
CPUID.01H:EDX.FXSR[bit 24] = 1).

The CPUID feature bit for SSE extensions does not indicate the
presence of FXSAVE and FXRSTOR.

10.6 HANDLING SSE INSTRUCTION EXCEPTIONS

See Section 11.5, “SSE, SSE2, and SSE3 Exceptions,” for a detailed discussion of the
general and SIMD floating-point exceptions that can be generated with the SSE
instructions and for guidelines for handling these exceptions when they occur.

10.7 WRITING APPLICATIONS WITH THE SSE EXTENSIONS

See Section 11.6, “Writing Applications with SSE/SSE2 Extensions,” for additional
information about writing applications and operating-system code using the SSE
extensions.

Vol. 1 10-21
PROGRAMMING WITH STREAMING SIMD EXTENSIONS (SSE)

10-22 Vol. 1
 CHAPTER 11
PROGRAMMING WITH
STREAMING SIMD EXTENSIONS 2 (SSE2)

The streaming SIMD extensions 2 (SSE2) were introduced into the IA-32 architecture
in the Pentium 4 and Intel Xeon processors. These extensions enhance the perfor-
mance of IA-32 processors for advanced 3-D graphics, video decoding/encoding,
speech recognition, E-commerce, Internet, scientific, and engineering applications.
This chapter describes the SSE2 extensions and provides information to assist in
writing application programs that use these and the SSE extensions.

11.1 OVERVIEW OF SSE2 EXTENSIONS

SSE2 extensions use the single instruction multiple data (SIMD) execution model
that is used with MMX technology and SSE extensions. They extend this model with
support for packed double-precision floating-point values and for 128-bit packed
integers.
If CPUID.01H:EDX.SSE2[bit 26] = 1, SSE2 extensions are present.
SSE2 extensions add the following features to the IA-32 architecture, while main-
taining backward compatibility with all existing IA-32 processors, applications and
operating systems.
• Six data types:
— 128-bit packed double-precision floating-point (two IEEE Standard 754
double-precision floating-point values packed into a double quadword)
— 128-bit packed byte integers
— 128-bit packed word integers
— 128-bit packed doubleword integers
— 128-bit packed quadword integers
• Instructions to support the additional data types and extend existing SIMD
integer operations:
— Packed and scalar double-precision floating-point instructions
— Additional 64-bit and 128-bit SIMD integer instructions
— 128-bit versions of SIMD integer instructions introduced with the MMX
technology and the SSE extensions
— Additional cacheability-control and instruction-ordering instructions
• Modifications to existing IA-32 instructions to support SSE2 features:
— Extensions and modifications to the CPUID instruction
— Modifications to the RDPMC instruction

Vol. 1 11-1
PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

These new features extend the IA-32 architecture’s SIMD programming model in
three important ways:
• They provide the ability to perform SIMD operations on pairs of packed double-
precision floating-point values. This permits higher precision computations to be
carried out in XMM registers, which enhances processor performance in scientific
and engineering applications and in applications that use advanced 3-D geometry
techniques (such as ray tracing). Additional flexibility is provided with instruc-
tions that operate on single (scalar) double-precision floating-point values
located in the low quadword of an XMM register.
• They provide the ability to operate on 128-bit packed integers (bytes, words,
doublewords, and quadwords) in XMM registers. This provides greater flexibility
and greater throughput when performing SIMD operations on packed integers.
The capability is particularly useful for applications such as RSA authentication
and RC5 encryption. Using the full set of SIMD registers, data types, and instruc-
tions provided with the MMX technology and SSE/SSE2 extensions, programmers
can develop algorithms that finely mix packed single- and double-precision
floating-point data and 64- and 128-bit packed integer data.
• SSE2 extensions enhance the support introduced with SSE extensions for
controlling the cacheability of SIMD data. SSE2 cache control instructions provide
the ability to stream data in and out of the XMM registers without polluting the
caches and the ability to prefetch data before it is actually used.
SSE2 extensions are fully compatible with all software written for IA-32 processors.
All existing software continues to run correctly, without modification, on processors
that incorporate SSE2 extensions, as well as in the presence of applications that
incorporate these extensions. Enhancements to the CPUID instruction permit detec-
tion of the SSE2 extensions. Also, because the SSE2 extensions use the same regis-
ters as the SSE extensions, no new operating-system support is required for saving
and restoring program state during a context switch beyond that provided for the
SSE extensions.
SSE2 extensions are accessible from all IA-32 execution modes: protected mode,
real address mode, virtual 8086 mode.
The following sections in this chapter describe the programming environment for
SSE2 extensions including: the 128-bit XMM floating-point register set, data types,
and SSE2 instructions. It also describes exceptions that can be generated with the
SSE and SSE2 instructions and gives guidelines for writing applications with SSE and
SSE2 extensions.
For additional information about SSE2 extensions, see:
• Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volumes
2A & 2B, provide a detailed description of individual SSE3 instructions.
• Chapter 12, “System Programming for Streaming SIMD Instruction Sets,” in the
Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volume 3A,
gives guidelines for integrating the SSE and SSE2 extensions into an operating-
system environment.

11-2 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

11.2 SSE2 PROGRAMMING ENVIRONMENT

Figure 11-1 shows the programming environment for SSE2 extensions. No new
registers or other instruction execution state are defined with SSE2 extensions. SSE2
instructions use the XMM registers, the MMX registers, and/or IA-32 general-purpose
registers, as follows:
• XMM registers — These eight registers (see Figure 10-2) are used to operate on
packed or scalar double-precision floating-point data. Scalar operations are
operations performed on individual (unpacked) double-precision floating-point
values stored in the low quadword of an XMM register. XMM registers are also
used to perform operations on 128-bit packed integer data. They are referenced
by the names XMM0 through XMM7.

Address Space

XMM Registers 232 -1

Eight 128-Bit

MXCSR Register 32 Bits

MMX Registers
Eight 64-Bit

General-Purpose
Registers
Eight 32-Bit

0
EFLAGS Register 32 Bits

Figure 11-1. Steaming SIMD Extensions 2 Execution Environment

• MXCSR register — This 32-bit register (see Figure 10-3) provides status and
control bits used in floating-point operations. The denormals-are-zeros and
flush-to-zero flags in this register provide a higher performance alternative for
the handling of denormal source operands and denormal (underflow) results. For
more information on the functions of these flags see Section 10.2.3.4,
“Denormals-Are-Zeros,” and Section 10.2.3.3, “Flush-To-Zero.”
• MMX registers — These eight registers (see Figure 9-2) are used to perform
operations on 64-bit packed integer data. They are also used to hold operands for
some operations performed between MMX and XMM registers. MMX registers are
referenced by the names MM0 through MM7.

Vol. 1 11-3
PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

• General-purpose registers — The eight general-purpose registers (see

Figure 3-5) are used along with the existing IA-32 addressing modes to address
operands in memory. MMX and XMM registers cannot be used to address
memory. The general-purpose registers are also used to hold operands for some
SSE2 instructions. These registers are referenced by the names EAX, EBX, ECX,
EDX, EBP, ESI, EDI, and ESP.
• EFLAGS register — This 32-bit register (see Figure 3-8) is used to record the
results of some compare operations.

11.2.1 SSE2 in 64-Bit Mode and Compatibility Mode

In compatibility mode, SSE2 extensions function like they do in protected mode. In
64-bit mode, eight additional XMM registers are accessible. Registers XMM8-XMM15
are accessed by using REX prefixes.
Memory operands are specified using the ModR/M, SIB encoding described in Section
3.7.5.
Some SSE2 instructions may be used to operate on general-purpose registers. Use
the REX.W prefix to access 64-bit general-purpose registers. Note that if a REX prefix
is used when it has no meaning, the prefix is ignored.

11.2.2 Compatibility of SSE2 Extensions with SSE, MMX

Technology and x87 FPU Programming Environment
SSE2 extensions do not introduce any new state to the IA-32 execution environment
beyond that of SSE. SSE2 extensions represent an enhancement of SSE extensions;
they are fully compatible and share the same state information. SSE and SSE2
instructions can be executed together in the same instruction stream without the
need to save state when switching between instruction sets.
XMM registers are independent of the x87 FPU and MMX registers; so SSE and SSE2
operations performed on XMM registers can be performed in parallel with x87 FPU or
MMX technology operations (see Section 11.6.7, “Interaction of SSE/SSE2 Instruc-
tions with x87 FPU and MMX Instructions”).
The FXSAVE and FXRSTOR instructions save and restore the SSE and SSE2 states
along with the x87 FPU and MMX states.

11.2.3 Denormals-Are-Zeros Flag

The denormals-are-zeros flag (bit 6 in the MXCSR register) was introduced into the
IA-32 architecture with the SSE2 extensions. See Section 10.2.3.4, “Denormals-Are-
Zeros,” for a description of this flag.

11-4 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

11.3 SSE2 DATA TYPES

SSE2 extensions introduced one 128-bit packed floating-point data type and four
128-bit SIMD integer data types to the IA-32 architecture (see Figure 11-2).
• Packed double-precision floating-point — This 128-bit data type consists of
two IEEE 64-bit double-precision floating-point values packed into a double
quadword. (See Figure 4-3 for the layout of a 64-bit double-precision floating-
point value; refer to Section 4.2.2, “Floating-Point Data Types,” for a detailed
description of double-precision floating-point values.)
• 128-bit packed integers — The four 128-bit packed integer data types can
contain 16 byte integers, 8 word integers, 4 doubleword integers, or 2 quadword
integers. (Refer to Section 4.6.2, “128-Bit Packed SIMD Data Types,” for a
detailed description of the 128-bit packed integers.)

128-Bit Packed Double-

Precision Floating-Point
127 64 63 0

128-Bit Packed Byte Integers

127 0

128-Bit Packed Word Integers

127 0

128-Bit Packed Doubleword

Integers
127 0

128-Bit Packed Quadword

Integers
127 0

Figure 11-2. Data Types Introduced with the SSE2 Extensions

All of these data types are operated on in XMM registers or memory. Instructions are
provided to convert between these 128-bit data types and the 64-bit and 32-bit data
types.
The address of a 128-bit packed memory operand must be aligned on a 16-byte
boundary, except in the following cases:
• a MOVUPD instruction which supports unaligned accesses
• scalar instructions that use an 8-byte memory operand that is not subject to
alignment requirements
Figure 4-2 shows the byte order of 128-bit (double quadword) and 64-bit (quad-
word) data types in memory.

Vol. 1 11-5
PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

11.4 SSE2 INSTRUCTIONS

The SSE2 instructions are divided into four functional groups:
• Packed and scalar double-precision floating-point instructions
• 64-bit and 128-bit SIMD integer instructions
• 128-bit extensions of SIMD integer instructions introduced with the MMX
technology and the SSE extensions
• Cacheability-control and instruction-ordering instructions
The following sections provide more information about each group.

11.4.1 Packed and Scalar Double-Precision Floating-Point

Instructions
The packed and scalar double-precision floating-point instructions are divided into
the following sub-groups:
• Data movement instructions
• Arithmetic instructions
• Comparison instructions
• Conversion instructions
• Logical instructions
• Shuffle instructions
The packed double-precision floating-point instructions perform SIMD operations
similarly to the packed single-precision floating-point instructions (see Figure 11-3).
Each source operand contains two double-precision floating-point values, and the
destination operand contains the results of the operation (OP) performed in parallel
on the corresponding values (X0 and Y0, and X1 and Y1) in each operand.

X1 X0

Y1 Y0

OP OP

X1 OP Y1 X0 OP Y0

Figure 11-3. Packed Double-Precision Floating-Point Operations

11-6 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

The scalar double-precision floating-point instructions operate on the low (least

significant) quadwords of two source operands (X0 and Y0), as shown in Figure 11-4.
The high quadword (X1) of the first source operand is passed through to the destina-
tion. The scalar operations are similar to the floating-point operations performed in
x87 FPU data registers with the precision control field in the x87 FPU control word set
for double precision (53-bit significand), except that x87 stack operations use a
15-bit exponent range for the result while SSE2 operations use an 11-bit exponent
range.
See Section 11.6.8, “Compatibility of SIMD and x87 FPU Floating-Point Data Types,”
for more information about obtaining compatible results when performing both
scalar double-precision floating-point operations in XMM registers and in x87 FPU
data registers.

X1 X0

Y1 Y0

OP

X1 X0 OP Y0

Figure 11-4. Scalar Double-Precision Floating-Point Operations

11.4.1.1 Data Movement Instructions

Data movement instructions move double-precision floating-point data between
XMM registers and between XMM registers and memory.
The MOVAPD (move aligned packed double-precision floating-point) instruction
transfers a 128-bit packed double-precision floating-point operand from memory to
an XMM register or vice versa, or between XMM registers. The memory address must
be aligned to a 16-byte boundary; if not, a general-protection exception (GP#) is
generated.
The MOVUPD (move unaligned packed double-precision floating-point) instruction
transfers a 128-bit packed double-precision floating-point operand from memory to
an XMM register or vice versa, or between XMM registers. Alignment of the memory
address is not required.
The MOVSD (move scalar double-precision floating-point) instruction transfers a
64-bit double-precision floating-point operand from memory to the low quadword of

Vol. 1 11-7
PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

an XMM register or vice versa, or between XMM registers. Alignment of the memory
address is not required, unless alignment checking is enabled.
The MOVHPD (move high packed double-precision floating-point) instruction trans-
fers a 64-bit double-precision floating-point operand from memory to the high quad-
word of an XMM register or vice versa. The low quadword of the register is left
unchanged. Alignment of the memory address is not required, unless alignment
checking is enabled.
The MOVLPD (move low packed double-precision floating-point) instruction transfers
a 64-bit double-precision floating-point operand from memory to the low quadword
of an XMM register or vice versa. The high quadword of the register is left unchanged.
Alignment of the memory address is not required, unless alignment checking is
enabled.
The MOVMSKPD (move packed double-precision floating-point mask) instruction
extracts the sign bit of each of the two packed double-precision floating-point
numbers in an XMM register and saves them in a general-purpose register. This 2-bit
value can then be used as a condition to perform branching.

11.4.1.2 SSE2 Arithmetic Instructions

SSE2 arithmetic instructions perform addition, subtraction, multiply, divide, square
root, and maximum/minimum operations on packed and scalar double-precision
floating-point values.
The ADDPD (add packed double-precision floating-point values) and SUBPD
(subtract packed double-precision floating-point values) instructions add and
subtract, respectively, two packed double-precision floating-point operands.
The ADDSD (add scalar double-precision floating-point values) and SUBSD (subtract
scalar double-precision floating-point values) instructions add and subtract, respec-
tively, the low double-precision floating-point values of two operands and stores the
result in the low quadword of the destination operand.
The MULPD (multiply packed double-precision floating-point values) instruction
multiplies two packed double-precision floating-point operands.
The MULSD (multiply scalar double-precision floating-point values) instruction multi-
plies the low double-precision floating-point values of two operands and stores the
result in the low quadword of the destination operand.
The DIVPD (divide packed double-precision floating-point values) instruction divides
two packed double-precision floating-point operands.
The DIVSD (divide scalar double-precision floating-point values) instruction divides
the low double-precision floating-point values of two operands and stores the result
in the low quadword of the destination operand.
The SQRTPD (compute square roots of packed double-precision floating-point
values) instruction computes the square roots of the values in a packed double-preci-
sion floating-point operand.

11-8 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

The SQRTSD (compute square root of scalar double-precision floating-point values)

instruction computes the square root of the low double-precision floating-point value
in the source operand and stores the result in the low quadword of the destination
operand.
The MAXPD (return maximum of packed double-precision floating-point values)
instruction compares the corresponding values in two packed double-precision
floating-point operands and returns the numerically greater value from each compar-
ison to the destination operand.
The MAXSD (return maximum of scalar double-precision floating-point values)
instruction compares the low double-precision floating-point values from two packed
double-precision floating-point operands and returns the numerically higher value
from the comparison to the low quadword of the destination operand.
The MINPD (return minimum of packed double-precision floating-point values)
instruction compares the corresponding values from two packed double-precision
floating-point operands and returns the numerically lesser value from each compar-
ison to the destination operand.
The MINSD (return minimum of scalar double-precision floating-point values)
instruction compares the low values from two packed double-precision floating-point
operands and returns the numerically lesser value from the comparison to the low
quadword of the destination operand.

11.4.1.3 SSE2 Logical Instructions

SSE2 logical instructions perform AND, AND NOT, OR, and XOR operations on packed
double-precision floating-point values.
The ANDPD (bitwise logical AND of packed double-precision floating-point values)
instruction returns the logical AND of two packed double-precision floating-point
operands.
The ANDNPD (bitwise logical AND NOT of packed double-precision floating-point
values) instruction returns the logical AND NOT of two packed double-precision
floating-point operands.
The ORPD (bitwise logical OR of packed double-precision floating-point values)
instruction returns the logical OR of two packed double-precision floating-point oper-
ands.
The XORPD (bitwise logical XOR of packed double-precision floating-point values)
instruction returns the logical XOR of two packed double-precision floating-point
operands.

11.4.1.4 SSE2 Comparison Instructions

SSE2 compare instructions compare packed and scalar double-precision floating-
point values and return the results of the comparison either to the destination
operand or to the EFLAGS register.

Vol. 1 11-9
PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

The CMPPD (compare packed double-precision floating-point values) instruction

compares the corresponding values from two packed double-precision floating-point
operands, using an immediate operand as a predicate, and returns a 64-bit mask
result of all 1s or all 0s for each comparison to the destination operand. The value of
the immediate operand allows the selection of any of eight compare conditions:
equal, less than, less than equal, unordered, not equal, not less than, not less than
or equal, or ordered.
The CMPSD (compare scalar double-precision floating-point values) instruction
compares the low values from two packed double-precision floating-point operands,
using an immediate operand as a predicate, and returns a 64-bit mask result of all 1s
or all 0s for the comparison to the low quadword of the destination operand. The
immediate operand selects the compare condition as with the CMPPD instruction.
The COMISD (compare scalar double-precision floating-point values and set EFLAGS)
and UCOMISD (unordered compare scalar double-precision floating-point values and
set EFLAGS) instructions compare the low values of two packed double-precision
floating-point operands and set the ZF, PF, and CF flags in the EFLAGS register to
show the result (greater than, less than, equal, or unordered). These two instruc-
tions differ as follows: the COMISD instruction signals a floating-point invalid-opera-
tion (#I) exception when a source operand is either a QNaN or an SNaN; the
UCOMISD instruction only signals an invalid-operation exception when a source
operand is an SNaN.

11.4.1.5 SSE2 Shuffle and Unpack Instructions

SSE2 shuffle instructions shuffle the contents of two packed double-precision
floating-point values and store the results in the destination operand.
The SHUFPD (shuffle packed double-precision floating-point values) instruction
places either of the two packed double-precision floating-point values from the desti-
nation operand in the low quadword of the destination operand, and places either of
the two packed double-precision floating-point values from source operand in the
high quadword of the destination operand (see Figure 11-5). By using the same
register for the source and destination operands, the SHUFPD instruction can swap
two packed double-precision floating-point values.

11-10 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

DEST X1 X0

SRC Y1 Y0

DEST Y1 or Y0 X1 or X0

Figure 11-5. SHUFPD Instruction, Packed Shuffle Operation

The UNPCKHPD (unpack and interleave high packed double-precision floating-point

values) instruction performs an interleaved unpack of the high values from the
source and destination operands and stores the result in the destination operand
(see Figure 11-6).
The UNPCKLPD (unpack and interleave low packed double-precision floating-point
values) instruction performs an interleaved unpack of the low values from the source
and destination operands and stores the result in the destination operand (see
Figure 11-7).

DEST X1 X0

SRC Y1 Y0

DEST Y1 X1

Figure 11-6. UNPCKHPD Instruction, High Unpack and Interleave Operation

Vol. 1 11-11
PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

DEST X1 X0

SRC Y1 Y0

DEST Y0 X0

Figure 11-7. UNPCKLPD Instruction, Low Unpack and Interleave Operation

11.4.1.6 SSE2 Conversion Instructions

SSE2 conversion instructions (see Figure 11-8) support packed and scalar conver-
sions between:
• Double-precision and single-precision floating-point formats
• Double-precision floating-point and doubleword integer formats
• Single-precision floating-point and doubleword integer formats
Conversion between double-precision and single-precision floating-points
values — The following instructions convert operands between double-precision and
single-precision floating-point formats. The operands being operated on are
contained in XMM registers or memory (at most, one operand can reside in memory;
the destination is always an MMX register).
The CVTPS2PD (convert packed single-precision floating-point values to packed
double-precision floating-point values) instruction converts two packed single-
precision floating-point values to two double-precision floating-point values.
The CVTPD2PS (convert packed double-precision floating-point values to packed
single-precision floating-point values) instruction converts two packed double-
precision floating-point values to two single-precision floating-point values. When a
conversion is inexact, the result is rounded according to the rounding mode selected
in the MXCSR register.
The CVTSS2SD (convert scalar single-precision floating-point value to scalar double-
precision floating-point value) instruction converts a single-precision floating-point
value to a double-precision floating-point value.
The CVTSD2SS (convert scalar double-precision floating-point value to scalar single-
precision floating-point value) instruction converts a double-precision floating-point
value to a single-precision floating-point value. When the conversion is inexact, the
result is rounded according to the rounding mode selected in the MXCSR register.

11-12 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

Single-Precision
Floating Point
SI I (XMM/mem)
SS2 2S
T SS CV
CV TT CV TPS
TT 2D
CV S PI PS Q
2S S2 2PI 2D
SI

C
T P
T S Q

VT
CV CV TTP

D
Q
CV

2P
S
2P

S
PI
4 Doubleword

VT
Integer
C

CVTSD2SS
CVTPD2PS

CVTPS2PD
CVTSS2SD
(XMM/mem)
Doubleword 2 Doubleword
Integer Integer 2 Doubleword
(r32/mem) (MMX/mem) Integer
(XMM/mem)
C
VT

D
2P
PI

Q
2P

D
D

VT

D DQ
C TT
VT S
C

Q
C
C

TP D2
V

2D
CV VT
SD D2

VT P
TT PD

C VT
2S SI

PD 2P

C
I
C

2P I
VT

I
SI
2S

Double-Precision
D

Floating-Point
(XMM/mem)

Figure 11-8. SSE and SSE2 Conversion Instructions

Conversion between double-precision floating-point values and doubleword

integers — The following instructions convert operands between double-precision
floating-point and doubleword integer formats. Operands are housed in XMM regis-
ters, MMX registers, general registers or memory (at most one operand can reside in
memory; the destination is always an XMM, MMX, or general register).
The CVTPD2PI (convert packed double-precision floating-point values to packed
doubleword integers) instruction converts two packed double-precision floating-point
numbers to two packed signed doubleword integers, with the result stored in an MMX
register. When rounding to an integer value, the source value is rounded according to
the rounding mode in the MXCSR register. The CVTTPD2PI (convert with truncation
packed double-precision floating-point values to packed doubleword integers)
instruction is similar to the CVTPD2PI instruction except that truncation is used to
round a source value to an integer value (see Section 4.8.4.2, “Truncation with SSE
and SSE2 Conversion Instructions”).
The CVTPI2PD (convert packed doubleword integers to packed double-precision
floating-point values) instruction converts two packed signed doubleword integers to
two double-precision floating-point values.

Vol. 1 11-13
PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

The CVTPD2DQ (convert packed double-precision floating-point values to packed

doubleword integers) instruction converts two packed double-precision floating-point
numbers to two packed signed doubleword integers, with the result stored in the low
quadword of an XMM register. When rounding an integer value, the source value is
rounded according to the rounding mode selected in the MXCSR register. The
CVTTPD2DQ (convert with truncation packed double-precision floating-point values
to packed doubleword integers) instruction is similar to the CVTPD2DQ instruction
except that truncation is used to round a source value to an integer value (see
Section 4.8.4.2, “Truncation with SSE and SSE2 Conversion Instructions”).
The CVTDQ2PD (convert packed doubleword integers to packed double-precision
floating-point values) instruction converts two packed signed doubleword integers
located in the low-order doublewords of an XMM register to two double-precision
floating-point values.
The CVTSD2SI (convert scalar double-precision floating-point value to doubleword
integer) instruction converts a double-precision floating-point value to a doubleword
integer, and stores the result in a general-purpose register. When rounding an
integer value, the source value is rounded according to the rounding mode selected
in the MXCSR register. The CVTTSD2SI (convert with truncation scalar double-preci-
sion floating-point value to doubleword integer) instruction is similar to the
CVTSD2SI instruction except that truncation is used to round the source value to an
integer value (see Section 4.8.4.2, “Truncation with SSE and SSE2 Conversion
Instructions”).
The CVTSI2SD (convert doubleword integer to scalar double-precision floating-point
value) instruction converts a signed doubleword integer in a general-purpose register
to a double-precision floating-point number, and stores the result in an XMM register.
Conversion between single-precision floating-point and doubleword integer
formats — These instructions convert between packed single-precision floating-
point and packed doubleword integer formats. Operands are housed in XMM regis-
ters, MMX registers, general registers, or memory (the latter for at most one source
operand). The destination is always an XMM, MMX, or general register. These SSE2
instructions supplement conversion instructions (CVTPI2PS, CVTPS2PI, CVTTPS2PI,
CVTSI2SS, CVTSS2SI, and CVTTSS2SI) introduced with SSE extensions.
The CVTPS2DQ (convert packed single-precision floating-point values to packed
doubleword integers) instruction converts four packed single-precision floating-point
values to four packed signed doubleword integers, with the source and destination
operands in XMM registers or memory (the latter for at most one source operand).
When the conversion is inexact, the rounded value according to the rounding mode
selected in the MXCSR register is returned. The CVTTPS2DQ (convert with truncation
packed single-precision floating-point values to packed doubleword integers)
instruction is similar to the CVTPS2DQ instruction except that truncation is used to
round a source value to an integer value (see Section 4.8.4.2, “Truncation with SSE
and SSE2 Conversion Instructions”).
The CVTDQ2PS (convert packed doubleword integers to packed single-precision
floating-point values) instruction converts four packed signed doubleword integers to
four packed single-precision floating-point numbers, with the source and destination

11-14 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

operands in XMM registers or memory (the latter for at most one source operand).
When the conversion is inexact, the rounded value according to the rounding mode
selected in the MXCSR register is returned.

11.4.2 SSE2 64-Bit and 128-Bit SIMD Integer Instructions

SSE2 extensions add several 128-bit packed integer instructions to the IA-32 archi-
tecture. Where appropriate, a 64-bit version of each of these instructions is also
provided. The 128-bit versions of instructions operate on data in XMM registers;
64-bit versions operate on data in MMX registers. The instructions follow.
The MOVDQA (move aligned double quadword) instruction transfers a double quad-
word operand from memory to an XMM register or vice versa; or between XMM regis-
ters. The memory address must be aligned to a 16-byte boundary; otherwise, a
general-protection exception (#GP) is generated.
The MOVDQU (move unaligned double quadword) instruction performs the same
operations as the MOVDQA instruction, except that 16-byte alignment of a memory
address is not required.
The PADDQ (packed quadword add) instruction adds two packed quadword integer
operands or two single quadword integer operands, and stores the results in an XMM
or MMX register, respectively. This instruction can operate on either unsigned or
signed (two’s complement notation) integer operands.
The PSUBQ (packed quadword subtract) instruction subtracts two packed quadword
integer operands or two single quadword integer operands, and stores the results in
an XMM or MMX register, respectively. Like the PADDQ instruction, PSUBQ can
operate on either unsigned or signed (two’s complement notation) integer operands.
The PMULUDQ (multiply packed unsigned doubleword integers) instruction performs
an unsigned multiply of unsigned doubleword integers and returns a quadword
result. Both 64-bit and 128-bit versions of this instruction are available. The 64-bit
version operates on two doubleword integers stored in the low doubleword of each
source operand, and the quadword result is returned to an MMX register. The 128-bit
version performs a packed multiply of two pairs of doubleword integers. Here, the
doublewords are packed in the first and third doublewords of the source operands,
and the quadword results are stored in the low and high quadwords of an XMM
register.
The PSHUFLW (shuffle packed low words) instruction shuffles the word integers
packed into the low quadword of the source operand and stores the shuffled result in
the low quadword of the destination operand. An 8-bit immediate operand specifies
the shuffle order.
The PSHUFHW (shuffle packed high words) instruction shuffles the word integers
packed into the high quadword of the source operand and stores the shuffled result
in the high quadword of the destination operand. An 8-bit immediate operand speci-
fies the shuffle order.

Vol. 1 11-15
PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

The PSHUFD (shuffle packed doubleword integers) instruction shuffles the double-
word integers packed into the source operand and stores the shuffled result in the
destination operand. An 8-bit immediate operand specifies the shuffle order.
The PSLLDQ (shift double quadword left logical) instruction shifts the contents of the
source operand to the left by the amount of bytes specified by an immediate
operand. The empty low-order bytes are cleared (set to 0).
The PSRLDQ (shift double quadword right logical) instruction shifts the contents of
the source operand to the right by the amount of bytes specified by an immediate
operand. The empty high-order bytes are cleared (set to 0).
The PUNPCKHQDQ (Unpack high quadwords) instruction interleaves the high quad-
word of the source operand and the high quadword of the destination operand and
writes them to the destination register.
The PUNPCKLQDQ (Unpack low quadwords) instruction interleaves the low quad-
words of the source operand and the low quadwords of the destination operand and
writes them to the destination register.
Two additional SSE instructions enable data movement from the MMX registers to the
XMM registers.
The MOVQ2DQ (move quadword integer from MMX to XMM registers) instruction
moves the quadword integer from an MMX source register to an XMM destination
register.
The MOVDQ2Q (move quadword integer from XMM to MMX registers) instruction
moves the low quadword integer from an XMM source register to an MMX destination
register.

11.4.3 128-Bit SIMD Integer Instruction Extensions

All of 64-bit SIMD integer instructions introduced with MMX technology and SSE
extensions (with the exception of the PSHUFW instruction) have been extended by
SSE2 extensions to operate on 128-bit packed integer operands located in XMM
registers. The 128-bit versions of these instructions follow the same SIMD conven-
tions regarding packed operands as the 64-bit versions. For example, where the
64-bit version of the PADDB instruction operates on 8 packed bytes, the 128-bit
version operates on 16 packed bytes.

11.4.4 Cacheability Control and Memory Ordering Instructions

SSE2 extensions that give programs more control over the caching, loading, and
storing of data. are described below.

11-16 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

11.4.4.1 FLUSH Cache Line

The CLFLUSH (flush cache line) instruction writes and invalidates the cache line asso-
ciated with a specified linear address. The invalidation is for all levels of the
processor’s cache hierarchy, and it is broadcast throughout the cache coherency
domain.

NOTE
CLFLUSH was introduced with the SSE2 extensions. However, the
instruction can be implemented in IA-32 processors that do not
implement the SSE2 extensions. Detect CLFLUSH using the feature
bit (if CPUID.01H:EDX.CLFSH[bit 19] = 1).

11.4.4.2 Cacheability Control Instructions

The following four instructions enable data from XMM and general-purpose registers
to be stored to memory using a non-temporal hint. The non-temporal hint directs the
processor to store data to memory without writing the data into the cache hierarchy
whenever this is possible. See Section 10.4.6.2, “Caching of Temporal vs. Non-
Temporal Data,” for more information about non-temporal stores and hints.
The MOVNTDQ (store double quadword using non-temporal hint) instruction stores
packed integer data from an XMM register to memory, using a non-temporal hint.
The MOVNTPD (store packed double-precision floating-point values using non-
temporal hint) instruction stores packed double-precision floating-point data from an
XMM register to memory, using a non-temporal hint.
The MOVNTI (store doubleword using non-temporal hint) instruction stores integer
data from a general-purpose register to memory, using a non-temporal hint.
The MASKMOVDQU (store selected bytes of double quadword) instruction stores
selected byte integers from an XMM register to memory, using a byte mask to selec-
tively write the individual bytes. The memory location does not need to be aligned on
a natural boundary. This instruction also uses a non-temporal hint.

11.4.4.3 Memory Ordering Instructions

SSE2 extensions introduce two new fence instructions (LFENCE and MFENCE) as
companions to the SFENCE instruction introduced with SSE extensions.
The LFENCE instruction establishes a memory fence for loads. It guarantees ordering
between two loads and prevents speculative loads from passing the load fence (that
is, no speculative loads are allowed until all loads specified before the load fence have
been carried out).
The MFENCE instruction combines the functions of LFENCE and SFENCE by estab-
lishing a memory fence for both loads and stores. It guarantees that all loads and
stores specified before the fence are globally observable prior to any loads or stores
being carried out after the fence.

Vol. 1 11-17
PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

11.4.4.4 Pause
The PAUSE instruction is provided to improve the performance of “spin-wait loops”
executed on a Pentium 4 or Intel Xeon processor. On a Pentium 4 processor, it also
provides the added benefit of reducing processor power consumption while executing
a spin-wait loop. It is recommended that a PAUSE instruction always be included in
the code sequence for a spin-wait loop.

11.4.5 Branch Hints

SSE2 extensions designate two instruction prefixes (2EH and 3EH) to provide branch
hints to the processor (see “Instruction Prefixes” in Chapter 2 of the Intel® 64 and
IA-32 Architectures Software Developer’s Manual, Volume 2A). These prefixes can
only be used with the Jcc instruction and only at the machine code level (that is,
there are no mnemonics for the branch hints).

11.5 SSE, SSE2, AND SSE3 EXCEPTIONS

SSE/SSE2/SSE3 extensions generate two general types of exceptions:
• Non-numeric exceptions
• SIMD floating-point exceptions1
SSE/SSE2/SSE3 instructions can generate the same type of memory-access and
non-numeric exceptions as other IA-32 architecture instructions. Existing exception
handlers can generally handle these exceptions without any code modification. See
“Providing Non-Numeric Exception Handlers for Exceptions Generated by the SSE,
SSE2 and SSE3 Instructions” in Chapter 12 of the Intel® 64 and IA-32 Architectures
Software Developer’s Manual, Volume 3A, for a list of the non-numeric exceptions
that can be generated by SSE/SSE2/SSE3 instructions and for guidelines for handling
these exceptions.
SSE/SSE2/SSE3 instructions do not generate numeric exceptions on packed integer
operations; however, they can generate numeric (SIMD floating-point) exceptions on
packed single-precision and double-precision floating-point operations. These SIMD
floating-point exceptions are defined in the IEEE Standard 754 for Binary Floating-
Point Arithmetic and are the same exceptions that are generated for x87 FPU instruc-
tions. See Section 11.5.1, “SIMD Floating-Point Exceptions,” for a description of
these exceptions.

1. The FISTTP instruction in SSE3 does not generate SIMD floating-point exceptions, but it can gen-
erate x87 FPU floating-point exceptions.

11-18 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

11.5.1 SIMD Floating-Point Exceptions

SIMD floating-point exceptions are those exceptions that can be generated by
SSE/SSE2/SSE3 instructions that operate on packed or scalar floating-point operands.
Six classes of SIMD floating-point exceptions can be generated:
• Invalid operation (#I)
• Divide-by-zero (#Z)
• Denormal operand (#D)
• Numeric overflow (#O)
• Numeric underflow (#U)
• Inexact result (Precision) (#P)
All of these exceptions (except the denormal operand exception) are defined in IEEE
Standard 754, and they are the same exceptions that are generated with the x87
floating-point instructions. Section 4.9, “Overview of Floating-Point Exceptions,”
gives a detailed description of these exceptions and of how and when they are gener-
ated. The following sections discuss the implementation of these exceptions in
SSE/SSE2/SSE3 extensions.
All SIMD floating-point exceptions are precise and occur as soon as the instruction
completes execution.
Each of the six exception conditions has a corresponding flag (IE, DE, ZE, OE, UE,
and PE) and mask bit (IM, DM, ZM, OM, UM, and PM) in the MXCSR register (see
Figure 10-3). The mask bits can be set with the LDMXCSR or FXRSTOR instruction;
the mask and flag bits can be read with the STMXCSR or FXSAVE instruction.
The OSXMMEXCEPT flag (bit 10) of control register CR4 provides additional control
over generation of SIMD floating-point exceptions by allowing the operating system
to indicate whether or not it supports software exception handlers for SIMD floating-
point exceptions. If an unmasked SIMD floating-point exception is generated and the
OSXMMEXCEPT flag is set, the processor invokes a software exception handler by
generating a SIMD floating-point exception (#XM). If the OSXMMEXCEPT bit is clear,
the processor generates an invalid-opcode exception (#UD) on the first SSE or SSE2
instruction that detects a SIMD floating-point exception condition. See Section
11.6.2, “Checking for SSE/SSE2 Support.”

11.5.2 SIMD Floating-Point Exception Conditions

The following sections describe the conditions that cause a SIMD floating-point
exception to be generated and the masked response of the processor when these
conditions are detected.
See Section 4.9.2, “Floating-Point Exception Priority,” for a description of the rules for
exception precedence when more than one floating-point exception condition is
detected for an instruction.

Vol. 1 11-19
PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

11.5.2.1 Invalid Operation Exception (#I)

The floating-point invalid-operation exception (#I) occurs in response to an invalid
arithmetic operand. The flag (IE) and mask (IM) bits for the invalid operation excep-
tion are bits 0 and 7, respectively, in the MXCSR register.
If the invalid-operation exception is masked, the processor returns a QNaN, QNaN
floating-point indefinite, integer indefinite, one of the source operands to the destina-
tion operand, or it sets the EFLAGS, depending on the operation being performed.
When a value is returned to the destination operand, it overwrites the destination
register specified by the instruction. Table 11-1 lists the invalid-arithmetic operations
that the processor detects for instructions and the masked responses to these opera-
tions.

Table 11-1. Masked Responses of SSE/SSE2/SSE3 Instructions to Invalid Arithmetic

Operations
Condition Masked Response
ADDPS, ADDSS, ADDPD, ADDSD, SUBPS, SUBSS, Return the SNaN converted to a QNaN; Refer to
SUBPD, SUBSD, MULPS, MULSS, MULPD, Table 4-7 for more details
MULSD, DIVPS, DIVSS, DIVPD, DIVSD,
ADDSUBPD, ADDSUBPD, HADDPD, HADDPS,
HSUBPD or HSUBPS instruction with an SNaN
operand
SQRTPS, SQRTSS, SQRTPD, or SQRTSD with Return the SNaN converted to a QNaN
SNaN operands
SQRTPS, SQRTSS, SQRTPD, or SQRTSD with Return the QNaN floating-point Indefinite
negative operands (except zero)
MAXPS, MAXSS, MAXPD, MAXSD, MINPS, Return the source 2 operand value
MINSS, MINPD, or MINSD instruction with QNaN
or SNaN operands
CMPPS, CMPSS, CMPPD or CMPSD instruction Return a mask of all 0s (except for the
with QNaN or SNaN operands predicates “not-equal,” “unordered,” “not-less-
than,” or “not-less-than-or-equal,” which returns
a mask of all 1s)
CVTPD2PS, CVTSD2SS, CVTPS2PD, CVTSS2SD Return the SNaN converted to a QNaN
with SNaN operands
COMISS or COMISD with QNaN or SNaN Set EFLAGS values to “not comparable”
operand(s)
Addition of opposite signed infinities or Return the QNaN floating-point Indefinite
subtraction of like-signed infinities
Multiplication of infinity by zero Return the QNaN floating-point Indefinite
Divide of (0/0) or ( ∞ / ∞ ) Return the QNaN floating-point Indefinite

11-20 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

Table 11-1. Masked Responses of SSE/SSE2/SSE3 Instructions to Invalid Arithmetic

Operations (Contd.)
Condition Masked Response
Conversion to integer when the value in the Return the integer Indefinite
source register is a NaN, ∞, or exceeds the
representable range for CVTPS2PI, CVTTPS2PI,
CVTSS2SI, CVTTSS2SI, CVTPD2PI, CVTSD2SI,
CVTPD2DQ, CVTTPD2PI, CVTTSD2SI,
CVTTPD2DQ, CVTPS2DQ, or CVTTPS2DQ

If the invalid operation exception is not masked, a software exception handler is

invoked and the operands remain unchanged. See Section 11.5.4, “Handling SIMD
Floating-Point Exceptions in Software.”
Normally, when one or more of the source operands are QNaNs (and neither is an
SNaN or in an unsupported format), an invalid-operation exception is not generated.
The following instructions are exceptions to this rule: the COMISS and COMISD
instructions; and the CMPPS, CMPSS, CMPPD, and CMPSD instructions (when the
predicate is less than, less-than or equal, not less-than, or not less-than or equal).
With these instructions, a QNaN source operand will generate an invalid-operation
exception.
The invalid-operation exception is not affected by the flush-to-zero mode or by the
denormals-are-zeros mode.

11.5.2.2 Denormal-Operand Exception (#D)

The processor signals the denormal-operand exception if an arithmetic instruction
attempts to operate on a denormal operand. The flag (DE) and mask (DM) bits for
the denormal-operand exception are bits 1 and 8, respectively, in the MXCSR
register.
The CVTPI2PD, CVTPD2PI, CVTTPD2PI, CVTDQ2PD, CVTPD2DQ, CVTTPD2DQ,
CVTSI2SD, CVTSD2SI, CVTTSD2SI, CVTPI2PS, CVTPS2PI, CVTTPS2PI, CVTSS2SI,
CVTTSS2SI, CVTSI2SS, CVTDQ2PS, CVTPS2DQ, and CVTTPS2DQ conversion instruc-
tions do not signal denormal exceptions. The RCPSS, RCPPS, RSQRTSS, and
RSQRTPS instructions do not signal any kind of floating-point exception.
The denormals-are-zero flag (bit 6) of the MXCSR register provides an additional
option for handling denormal-operand exceptions. When this flag is set, denormal
source operands are automatically converted to zeros with the sign of the source
operand (see Section 10.2.3.4, “Denormals-Are-Zeros”). The denormal operand
exception is not affected by the flush-to-zero mode.
See Section 4.9.1.2, “Denormal Operand Exception (#D),” for more information
about the denormal exception. See Section 11.5.4, “Handling SIMD Floating-Point
Exceptions in Software,” for information on handling unmasked exceptions.

Vol. 1 11-21
PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

11.5.2.3 Divide-By-Zero Exception (#Z)

The processor reports a divide-by-zero exception when a DIVPS, DIVSS, DIVPD or
DIVSD instruction attempts to divide a finite non-zero operand by 0. The flag (ZE)
and mask (ZM) bits for the divide-by-zero exception are bits 2 and 9, respectively, in
the MXCSR register.
See Section 4.9.1.3, “Divide-By-Zero Exception (#Z),” for more information about
the divide-by-zero exception. See Section 11.5.4, “Handling SIMD Floating-Point
Exceptions in Software,” for information on handling unmasked exceptions.
The divide-by-zero exception is not affected by the flush-to-zero mode or by the
denormals-are-zeros mode.

11.5.2.4 Numeric Overflow Exception (#O)

The processor reports a numeric overflow exception whenever the rounded result of
an arithmetic instruction exceeds the largest allowable finite value that fits in the
destination operand. This exception can be generated with the ADDPS, ADDSS,
ADDPD, ADDSD, SUBPS, SUBSS, SUBPD, SUBSD, MULPS, MULSS, MULPD, MULSD,
DIVPS, DIVSS, DIVPD, DIVSD, CVTPD2PS, CVTSD2SS, ADDSUBPD, ADDSUBPS,
HADDPD, HADDPS, HSUBPD and HSUBPS instructions. The flag (OE) and mask (OM)
bits for the numeric overflow exception are bits 3 and 10, respectively, in the MXCSR
register.
See Section 4.9.1.4, “Numeric Overflow Exception (#O),” for more information about
the numeric-overflow exception. See Section 11.5.4, “Handling SIMD Floating-Point
Exceptions in Software,” for information on handling unmasked exceptions.
The numeric overflow exception is not affected by the flush-to-zero mode or by the
denormals-are-zeros mode.

11.5.2.5 Numeric Underflow Exception (#U)

The processor reports a numeric underflow exception whenever the rounded result of
an arithmetic instruction is less than the smallest possible normalized, finite value
that will fit in the destination operand and the numeric-underflow exception is not
masked. If the numeric underflow exception is masked, both underflow and the
inexact-result condition must be detected before numeric underflow is reported. This
exception can be generated with the ADDPS, ADDSS, ADDPD, ADDSD, SUBPS,
SUBSS, SUBPD, SUBSD, MULPS, MULSS, MULPD, MULSD, DIVPS, DIVSS, DIVPD,
DIVSD, CVTPD2PS, CVTSD2SS, ADDSUBPD, ADDSUBPS, HADDPD, HADDPS,
HSUBPD, and HSUBPS instructions. The flag (UE) and mask (UM) bits for the numeric
underflow exception are bits 4 and 11, respectively, in the MXCSR register.
The flush-to-zero flag (bit 15) of the MXCSR register provides an additional option for
handling numeric underflow exceptions. When this flag is set and the numeric under-
flow exception is masked, tiny results (results that trigger the underflow exception)
are returned as a zero with the sign of the true result (see Section 10.2.3.3, “Flush-

11-22 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

To-Zero”). The numeric underflow exception is not affected by the denormals-are-

zero mode.
See Section 4.9.1.5, “Numeric Underflow Exception (#U),” for more information
about the numeric underflow exception. See Section 11.5.4, “Handling SIMD
Floating-Point Exceptions in Software,” for information on handling unmasked
exceptions.

11.5.2.6 Inexact-Result (Precision) Exception (#P)

The inexact-result exception (also called the precision exception) occurs if the result
of an operation is not exactly representable in the destination format. For example,
the fraction 1/3 cannot be precisely represented in binary form. This exception
occurs frequently and indicates that some (normally acceptable) accuracy has been
lost. The exception is supported for applications that need to perform exact arith-
metic only. Because the rounded result is generally satisfactory for most applica-
tions, this exception is commonly masked.
The flag (PE) and mask (PM) bits for the inexact-result exception are bits 2 and 12,
respectively, in the MXCSR register.
See Section 4.9.1.6, “Inexact-Result (Precision) Exception (#P),” for more informa-
tion about the inexact-result exception. See Section 11.5.4, “Handling SIMD
Floating-Point Exceptions in Software,” for information on handling unmasked excep-
tions.
In flush-to-zero mode, the inexact result exception is reported. The inexact result
exception is not affected by the denormals-are-zero mode.

11.5.3 Generating SIMD Floating-Point Exceptions

When the processor executes a packed or scalar floating-point instruction, it looks for
and reports on SIMD floating-point exception conditions using two sequential steps:
1. Looks for, reports on, and handles pre-computation exception conditions (invalid-
operand, divide-by-zero, and denormal operand)
2. Looks for, reports on, and handles post-computation exception conditions
(numeric overflow, numeric underflow, and inexact result)
If both pre- and post-computational exceptions are unmasked, it is possible for the
processor to generate a SIMD floating-point exception (#XM) twice during the execu-
tion of an SSE, SSE2 or SSE3 instruction: once when it detects and handles a pre-
computational exception and when it detects a post-computational exception.

11.5.3.1 Handling Masked Exceptions

If all exceptions are masked, the processor handles the exceptions it detects by
placing the masked result (or results for packed operands) in a destination operand

Vol. 1 11-23
PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

and continuing program execution. The masked result may be a rounded normalized
value, signed infinity, a denormal finite number, zero, a QNaN floating-point indefi-
nite, or a QNaN depending on the exception condition detected. In most cases, the
corresponding exception flag bit in MXCSR is also set. The one situation where an
exception flag is not set is when an underflow condition is detected and it is not
accompanied by an inexact result.
When operating on packed floating-point operands, the processor returns a masked
result for each of the sub-operand computations and sets a separate set of internal
exception flags for each computation. It then performs a logical-OR on the internal
exception flag settings and sets the exception flags in the MXCSR register according
to the results of OR operations.
For example, Figure 11-9 shows the results of an MULPS instruction. In the example,
all SIMD floating-point exceptions are masked. Assume that a denormal exception
condition is detected prior to the multiplication of sub-operands X0 and Y0, no excep-
tion condition is detected for the multiplication of X1 and Y1, a numeric overflow
exception condition is detected for the multiplication of X2 and Y2, and another
denormal exception is detected prior to the multiplication of sub-operands X3 and
Y3. Because denormal exceptions are masked, the processor uses the denormal
source values in the multiplications of (X0 and Y0) and of (X3 and Y3) passing the
results of the multiplications through to the destination operand. With the denormal
operand, the result of the X0 and Y0 computation is a normalized finite value, with no
exceptions detected. However, the X3 and Y3 computation produces a tiny and
inexact result. This causes the corresponding internal numeric underflow and
inexact-result exception flags to be set.

X3 X2 X1 X0 (Denormal)

Y3 (Denormal) Y2 Y1 Y0

MULPS MULPS MULPS MULPS

Tiny, Inexact, Finite ∞ Normalized Finite Normalized Finite

Figure 11-9. Example Masked Response for Packed Operations

For the multiplication of X2 and Y2, the processor stores the floating-point ∞ in the
destination operand, and sets the corresponding internal sub-operand numeric over-
flow flag. The result of the X1 and Y1 multiplication is passed through to the destina-
tion operand, with no internal sub-operand exception flags being set. Following the

11-24 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

computations, the individual sub-operand exceptions flags for denormal operand,

numeric underflow, inexact result, and numeric overflow are OR’d and the corre-
sponding flags are set in the MXCSR register.
The net result of this computation is that:
• Multiplication of X0 and Y0 produces a normalized finite result
• Multiplication of X1 and Y1 produces a normalized finite result
• Multiplication of X2 and Y2 produces a floating-point ∞ result
• Multiplication of X3 and Y3 produces a tiny, inexact, finite result
• Denormal operand, numeric underflow, numeric underflow, and inexact result
flags are set in the MXCSR register

11.5.3.2 Handling Unmasked Exceptions

If all exceptions are unmasked, the processor:
1. First detects any pre-computation exceptions: it ORs those exceptions, sets the
appropriate exception flags, leaves the source and destination operands
unaltered, and goes to step 2. If it does not detect any pre-computation
exceptions, it goes to step 5.
2. Checks CR4.OSXMMEXCPT[bit 10]. If this flag is set, the processor goes to step
3; if the flag is clear, it generates an invalid-opcode exception (#UD) and makes
an implicit call to the invalid-opcode exception handler.
3. Generates a SIMD floating-point exception (#XM) and makes an implicit call to
the SIMD floating-point exception handler.
4. If the exception handler is able to fix the source operands that generated the pre-
computation exceptions or mask the condition in such a way as to allow the
processor to continue executing the instruction, the processor resumes
instruction execution as described in step 5.
5. Upon returning from the exception handler (or if no pre-computation exceptions
were detected), the processor checks for post-computation exceptions. If the
processor detects any post-computation exceptions: it ORs those exceptions,
sets the appropriate exception flags, leaves the source and destination operands
unaltered, and repeats steps 2, 3, and 4.
6. Upon returning from the exceptions handler in step 4 (or if no post-computation
exceptions were detected), the processor completes the execution of the
instruction.
The implication of this procedure is that for unmasked exceptions, the processor can
generate a SIMD floating-point exception (#XM) twice: once if it detects pre-compu-
tation exception conditions and a second time if it detects post-computation excep-
tion conditions. For example, if SIMD floating-point exceptions are unmasked for the
computation shown in Figure 11-9, the processor would generate one SIMD floating-
point exception for denormal operand conditions and a second SIMD floating-point

Vol. 1 11-25
PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

exception for overflow and underflow (no inexact result exception would be gener-
ated because the multiplications of X0 and Y0 and of X1 and Y1 are exact).

11.5.3.3 Handling Combinations of Masked and Unmasked Exceptions

In situations where both masked and unmasked exceptions are detected, the
processor will set exception flags for the masked and the unmasked exceptions.
However, it will not return masked results until after the processor has detected and
handled unmasked post-computation exceptions and returned from the exception
handler (as in step 6 above) to finish executing the instruction.

11.5.4 Handling SIMD Floating-Point Exceptions in Software

Section 4.9.3, “Typical Actions of a Floating-Point Exception Handler,” shows actions
that may be carried out by a SIMD floating-point exception handler. The
SSE/SSE2/SSE3 state is saved with the FXSAVE instruction (see Section 11.6.5,
“Saving and Restoring the SSE/SSE2 State”).

11.5.5 Interaction of SIMD and x87 FPU Floating-Point Exceptions

SIMD floating-point exceptions are generated independently from x87 FPU floating-
point exceptions. SIMD floating-point exceptions do not cause assertion of the
FERR# pin (independent of the value of CR0.NE[bit 5]). They ignore the assertion
and deassertion of the IGNNE# pin.
If applications use SSE/SSE2/SSE3 instructions along with x87 FPU instructions (in
the same task or program), consider the following:
• SIMD floating-point exceptions are reported independently from the x87 FPU
floating-point exceptions. SIMD and x87 FPU floating-point exceptions can be
unmasked independently. Separate x87 FPU and SIMD floating-point exception
handlers must be provided if the same exception is unmasked for x87 FPU and for
SSE/SSE2/SSE3 operations.
• The rounding mode specified in the MXCSR register does not affect x87 FPU
instructions. Likewise, the rounding mode specified in the x87 FPU control word
does not affect the SSE/SSE2/SSE3 instructions. To use the same rounding
mode, the rounding control bits in the MXCSR register and in the x87 FPU control
word must be set explicitly to the same value.
• The flush-to-zero mode set in the MXCSR register for SSE/SSE2/SSE3 instruc-
tions has no counterpart in the x87 FPU. For compatibility with the x87 FPU, set
the flush-to-zero bit to 0.
• The denormals-are-zeros mode set in the MXCSR register for SSE/SSE2/SSE3
instructions has no counterpart in the x87 FPU. For compatibility with the x87
FPU, set the denormals-are-zeros bit to 0.

11-26 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

• An application that expects to detect x87 FPU exceptions that occur during the
execution of x87 FPU instructions will not be notified if exceptions occurs during
the execution of corresponding SSE/SSE2/SSE31 instructions, unless the
exception masks that are enabled in the x87 FPU control word have also been
enabled in the MXCSR register and the application is capable of handling SIMD
floating-point exceptions (#XM).
— Masked exceptions that occur during an SSE/SSE2/SSE3 library call cannot
be detected by unmasking the exceptions after the call (in an attempt to
generate the fault based on the fact that an exception flag is set). A SIMD
floating-point exception flag that is set when the corresponding exception is
unmasked will not generate a fault; only the next occurrence of that
unmasked exception will generate a fault.
— An application which checks the x87 FPU status word to determine if any
masked exception flags were set during an x87 FPU library call will also need
to check the MXCSR register to detect a similar occurrence of a masked
exception flag being set during an SSE/SSE2/SSE3 library call.

11.6 WRITING APPLICATIONS WITH SSE/SSE2

EXTENSIONS
The following sections give some guidelines for writing application programs and
operating-system code that uses the SSE and SSE2 extensions. Because SSE and
SSE2 extensions share the same state and perform companion operations, these
guidelines apply to both sets of extensions.
Chapter 12 in the Intel® 64 and IA-32 Architectures Software Developer’s Manual,
Volume 3A, discusses the interface to the processor for context switching as well as
other operating system considerations when writing code that uses SSE/SSE2/SSE3
extensions.

11.6.1 General Guidelines for Using SSE/SSE2 Extensions

The following guidelines describe how to take full advantage of the performance
gains available with the SSE and SSE2 extensions:
• Ensure that the processor supports the SSE and SSE2 extensions.
• Ensure that your operating system supports the SSE and SSE2 extensions.
(Operating system support for the SSE extensions implies support for SSE2
extension and vice versa.)

1. SSE3 refers to ADDSUBPD, ADDSUBPS, HADDPD, HADDPS, HSUBPD and HSUBPS; the only other
SSE3 instruction that can raise floating-point exceptions is FISTTP: it can generate x87 FPU
invalid operation and inexact result exceptions.

Vol. 1 11-27
PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

• Use stack and data alignment techniques to keep data properly aligned for
efficient memory use.
• Use the non-temporal store instructions offered with the SSE and SSE2
extensions.
• Employ the optimization and scheduling techniques described in the Intel
Pentium 4 Optimization Reference Manual (see Section 1.4, “Related Literature,”
for the order number for this manual).

11.6.2 Checking for SSE/SSE2 Support

Before an application attempts to use the SSE and/or SSE2 extensions, it should
check that they are present on the processor:
1. Check that the processor supports the CPUID instruction. Bit 21 of the EFLAGS
register can be used to check processor’s support the CPUID instruction.
2. Check that the processor supports the SSE and/or SSE2 extensions (true if
CPUID.01H:EDX.SSE[bit 25] = 1 and/or CPUID.01H:EDX.SSE2[bit 26] = 1).
Operating system must provide system level support for handling SSE state, excep-
tions before an application can use the SSE and/or SSE2 extensions (see Chapter 12
in the Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volume 3A,).
If the processor attempts to execute an unsupported SSE or SSE2 instruction, the
processor will generate an invalid-opcode exception (#UD). If an operating system
did not provide adequate system level support for SSE, executing an SSE or SSE2
instructions can also generate #UD.

11.6.3 Checking for the DAZ Flag in the MXCSR Register

The denormals-are-zero flag in the MXCSR register is available in most of the
Pentium 4 processors and in the Intel Xeon processor, with the exception of some
early steppings. To check for the presence of the DAZ flag in the MXCSR register, do
the following:
1. Establish a 512-byte FXSAVE area in memory.
2. Clear the FXSAVE area to all 0s.
3. Execute the FXSAVE instruction, using the address of the first byte of the cleared
FXSAVE area as a source operand. See “FXSAVE—Save x87 FPU, MMX, SSE, and
SSE2 State” in Chapter 3 of the Intel® 64 and IA-32 Architectures Software
Developer’s Manual, Volume 2A, for a description of the FXSAVE instruction and
the layout of the FXSAVE image.
4. Check the value in the MXCSR_MASK field in the FXSAVE image (bytes 28
through 31).
— If the value of the MXCSR_MASK field is 00000000H, the DAZ flag and
denormals-are-zero mode are not supported.

11-28 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

— If the value of the MXCSR_MASK field is non-zero and bit 6 is set, the DAZ
flag and denormals-are-zero mode are supported.
If the DAZ flag is not supported, then it is a reserved bit and attempting to write a 1
to it will cause a general-protection exception (#GP). See Section 11.6.6, “Guidelines
for Writing to the MXCSR Register,” for general guidelines for preventing general-
protection exceptions when writing to the MXCSR register.

11.6.4 Initialization of SSE/SE2 Extensions

The SSE and SSE2 state is contained in the XMM and MXCSR registers. Upon a hard-
ware reset of the processor, this state is initialized as follows (see Table 11-2):
• All SIMD floating-point exceptions are masked (bits 7 through 12 of the MXCSR
register is set to 1).
• All SIMD floating-point exception flags are cleared (bits 0 through 5 of the MXCSR
register is set to 0).
• The rounding control is set to round-nearest (bits 13 and 14 of the MXCSR
register are set to 00B).
• The flush-to-zero mode is disabled (bit 15 of the MXCSR register is set to 0).
• The denormals-are-zeros mode is disabled (bit 6 of the MXCSR register is set to
0). If the denormals-are-zeros mode is not supported, this bit is reserved and will
be set to 0 on initialization.
• Each of the XMM registers is cleared (set to all zeros).

Vol. 1 11-29
PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

Table 11-2. SSE and SSE2 State Following a Power-up/Reset or INIT

Registers Power-Up or INIT
Reset
XMM0 through XMM7 +0.0 Unchanged
MXCSR 1F80H Unchanged

If the processor is reset by asserting the INIT# pin, the SSE and SSE2 state is not
changed.

11.6.5 Saving and Restoring the SSE/SSE2 State

The FXSAVE instruction saves the x87 FPU, MMX, SSE and SSE2 states (which
includes the contents of eight XMM registers and the MXCSR registers) in a 512-byte
block of memory. The FXRSTOR instruction restores the saved SSE and SSE2 state
from memory. See the FXSAVE instruction in Chapter 3 of the Intel® 64 and IA-32
Architectures Software Developer’s Manual, Volume 2A, for the layout of the
512-byte state block.
In addition to saving and restoring the SSE and SSE2 state, FXSAVE and FXRSTOR
also save and restore the x87 FPU state (because MMX registers are aliased to the
x87 FPU data registers this includes saving and restoring the MMX state). For greater
code efficiency, it is suggested that FXSAVE and FXRSTOR be substituted for the
FSAVE, FNSAVE and FRSTOR instructions in the following situations:
• When a context switch is being made in a multitasking environment
• During calls and returns from interrupt and exception handlers
In situations where the code is switching between x87 FPU and MMX technology
computations (without a context switch or a call to an interrupt or exception), the
FSAVE/FNSAVE and FRSTOR instructions are more efficient than the FXSAVE and
FXRSTOR instructions.

11.6.6 Guidelines for Writing to the MXCSR Register

The MXCSR has several reserved bits, and attempting to write a 1 to any of these bits
will cause a general-protection exception (#GP) to be generated. To allow software to
identify these reserved bits, the MXCSR_MASK value is provided. Software can deter-
mine this mask value as follows:
1. Establish a 512-byte FXSAVE area in memory.
2. Clear the FXSAVE area to all 0s.
3. Execute the FXSAVE instruction, using the address of the first byte of the cleared
FXSAVE area as a source operand. See “FXSAVE—Save x87 FPU, MMX, SSE, and
SSE2 State” in Chapter 3 of the Intel® 64 and IA-32 Architectures Software

11-30 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

Developer’s Manual, Volume 2A, for a description of FXSAVE and the layout of the
FXSAVE image.
4. Check the value in the MXCSR_MASK field in the FXSAVE image (bytes 28
through 31).
— If the value of the MXCSR_MASK field is 00000000H, then the MXCSR_MASK
value is the default value of 0000FFBFH. Note that this value indicates that bit
6 of the MXCSR register is reserved; this setting indicates that the
denormals-are-zero mode is not supported on the processor.
— If the value of the MXCSR_MASK field is non-zero, the MXCSR_MASK value
should be used as the MXCSR_MASK.
All bits set to 0 in the MXCSR_MASK value indicate reserved bits in the MXCSR
register. Thus, if the MXCSR_MASK value is AND’d with a value to be written into the
MXCSR register, the resulting value will be assured of having all its reserved bits set
to 0, preventing the possibility of a general-protection exception being generated
when the value is written to the MXCSR register.
For example, the default MXCSR_MASK value when 00000000H is returned in the
FXSAVE image is 0000FFBFH. If software AND’s a value to be written to MXCSR
register with 0000FFBFH, bit 6 of the result (the DAZ flag) will be ensured of being
set to 0, which is the required setting to prevent general-protection exceptions on
processors that do not support the denormals-are-zero mode.
To prevent general-protection exceptions, the MXCSR_MASK value should be AND’d
with the value to be written into the MXCSR register in the following situations:
• Operating system routines that receive a parameter from an application program
and then write that value to the MXCSR register (either with an FXRSTOR or
LDMXCSR instruction)
• Any application program that writes to the MXCSR register and that needs to run
robustly on several different IA-32 processors
Note that all bits in the MXCSR_MASK value that are set to 1 indicate features that
are supported by the MXCSR register; they can be treated as feature flags for identi-
fying processor capabilities.

11.6.7 Interaction of SSE/SSE2 Instructions with x87 FPU and MMX

Instructions
The XMM registers and the x87 FPU and MMX registers represent separate execution
environments, which has certain ramifications when executing SSE, SSE2, MMX, and
x87 FPU instructions in the same code module or when mixing code modules that
contain these instructions:
• Those SSE and SSE2 instructions that operate only on XMM registers (such as the
packed and scalar floating-point instructions and the 128-bit SIMD integer
instructions) in the same instruction stream with 64-bit SIMD integer or x87 FPU
instructions without any restrictions. For example, an application can perform the

Vol. 1 11-31
PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

majority of its floating-point computations in the XMM registers, using the packed
and scalar floating-point instructions, and at the same time use the x87 FPU to
perform trigonometric and other transcendental computations. Likewise, an
application can perform packed 64-bit and 128-bit SIMD integer operations
together without restrictions.
• Those SSE and SSE2 instructions that operate on MMX registers (such as the
CVTPS2PI, CVTTPS2PI, CVTPI2PS, CVTPD2PI, CVTTPD2PI, CVTPI2PD,
MOVDQ2Q, MOVQ2DQ, PADDQ, and PSUBQ instructions) can also be executed in
the same instruction stream as 64-bit SIMD integer or x87 FPU instructions,
however, here they are subject to the restrictions on the simultaneous use of
MMX technology and x87 FPU instructions, which include:
— Transition from x87 FPU to MMX technology instructions or to SSE or SSE2
instructions that operate on MMX registers should be preceded by saving the
state of the x87 FPU.
— Transition from MMX technology instructions or from SSE or SSE2 instruc-
tions that operate on MMX registers to x87 FPU instructions should be
preceded by execution of the EMMS instruction.

11.6.8 Compatibility of SIMD and x87 FPU Floating-Point Data

Types
SSE and SSE2 extensions operate on the same single-precision and double-precision
floating-point data types that the x87 FPU operates on. However, when operating on
these data types, the SSE and SSE2 extensions operate on them in their native
format (single-precision or double-precision), in contrast to the x87 FPU which
extends them to double extended-precision floating-point format to perform compu-
tations and then rounds the result back to a single-precision or double-precision
format before writing results to memory. Because the x87 FPU operates on a higher
precision format and then rounds the result to a lower precision format, it may return
a slightly different result when performing the same operation on the same single-
precision or double-precision floating-point values than is returned by the SSE and
SSE2 extensions. The difference occurs only in the least-significant bits of the signif-
icand.

11.6.9 Mixing Packed and Scalar Floating-Point and 128-Bit SIMD

Integer Instructions and Data
SSE and SSE2 extensions define typed operations on packed and scalar floating-
point data types and on 128-bit SIMD integer data types, but IA-32 processors do not
enforce this typing at the architectural level. They only enforce it at the microarchi-
tectural level. Therefore, when a Pentium 4 or Intel Xeon processor loads a packed or
scalar floating-point operand or a 128-bit packed integer operand from memory into
an XMM register, it does not check that the actual data being loaded matches the
data type specified in the instruction. Likewise, when the processor performs an

11-32 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

arithmetic operation on the data in an XMM register, it does not check that the data
being operated on matches the data type specified in the instruction.
As a general rule, because data typing of SIMD floating-point and integer data types
is not enforced at the architectural level, it is the responsibility of the programmer,
assembler, or compiler to insure that code enforces data typing. Failure to enforce
correct data typing can lead to computations that return unexpected results.
For example, in the following code sample, two packed single-precision floating-point
operands are moved from memory into XMM registers (using MOVAPS instructions);
then a double-precision packed add operation (using the ADDPD instruction) is
performed on the operands:
movaps xmm0, [eax] ; EAX register contains pointer to packed
; single-precision floating-point operand
movaps xmm1, [ebx]
addpd xmm0, xmm1
Pentium 4 and Intel Xeon processors execute these instructions without generating
an invalid-operand exception (#UD) and will produce the expected results in register
XMM0 (that is, the high and low 64-bits of each register will be treated as a double-
precision floating-point value and the processor will operate on them accordingly).
Because the data types operated on and the data type expected by the ADDPD
instruction were inconsistent, the instruction may result in a SIMD floating-point
exception (such as numeric overflow [#O] or invalid operation [#I]) being gener-
ated, but the actual source of the problem (inconsistent data types) is not detected.
The ability to operate on an operand that contains a data type that is inconsistent
with the typing of the instruction being executed, permits some valid operations to be
performed. For example, the following instructions load a packed double-precision
floating-point operand from memory to register XMM0, and a mask to register
XMM1; then they use XORPD to toggle the sign bits of the two packed values in
register XMM0.
movapd xmm0, [eax] ; EAX register contains pointer to packed
; double-precision floating-point operand
movaps xmm1, [ebx] ; EBX register contains pointer to packed
; double-precision floating-point mask
xorpd xmm0, xmm1 ; XOR operation toggles sign bits using
; the mask in xmm1
In this example: XORPS or PXOR can be used in place of XORPD and yield the same
correct result. However, because of the type mismatch between the operand data
type and the instruction data type, a latency penalty will be incurred due to imple-
mentations of the instructions at the microarchitecture level.
Latency penalties can also be incurred by using move instructions of the wrong type.
For example, MOVAPS and MOVAPD can both be used to move a packed single-preci-

Vol. 1 11-33
PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

sion operand from memory to an XMM register. However, if MOVAPD is used, a

latency penalty will be incurred when a correctly typed instruction attempts to use
the data in the register.
Note that these latency penalties are not incurred when moving data from XMM
registers to memory.

11.6.10 Interfacing with SSE/SSE2 Procedures and Functions

SSE and SSE2 extensions allow direct access to XMM registers. This means that all
existing interface conventions between procedures and functions that apply to the
use of the general-purpose registers (EAX, EBX, etc.) also apply to XMM register
usage.

11.6.10.1 Passing Parameters in XMM Registers

The state of XMM registers is preserved across procedure (or function) boundaries.
Parameters can be passed from one procedure to another using XMM registers.

11.6.10.2 Saving XMM Register State on a Procedure or Function Call

The state of XMM registers can be saved in two ways: using an FXSAVE instruction or
a move instruction. FXSAVE saves the state of all XMM registers (along with the state
of MXCSR and the x87 FPU registers). This instruction is typically used for major
changes in the context of the execution environment, such as a task switch.
FXRSTOR restores the XMM, MXCSR, and x87 FPU registers stored with FXSAVE.
In cases where only XMM registers must be saved, or where selected XMM registers
need to be saved, move instructions (MOVAPS, MOVUPS, MOVSS, MOVAPD,
MOVUPD, MOVSD, MOVDQA, and MOVDQU) can be used. These instructions can also
be used to restore the contents of XMM registers. To avoid performance degradation
when saving XMM registers to memory or when loading XMM registers from memory,
be sure to use the appropriately typed move instructions.
The move instructions can also be used to save the contents of XMM registers on the
stack. Here, the stack pointer (in the ESP register) can be used as the memory
address to the next available byte in the stack. Note that the stack pointer is not
automatically incremented when using a move instruction (as it is with PUSH).
A move-instruction procedure that saves the contents of an XMM register to the stack
is responsible for decrementing the value in the ESP register by 16. Likewise, a
move-instruction procedure that loads an XMM register from the stack needs also to
increment the ESP register by 16. To avoid performance degradation when moving
the contents of XMM registers, use the appropriately typed move instructions.
Use the LDMXCSR and STMXCSR instructions to save and restore, respectively, the
contents of the MXCSR register on a procedure call and return.

11-34 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

11.6.10.3 Caller-Save Requirement for Procedure and Function Calls

When making procedure (or function) calls from SSE or SSE2 code, a caller-save
convention is recommended for saving the state of the calling procedure. Using this
convention, any register whose content must survive intact across a procedure call
must be stored in memory by the calling procedure prior to executing the call.
The primary reason for using the caller-save convention is to prevent performance
degradation. XMM registers can contain packed or scalar double-precision floating-
point, packed single-precision floating-point, and 128-bit packed integer data types.
The called procedure has no way of knowing the data types in XMM registers
following a call; so it is unlikely to use the correctly typed move instruction to store
the contents of XMM registers in memory or to restore the contents of XMM registers
from memory.
As described in Section 11.6.9, “Mixing Packed and Scalar Floating-Point and 128-Bit
SIMD Integer Instructions and Data,” executing a move instruction that does not
match the type for the data being moved to/from XMM registers will be carried out
correctly, but can lead to a greater instruction latency.

11.6.11 Updating Existing MMX Technology Routines

Using 128-Bit SIMD Integer Instructions
SSE2 extensions extend all 64-bit MMX SIMD integer instructions to operate on 128-
bit SIMD integers using XMM registers. The extended 128-bit SIMD integer instruc-
tions operate like the 64-bit SIMD integer instructions; this simplifies the porting of
MMX technology applications. However, there are considerations:
• To take advantage of wider 128-bit SIMD integer instructions, MMX technology
code must be recompiled to reference the XMM registers instead of MMX
registers.
• Computation instructions that reference memory operands that are not aligned
on 16-byte boundaries should be replaced with an unaligned 128-bit load
(MOVUDQ instruction) followed by a version of the same computation operation
that uses register instead of memory operands. Use of 128-bit packed integer
computation instructions with memory operands that are not 16-byte aligned
results in a general protection exception (#GP).
• Extension of the PSHUFW instruction (shuffle word across 64-bit integer
operand) across a full 128-bit operand is emulated by a combination of the
following instructions: PSHUFHW, PSHUFLW, and PSHUFD.
• Use of the 64-bit shift by bit instructions (PSRLQ, PSLLQ) can be extended to 128
bits in either of two ways:
— Use of PSRLQ and PSLLQ, along with masking logic operations.
— Rewriting the code sequence to use PSRLDQ and PSLLDQ (shift double
quadword operand by bytes)

Vol. 1 11-35
PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

• Loop counters need to be updated, since each 128-bit SIMD integer instruction
operates on twice the amount of data as its 64-bit SIMD integer counterpart.

11.6.12 Branching on Arithmetic Operations

There are no condition codes in SSE or SSE2 states. A packed-data comparison
instruction generates a mask which can then be transferred to an integer register.
The following code sequence provides an example of how to perform a conditional
branch, based on the result of an SSE2 arithmetic operation.
cmppd XMM0, XMM1 ; generates a mask in XMM0
movmskpd EAX, XMM0 ; moves a 2 bit mask to eax
test EAX, 0 ; compare with desired result
jne BRANCH TARGET
The COMISD and UCOMISD instructions update the EFLAGS as the result of a scalar
comparison. A conditional branch can then be scheduled immediately following
COMISD/UCOMISD.

11.6.13 Cacheability Hint Instructions

SSE and SSE2 cacheability control instructions enable the programmer to control
prefetching, caching, loading and storing of data. When correctly used, these instruc-
tions improve application performance.
To make efficient use of the processor’s super-scalar microarchitecture, a program
needs to provide a steady stream of data to the executing program to avoid stalling
the processor. PREFETCHh instructions minimize the latency of data accesses in
performance-critical sections of application code by allowing data to be fetched into
the processor cache hierarchy in advance of actual usage.
PREFETCHh instructions do not change the user-visible semantics of a program,
although they may affect performance. The operation of these instructions is imple-
mentation-dependent. Programmers may need to tune code for each IA-32
processor implementation. Excessive usage of PREFETCHh instructions may waste
memory bandwidth and reduce performance. For more detailed information on the
use of prefetch hints, refer to Chapter 6, “Optimizing Cache Usage”, in the Intel® 64
and IA-32 Architectures Optimization Reference Manual.
The non-temporal store instructions (MOVNTI, MOVNTPD, MOVNTPS, MOVNTDQ,
MOVNTQ, MASKMOVQ, and MASKMOVDQU) minimize cache pollution when writing
non-temporal data to memory (see Section 10.4.6.2, “Caching of Temporal vs. Non-
Temporal Data,” and Section 10.4.6.1, “Cacheability Control Instructions”). They
prevent non-temporal data from being written into processor caches on a store oper-
ation. These instructions are implementation specific. Programmers may have to
tune their applications for each IA-32 processor implementation to take advantage of
these instructions.

11-36 Vol. 1
 PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

Besides reducing cache pollution, the use of weakly-ordered memory types can be
important under certain data sharing relationships, such as a producer-consumer
relationship. The use of weakly ordered memory can make the assembling of data
more efficient; but care must be taken to ensure that the consumer obtains the data
that the producer intended. Some common usage models that may be affected in this
way by weakly-ordered stores are:
• Library functions that use weakly ordered memory to write results
• Compiler-generated code that writes weakly-ordered results
• Hand-crafted code
The degree to which a consumer of data knows that the data is weakly ordered can
vary for these cases. As a result, the SFENCE or MFENCE instruction should be used
to ensure ordering between routines that produce weakly-ordered data and routines
that consume the data. SFENCE and MFENCE provide a performance-efficient way to
ensure ordering by guaranteeing that every store instruction that precedes
SFENCE/MFENCE in program order is globally visible before a store instruction that
follows the fence.

11.6.14 Effect of Instruction Prefixes on the SSE/SSE2 Instructions

Table 11-3 describes the effects of instruction prefixes on SSE and SSE2 instructions.
(Table 11-3 also applies to SIMD integer and SIMD floating-point instructions in
SSE3.) Unpredictable behavior can range from prefixes being treated as a reserved
operation on one generation of IA-32 processors to generating an invalid opcode
exception on another generation of processors.

See also “Instruction Prefixes” in Chapter 2 of the Intel® 64 and IA-32 Architectures
Software Developer’s Manual, Volume 2A, for complete description of instruction
prefixes.

NOTE
Some SSE/SSE2/SSE3 instructions have two-byte opcodes that are
either 2 bytes or 3 bytes in length. Two-byte opcodes that are 3 bytes
in length consist of: a mandatory prefix (F2H, F3H, or 66H), 0FH, and
an opcode byte. See Table 11-3.

Table 11-3. Effect of Prefixes on SSE, SSE2, and SSE3 Instructions

Prefix Type Effect on SSE, SSE2 and SSE3 Instructions
Address Size Prefix (67H) Affects instructions with a memory operand.
Reserved for instructions without a memory operand and
may result in unpredictable behavior.
Operand Size (66H) Reserved and may result in unpredictable behavior.

Vol. 1 11-37
PROGRAMMING WITH STREAMING SIMD EXTENSIONS 2 (SSE2)

Table 11-3. Effect of Prefixes on SSE, SSE2, and SSE3 Instructions

Prefix Type Effect on SSE, SSE2 and SSE3 Instructions
Segment Override Affects instructions with a memory operand.
(2EH,36H,3EH,26H,64H,65H)
Reserved for instructions without a memory operand and
may result in unpredictable behavior.
Repeat Prefixes (F2H and F3H) Reserved and may result in unpredictable behavior.
Lock Prefix (F0H) Reserved; generates invalid opcode exception (#UD).
Branch Hint Prefixes(E2H and Reserved and may result in unpredictable behavior.
E3H)

11-38 Vol. 1
 CHAPTER 12
PROGRAMMING WITH SSE3, SSSE3, AND SSE4

The Pentium 4 processor supporting Hyper-Threading Technology (HT Technology)

introduces Streaming SIMD Extensions 3 (SSE3). The Intel Xeon processor 5100
series, Intel Core 2 processor families introduced Supplemental Streaming SIMD
Extensions 3 (SSSE3). SSE4 are introduced in Intel processor generations built from
45nm process technology. This chapter describes SSE3, SSSE3, SSE4, and provides
information to assist in writing application programs that use these extensions.

12.1 PROGRAMMING ENVIRONMENT AND DATA TYPES

The programming environment for using SSE3, SSSE3, and SSE4 is unchanged from
those shown in Figure 3-1 and Figure 3-2. SSE3, SSSE3, and SSE4 do not introduce
new data types. XMM registers are used to operate on packed integer data, single-
precision floating-point data, or double-precision floating-point data.
One SSE3 instruction uses the x87 FPU for x87-style programming. There are two
SSE3 instructions that use the general registers for thread synchronization. The
MXCSR register governs SIMD floating-point operations. Note, however, that the
x87FPU control word does not affect the SSE3 instruction that is executed by the x87
FPU (FISTTP), other than by unmasking an invalid operand or inexact result excep-
tion.
SSE4 instructions do not use MMX registers. Two of the SSE4.2 instructions operate
on general-purpose registers; the rest of SSE4.2 instruction and SSE4.1 instructions
operate on XMM registers.

12.1.1 SSE3, SSSE3, SSE4 in 64-Bit Mode and Compatibility Mode

In compatibility mode, SSE3, SSSE3, and SSE4 function like they do in protected
mode. In 64-bit mode, eight additional XMM registers are accessible. Registers
XMM8-XMM15 are accessed by using REX prefixes.
Memory operands are specified using the ModR/M, SIB encoding described in Section
3.7.5.
Some SSE3, SSSE3, and SSE4 instructions may be used to operate on general-
purpose registers. Use the REX.W prefix to access 64-bit general-purpose registers.
Note that if a REX prefix is used when it has no meaning, the prefix is ignored.

Vol. 1 12-1
PROGRAMMING WITH SSE3, SSSE3, AND SSE4

12.1.2 Compatibility of SSE3/SSSE3 with MMX Technology, the x87

FPU Environment, and SSE/SSE2 Extensions
SSE3, SSSE3, and SSE4 do not introduce any new state to the Intel 64 and IA-32
execution environments.
For SIMD and x87 programming, the FXSAVE and FXRSTOR instructions save and
restore the architectural states of XMM, MXCSR, x87 FPU, and MMX registers. The
MONITOR and MWAIT instructions use general purpose registers on input, they do
not modify the content of those registers.

12.1.3 Horizontal and Asymmetric Processing

Many SSE/SSE2/SSE3/SSSE3 instructions accelerate SIMD data processing using a
model referred to as vertical computation. Using this model, data flow is vertical
between the data elements of the inputs and the output.
Figure 12-1 illustrates the asymmetric processing of the SSE3 instruction
ADDSUBPD. Figure 12-2 illustrates the horizontal data movement of the SSE3
instruction HADDPD.

X1 X0

Y1 Y0

ADD SUB

X1 + Y1 X0 -Y0

Figure 12-1. Asymmetric Processing in ADDSUBPD

12-2 Vol. 1
 PROGRAMMING WITH SSE3, SSSE3, AND SSE4

X1 X0

Y1 Y0

ADD ADD

Y0 + Y1 X0 + X1

Figure 12-2. Horizontal Data Movement in HADDPD

12.2 OVERVIEW OF SSE3 INSTRUCTIONS

SSE3 extensions include 13 instructions. See:
• Section 12.3, “SSE3 Instructions,” provides an introduction to individual SSE3
instructions.
• Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volumes
2A & 2B, provide detailed information on individual instructions.
• Chapter 12, “System Programming for Streaming SIMD Instruction Sets,” in the
Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volume 3A,
gives guidelines for integrating SSE/SSE2/SSE3 extensions into an operating-
system environment.

12.3 SSE3 INSTRUCTIONS

SSE3 instructions are grouped as follows:
• x87 FPU instruction
— One instruction that improves x87 FPU floating-point to integer conversion
• SIMD integer instruction
— One instruction that provides a specialized 128-bit unaligned data load
• SIMD floating-point instructions
— Three instructions that enhance LOAD/MOVE/DUPLICATE performance
— Two instructions that provide packed addition/subtraction
— Four instructions that provide horizontal addition/subtraction

Vol. 1 12-3
PROGRAMMING WITH SSE3, SSSE3, AND SSE4

• Thread synchronization instructions

— Two instructions that improve synchronization between multi-threaded
agents
The instructions are discussed in more detail in the following paragraphs.

12.3.1 x87 FPU Instruction for Integer Conversion

The FISTTP instruction (x87 FPU Store Integer and Pop with Truncation) behaves like
FISTP, but uses truncation regardless of what rounding mode is specified in the x87
FPU control word. The instruction converts the top of stack (ST0) to integer with
rounding to and pops the stack.
The FISTTP instruction is available in three precisions: short integer (word or 16-bit),
integer (double word or 32-bit), and long integer (64-bit). With FISTTP, applications
no longer need to change the FCW when truncation is required.

12.3.2 SIMD Integer Instruction for Specialized 128-bit Unaligned

Data Load
The LDDQU instruction is a special 128-bit unaligned load designed to avoid cache
line splits. If the address of a 16-byte load is on a 16-byte boundary, LDQQU loads
the bytes requested. If the address of the load is not aligned on a 16-byte boundary,
LDDQU loads a 32-byte block starting at the 16-byte aligned address immediately
below the load request. It then extracts the requested 16 bytes.
The instruction provides significant performance improvement on 128-bit unaligned
memory accesses at the cost of some usage model restrictions.

12.3.3 SIMD Floating-Point Instructions That Enhance

LOAD/MOVE/DUPLICATE Performance
The MOVSHDUP instruction loads/moves 128-bits, duplicating the second and fourth
32-bit data elements.
• MOVSHDUP OperandA, OperandB
— OperandA (128 bits, four data elements): 3a, 2a, 1a, 0a
— OperandB (128 bits, four data elements): 3b, 2b, 1b, 0b
— Result (stored in OperandA): 3b, 3b, 1b, 1b
The MOVSLDUP instruction loads/moves 128-bits, duplicating the first and third
32-bit data elements.
• MOVSLDUP OperandA, OperandB
— OperandA (128 bits, four data elements): 3a, 2a, 1a, 0a

12-4 Vol. 1
 PROGRAMMING WITH SSE3, SSSE3, AND SSE4

— OperandB (128 bits, four data elements): 3b, 2b, 1b, 0b

— Result (stored in OperandA): 2b, 2b, 0b, 0b
The MOVDDUP instruction loads/moves 64-bits; duplicating the 64 bits from the
source.
• MOVDDUP OperandA, OperandB
— OperandA (128 bits, two data elements): 1a, 0a
— OperandB (64 bits, one data element): 0b
— Result (stored in OperandA): 0b, 0b

12.3.4 SIMD Floating-Point Instructions Provide Packed

Addition/Subtraction
The ADDSUBPS instruction has two 128-bit operands. The instruction performs
single-precision addition on the second and fourth pairs of 32-bit data elements
within the operands; and single-precision subtraction on the first and third pairs.
• ADDSUBPS OperandA, OperandB
— OperandA (128 bits, four data elements): 3a, 2a, 1a, 0a
— OperandB (128 bits, four data elements): 3b, 2b, 1b, 0b
— Result (stored in OperandA): 3a+3b, 2a-2b, 1a+1b, 0a-0b
The ADDSUBPD instruction has two 128-bit operands. The instruction performs
double-precision addition on the second pair of quadwords, and double-precision
subtraction on the first pair.
• ADDSUBPD OperandA, OperandB
— OperandA (128 bits, two data elements): 1a, 0a
— OperandB (128 bits, two data elements): 1b, 0b
— Result (stored in OperandA): 1a+1b, 0a-0b

12.3.5 SIMD Floating-Point Instructions Provide Horizontal

Addition/Subtraction
Most SIMD instructions operate vertically. This means that the result in position i is a
function of the elements in position i of both operands. Horizontal addition/subtrac-
tion operates horizontally. This means that contiguous data elements in the same
source operand are used to produce a result.
The HADDPS instruction performs a single-precision addition on contiguous data
elements. The first data element of the result is obtained by adding the first and
second elements of the first operand; the second element by adding the third and
fourth elements of the first operand; the third by adding the first and second

Vol. 1 12-5
PROGRAMMING WITH SSE3, SSSE3, AND SSE4

elements of the second operand; and the fourth by adding the third and fourth
elements of the second operand.
• HADDPS OperandA, OperandB
— OperandA (128 bits, four data elements): 3a, 2a, 1a, 0a
— OperandB (128 bits, four data elements): 3b, 2b, 1b, 0b
— Result (Stored in OperandA): 3b+2b, 1b+0b, 3a+2a, 1a+0a
The HSUBPS instruction performs a single-precision subtraction on contiguous data
elements. The first data element of the result is obtained by subtracting the second
element of the first operand from the first element of the first operand; the second
element by subtracting the fourth element of the first operand from the third element
of the first operand; the third by subtracting the second element of the second
operand from the first element of the second operand; and the fourth by subtracting
the fourth element of the second operand from the third element of the second
operand.
• HSUBPS OperandA, OperandB
— OperandA (128 bits, four data elements): 3a, 2a, 1a, 0a
— OperandB (128 bits, four data elements): 3b, 2b, 1b, 0b
— Result (Stored in OperandA): 2b-3b, 0b-1b, 2a-3a, 0a-1a
The HADDPD instruction performs a double-precision addition on contiguous data
elements. The first data element of the result is obtained by adding the first and
second elements of the first operand; the second element by adding the first and
second elements of the second operand.
• HADDPD OperandA, OperandB
— OperandA (128 bits, two data elements): 1a, 0a
— OperandB (128 bits, two data elements): 1b, 0b
— Result (Stored in OperandA): 1b+0b, 1a+0a
The HSUBPD instruction performs a double-precision subtraction on contiguous data
elements. The first data element of the result is obtained by subtracting the second
element of the first operand from the first element of the first operand; the second
element by subtracting the second element of the second operand from the first
element of the second operand.
• HSUBPD OperandA OperandB
— OperandA (128 bits, two data elements): 1a, 0a
— OperandB (128 bits, two data elements): 1b, 0b
— Result (Stored in OperandA): 0b-1b, 0a-1a

12-6 Vol. 1
 PROGRAMMING WITH SSE3, SSSE3, AND SSE4

12.3.6 Two Thread Synchronization Instructions

The MONITOR instruction sets up an address range that is used to monitor write-
back-stores.
MWAIT enables a logical processor to enter into an optimized state while waiting for
a write-back-store to the address range set up by MONITOR. MONITOR and MWAIT
require the use of general purpose registers for its input. The registers used by
MONITOR and MWAIT must be initialized properly; register content is not modified by
these instructions.

12.4 WRITING APPLICATIONS WITH SSE3 EXTENSIONS

The following sections give guidelines for writing application programs and oper-
ating-system code that use SSE3 instructions.

12.4.1 Guidelines for Using SSE3 Extensions

The following guidelines describe how to maximize the benefits of using SSE3 exten-
sions:
• Check that the processor supports SSE3 extensions.
— Application may need to ensure that the target operating system supports
SSE3. (Operating system support for the SSE extensions implies sufficient
support for SSE2 extensions and SSE3 extensions.)
• Ensure your operating system supports MONITOR and MWAIT.
• Employ the optimization and scheduling techniques described in the Intel® 64
and IA-32 Architectures Optimization Reference Manual (see Section 1.4,
“Related Literature”).

12.4.2 Checking for SSE3 Support

Before an application attempts to use the SIMD subset of SSE3 extensions, the appli-
cation should follow the steps illustrated in Section 11.6.2, “Checking for SSE/SSE2
Support.” Next, use the additional step provided below:
• Check that the processor supports the SIMD and x87 SSE3 extensions (if
CPUID.01H:ECX.SSE3[bit 0] = 1).
An operating systems that provides application support for SSE, SSE2 also provides
sufficient application support for SSE3. To use FISTTP, software only needs to check
support for SSE3.
In the initial implementation of MONITOR and MWAIT, these two instructions are
available to ring 0 and conditionally available at ring level greater than 0. Before an

Vol. 1 12-7
PROGRAMMING WITH SSE3, SSSE3, AND SSE4

application attempts to use the MONITOR and MWAIT instructions, the application
should use the following steps:
1. Check that the processor supports MONITOR and MWAIT. If
CPUID.01H:ECX.MONITOR[bit 3] = 1, MONITOR and MWAIT are available at
ring 0.
2. Query the smallest and largest line size that MONITOR uses. Use
CPUID.05H:EAX.smallest[bits 15:0];EBX.largest[bits15:0]. Values are returned
in bytes in EAX and EBX.
3. Ensure the memory address range(s) that will be supplied to MONITOR meets
memory type requirements.
MONITOR and MWAIT are targeted for system software that supports efficient thread
synchronization, See Chapter 12 in the Intel® 64 and IA-32 Architectures Software
Developer’s Manual, Volume 3A for details.

12.4.3 Enable FTZ and DAZ for SIMD Floating-Point Computation

Enabling the FTZ and DAZ flags in the MXCSR register is likely to accelerate SIMD
floating-point computation where strict compliance to the IEEE standard 754-1985 is
not required. The FTZ flag is available to Intel 64 and IA-32 processors that support
the SSE; DAZ is available to Intel 64 processors and to most IA-32 processors that
support SSE/SSE2/SSE3.
Software can detect the presence of DAZ, modify the MXCSR register, and save and
restore state information by following the techniques discussed in Section 11.6.3
through Section 11.6.6.

12.4.4 Programming SSE3 with SSE/SSE2 Extensions

SIMD instructions in SSE3 extensions are intended to complement the use of
SSE/SSE2 in programming SIMD applications. Application software that intends to
use SSE3 instructions should also check for the availability of SSE/SSE2 instructions.
The FISTTP instruction in SSE3 is intended to accelerate x87 style programming
where performance is limited by frequent floating-point conversion to integers; this
happens when the x87 FPU control word is modified frequently. Use of FISTTP can
eliminate the need to access the x87 FPU control word.

12.5 OVERVIEW OF SSSE3 INSTRUCTIONS

SSSE3 provides 32 instructions to accelerate a variety of multimedia and signal
processing applications employing SIMD integer data. See:
• Section 12.6, “SSSE3 Instructions,” provides an introduction to individual SSE3
instructions.

12-8 Vol. 1
 PROGRAMMING WITH SSE3, SSSE3, AND SSE4

• Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volumes

2A & 2B, provide detailed information on individual instructions.
• Chapter 12, “System Programming for Streaming SIMD Instruction Sets,” in the
Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volume 3A,
gives guidelines for integrating SSE/SSE2/SSE3/SSSE3 extensions into an
operating-system environment.

12.6 SSSE3 INSTRUCTIONS

SSSE3 instructions include:
• Twelve instructions that perform horizontal addition or subtraction operations.
• Six instructions that evaluate the absolute values.
• Two instructions that perform multiply and add operations and speed up the
evaluation of dot products.
• Two instructions that accelerate packed-integer multiply operations and produce
integer values with scaling.
• Two instructions that perform a byte-wise, in-place shuffle according to the
second shuffle control operand.
• Six instructions that negate packed integers in the destination operand if the
signs of the corresponding element in the source operand is less than zero.
• Two instructions that align data from the composite of two operands.
The operands of these instructions are packed integers of byte, word, or double word
sizes. The operands are stored as 64 or 128 bit data in MMX registers, XMM registers,
or memory.
The instructions are discussed in more detail in the following paragraphs.

12.6.1 Horizontal Addition/Subtraction

In analogy to the packed, floating-point horizontal add and subtract instructions in
SSE3, SSSE3 offers similar capabilities on packed integer data. Data elements of
signed words, doublewords are supported. Saturated version for horizontal add and
subtract on signed words are also supported. The horizontal data movement of
PHADD is shown in Figure 12-3.

Vol. 1 12-9
PROGRAMMING WITH SSE3, SSSE3, AND SSE4

X3 X2 X1 X0

Y3 Y2 Y1 Y0

ADD ADD ADD ADD

Y2 + Y3 Y0 + Y1 X2 + X3 X0 + X1

Figure 12-3. Horizontal Data Movement in PHADDD

There are six horizontal add instructions (represented by three mnemonics); three
operate on 128-bit operands and three operate on 64-bit operands. The width of
each data element is either 16 bits or 32 bits. The mnemonics are listed below.
• PHADDW adds two adjacent, signed 16-bit integers horizontally from the source
and destination operands and packs the signed 16-bit results to the destination
operand.
• PHADDSW adds two adjacent, signed 16-bit integers horizontally from the source
and destination operands and packs the signed, saturated 16-bit results to the
destination operand.
• PHADDD adds two adjacent, signed 32-bit integers horizontally from the source
and destination operands and packs the signed 32-bit results to the destination
operand.
There are six horizontal subtract instructions (represented by three mnemonics);
three operate on 128-bit operands and three operate on 64-bit operands. The width
of each data element is either 16 bits or 32 bits. These are listed below.
• PHSUBW performs horizontal subtraction on each adjacent pair of 16-bit signed
integers by subtracting the most significant word from the least significant word
of each pair in the source and destination operands. The signed 16-bit results are
packed and written to the destination operand.
• PHSUBSW performs horizontal subtraction on each adjacent pair of 16-bit signed
integers by subtracting the most significant word from the least significant word
of each pair in the source and destination operands. The signed, saturated 16-bit
results are packed and written to the destination operand.
• PHSUBD performs horizontal subtraction on each adjacent pair of 32-bit signed
integers by subtracting the most significant doubleword from the least significant

12-10 Vol. 1
 PROGRAMMING WITH SSE3, SSSE3, AND SSE4

double word of each pair in the source and destination operands. The signed
32-bit results are packed and written to the destination operand.

12.6.2 Packed Absolute Values

There are six packed-absolute-value instructions (represented by three mnemonics).
Three operate on 128-bit operands and three operate on 64-bit operands. The widths
of data elements are 8 bits, 16 bits or 32 bits. The absolute value of each data
element of the source operand is stored as an UNSIGNED result in the destination
operand.
• PABSB computes the absolute value of each signed byte data element.
• PABSW computes the absolute value of each signed 16-bit data element.
• PABSD computes the absolute value of each signed 32-bit data element.

12.6.3 Multiply and Add Packed Signed and Unsigned Bytes

There are two multiply-and-add-packed-signed-unsigned-byte instructions (repre-
sented by one mnemonic). One operates on 128-bit operands and the other operates
on 64-bit operands. Multiplications are performed on each vertical pair of data
elements. The data elements in the source operand are signed byte values, the input
data elements of the destination operand are unsigned byte values.
• PMADDUBSW multiplies each unsigned byte value with the corresponding signed
byte value to produce an intermediate, 16-bit signed integer. Each adjacent pair
of 16-bit signed values are added horizontally. The signed, saturated 16-bit
results are packed to the destination operand.

12.6.4 Packed Multiply High with Round and Scale

There are two packed-multiply-high-with-round-and-scale instructions (represented
by one mnemonic). One operates on 128-bit operands and the other operates on
64-bit operands.
• PMULHRSW multiplies vertically each signed 16-bit integer from the destination
operand with the corresponding signed 16-bit integer of the source operand,
producing intermediate, signed 32-bit integers. Each intermediate 32-bit integer
is truncated to the 18 most significant bits. Rounding is always performed by
adding 1 to the least significant bit of the 18-bit intermediate result. The final
result is obtained by selecting the 16 bits immediately to the right of the most
significant bit of each 18-bit intermediate result and packed to the destination
operand.

Vol. 1 12-11
PROGRAMMING WITH SSE3, SSSE3, AND SSE4

12.6.5 Packed Shuffle Bytes

There are two packed-shuffle-bytes instructions (represented by one mnemonic).
One operates on 128-bit operands and the other operates on 64-bit operands. The
shuffle operations are performed bytewise on the destination operand using the
source operand as a control mask.
• PSHUFB permutes each byte in place, according to a shuffle control mask. The
least significant three or four bits of each shuffle control byte of the control mask
form the shuffle index. The shuffle mask is unaffected. If the most significant bit
(bit 7) of a shuffle control byte is set, the constant zero is written in the result
byte.

12.6.6 Packed Sign

There are six packed-sign instructions (represented by three mnemonics). Three
operate on 128-bit operands and three operate on 64-bit operands. The widths of
each data element for these instructions are 8 bit, 16 bit or 32 bit signed integers.
• PSIGNB/W/D negates each signed integer element of the destination operand if
the sign of the corresponding data element in the source operand is less than
zero.

12.6.7 Packed Align Right

There are two packed-align-right instructions (represented by one mnemonic). One
operates on 128-bit operands and the other operates on 64-bit operands. These
instructions concatenate the destination and source operand into a composite, and
extract the result from the composite according to an immediate constant.
• PALIGNR’s source operand is appended after the destination operand forming an
intermediate value of twice the width of an operand. The result is extracted from
the intermediate value into the destination operand by selecting the 128-bit or
64-bit value that are right-aligned to the byte offset specified by the immediate
value.

12.7 WRITING APPLICATIONS WITH SSSE3 EXTENSIONS

The following sections give guidelines for writing application programs and oper-
ating-system code that use SSSE3 instructions.

12.7.1 Guidelines for Using SSSE3 Extensions

The following guidelines describe how to maximize the benefits of using SSSE3
extensions:

12-12 Vol. 1
 PROGRAMMING WITH SSE3, SSSE3, AND SSE4

• Check that the processor supports SSSE3 extensions.

• Ensure that your operating system supports SSE/SSE2/SSE3/SSSE3 extensions.
(Operating system support for the SSE extensions implies sufficient support for
SSE2, SSE3, and SSSE3.)
• Employ the optimization and scheduling techniques described in the Intel® 64
and IA-32 Architectures Optimization Reference Manual (see Section 1.4,
“Related Literature”).

12.7.2 Checking for SSSE3 Support

Before an application attempts to use the SSSE3 extensions, the application should
follow the steps illustrated in Section 11.6.2, “Checking for SSE/SSE2 Support.”
Next, use the additional step provided below:
• Check that the processor supports SSSE3 (if CPUID.01H:ECX.SSSE3[bit 9] = 1).

12.8 SSE3/SSSE3 AND SSE4 EXCEPTIONS

SSE3, SSSE3, and SSE4 instructions can generate the same type of memory-access
and non-numeric exceptions as other Intel 64 or IA-32 instructions. Existing excep-
tion handlers generally handle these exceptions without code modification.
FISTTP can generate floating-point exceptions. Some SSE3 instructions can also
generate SIMD floating-point exceptions.
SSE3 additions and changes are noted in the following sections. See also: Section
11.5, “SSE, SSE2, and SSE3 Exceptions”.

12.8.1 Device Not Available (DNA) Exceptions

SSE3, SSSE3, and SSE4 will cause a DNA Exception (#NM) if the processor attempts
to execute an SSE3 instruction while CR0.TS[bit 3] = 1. If
CPUID.01H:ECX.SSE3[bit 0] = 0, execution of an SSE3 extension will cause an
invalid opcode fault regardless of the state of CR0.TS[bit 3].
Similarly, an attempt to execute an SSSE3 instruction on a processor that reports
CPUID.01H:ECX.SSSE3[bit 9] = 0 will cause an invalid opcode fault regardless of the
state of CR0.TS[bit 3]. An attempt to execute an SSE4.1 instruction on a processor
that reports CPUID.01H:ECX.SSE4_1[bit 19] = 0 will cause an invalid opcode fault
regardless of the state of CR0.TS[bit 3].
An attempt to execute PCMPGTQ or any one of the four string processing instructions
in SSE4.2 on a processor that reports CPUID.01H:ECX.SSSE3[bit 20] = 0 will cause
an invalid opcode fault regardless of the state of CR0.TS[bit 3]. CRC32 and POPCNT
do not cause #NM.

Vol. 1 12-13
PROGRAMMING WITH SSE3, SSSE3, AND SSE4

12.8.2 Numeric Error flag and IGNNE#

Most SSE3 instructions ignore CR0.NE[bit 5] (treats it as if it were always set) and
the IGNNE# pin. With one exception, all use the vector 19 software exception for
error reporting. The exception is FISTTP; it behaves like other x87-FP instructions.
SSSE3 instructions ignore CR0.NE[bit 5] (treats it as if it were always set) and the
IGNNE# pin.
SSSE3 instructions do not cause floating-point errors. Floating-point numeric errors
for SSE4.1 are described in Section 12.8.4. SSE4.2 instructions do not cause
floating-point errors.

12.8.3 Emulation
CR0.EM is used by some software to emulate x87 floating-point instructions,
CR0.EM[bit 2] cannot be used for emulation of SSE, SSE2, SSE3, SSSE3, and SSE4.
If an SSE3, SSSE3, and SSE4 instruction executes with CR0.EM[bit 2] set, an invalid
opcode exception (INT 6) is generated instead of a device not available exception
(INT 7).

12.8.4 IEEE 754 Compliance of SSE4.1 Floating-Point Instructions

The six SSE4.1 instructions that perform floating-point arithmetic are:
• DPPS
• DPPD
• ROUNDPS
• ROUNDPD
• ROUNDSS
• ROUNDSD
Dot Product operations are not specified in IEEE-754. When neither FTZ nor DAZ are
enabled, the dot product instructions resemble sequences of IEEE-754 multiplies and
adds (with rounding at each stage), except that the treatment of input NaN’s is
implementation specific (there will be at least one NaN in the output). The input
select fields (bits imm8[4:7]) force input elements to +0.0f prior to the first multiply
and will suppress input exceptions that would otherwise have been be generated.
As a convenience to the exception handler, any exceptions signaled from DPPS or
DPPD leave the destination unmodified.
Round operations signal invalid and precision only.

12-14 Vol. 1
 PROGRAMMING WITH SSE3, SSSE3, AND SSE4

Table 12-1. SIMD numeric exceptions signaled by SSE4.1

DPPS DPPD ROUNDPS ROUNDPD
ROUNDSS ROUNDSD
Overflow X X
Underflow X X
Invalid X X X (1) X (1)
Inexact Precision X X X (2)
X (2)
Denormal X X
NOTE:
1. Invalid is signaled only if Src = SNaN.
2. Precision is ignored (regardless of the MXCSR precision mask) if if imm8[3] = ‘1’.

The other SSE4.1 instructions with floating-point arguments (BLENDPS, BLENDPD,

BLENDVPS, BLENDVPD, INSERTPS, EXTRACTPS) do not signal any SIMD numeric
exceptions.

12.9 SSE4 OVERVIEW

SSE4 comprises of two sets of extensions: SSE4.1 and SSE4.2. SSE4.1 is targeted to
improve the performance of media, imaging, and 3D workloads. SSE4.1 adds
instructions that improve compiler vectorization and significantly increase support
for packed dword computation. The technology also provides a hint that can improve
memory throughput when reading from uncacheable WC memory type.
The 47 SSE4.1 instructions include:
• Two instructions perform packed dword multiplies.
• Two instructions perform floating-point dot products with input/output selects.
• One instruction performs a load with a streaming hint.
• Six instructions simplify packed blending.
• Eight instructions expand support for packed integer MIN/MAX.
• Four instructions support floating-point round with selectable rounding mode and
precision exception override.
• Seven instructions improve data insertion and extractions from XMM registers
• Twelve instructions improve packed integer format conversions (sign and zero
extensions).
• One instruction improves SAD (sum absolute difference) generation for small
block sizes.
• One instruction aids horizontal searching operations.

Vol. 1 12-15
PROGRAMMING WITH SSE3, SSSE3, AND SSE4

• One instruction improves masked comparisons.

• One instruction adds qword packed equality comparisons.
• One instruction adds dword packing with unsigned saturation.
The seven SSE4.2 instructions improve performance in the following areas:
• String and text processing that can take advantage of single-instruction multiple-
data programming techniques.
• Application-targeted accelerator (ATA) instructions.
• A SIMD integer instruction that enhances the capability of the 128-bit integer
SIMD capability in SSE4.1.

12.10 SSE4.1 INSTRUCTION SET

12.10.1 Dword Multiply Instructions

SSE4.1 adds two dword multiply instructions that aid vectorization. They allow four
simultaneous 32 bit by 32 bit multiplies. PMULLD returns a low 32-bits of the result
and PMULDQ returns a 64-bit signed result. These represent the most common
integer multiply operation. See Table 12-2.

Table 12-2. Enhanced 32-bit SIMD Multiply Supported by SSE4.1

32 bit Integer Operation
unsigned x unsigned signed x signed
Low 32-bit (not available) PMULLD
Result

High 32-bit (not available) (not available)

64-bit PMULUDQ* PMULDQ
NOTE:
* Available prior to SSE4.1.

12.10.2 Floating-Point Dot Product Instructions

SSE4.1 adds two instructions for double-precision (for up to 2 elements; DPPD) and
single-precision dot products (for up to 4 elements; DPPS).
These dot-product instructions include source select and destination broadcast which
generally improves the flexibility. For example, a single DPPS instruction can be used
for a 2, 3, or 4 element dot product.

12-16 Vol. 1
 PROGRAMMING WITH SSE3, SSSE3, AND SSE4

12.10.3 Streaming Load Hint Instruction

Historically, CPU read accesses of WC memory type regions have significantly lower
throughput than accesses to cacheable memory.
The streaming load instruction in SSE4.1, MOVNTDQA, provides a non-temporal hint
that can cause adjacent 16-byte items within an aligned 64-byte region of WC
memory type (a streaming line) to be fetched and held in a small set of temporary
buffers (“streaming load buffers”). Subsequent streaming loads to other aligned 16-
byte items in the same streaming line may be satisfied from the streaming load
buffer and can improve throughput.
Programmers are advised to use the following practices to improve the efficiency of
MOVNTDQA streaming loads from WC memory:
• Streaming loads must be 16-byte aligned.
• Temporally group streaming loads of the same streaming cache line for effective
use of the small number of streaming load buffers. If loads to the same streaming
line are excessively spaced apart, it may cause the streaming line to be re-
fetched from memory.
• Temporally group streaming loads from at most a few streaming lines together.
The number of streaming load buffers is small; grouping a modest number of
streams will avoid running out of streaming load buffers and the resultant re-
fetching of streaming lines from memory.
• Avoid writing to a streaming line until all 16-byte-aligned reads from the
streaming line have occurred. Reading a 16-byte item from a streaming line that
has been written, may cause the streaming line to be re-fetched.
• Avoid reading a given 16-byte item within a streaming line more than once;
repeated loads of a particular 16-byte item are likely to cause the streaming line
to be re-fetched.
• The streaming load buffers, reflecting the WC memory type characteristics, are
not required to be snooped by operations from other agents. Software should not
rely upon such coherency actions to provide any data coherency with respect to
other logical processors or bus agents. Rather, software must insure the
consistency of WC memory accesses between producers and consumers.
• Streaming loads may be weakly ordered and may appear to software to execute
out of order with respect to other memory operations. Software must explicitly
use fences (e.g. MFENCE) if it needs to preserve order among streaming loads or
between streaming loads and other memory operations.
• Streaming loads must not be used to reference memory addresses that are
mapped to I/O devices having side effects or when reads to these devices are
destructive. This is because MOVNTDQA is speculative in nature.
Example 12-1 and Example 12-2 give two sketches of the basic assembly sequences
that illustrate the principles of using MOVNTDQA in a situation of a pair of producer-
consumer accessing a WC memory region.

Vol. 1 12-17
PROGRAMMING WITH SSE3, SSSE3, AND SSE4

Example 12-1. Sketch of MOVNTDQA Usage of a Consumer and a PCI Producer

// P0: producer is a PCI device writing into the WC space
# the PCI device updates status through a UC flag, "u_dev_status" .
# the protocol for "u_dev_status" : 0: produce; 1: consume; 2: all done

mov eax, $0
mov [u_dev_status], eax
producerStart:
mov eax, [u_dev_status] # poll status flag to see if consumer is requestion data
cmp eax, $0 #
jne done # I no longer need to produce
commence PCI writes to WC region..

mov eax, $1 # producer ready to notify the consumer via status flag
mov [u_dev_status], eax
# now wait for consumer to signal its status
spinloop:
cmp [u_dev_status], $1 # did I get a signal from the consumer ?
jne producerStart # yes I did
jmp spinloop # check again
done:
// producer is finished at this point

12-18 Vol. 1
 PROGRAMMING WITH SSE3, SSSE3, AND SSE4

Example 12-1. Sketch of MOVNTDQA Usage of a Consumer and a PCI Producer (Contd.)
// P1: consumer check PCI status flag to consume WC data
mov eax, $0 # request to the producer
mov [u_dev_status], eax
consumerStart:
mov; eax, [u_dev_status] # reads the value of the PCI status
cmp eax, $1 # has producer written
jne consumerStart # tight loop; make it more efficient with pause, etc.
mfence # producer finished device writes to WC, ensure WC region is coherent
ntread:
movntdqa xmm0, [addr]
movntdqa xmm1, [addr + 16]
movntdqa xmm2, [addr + 32]
movntdqa xmm3, [addr + 48]
… # do any more NT reads as needed
mfence # ensure PCI device reads the correct value of [u_dev_status]
# now decide whether we are done or we need the producer to produce more data
# if we are done write a 2 into the variable, otherwise write a 0 into the variable
mov eax, $0/$2 # end or continue producing
mov [u_dev_status], eax
# if I want to consume again I will jump back to consumerStart after storing a 0 into eax
# otherwise I am done

Vol. 1 12-19
PROGRAMMING WITH SSE3, SSSE3, AND SSE4

Example 12-2. Sketch of MOVNTDQA Usage of Producer-Consumer Threads

// P0: producer writes into the WC space
# xchg is an implicitly locked operation.

producerStart:
# We use a locked operation to prevent any races between the producer and the consumer
# updating this variable. Assume initial value is 0
mov eax, $0
xchg eax, [signalVariable] # signalVariable is used for communicating
cmp eax, $0 # am I supposed to be writing for the consumer
jne done # I no longer need to produce
movntdq [addr1], xmm0 # producer writes the data
movntdq [addr2], xmm1 # ..
.
# We will again use a locked instruction. Serves 2 purposes. Updated value signals to the consumer
and
# the serialization of the lock flushes all the WC stores to memory
mov eax, $1
xchg [signalVariable], eax # signal to the consumer
# For simplicity, we show a spin loop, more efficient spin loop can be done using PAUSE
spinloop:
cmp [signalVariable], $1 # did I get a signal from the consumer ?
jne producerStart # yes I did
jmp spinloop # check again
done:
// producer is finished at this point

12-20 Vol. 1
 PROGRAMMING WITH SSE3, SSSE3, AND SSE4

Example 12-2. Sketch of MOVNTDQA Usage of Producer-Consumer Threads (Contd.)

// P1: consumer reads from write combining space
mov eax, $0
consumerStart:
lock; xadd [signalVariable], eax # reads the value of the signal variable in
cmp eax, $1 # has producer written to signal its state?
jne consumerStart # simple loop; replace with PAUSE to make it more efficient.
# read the data from the WC memory space with MOVNTDQA to achieve higher throughput
ntread: # keep reads from the same cache line as close together as possible
movntdqa xmm0, [addr]
movntdqa xmm1, [addr + 16]
movntdqa xmm2, [addr + 32]
movntdqa xmm3, [addr + 48]
# since a lock prevents younger MOVNTDQA from passing it, the
# above non temporal loads will happen only after the producer has signaled
… # do any more NT reads as needed

# now decide whether we are done or we need the producer to produce more data
# if we are done write a 2 into the variable, otherwise write a 0 into the variable
mov eax, $0/$2 # end or continue producing
xchg [signalVariable], eax
# if I want to consume again I will jump back to consumerStart after storing a 0 into eax
# otherwise I am done

12.10.4 Packed Blending Instructions

SSE4.1 adds 6 instructions used for blending (BLENDPS, BLENDPD, BLENDVPS,
BLENDVPD, PBLENDVB, PBLENDW).
Blending conditionally copies a data element in a source operand to the same
element in the destination. SSE4.1 instructions improve blending operations for most
field sizes. A single new SSE4.1 instruction can generally replace a sequence of 2 to
4 operations using previous architectures.
The variable blend instructions (BLENDVPS, PBLENDVPD, PBLENDW) introduce the
use of control bits stored in an implicit XMM register (XMM0). The most significant bit
in each field (the sign bit, for 2’s compliment integer or floating-point) is used as a
selector. See Table 12-3.

Vol. 1 12-21
PROGRAMMING WITH SSE3, SSSE3, AND SSE4

Table 12-3. Blend Field Size and Control Modes Supported by SSE4.1
Packed Packed
Double Single Packed Packed Packed Packed Blend
Instructions FP FP QWord DWord Word Byte Control
BLENDPS X Imm8
BLENDPD X Imm8
BLENDVPS X X(1) XMM0
BLENDVPD X X(1) XMM0
(2) (2) (2)
PBLENDVB X XMM0
PBLENDW X X X Imm8
NOTE:
1. Use of floating-point SIMD instructions on integer data types may incur performance penalties.
2. Byte variable blend can be used for larger sized fields by reformatting (or shuffling) the blend
control.

12.10.5 Packed Integer MIN/MAX Instructions

SSE4.1 adds 8 packed integer MIN and MAX instructions (PMINUW, PMINUD,
PMINSB, PMINSD; PMAXUW, PMAXUD, PMAXSB, PMAXSD).
Four 32-bit integer packed MIN and MAX instructions operate on unsigned and signed
dwords. Two instructions operate on signed bytes. Two instructions operate on
unsigned words. See Table 12-4.

Table 12-4. Enhanced SIMD Integer MIN/MAX Instructions Supported by SSE4.1

Integer Width
Byte Word DWord
Integer PMINUB* PMINUW PMINUD
Format Unsigned PMAXUB* PMAXUW PMAXUD
PMINSB PMINSW* PMINSD
Signed PMAXSB PMAXSW* PMAXSD
NOTE:
* Available prior to SSE4.1.

12-22 Vol. 1
 PROGRAMMING WITH SSE3, SSSE3, AND SSE4

12.10.6 Floating-Point Round Instructions with Selectable Rounding

Mode
High level languages and libraries often expose rounding operations having a variety
of numeric rounding and exception behaviors. Using SSE/SSE2/SSE3 instructions to
mitigate the rounding-mode-related problem is sometimes not straight forward.
SSE4.1 introduces four rounding instructions (ROUNDPS, ROUNDPD, ROUNDSS,
ROUNDSD) that cover scalar and packed single- and double-precision floating-point
operands. The rounding mode can be selected using an immediate from one of the
IEEE-754 modes (Nearest, -Inf, +Inf, and Truncate) without changing the current
rounding mode; or the the instruction can be forced to use the current rounding
mode. Another bit in the immediate is used to suppress inexact precision exceptions.
Rounding instructions in SSE4.1 generally permit single-instruction solutions to C99
functions ceil(), floor(), trunc(), rint(), nearbyint(). These instructions simplify the
implementations of half-way-away-from-zero rounding modes as used by C99
round() and F90’s nint().

12.10.7 Insertion and Extractions from XMM Registers

SSE4.1 adds 7 instructions (corresponding to 9 assembly instruction mnemonics)
that simplify data insertion and extraction between general-purpose register (GPR)
and XMM registers (EXTRACTPS, INSERTPS, PINSRB, PINSRD, PINSRQ, PEXTRB,
PEXTRW, PEXTRD, and PEXTRQ). When accessing memory, no alignment is required
for any of these instructions (unless alignment checking is enabled).
EXTRACTPS extracts a single-precision floating-point value from any dword offset in
an XMM register and stores the result to memory or a general-purpose register.
INSERTPS inserts a single floating-point value from either a 32-bit memory location
or from specified element in an XMM register to a selected element in the destination
XMM register. In addition, INSERTPS allows the insertion of +0.0f into any destina-
tion elements using a mask.
PINSRB, PINSRD, and PINSRQ insert byte, dword, or qword integer values from a
register or memory into an XMM register. Insertion of integer word values were
already supported by SSE2 (PINSRW).
PEXTRB, PEXTRW, PEXTRD, and PEXTRQ extract byte, word, dword, and qword from
an XMM register and insert the values into a general-purpose register or memory.

12.10.8 Packed Integer Format Conversions

A common type of operation on packed integers is the conversion by zero- or sign-
extension of packed integers into wider data types. SSE4.1 adds 12 instructions that
convert from a smaller packed integer type to a larger integer type (PMOVSXBW,
PMOVZXBW, PMOVSXBD, PMOVZXBD, PMOVSXWD, PMOVZXWD, PMOVSXBQ,
PMOVZXBQ, PMOVSXWQ, PMOVZXWQ, PMOVSXDQ, PMOVZXDQ).

Vol. 1 12-23
PROGRAMMING WITH SSE3, SSSE3, AND SSE4

The source operand is from either an XMM register or memory; the destination is an
XMM register. See Table 12-5.
When accessing memory, no alignment is required for any of the instructions unless
alignment checking is enabled. In which case, all conversions must be aligned to the
width of the memory reference. The number of elements converted (and width of
memory reference) is illustrated in Table 12-6. The alignment requirement is shown
in parenthesis.

Table 12-5. New SIMD Integer conversions supported by SSE4.1

Source Type
Byte Word Dword
Signed Word PMOVSXBW
Unsigned Word PMOVZXBW
Destination

Signed Dword PMOVSXBD PMOVSXWD

Unsigned Dword PMOVZXBD PMOVZXWD
Type

Signed Qword PMOVSXBQ PMOVSXWQ PMOVSXDQ

Unsigned Qword PMOVZXBQ PMOVZXWQ PMOVZXDQ

Table 12-6. New SIMD Integer Conversions Supported by SSE4.1

Source Type
Byte Word Dword
Word 8 (64 bits)
Destination

Dword 4 (32 bits) 4 (64 bits)

Qword 2 (16 bits) 2 (32 bits) 2 (64 bits)
Type

12.10.9 Improved Sums of Absolute Differences (SAD) for 4-Byte

Blocks
SSE4.1 adds an instruction (MPSADBW) that performs eight 4-byte wide SAD opera-
tions per instruction to produce eight results. Compared to PSADBW, MPSADBW
operates on smaller chunks (4-byte instead of 8-byte chunks); this makes the
instruction better suited to video coding standards such as VC.1 and H.264.
MPSADBW performs four times the number of absolute difference operations than
that of PSADBW (per instruction). This can improve performance for dense motion
searches.
MPSADBW uses a 4-byte wide field from a source operand; the offset of the 4-byte
field within the 128-bit source operand is specified by two immediate control bits.
MPSADBW produces eight 16-bit SAD results. Each 16-bit SAD result is formed from

12-24 Vol. 1
 PROGRAMMING WITH SSE3, SSSE3, AND SSE4

overlapping pairs of 4 bytes in the destination with the 4-byte field from the source
operand. MPSADBW uses eleven consecutive bytes in the destination operand, its
offset is specified by a control bit in the immediate byte (i.e. the offset can be from
byte 0 or from byte 4). Figure 12-4 illustrates the operation of MPSADBW. MPSADBW
can simplify coding of dense motion estimation by providing source and destination
offset control, higher throughput of SAD operations, and the smaller chunk size.

Imm[1:0]*32
127 96 64 0

Source Abs. Diff. Imm[2]*32

Destination
Sum

127 16 0

Figure 12-4. MPSADBW Operation

12.10.10 Horizontal Search

SSE4.1 adds a search instruction (PHMINPOSUW) that finds the value and location of
the minimum unsigned word from one of 8 horizontally packed unsigned words. The
resulting value and location (offset within the source) are packed into the low dword
of the destination XMM register.
Rapid search is often a significant component of motion estimation. MPSADBW and
PHMINPOSUW can be used together to improve video encode.

12.10.11 Packed Test

The packed test instruction PTEST is similar to a 128-bit equivalent to the legacy
instruction TEST. With PTEST, the source argument is typically used like a bit mask.
PTEST performs a logical AND between the destination with this mask and sets the ZF
flag if the result is zero. The CF flag (zero for TEST) is set if the inverted mask AND’d
with the destination is all zero. Because the destination is not modified, PTEST
simplifies branching operations (such as branching on signs of packed floating-point
numbers, or branching on zero fields).

Vol. 1 12-25
PROGRAMMING WITH SSE3, SSSE3, AND SSE4

12.10.12 Packed Qword Equality Comparisons

SSE4.1 adds a 128-bit packed qword equality test. The new instruction (PCMPEQQ)
is identical to PCMPEQD, but has qword granularity.

12.10.13 Dword Packing With Unsigned Saturation

SSE4.1 adds a new instruction PACKUSDW to complete the set of small integer pack
instructions in the family of SIMD instruction extensions. PACKUSDW packs dword to
word with unsigned saturation. See Table 12-7 for the complete set of packing
instructions for small integers.

Table 12-7. Enhanced SIMD Pack support by SSE4.1

Pack Type
DWord -> word Word -> Byte
Unsigned PACKUSDW (new!) PACKUSWB
Saturation

Signed PACKSSDW PACKSSWB

Type

12.11 SSE4.2 INSTRUCTION SET

Five of the seven SSE4.2 instructions can use an XMM register as a source or desti-
nation. These include four text/string processing instructions and one packed quad-
word compare SIMD instruction. Programming these five SSE4.2 instructions is
similar to programming 128-bit Integer SIMD in SSE2/SSSE3. SSE4.2 does not
provide any 64-bit integer SIMD instructions.
The remaining two SSE4.2 instructions uses general-purpose registers to perform
accelerated processing functions in specific application areas.

12.11.1 String and Text Processing Instructions

String and text processing instructions in SSE4.2 allocates 4 opcodes to provide a
rich set of string and text processing capabilities that traditionally required many
more opcodes. These 4 instructions use XMM registers to process string or text
elements of up to 128-bits (16 bytes or 8 words). Each instruction uses an immediate
byte to support a rich set of programmable controls. A string-processing SSE4.2
instruction returns the result of processing each pair of string elements using either
an index or a mask.
The capabilities of the string/text processing instructions include:

12-26 Vol. 1
 PROGRAMMING WITH SSE3, SSSE3, AND SSE4

• Handling string/text fragments consisting of bytes or words, either signed or

unsigned
• Support for partial string or fragments less than 16 bytes in length, using either
explicit length or implicit null-termination
• Four types of string compare operations on word/byte elements
• Up to 256 compare operations performed in a single instruction on all string/text
element pairs
• Built-in aggregation of intermediate results from comparisons
• Programmable control of processing on intermediate results
• Programmable control of output formats in terms of an index or mask
• Bi-directional support for the index format
• Support for two mask formats: bit or natural element width
• Not requiring 16-byte alignment for memory operand
The four SSE4.2 instructions that process text/string fragments are:
• PCMPESTRI — Packed compare explicit-length strings, return index in ECX/RCX
• PCMPESTRM — Packed compare explicit-length strings, return mask in XMM0
• PCMPISTRI — Packed compare implicit-length strings, return index in ECX/RCX
• PCMPISTRM — Packed compare implicit-length strings, return mask in XMM0
All four require the use of an immediate byte to control operation. The two source
operands can be XMM registers or a combination of XMM register and memory
address. The immediate byte provides programmable control with the following
attributes:
• Input data format
• Compare operation mode
• Intermediate result processing
• Output selection
Depending on the output format associated with the instruction, the text/string
processing instructions implicitly uses either a general-purpose register (ECX/RCX)
or an XMM register (XMM0) to return the final result.
Two of the four text-string processing instructions specify string length explicitly.
They use two general-purpose registers (EDX, EAX) to specify the number of valid
data elements (either word or byte) in the source operands. The other two instruc-
tions specify valid string elements using null termination. A data element is consid-
ered valid only if it has a lower index than the least significant null data element.

12.11.1.1 Memory Operand Alignment

The text and string processing instructions in SSE4.2 do not perform alignment
checking on memory operands. This is different from most other 128-bit SIMD

Vol. 1 12-27
PROGRAMMING WITH SSE3, SSSE3, AND SSE4

instructions accessing the XMM registers. The absence of an alignment check for
these four instructions does not imply any modification to the existing definitions of
other instructions.

12.11.2 Packed Comparison SIMD Integer Instruction

SSE4.2 also provides a 128-bit integer SIMD instruction PCMPGTQ that performs
logical compare of greater-than on packed integer quadwords.

12.11.3 Application-Targeted Accelerator Instructions

There are two application-targeted accelerator instructions in SSE4.2:
• CRC32 — Provides hardware acceleration to calculate cyclic redundancy checks
for fast and efficient implementation of data integrity protocols.
• POPCNT — Accelerates software performance in the searching of bit patterns.

12.12 WRITING APPLICATIONS WITH SSE4 EXTENSIONS

12.12.1 Guidelines for Using SSE4 Extensions

The following guidelines describe how to maximize the benefits of using SSE4 exten-
sions:
• Check that the processor supports SSE4 extensions.
• Ensure that your operating system supports SSE/SSE2/SSE3/SSSE3 extensions.
(Operating system support for the SSE extensions implies sufficient support for
SSE2, SSE3, SSSE3, and SSE4.)
• Employ the optimization and scheduling techniques described in the Intel® 64
and IA-32 Architectures Optimization Reference Manual (see Section 1.4,
“Related Literature”).

12.12.2 Checking for SSE4.1 Support

Before an application attempts to use SSE4.1 instructions, the application should
follow the steps illustrated in Section 11.6.2, “Checking for SSE/SSE2 Support.”
Next, use the additional step provided below:
Check that the processor supports SSE4.1 (if CPUID.01H:ECX.SSE4_1[bit 19] = 1),
SSE3 (if CPUID.01H:ECX.SSE3[bit 0] = 1), and SSSE3 (if CPUID.01H:ECX.SSSE3[bit
9] = 1).

12-28 Vol. 1
 PROGRAMMING WITH SSE3, SSSE3, AND SSE4

12.12.3 Checking for SSE4.2 Support

Before an application attempts to use the following SSE4.2 instructions:
PCMPESTRI/PCMPESTRM/PCMPISTRI/PCMPISTRM, PCMPGTQ;the application should
follow the steps illustrated in Section 11.6.2, “Checking for SSE/SSE2 Support.”
Next, use the additional step provided below:
Check that the processor supports SSE4.2 (if CPUID.01H:ECX.SSE4_2[bit 20] = 1),
SSE4.1 (if CPUID.01H:ECX.SSE4_1[bit 19] = 1), and SSSE3 (if
CPUID.01H:ECX.SSSE3[bit 9] = 1).
Before an application attempts to use the CRC32 instruction, it must check that the
processor supports SSE4.2 (if CPUID.01H:ECX.SSE4_2[bit 20] = 1).
Before an application attempts to use the POPCNT instruction, it must check that the
processor supports SSE4.2 (if CPUID.01H:ECX.POPCNT[bit 23] = 1).

Vol. 1 12-29
PROGRAMMING WITH SSE3, SSSE3, AND SSE4

12-30 Vol. 1
 CHAPTER 13
INPUT/OUTPUT

In addition to transferring data to and from external memory, IA-32 processors can
also transfer data to and from input/output ports (I/O ports). I/O ports are created in
system hardware by circuity that decodes the control, data, and address pins on the
processor. These I/O ports are then configured to communicate with peripheral
devices. An I/O port can be an input port, an output port, or a bidirectional port.
Some I/O ports are used for transmitting data, such as to and from the transmit and
receive registers, respectively, of a serial interface device. Other I/O ports are used
to control peripheral devices, such as the control registers of a disk controller.
This chapter describes the processor’s I/O architecture. The topics discussed include:
• I/O port addressing
• I/O instructions
• I/O protection mechanism

13.1 I/O PORT ADDRESSING

The processor permits applications to access I/O ports in either of two ways:
• Through a separate I/O address space
• Through memory-mapped I/O
Accessing I/O ports through the I/O address space is handled through a set of I/O
instructions and a special I/O protection mechanism. Accessing I/O ports through
memory-mapped I/O is handled with the processors general-purpose move and
string instructions, with protection provided through segmentation or paging. I/O
ports can be mapped so that they appear in the I/O address space or the physical-
memory address space (memory mapped I/O) or both.
One benefit of using the I/O address space is that writes to I/O ports are guaranteed
to be completed before the next instruction in the instruction stream is executed.
Thus, I/O writes to control system hardware cause the hardware to be set to its new
state before any other instructions are executed. See Section 13.6, “Ordering I/O,”
for more information on serializing of I/O operations.

13.2 I/O PORT HARDWARE

From a hardware point of view, I/O addressing is handled through the processor’s
address lines. For the P6 family, Pentium 4, and Intel Xeon processors, the request
command lines signal whether the address lines are being driven with a memory
address or an I/O address; for Pentium processors and earlier IA-32 processors, the

Vol. 1 13-1
INPUT/OUTPUT

M/IO# pin indicates a memory address (1) or an I/O address (0). When the separate
I/O address space is selected, it is the responsibility of the hardware to decode the
memory-I/O bus transaction to select I/O ports rather than memory. Data is trans-
mitted between the processor and an I/O device through the data lines.

13.3 I/O ADDRESS SPACE

The processor’s I/O address space is separate and distinct from the physical-memory
address space. The I/O address space consists of 216 (64K) individually addressable
8-bit I/O ports, numbered 0 through FFFFH. I/O port addresses 0F8H through 0FFH
are reserved. Do not assign I/O ports to these addresses. The result of an attempt to
address beyond the I/O address space limit of FFFFH is implementation-specific; see
the Developer’s Manuals for specific processors for more details.
Any two consecutive 8-bit ports can be treated as a 16-bit port, and any four consec-
utive ports can be a 32-bit port. In this manner, the processor can transfer 8, 16, or
32 bits to or from a device in the I/O address space. Like words in memory, 16-bit
ports should be aligned to even addresses (0, 2, 4, ...) so that all 16 bits can be
transferred in a single bus cycle. Likewise, 32-bit ports should be aligned to
addresses that are multiples of four (0, 4, 8, ...). The processor supports data trans-
fers to unaligned ports, but there is a performance penalty because one or more
extra bus cycle must be used.
The exact order of bus cycles used to access unaligned ports is undefined and is not
guaranteed to remain the same in future IA-32 processors. If hardware or software
requires that I/O ports be written to in a particular order, that order must be specified
explicitly. For example, to load a word-length I/O port at address 2H and then
another word port at 4H, two word-length writes must be used, rather than a single
doubleword write at 2H.
Note that the processor does not mask parity errors for bus cycles to the I/O address
space. Accessing I/O ports through the I/O address space is thus a possible source of
parity errors.

13.3.1 Memory-Mapped I/O

I/O devices that respond like memory components can be accessed through the
processor’s physical-memory address space (see Figure 13-1). When using memory-
mapped I/O, any of the processor’s instructions that reference memory can be used
to access an I/O port located at a physical-memory address. For example, the MOV
instruction can transfer data between any register and a memory-mapped I/O port.
The AND, OR, and TEST instructions may be used to manipulate bits in the control
and status registers of a memory-mapped peripheral devices.
When using memory-mapped I/O, caching of the address space mapped for I/O
operations must be prevented. With the Pentium 4, Intel Xeon, and P6 family proces-
sors, caching of I/O accesses can be prevented by using memory type range regis-

13-2 Vol. 1
 INPUT/OUTPUT

ters (MTRRs) to map the address space used for the memory-mapped I/O as
uncacheable (UC). See Chapter 10, “Memory Cache Control,” in the Intel® 64 and
IA-32 Architectures Software Developer’s Manual, Volume 3A, for a complete discus-
sion of the MTRRs.
The Pentium and Intel486 processors do not support MTRRs. Instead, they provide
the KEN# pin, which when held inactive (high) prevents caching of all addresses sent
out on the system bus. To use this pin, external address decoding logic is required to
block caching in specific address spaces.

Physical Memory
FFFF

EPROM

I/O Port
I/O Port
I/O Port

RAM

0
Figure 13-1. Memory-Mapped I/O

All the IA-32 processors that have on-chip caches also provide the PCD (page-level
cache disable) flag in page table and page directory entries. This flag allows caching
to be disabled on a page-by-page basis. See “Page-Directory and Page-Table Entries”
in Chapter 3 of in the Intel® 64 and IA-32 Architectures Software Developer’s
Manual, Volume 3A.

13.4 I/O INSTRUCTIONS

The processor’s I/O instructions provide access to I/O ports through the I/O address
space. (These instructions cannot be used to access memory-mapped I/O ports.)
There are two groups of I/O instructions:
• Those that transfer a single item (byte, word, or doubleword) between an I/O
port and a general-purpose register

Vol. 1 13-3
INPUT/OUTPUT

• Those that transfer strings of items (strings of bytes, words, or doublewords)

between an I/O port and memory
The register I/O instructions IN (input from I/O port) and OUT (output to I/O port)
move data between I/O ports and the EAX register (32-bit I/O), the AX register
(16-bit I/O), or the AL (8-bit I/O) register. The address of the I/O port can be given
with an immediate value or a value in the DX register.
The string I/O instructions INS (input string from I/O port) and OUTS (output string
to I/O port) move data between an I/O port and a memory location. The address of
the I/O port being accessed is given in the DX register; the source or destination
memory address is given in the DS:ESI or ES:EDI register, respectively.
When used with one of the repeat prefixes (such as REP), the INS and OUTS instruc-
tions perform string (or block) input or output operations. The repeat prefix REP
modifies the INS and OUTS instructions to transfer blocks of data between an I/O
port and memory. Here, the ESI or EDI register is incremented or decremented
(according to the setting of the DF flag in the EFLAGS register) after each byte, word,
or doubleword is transferred between the selected I/O port and memory.
See the references for IN, INS, OUT, and OUTS in Chapter 3 and Chapter 4 of the
Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volumes 3A & 3B,
for more information on these instructions.

13.5 PROTECTED-MODE I/O

When the processor is running in protected mode, the following protection mecha-
nisms regulate access to I/O ports:
• When accessing I/O ports through the I/O address space, two protection devices
control access:
— The I/O privilege level (IOPL) field in the EFLAGS register
— The I/O permission bit map of a task state segment (TSS)
• When accessing memory-mapped I/O ports, the normal segmentation and
paging protection and the MTRRs (in processors that support them) also affect
access to I/O ports. See Chapter 4, “Protection,” and Chapter 10, “Memory Cache
Control,” in the Intel® 64 and IA-32 Architectures Software Developer’s Manual,
Volume 3A, for a complete discussion of memory protection.
The following sections describe the protection mechanisms available when accessing
I/O ports in the I/O address space with the I/O instructions.

13.5.1 I/O Privilege Level

In systems where I/O protection is used, the IOPL field in the EFLAGS register
controls access to the I/O address space by restricting use of selected instructions.
This protection mechanism permits the operating system or executive to set the priv-

13-4 Vol. 1
 INPUT/OUTPUT

ilege level needed to perform I/O. In a typical protection ring model, access to the
I/O address space is restricted to privilege levels 0 and 1. Here, kernel and the device
drivers are allowed to perform I/O, while less privileged device drivers and applica-
tion programs are denied access to the I/O address space. Application programs
must then make calls to the operating system to perform I/O.
The following instructions can be executed only if the current privilege level (CPL) of
the program or task currently executing is less than or equal to the IOPL: IN, INS,
OUT, OUTS, CLI (clear interrupt-enable flag), and STI (set interrupt-enable flag).
These instructions are called I/O sensitive instructions, because they are sensitive
to the IOPL field. Any attempt by a less privileged program or task to use an I/O
sensitive instruction results in a general-protection exception (#GP) being signaled.
Because each task has its own copy of the EFLAGS register, each task can have a
different IOPL.
The I/O permission bit map in the TSS can be used to modify the effect of the IOPL
on I/O sensitive instructions, allowing access to some I/O ports by less privileged
programs or tasks (see Section 13.5.2, “I/O Permission Bit Map”).
A program or task can change its IOPL only with the POPF and IRET instructions;
however, such changes are privileged. No procedure may change the current IOPL
unless it is running at privilege level 0. An attempt by a less privileged procedure to
change the IOPL does not result in an exception; the IOPL simply remains
unchanged.
The POPF instruction also may be used to change the state of the IF flag (as can the
CLI and STI instructions); however, the POPF instruction in this case is also I/O sensi-
tive. A procedure may use the POPF instruction to change the setting of the IF flag
only if the CPL is less than or equal to the current IOPL. An attempt by a less privi-
leged procedure to change the IF flag does not result in an exception; the IF flag
simply remains unchanged.

13.5.2 I/O Permission Bit Map

The I/O permission bit map is a device for permitting limited access to I/O ports by
less privileged programs or tasks and for tasks operating in virtual-8086 mode. The
I/O permission bit map is located in the TSS (see Figure 13-2) for the currently
running task or program. The address of the first byte of the I/O permission bit map
is given in the I/O map base address field of the TSS. The size of the I/O permission
bit map and its location in the TSS are variable.

Vol. 1 13-5
INPUT/OUTPUT

Task State Segment (TSS)

31 24 23 0
Last byte of 1 1 1 1 1 1 1 1
bit
map must be
followed by a I/O Permission Bit Map

I/O map I/O Map Base 64H

base must
not exceed
DFFFH.
0

Figure 13-2. I/O Permission Bit Map

Because each task has its own TSS, each task has its own I/O permission bit map.
Access to individual I/O ports can thus be granted to individual tasks.
If in protected mode and the CPL is less than or equal to the current IOPL, the
processor allows all I/O operations to proceed. If the CPL is greater than the IOPL or
if the processor is operating in virtual-8086 mode, the processor checks the I/O
permission bit map to determine if access to a particular I/O port is allowed. Each bit
in the map corresponds to an I/O port byte address. For example, the control bit for
I/O port address 29H in the I/O address space is found at bit position 1 of the sixth
byte in the bit map. Before granting I/O access, the processor tests all the bits corre-
sponding to the I/O port being addressed. For a doubleword access, for example, the
processors tests the four bits corresponding to the four adjacent 8-bit port
addresses. If any tested bit is set, a general-protection exception (#GP) is signaled.
If all tested bits are clear, the I/O operation is allowed to proceed.
Because I/O port addresses are not necessarily aligned to word and doubleword
boundaries, the processor reads two bytes from the I/O permission bit map for every
access to an I/O port. To prevent exceptions from being generated when the ports
with the highest addresses are accessed, an extra byte needs to included in the TSS
immediately after the table. This byte must have all of its bits set, and it must be
within the segment limit.
It is not necessary for the I/O permission bit map to represent all the I/O addresses.
I/O addresses not spanned by the map are treated as if they had set bits in the map.
For example, if the TSS segment limit is 10 bytes past the bit-map base address, the
map has 11 bytes and the first 80 I/O ports are mapped. Higher addresses in the I/O
address space generate exceptions.

13-6 Vol. 1
 INPUT/OUTPUT

If the I/O bit map base address is greater than or equal to the TSS segment limit,
there is no I/O permission map, and all I/O instructions generate exceptions when
the CPL is greater than the current IOPL.

13.6 ORDERING I/O

When controlling I/O devices it is often important that memory and I/O operations be
carried out in precisely the order programmed. For example, a program may write a
command to an I/O port, then read the status of the I/O device from another I/O
port. It is important that the status returned be the status of the device after it
receives the command, not before.
When using memory-mapped I/O, caution should be taken to avoid situations in
which the programmed order is not preserved by the processor. To optimize perfor-
mance, the processor allows cacheable memory reads to be reordered ahead of buff-
ered writes in most situations. Internally, processor reads (cache hits) can be
reordered around buffered writes. When using memory-mapped I/O, therefore, is
possible that an I/O read might be performed before the memory write of a previous
instruction. The recommended method of enforcing program ordering of memory-
mapped I/O accesses with the Pentium 4, Intel Xeon, and P6 family processors is to
use the MTRRs to make the memory mapped I/O address space uncacheable; for the
Pentium and Intel486 processors, either the #KEN pin or the PCD flags can be used
for this purpose (see Section 13.3.1, “Memory-Mapped I/O”).
When the target of a read or write is in an uncacheable region of memory, memory
reordering does not occur externally at the processor’s pins (that is, reads and writes
appear in-order). Designating a memory mapped I/O region of the address space as
uncacheable insures that reads and writes of I/O devices are carried out in program
order. See Chapter 10, “Memory Cache Control,” in the Intel® 64 and IA-32 Architec-
tures Software Developer’s Manual, Volume 3A, for more information on using
MTRRs.
Another method of enforcing program order is to insert one of the serializing instruc-
tions, such as the CPUID instruction, between operations. See Chapter 7, “Multiple-
Processor Management,” in the Intel® 64 and IA-32 Architectures Software Devel-
oper’s Manual, Volume 3A, for more information on serialization of instructions.
It should be noted that the chip set being used to support the processor (bus
controller, memory controller, and/or I/O controller) may post writes to uncacheable
memory which can lead to out-of-order execution of memory accesses. In situations
where out-of-order processing of memory accesses by the chip set can potentially
cause faulty memory-mapped I/O processing, code must be written to force synchro-
nization and ordering of I/O operations. Serializing instructions can often be used for
this purpose.
When the I/O address space is used instead of memory-mapped I/O, the situation is
different in two respects:

Vol. 1 13-7
INPUT/OUTPUT

• The processor never buffers I/O writes. Therefore, strict ordering of I/O
operations is enforced by the processor. (As with memory-mapped I/O, it is
possible for a chip set to post writes in certain I/O ranges.)
• The processor synchronizes I/O instruction execution with external bus activity
(see Table 13-1).

Table 13-1. I/O Instruction Serialization

Processor Delays Execution of … Until Completion of …

Instruction Being Current Next

Executed Instruction? Instruction? Pending Stores? Current Store?
IN Yes Yes
INS Yes Yes
REP INS Yes Yes
OUT Yes Yes Yes
OUTS Yes Yes Yes
REP OUTS Yes Yes Yes

13-8 Vol. 1
 CHAPTER 14
PROCESSOR IDENTIFICATION AND
FEATURE DETERMINATION

When writing software intended to run on IA-32 processors, it is necessary to identify

the type of processor present in a system and the processor features that are avail-
able to an application.

14.1 USING THE CPUID INSTRUCTION

Use the CPUID instruction for processor identification in the Pentium M processor
family, Pentium 4 processor family, Intel Xeon processor family, P6 family, Pentium
processor, and later Intel486 processors. This instruction returns the family, model
and (for some processors) a brand string for the processor that executes the instruc-
tion. It also indicates the features that are present in the processor and give informa-
tion about the processors caches and TLB.
The ID flag (bit 21) in the EFLAGS register indicates support for the CPUID instruc-
tion. If a software procedure can set and clear this flag, the processor executing the
procedure supports the CPUID instruction. The CPUID instruction will cause the
invalid opcode exception (#UD) if executed on a processor that does not support it.
To obtain processor identification information, a source operand value is placed in the
EAX register to select the type of information to be returned. When the CPUID
instruction is executed, selected information is returned in the EAX, EBX, ECX, and
EDX registers. For a complete description of the CPUID instruction, tables indicating
values returned, and example code, see “CPUID—CPUID Identification” in Chapter 3
of the Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volume 2A.

14.1.1 Notes on Where to Start

For detailed application notes on the instruction, see AP-485, Intel Processor Identi-
fication and the CPUID Instruction (Order Number 241618). This publication provides
additional information and example source code for use in identifying IA-32 proces-
sors. It also contains guidelines for using the CPUID instruction to help maintain the
widest range of software compatibility. The following guidelines are among the most
important, and should always be followed when using the CPUID instruction to deter-
mine available features:
• Always begin by testing for the “GenuineIntel,” message in the EBX, EDX, and
ECX registers when the CPUID instruction is executed with EAX equal to 0. If the
processor is not genuine Intel, the feature identification flags may have different
meanings than are described in Intel documentation.

Vol. 1 14-1
PROCESSOR IDENTIFICATION AND FEATURE DETERMINATION

• Test feature identification flags individually and do not make assumptions about
undefined bits.

14.1.2 Identification of Earlier IA-32 Processors

The CPUID instruction is not available in earlier IA-32 processors up through the
earlier Intel486 processors. For these processors, several other architectural
features can be exploited to identify the processor.
The settings of bits 12 and 13 (IOPL), 14 (NT), and 15 (reserved) in the EFLAGS
register are different for Intel’s 32-bit processors than for the Intel 8086 and Intel
286 processors. By examining the settings of these bits (with the PUSHF/PUSHFD
and POP/POPFD instructions), an application program can determine whether the
processor is an 8086, Intel 286, or one of the Intel 32-bit processors:
• 8086 processor — Bits 12 through 15 of the EFLAGS register are always set.
• Intel 286 processor — Bits 12 through 15 are always clear in real-address mode.
• 32-bit processors — In real-address mode, bit 15 is always clear and bits 12
through 14 have the last value loaded into them. In protected mode, bit 15 is
always clear, bit 14 has the last value loaded into it, and the IOPL bits depends on
the current privilege level (CPL). The IOPL field can be changed only if the CPL
is 0.
Other EFLAG register bits that can be used to differentiate between the 32-bit
processors:
• Bit 18 (AC) — Implemented only on the Pentium 4, Intel Xeon, P6 family,
Pentium, and Intel486 processors. The inability to set or clear this bit distin-
guishes an Intel386 processor from the later IA-32 processors.
• Bit 21 (ID) — Determines if the processor is able to execute the CPUID
instruction. The ability to set and clear this bit indicates that it is a Pentium 4,
Intel Xeon, P6 family, Pentium, or later-version Intel486 processor.
To determine whether an x87 FPU or NPX is present in a system, applications can
write to the x87 FPU status and control registers using the FNINIT instruction and
then verify that the correct values are read back using the FNSTENV instruction.
After determining that an x87 FPU or NPX is present, its type can then be deter-
mined. In most cases, the processor type will determine the type of FPU or NPX;
however, an Intel386 processor is compatible with either an Intel 287 or Intel 387
math coprocessor.
The method the coprocessor uses to represent ∞ (after the execution of the FINIT,
FNINIT, or RESET instruction) indicates which coprocessor is present. The Intel 287
math coprocessor uses the same bit representation for +∞ and −∞; whereas, the
Intel 387 math coprocessor uses different representations for +∞ and −∞.

14-2 Vol. 1
 APPENDIX A
EFLAGS CROSS-REFERENCE

A.1 EFLAGS AND INSTRUCTIONS

Table A-2 summarizes how the instructions affect the flags in the EFLAGS register.
The following codes describe how the flags are affected.

Table A-1. Codes Describing Flags

T Instruction tests flag.
M Instruction modifies flag (either sets or resets depending on operands).
0 Instruction resets flag.
1 Instruction sets flag.
— Instruction's effect on flag is undefined.
R Instruction restores prior value of flag.
Blank Instruction does not affect flag.

Table A-2. EFLAGS Cross-Reference

Instruction OF SF ZF AF PF CF TF IF DF NT RF
AAA — — — TM — M
AAD — M M — M —
AAM — M M — M —
AAS — — — TM — M
ADC M M M M M TM
ADD M M M M M M
AND 0 M M — M 0
ARPL M
BOUND
BSF/BSR — — M — — —
BSWAP
BT/BTS/BTR/BTC — — — — — M
CALL

Vol. 1 A-1
EFLAGS CROSS-REFERENCE

Table A-2. EFLAGS Cross-Reference (Contd.)

Instruction OF SF ZF AF PF CF TF IF DF NT RF
CBW
CLC 0
CLD 0
CLI 0
CLTS
CMC M
CMOVcc T T T T T
CMP M M M M M M
CMPS M M M M M M T
CMPXCHG M M M M M M
CMPXCHG8B M
COMSID 0 0 M 0 M M
COMISS 0 0 M 0 M M
CPUID
CWD
DAA — M M TM M TM
DAS — M M TM M TM
DEC M M M M M
DIV — — — — — —
ENTER
ESC
FCMOVcc T T T
FCOMI, FCOMIP, FUCOMI, M M M
FUCOMIP
HLT
IDIV — — — — — —
IMUL M — — — — M
IN
INC M M M M M
INS T
INT 0 0

A-2 Vol. 1
 EFLAGS CROSS-REFERENCE

Table A-2. EFLAGS Cross-Reference (Contd.)

Instruction OF SF ZF AF PF CF TF IF DF NT RF
INTO T 0 0
INVD
INVLPG
UCOMSID 0 0 M 0 M M
UCOMISS 0 0 M 0 M M
IRET R R R R R R R R R T
Jcc T T T T T
JCXZ
JMP
LAHF
LAR M
LDS/LES/LSS/LFS/LGS
LEA
LEAVE
LGDT/LIDT/LLDT/LMSW
LOCK
LODS T
LOOP
LOOPE/LOOPNE T
LSL M
LTR
MONITOR
MWAIT
MOV
MOV control, debug, test — — — — — —
MOVS T
MOVSX/MOVZX
MUL M — — — — M
NEG M M M M M M
NOP
NOT

Vol. 1 A-3
EFLAGS CROSS-REFERENCE

Table A-2. EFLAGS Cross-Reference (Contd.)

Instruction OF SF ZF AF PF CF TF IF DF NT RF
OR 0 M M — M 0
OUT
OUTS T
POP/POPA
POPF R R R R R R R R R R
PUSH/PUSHA/PUSHF
RCL/RCR 1 M TM
RCL/RCR count — TM
RDMSR
RDPMC
RDTSC
REP/REPE/REPNE
RET
ROL/ROR 1 M M
ROL/ROR count — M
RSM M M M M M M M M M M M
SAHF R R R R R
SAL/SAR/SHL/SHR 1 M M M — M M
SAL/SAR/SHL/SHR — M M — M M
count
SBB M M M M M TM
SCAS M M M M M M T
SETcc T T T T T
SGDT/SIDT/SLDT/SMSW
SHLD/SHRD — M M — M M
STC 1
STD 1
STI 1
STOS T
STR
SUB M M M M M M

A-4 Vol. 1
 EFLAGS CROSS-REFERENCE

Table A-2. EFLAGS Cross-Reference (Contd.)

Instruction OF SF ZF AF PF CF TF IF DF NT RF
TEST 0 M M — M 0
UD2
VERR/VERRW M
WAIT
WBINVD
WRMSR
XADD M M M M M M
XCHG
XLAT
XOR 0 M M — M 0

Vol. 1 A-5
EFLAGS CROSS-REFERENCE

A-6 Vol. 1
 APPENDIX B
EFLAGS CONDITION CODES

B.1 CONDITION CODES

Table B-1 lists condition codes that can be queried using CMOVcc, FCMOVcc, Jcc, and
SETcc. Condition codes refer to the setting of one or more status flags (CF, OF, SF, ZF,
and PF) in the EFLAGS register. In the table below:
• The “Mnemonic” column provides the suffix (cc) added to the instruction to
specify a test condition.
• “Condition Tested For” describes the targeted condition.
• “Instruction Subcode” provides the opcode suffix added to the main opcode to
specify the test condition.
• “Status Flags Setting” describes the flag setting.

Table B-1. EFLAGS Condition Codes

Instruction
Mnemonic (cc) Condition Tested For Subcode Status Flags Setting
O Overflow 0000 OF = 1
NO No overflow 0001 OF = 0
B Below 0010 CF = 1
NAE Neither above nor equal
NB Not below 0011 CF = 0
AE Above or equal
E Equal 0100 ZF = 1
Z Zero
NE Not equal 0101 ZF = 0
NZ Not zero
BE Below or equal 0110 (CF OR ZF) = 1
NA Not above
NBE Neither below nor equal 0111 (CF OR ZF) = 0
A Above
S Sign 1000 SF = 1
NS No sign 1001 SF = 0
P Parity 1010 PF = 1
PE Parity even

Vol. 1 B-1
EFLAGS CONDITION CODES

Table B-1. EFLAGS Condition Codes (Contd.)

Instruction
Mnemonic (cc) Condition Tested For Subcode Status Flags Setting
NP No parity 1011 PF = 0
PO Parity odd
L Less 1100 (SF XOR OF) = 1
NGE Neither greater nor equal
NL Not less 1101 (SF XOR OF) = 0
GE Greater or equal
LE Less or equal 1110 ((SF XOR OF) OR ZF) = 1
NG Not greater
NLE Neither less nor equal 1111 ((SF XOR OF) OR ZF) = 0
G Greater

Many of the test conditions are described in two different ways. For example, LE (less
or equal) and NG (not greater) describe the same test condition. Alternate
mnemonics are provided to make code more intelligible.
The terms “above” and “below” are associated with the CF flag and refer to the rela-
tion between two unsigned integer values. The terms “greater” and “less” are asso-
ciated with the SF and OF flags and refer to the relation between two signed integer
values.

B-2 Vol. 1
 APPENDIX C
FLOATING-POINT EXCEPTIONS SUMMARY

C.1 OVERVIEW
This appendix shows which of the floating-point exceptions can be generated for:
• x87 FPU instructions — see Table C-2
• SSE instructions — see Table C-3
• SSE2 instructions — see Table C-4
• SSE3 instructions — see Table C-5
• SSE4 instructions — see Table C-6
Table C-1 lists types of floating-point exceptions that potentially can be generated by
the x87 FPU and by SSE/SSE2/SSE3 instructions.

Table C-1. x87 FPU and SIMD Floating-Point Exceptions

Floating-
point
Exception Description
#IS Invalid-operation exception for stack underflow or stack overflow (can only be
generated for x87 FPU instructions)*
#IA or #I Invalid-operation exception for invalid arithmetic operands and unsupported
formats*
#D Denormal-operand exception
#Z Divide-by-zero exception
#O Numeric-overflow exception
#U Numeric-underflow exception
#P Inexact-result (precision) exception
NOTE:
* The x87 FPU instruction set generates two types of invalid-operation exceptions: #IS (stack
underflow or stack overflow) and #IA (invalid arithmetic operation due to invalid arithmetic
operands or unsupported formats). SSE/SSE2/SSE3 instructions potentially generate #I (invalid
operation exceptions due to invalid arithmetic operands or unsupported formats).

The floating point exceptions shown in Table C-1 (except for #D and #IS) are defined
in IEEE Standard 754-1985 for Binary Floating-Point Arithmetic. See Section 4.9.1,
“Floating-Point Exception Conditions,” for a detailed discussion of floating-point
exceptions.

Vol. 1 C-1
FLOATING-POINT EXCEPTIONS SUMMARY

C.2 X87 FPU INSTRUCTIONS

Table C-2 lists the x87 FPU instructions in alphabetical order. For each instruction, it
summarizes the floating-point exceptions that the instruction can generate.

Table C-2. Exceptions Generated with x87 FPU Floating-Point Instructions

Mnemonic Instruction #IS #IA #D #Z #O #U #P
F2XM1 Exponential Y Y Y Y Y
FABS Absolute value Y
FADD(P) Add floating-point Y Y Y Y Y Y
FBLD BCD load Y
FBSTP BCD store and pop Y Y Y
FCHS Change sign Y
FCLEX Clear exceptions
FCMOVcc Floating-point conditional Y
move
FCOM, FCOMP, FCOMPP Compare floating-point Y Y Y
FCOMI, FCOMIP, FUCOMI, Compare floating-point and Y Y Y
FUCOMIP set EFLAGS
FCOS Cosine Y Y Y Y
FDECSTP Decrement stack pointer
FDIV(R)(P) Divide floating-point Y Y Y Y Y Y Y
FFREE Free register
FIADD Integer add Y Y Y Y Y Y
FICOM(P) Integer compare Y Y Y
FIDIV Integer divide Y Y Y Y Y Y
FIDIVR Integer divide reversed Y Y Y Y Y Y Y
FILD Integer load Y
FIMUL Integer multiply Y Y Y Y Y Y
FINCSTP Increment stack pointer
FINIT Initialize processor
FIST(P) Integer store Y Y Y
FISTTP Truncate to integer Y Y Y
(SSE3 instruction)
FISUB(R) Integer subtract Y Y Y Y Y Y

C-2 Vol. 1
 FLOATING-POINT EXCEPTIONS SUMMARY

Table C-2. Exceptions Generated with x87 FPU Floating-Point Instructions (Contd.)
Mnemonic Instruction #IS #IA #D #Z #O #U #P
FLD extended or stack Load floating-point Y
FLD single or double Load floating-point Y Y Y
FLD1 Load + 1.0 Y
FLDCW Load Control word Y Y Y Y Y Y Y
FLDENV Load environment Y Y Y Y Y Y Y
FLDL2E Load log2e Y
FLDL2T Load log210 Y
FLDLG2 Load log102 Y
FLDLN2 Load loge2 Y
FLDPI Load π Y
FLDZ Load + 0.0 Y
FMUL(P) Multiply floating-point Y Y Y Y Y Y
FNOP No operation
FPATAN Partial arctangent Y Y Y Y Y
FPREM Partial remainder Y Y Y Y
FPREM1 IEEE partial remainder Y Y Y Y
FPTAN Partial tangent Y Y Y Y Y
FRNDINT Round to integer Y Y Y Y
FRSTOR Restore state Y Y Y Y Y Y Y
FSAVE Save state
FSCALE Scale Y Y Y Y Y Y
FSIN Sine Y Y Y Y Y
FSINCOS Sine and cosine Y Y Y Y Y
FSQRT Square root Y Y Y Y
FST(P) stack or extended Store floating-point Y
FST(P) single or double Store floating-point Y Y Y Y Y
FSTCW Store control word
FSTENV Store environment
FSTSW (AX) Store status word
FSUB(R)(P) Subtract floating-point Y Y Y Y Y Y
FTST Test Y Y Y

Vol. 1 C-3
FLOATING-POINT EXCEPTIONS SUMMARY

Table C-2. Exceptions Generated with x87 FPU Floating-Point Instructions (Contd.)
Mnemonic Instruction #IS #IA #D #Z #O #U #P
FUCOM(P)(P) Unordered compare floating- Y Y Y
point
FWAIT CPU Wait
FXAM Examine
FXCH Exchange registers Y
FXTRACT Extract Y Y Y Y
FYL2X Logarithm Y Y Y Y Y Y Y
FYL2XP1 Logarithm epsilon Y Y Y Y Y Y

C.3 SSE INSTRUCTIONS

Table C-3 lists SSE instructions with at least one of the following characteristics:
• have floating-point operands
• generate floating-point results
• read or write floating-point status and control information
The table also summarizes the floating-point exceptions that each instruction can
generate.

Table C-3. Exceptions Generated with SSE Instructions

Mnemonic Instruction #I #D #Z #O #U #P
ADDPS Packed add. Y Y Y Y Y
ADDSS Scalar add. Y Y Y Y Y
ANDNPS Packed logical INVERT and
AND.
ANDPS Packed logical AND.
CMPPS Packed compare. Y Y
CMPSS Scalar compare. Y Y
COMISS Scalar ordered compare lower Y Y
SP FP numbers and set the
status flags.
CVTPI2PS Convert two 32-bit signed Y
integers from MM2/Mem to
two SP FP.

C-4 Vol. 1
 FLOATING-POINT EXCEPTIONS SUMMARY

Table C-3. Exceptions Generated with SSE Instructions (Contd.)

Mnemonic Instruction #I #D #Z #O #U #P
CVTPS2PI Convert lower two SP FP from Y Y
XMM/Mem to two 32-bit
signed integers in MM using
rounding specified by MXCSR.
CVTSI2SS Convert one 32-bit signed Y
integer from Integer Reg/Mem
to one SP FP.
CVTSS2SI Convert one SP FP from Y Y
XMM/Mem to one 32-bit
signed integer using rounding
mode specified by MXCSR, and
move the result to an integer
register.
CVTTPS2PI Convert two SP FP from Y Y
XMM2/Mem to two 32-bit
signed integers in MM1 using
truncate.
CVTTSS2SI Convert lowest SP FP from Y Y
XMM/Mem to one 32-bit
signed integer using truncate,
and move the result to an
integer register.
DIVPS Packed divide. Y Y Y Y Y Y
DIVSS Scalar divide. Y Y Y Y Y Y
LDMXCSR Load control/status word.
MAXPS Packed maximum. Y Y
MAXSS Scalar maximum. Y Y
MINPS Packed minimum. Y Y
MINSS Scalar minimum. Y Y
MOVAPS Move four packed SP values.
MOVHLPS Move packed SP high to low.
MOVHPS Move two packed SP values
between memory and the high
half of an XMM register.
MOVLHPS Move packed SP low to high.

Vol. 1 C-5
FLOATING-POINT EXCEPTIONS SUMMARY

Table C-3. Exceptions Generated with SSE Instructions (Contd.)

Mnemonic Instruction #I #D #Z #O #U #P
MOVLPS Move two packed SP values
between memory and the low
half of an XMM register.
MOVMSKPS Move sign mask to r32.
MOVSS Move scalar SP number
between an XMM register and
memory or a second XMM
register.
MOVUPS Move unaligned packed data.
MULPS Packed multiply. Y Y Y Y Y
MULSS Scalar multiply. Y Y Y Y Y
ORPS Packed OR.
RCPPS Packed reciprocal.
RCPSS Scalar reciprocal.
RSQRTPS Packed reciprocal square root.
RSQRTSS Scalar reciprocal square root.
SHUFPS Shuffle.
SQRTPS Square Root of the packed SP Y Y Y
FP numbers.
SQRTSS Scalar square roo. Y Y Y
STMXCSR Store control/status word.
SUBPS Packed subtract. Y Y Y Y Y
SUBSS Scalar subtract. Y Y Y Y Y
UCOMISS Unordered compare lower SP Y Y
FP numbers and set the status
flags.
UNPCKHPS Interleave SP FP numbers.
UNPCKLPS Interleave SP FP numbers.
XORPS Packed XOR.

C-6 Vol. 1
 FLOATING-POINT EXCEPTIONS SUMMARY

C.4 SSE2 INSTRUCTIONS

Table C-4 lists SSE2 instructions with at least one of the following characteristics:
• floating-point operands
• floating point results
For each instruction, the table summarizes the floating-point exceptions that the
instruction can generate.

Table C-4. Exceptions Generated with SSE2 Instructions

Instruction Description #I #D #Z #O #U #P
ADDPD Add two packed DP FP Y Y Y Y Y
numbers from XMM2/Mem to
XMM1.
ADDSD Add the lower DP FP number Y Y Y Y Y
from XMM2/Mem to XMM1.
ANDNPD Invert the 128 bits in
XMM1and then AND the result
with 128 bits from
XMM2/Mem.
ANDPD Logical And of 128 bits from
XMM2/Mem to XMM1 register.
CMPPD Compare packed DP FP Y Y
numbers from XMM2/Mem to
packed DP FP numbers in
XMM1 register using imm8 as
predicate.
CMPSD Compare lowest DP FP number Y Y
from XMM2/Mem to lowest DP
FP number in XMM1 register
using imm8 as predicate.
COMISD Compare lower DP FP number Y Y
in XMM1 register with lower
DP FP number in XMM2/Mem
and set the status flags
accordingly
CVTDQ2PS Convert four 32-bit signed Y
integers from XMM/Mem to
four SP FP.
CVTPS2DQ Convert four SP FP from Y Y
XMM/Mem to four 32-bit
signed integers in XMM using
rounding specified by MXCSR.

Vol. 1 C-7
FLOATING-POINT EXCEPTIONS SUMMARY

Table C-4. Exceptions Generated with SSE2 Instructions (Contd.)

Instruction Description #I #D #Z #O #U #P
CVTTPS2DQ Convert four SP FP from Y Y
XMM/Mem to four 32-bit
signed integers in XMM using
truncate.
CVTDQ2PD Convert two 32-bit signed
integers in XMM2/Mem to 2
DP FP in xmm1 using rounding
specified by MXCSR.
CVTPD2DQ Convert two DP FP from Y Y
XMM2/Mem to two 32-bit
signed integers in xmm1 using
rounding specified by MXCSR.
CVTPD2PI Convert lower two DP FP from Y Y
XMM/Mem to two 32-bit
signed integers in MM using
rounding specified by MXCSR.
CVTPD2PS Convert two DP FP to two SP Y Y Y Y Y
FP.
CVTPI2PD Convert two 32-bit signed
integers from MM2/Mem to
two DP FP.
CVTPS2PD Convert two SP FP to two DP Y Y
FP.
CVTSD2SI Convert one DP FP from Y Y
XMM/Mem to one 32 bit
signed integer using rounding
mode specified by MXCSR, and
move the result to an integer
register.
CVTSD2SS Convert scalar DP FP to scalar Y Y Y Y Y
SP FP.
CVTSI2SD Convert one 32-bit signed
integer from Integer Reg/Mem
to one DP FP.
CVTSS2SD Convert scalar SP FP to scalar Y Y
DP FP.

C-8 Vol. 1
 FLOATING-POINT EXCEPTIONS SUMMARY

Table C-4. Exceptions Generated with SSE2 Instructions (Contd.)

Instruction Description #I #D #Z #O #U #P
CVTTPD2DQ Convert two DP FP from Y Y
XMM2/Mem to two 32-bit
signed integers in XMM1 using
truncate.
CVTTPD2PI Convert two DP FP from Y Y
XMM2/Mem to two 32-bit
signed integers in MM1 using
truncate.
CVTTSD2SI Convert lowest DP FP from Y Y
XMM/Mem to one 32 bit
signed integer using truncate,
and move the result to an
integer register.
DIVPD Divide packed DP FP numbers Y Y Y Y Y Y
in XMM1 by XMM2/Mem
DIVSD Divide lower DP FP numbers in Y Y Y Y Y Y
XMM1 by XMM2/Mem
MAXPD Return the maximum DP FP Y Y
numbers between XMM2/Mem
and XMM1.
MAXSD Return the maximum DP FP Y Y
number between the lower DP
FP numbers from XMM2/Mem
and XMM1.
MINPD Return the minimum DP Y Y
numbers between XMM2/Mem
and XMM1.
MINSD Return the minimum DP FP Y Y
number between the lowest
DP FP numbers from
XMM2/Mem and XMM1.
MOVAPD Move 128 bits representing 2
packed DP data from
XMM2/Mem to XMM1 register.
Or Move 128 bits representing
2 packed DP from XMM1
register to XMM2/Mem.

Vol. 1 C-9
FLOATING-POINT EXCEPTIONS SUMMARY

Table C-4. Exceptions Generated with SSE2 Instructions (Contd.)

Instruction Description #I #D #Z #O #U #P
MOVHPD Move 64 bits representing one
DP operand from Mem to
upper field of XMM register.
Or move 64 bits representing
one DP operand from upper
field of XMM register to Mem.
MOVLPD Move 64 bits representing one
DP operand from Mem to
lower field of XMM register.
Or move 64 bits representing
one DP operand from lower
field of XMM register to Mem.
MOVMSKPD Move the sign mask to r32.
MOVSD Move 64 bits representing one
scalar DP operand from
XMM2/Mem to XMM1 register.
Or move 64 bits representing
one scalar DP operand from
XMM1 register to XMM2/Mem.
MOVUPD Move 128 bits representing 2
DP data from XMM2/Mem to
XMM1 register.
Or move 128 bits representing
2 DP data from XMM1 register
to XMM2/Mem.
MULPD Multiply packed DP FP Y Y Y Y Y
numbers in XMM2/Mem to
XMM1.
MULSD Multiply the lowest DP FP Y Y Y Y Y
number in XMM2/Mem to
XMM1.
ORPD OR 128 bits from XMM2/Mem
to XMM1 register.
SHUFPD Shuffle Double.
SQRTPD Square Root Packed Double- Y Y Y
Precision
SQRTSD Square Root Scaler Double- Y Y Y
Precision

C-10 Vol. 1
 FLOATING-POINT EXCEPTIONS SUMMARY

Table C-4. Exceptions Generated with SSE2 Instructions (Contd.)

Instruction Description #I #D #Z #O #U #P
SUBPD Subtract Packed Double- Y Y Y Y Y
Precision.
SUBSD Subtract Scaler Double- Y Y Y Y Y
Precision.
UCOMISD Compare lower DP FP number Y Y
in XMM1 register with lower
DP FP number in XMM2/Mem
and set the status flags
accordingly.
UNPCKHPD Interleaves DP FP numbers
from the high halves of XMM1
and XMM2/Mem into XMM1
register.
UNPCKLPD Interleaves DP FP numbers
from the low halves of XMM1
and XMM2/Mem into XMM1
register.
XORPD XOR 128 bits from
XMM2/Mem to XMM1 register.

C.5 SSE3 INSTRUCTIONS

Table C-5 lists the SSE3 instructions that have at least one of the following
characteristics:
• have floating-point operands
• generate floating-point results
For each instruction, the table summarizes the floating-point exceptions that the
instruction can generate.

Table C-5. Exceptions Generated with SSE3 Instructions

Instruction Description #I #D #Z #O #U #P
ADDSUBPD Add /Sub packed DP FP Y Y Y Y Y
numbers from XMM2/Mem to
XMM1.
ADDSUBPS Add /Sub packed SP FP Y Y Y Y Y
numbers from XMM2/Mem to
XMM1.

Vol. 1 C-11
FLOATING-POINT EXCEPTIONS SUMMARY

Table C-5. Exceptions Generated with SSE3 Instructions (Contd.)

Instruction Description #I #D #Z #O #U #P
FISTTP See Table C-2. Y Y
HADDPD Add horizontally packed DP Y Y Y Y Y
FP numbers XMM2/Mem to
XMM1.
HADDPS Add horizontally packed SP Y Y Y Y Y
FP numbers XMM2/Mem to
XMM1
HSUBPD Sub horizontally packed DP Y Y Y Y Y
FP numbers XMM2/Mem to
XMM1
HSUBPS Sub horizontally packed SP Y Y Y Y Y
FP numbers XMM2/Mem to
XMM1
LDDQU Load unaligned integer 128-
bit.
MOVDDUP Move 64 bits representing
one DP data from
XMM2/Mem to XMM1 and
duplicate.
MOVSHDUP Move 128 bits representing 4
SP data from XMM2/Mem to
XMM1 and duplicate high.
MOVSLDUP Move 128 bits representing 4
SP data from XMM2/Mem to
XMM1 and duplicate low.

C.6 SSSE3 INSTRUCTIONS

SSSE3 instructions operate on integer data elements. They do not generate floating-
point exceptions.

C.7 SSE4 INSTRUCTIONS

Table C-6 lists the SSE4.1 instructions that generate floating-point results.
For each instruction, the table summarizes the floating-point exceptions that the
instruction can generate.

C-12 Vol. 1
 FLOATING-POINT EXCEPTIONS SUMMARY

Table C-6. Exceptions Generated with SSE4 Instructions

Instruction Description #I #D #Z #O #U #P
DPPD DP FP dot product. Y Y Y Y Y
DPPS SP FP dot product. Y Y Y Y Y
ROUNDPD Round packed DP FP values Y Y1
to integer FP values.
ROUNDPS Round packed SP FP values Y Y1
to integer FP values.
ROUNDSD Round scalar DP FP value to Y Y1
integer FP value.
ROUNDSS Round scalar SP FP value to Y Y1
integer FP value.
NOTES:
1. If bit 3 of immediate operand is 0

Other SSE4.1 instructions and SSE4.2 instructions do not generate floating-point

exceptions.

Vol. 1 C-13
FLOATING-POINT EXCEPTIONS SUMMARY

C-14 Vol. 1
 APPENDIX D
GUIDELINES FOR WRITING X87 FPU
EXCEPTION HANDLERS

As described in Chapter 8, “Programming with the x87 FPU,” the IA-32 Architecture
supports two mechanisms for accessing exception handlers to handle unmasked x87
FPU exceptions: native mode and MS-DOS compatibility mode. The primary purpose
of this appendix is to provide detailed information to help software engineers design
and write x87 FPU exception-handling facilities to run on PC systems that use the
MS-DOS compatibility mode1 for handling x87 FPU exceptions. Some of the informa-
tion in this appendix will also be of interest to engineers who are writing native-mode
x87 FPU exception handlers. The information provided is as follows:
• Discussion of the origin of the MS-DOS x87 FPU exception handling mechanism
and its relationship to the x87 FPU’s native exception handling mechanism.
• Description of the IA-32 flags and processor pins that control the MS-DOS x87
FPU exception handling mechanism.
• Description of the external hardware typically required to support MS-DOS
exception handling mechanism.
• Description of the x87 FPU’s exception handling mechanism and the typical
protocol for x87 FPU exception handlers.
• Code examples that demonstrate various levels of x87 FPU exception handlers.
• Discussion of x87 FPU considerations in multitasking environments.
• Discussion of native mode x87 FPU exception handling.
The information given is oriented toward the most recent generations of IA-32
processors, starting with the Intel486. It is intended to augment the reference infor-
mation given in Chapter 8, “Programming with the x87 FPU.”
A more extensive version of this appendix is available in the application note AP-578,
Software and Hardware Considerations for x87 FPU Exception Handlers for Intel
Architecture Processors (Order Number 243291), which is available from Intel.

D.1 MS-DOS COMPATIBILITY SUB-MODE FOR HANDLING

X87 FPU EXCEPTIONS
The first generations of IA-32 processors (starting with the Intel 8086 and 8088
processors and going through the Intel 286 and Intel386 processors) did not have an

1 Microsoft Windows* 95 and Windows 3.1 (and earlier versions) operating systems use almost
the same x87 FPU exception handling interface as MS-DOS. The recommendations in this appen-
dix for a MS-DOS compatible exception handler thus apply to all three operating systems.

Vol. 1 D-1
GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

on-chip floating-point unit. Instead, floating-point capability was provided on a sepa-

rate numeric coprocessor chip. The first of these numeric coprocessors was the Intel
8087, which was followed by the Intel 287 and Intel 387 numeric coprocessors.
To allow the 8087 to signal floating-point exceptions to its companion 8086 or 8088,
the 8087 has an output pin, INT, which it asserts when an unmasked floating-point
exception occurs. The designers of the 8087 recommended that the output from this
pin be routed through a programmable interrupt controller (PIC) such as the Intel
8259A to the INTR pin of the 8086 or 8088. The accompanying interrupt vector
number could then be used to access the floating-point exception handler.
However, the original IBM* PC design and MS-DOS operating system used a different
mechanism for handling the INT output from the 8087. It connected the INT pin
directly to the NMI input pin of the 8086 or 8088. The NMI interrupt handler then had
to determine if the interrupt was caused by a floating-point exception or another NMI
event. This mechanism is the origin of what is now called the “MS-DOS compatibility
mode.” The decision to use this latter floating-point exception handling mechanism
came about because when the IBM PC was first designed, the 8087 was not available.
When the 8087 did become available, other functions had already been assigned to
the eight inputs to the PIC. One of these functions was a BIOS video interrupt, which
was assigned to interrupt number 16 for the 8086 and 8088.
The Intel 286 processor created the “native mode” for handling floating-point excep-
tions by providing a dedicated input pin (ERROR#) for receiving floating-point excep-
tion signals and a dedicated interrupt number, 16. Interrupt 16 was used to signal
floating-point errors (also called math faults). It was intended that the ERROR# pin
on the Intel 286 be connected to a corresponding ERROR# pin on the Intel 287
numeric coprocessor. When the Intel 287 signals a floating-point exception using this
mechanism, the Intel 286 generates an interrupt 16, to invoke the floating-point
exception handler.
To maintain compatibility with existing PC software, the native floating-point excep-
tion handling mode of the Intel 286 and 287 was not used in the IBM PC AT system
design. Instead, the ERROR# pin on the Intel 286 was tied permanently high, and
the ERROR# pin from the Intel 287 was routed to a second (cascaded) PIC. The
resulting output of this PIC was routed through an exception handler and eventually
caused an interrupt 2 (NMI interrupt). Here the NMI interrupt was shared with IBM
PC AT’s new parity checking feature. Interrupt 16 remained assigned to the BIOS
video interrupt handler. The external hardware for the MS-DOS compatibility mode
must prevent the Intel 286 processor from executing past the next x87 FPU instruc-
tion when an unmasked exception has been generated. To do this, it asserts the
BUSY# signal into the Intel 286 when the ERROR# signal is asserted by the Intel 287.
The Intel386 processor and its companion Intel 387 numeric coprocessor provided
the same hardware mechanism for signaling and handling floating-point exceptions
as the Intel 286 and 287 processors. And again, to maintain compatibility with
existing MS-DOS software, basically the same MS-DOS compatibility floating-point
exception handling mechanism that was used in the IBM PC AT was used in PCs based
on the Intel386 processor.

D-2 Vol. 1
 GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

D.2 IMPLEMENTATION OF THE MS-DOS* COMPATIBILITY

SUB-MODE IN THE INTEL486™, PENTIUM®, AND P6
PROCESSOR FAMILY, AND PENTIUM® 4 PROCESSORS
Beginning with the Intel486™ processor, the IA-32 architecture provided a dedicated
mechanism for enabling the MS-DOS compatibility mode for x87 FPU exceptions and
for generating external x87 FPU-exception signals while operating in this mode. The
following sections describe the implementation of the MS-DOS compatibility mode in
the Intel486 and Pentium processors and in the P6 family and Pentium 4 processors.
Also described is the recommended external hardware to support this mode of oper-
ation.

D.2.1 MS-DOS* Compatibility Sub-mode in the Intel486™ and

Pentium® Processors
In the Intel486 processor, several things were done to enhance and speed up the
numeric coprocessor, now called the floating-point unit (x87 FPU). The most impor-
tant enhancement was that the x87 FPU was included in the same chip as the
processor, for increased speed in x87 FPU computations and reduced latency for x87
FPU exception handling. Also, for the first time, the MS-DOS compatibility mode was
built into the chip design, with the addition of the NE bit in control register CR0 and
the addition of the FERR# (Floating-point ERRor) and IGNNE# (IGNore Numeric
Error) pins.
The NE bit selects the native x87 FPU exception handling mode (NE = 1) or the
MS-DOS compatibility mode (NE = 0). When native mode is selected, all signaling of
floating-point exceptions is handled internally in the Intel486 chip, resulting in the
generation of an interrupt 16.
When MS-DOS compatibility mode is selected, the FERRR# and IGNNE# pins are
used to signal floating-point exceptions. The FERR# output pin, which replaces the
ERROR# pin from the previous generations of IA-32 numeric coprocessors, is
connected to a PIC. A new input signal, IGNNE#, is provided to allow the x87 FPU
exception handler to execute x87 FPU instructions, if desired, without first clearing
the error condition and without triggering the interrupt a second time. This IGNNE#
feature is needed to replicate the capability that was provided on MS-DOS compat-
ible Intel 286 and Intel 287 and Intel386 and Intel 387 systems by turning off the
BUSY# signal, when inside the x87 FPU exception handler, before clearing the error
condition.
Note that Intel, in order to provide Intel486 processors for market segments that had
no need for an x87 FPU, created the “SX” versions. These Intel486 SX processors did
not contain the floating-point unit. Intel also produced Intel 487 SX processors for
end users who later decided to upgrade to a system with an x87 FPU. These Intel 487
SX processors are similar to standard Intel486 processors with a working x87 FPU on
board.

Vol. 1 D-3
GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

Thus, the external circuitry necessary to support the MS-DOS compatibility mode for
Intel 487 SX processors is the same as for standard Intel486 DX processors.
The Pentium, P6 family, and Pentium 4 processors offer the same mechanism (the NE
bit and the FERR# and IGNNE# pins) as the Intel486 processors for generating x87
FPU exceptions in MS-DOS compatibility mode. The actions of these mechanisms are
slightly different and more straightforward for the P6 family and Pentium 4 proces-
sors, as described in Section D.2.2, “MS-DOS* Compatibility Sub-mode in the P6
Family and Pentium® 4 Processors.”
For Pentium, P6 family, and Pentium 4 processors, it is important to note that the
special DP (Dual Processing) mode for Pentium processors and also the more general
Intel MultiProcessor Specification for systems with multiple Pentium, P6 family, or
Pentium 4 processors support x87 FPU exception handling only in the native mode.
Intel does not recommend using the MS-DOS compatibility x87 FPU mode for
systems using more than one processor.

D.2.1.1 Basic Rules: When FERR# Is Generated

When MS-DOS compatibility mode is enabled for the Intel486 or Pentium processors
(NE bit is set to 0) and the IGNNE# input pin is de-asserted, the FERR# signal is
generated as follows:
1. When an x87 FPU instruction causes an unmasked x87 FPU exception, the
processor (in most cases) uses a “deferred” method of reporting the error. This
means that the processor does not respond immediately, but rather freezes just
before executing the next WAIT or x87 FPU instruction (except for “no-wait”
instructions, which the x87 FPU executes regardless of an error condition).
2. When the processor freezes, it also asserts the FERR# output.
3. The frozen processor waits for an external interrupt, which must be supplied by
external hardware in response to the FERR# assertion.
4. In MS-DOS compatibility systems, FERR# is fed to the IRQ13 input in the
cascaded PIC. The PIC generates interrupt 75H, which then branches to interrupt
2, as described earlier in this appendix for systems using the Intel 286 and Intel
287 or Intel386 and Intel 387 processors.
The deferred method of error reporting is used for all exceptions caused by the basic
arithmetic instructions (including FADD, FSUB, FMUL, FDIV, FSQRT, FCOM and
FUCOM), for precision exceptions caused by all types of x87 FPU instructions, and for
numeric underflow and overflow exceptions caused by all types of x87 FPU instruc-
tions except stores to memory.
Some x87 FPU instructions with some x87 FPU exceptions use an “immediate”
method of reporting errors. Here, the FERR# is asserted immediately, at the time
that the exception occurs. The immediate method of error reporting is used for x87
FPU stack fault, invalid operation and denormal exceptions caused by all transcen-
dental instructions, FSCALE, FXTRACT, FPREM and others, and all exceptions (except
precision) when caused by x87 FPU store instructions. Like deferred error reporting,

D-4 Vol. 1
 GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

immediate error reporting will cause the processor to freeze just before executing
the next WAIT or x87 FPU instruction if the error condition has not been cleared by
that time.
Note that in general, whether deferred or immediate error reporting is used for an
x87 FPU exception depends both on which exception occurred and which instruction
caused that exception. A complete specification of these cases, which applies to both
the Pentium and the Intel486 processors, is given in Section 5.1.21 in the Pentium
Processor Family Developer’s Manual: Volume 1.
If NE = 0 but the IGNNE# input is active while an unmasked x87 FPU exception is in
effect, the processor disregards the exception, does not assert FERR#, and
continues. If IGNNE# is then de-asserted and the x87 FPU exception has not been
cleared, the processor will respond as described above. (That is, an immediate
exception case will assert FERR# immediately. A deferred exception case will assert
FERR# and freeze just before the next x87 FPU or WAIT instruction.) The assertion of
IGNNE# is intended for use only inside the x87 FPU exception handler, where it is
needed if one wants to execute non-control x87 FPU instructions for diagnosis,
before clearing the exception condition. When IGNNE# is asserted inside the excep-
tion handler, a preceding x87 FPU exception has already caused FERR# to be
asserted, and the external interrupt hardware has responded, but IGNNE# assertion
still prevents the freeze at x87 FPU instructions. Note that if IGNNE# is left active
outside of the x87 FPU exception handler, additional x87 FPU instructions may be
executed after a given instruction has caused an x87 FPU exception. In this case, if
the x87 FPU exception handler ever did get invoked, it could not determine which
instruction caused the exception.
To properly manage the interface between the processor’s FERR# output, its IGNNE#
input, and the IRQ13 input of the PIC, additional external hardware is needed. A
recommended configuration is described in the following section.

D.2.1.2 Recommended External Hardware to Support the MS-DOS*

Compatibility Sub-mode
Figure D-1 provides an external circuit that will assure proper handling of FERR# and
IGNNE# when an x87 FPU exception occurs. In particular, it assures that IGNNE# will
be active only inside the x87 FPU exception handler without depending on the order
of actions by the exception handler. Some hardware implementations have been less
robust because they have depended on the exception handler to clear the x87 FPU
exception interrupt request to the PIC (FP_IRQ signal) before the handler causes
FERR# to be de-asserted by clearing the exception from the x87 FPU itself.
Figure D-2 shows the details of how IGNNE# will behave when the circuit in
Figure D-1 is implemented. The temporal regions within the x87 FPU exception
handler activity are described as follows:
1. The FERR# signal is activated by an x87 FPU exception and sends an interrupt
request through the PIC to the processor’s INTR pin.

Vol. 1 D-5
GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

2. During the x87 FPU interrupt service routine (exception handler) the processor
will need to clear the interrupt request latch (Flip Flop #1). It may also want to
execute non-control x87 FPU instructions before the exception is cleared from the
x87 FPU. For this purpose the IGNNE# must be driven low. Typically in the PC
environment an I/O access to Port 0F0H clears the external x87 FPU exception
interrupt request (FP_IRQ). In the recommended circuit, this access also is used
to activate IGNNE#. With IGNNE# active, the x87 FPU exception handler may
execute any x87 FPU instruction without being blocked by an active x87 FPU
exception.
3. Clearing the exception within the x87 FPU will cause the FERR# signal to be
deactivated and then there is no further need for IGNNE# to be active. In the
recommended circuit, the deactivation of FERR# is used to deactivate IGNNE#. If
another circuit is used, the software and circuit together must assure that
IGNNE# is deactivated no later than the exit from the x87 FPU exception handler.

D-6 Vol. 1
 GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

5(6(7

,23RUW)+
$GGUHVV'HFRGH

9

))
)(55
35
3HQWLXPŠ3UR3URFHVVRU

9
,QWHOŒ3URFHVVRU
3HQWLXPŠ3URFHVVRU

9 &/5

))

35

9

,*11(

,175
,QWHUUXSW
)3B,54
&RQWUROOHU
/(*(1'
))Q)OLS)ORSQ
&/5&OHDURU5HVHW

Figure D-1. Recommended Circuit for MS-DOS Compatibility x87 FPU

Exception Handling

In the circuit in Figure D-1, when the x87 FPU exception handler accesses I/O port
0F0H it clears the IRQ13 interrupt request output from Flip Flop #1 and also clocks
out the IGNNE# signal (active) from Flip Flop #2. So the handler can activate
IGNNE#, if needed, by doing this 0F0H access before clearing the x87 FPU exception
condition (which de-asserts FERR#).
However, the circuit does not depend on the order of actions by the x87 FPU excep-
tion handler to guarantee the correct hardware state upon exit from the handler. Flip
Flop #2, which drives IGNNE# to the processor, has its CLEAR input attached to the
inverted FERR#. This ensures that IGNNE# can never be active when FERR# is inac-

Vol. 1 D-7
GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

tive. So if the handler clears the x87 FPU exception condition before the 0F0H
access, IGNNE# does not get activated and left on after exit from the handler.

0F0H Address
Decode

Figure D-2. Behavior of Signals During x87 FPU Exception Handling

D.2.1.3 No-Wait x87 FPU Instructions Can Get x87 FPU Interrupt in
Window
The Pentium and Intel486 processors implement the “no-wait” floating-point instruc-
tions (FNINIT, FNCLEX, FNSTENV, FNSAVE, FNSTSW, FNSTCW, FNENI, FNDISI or
FNSETPM) in the MS-DOS compatibility mode in the following manner. (See Section
8.3.11, “x87 FPU Control Instructions,” and Section 8.3.12, “Waiting vs. Non-waiting
Instructions,” for a discussion of the no-wait instructions.)
If an unmasked numeric exception is pending from a preceding x87 FPU instruction,
a member of the no-wait class of instructions will, at the beginning of its execution,
assert the FERR# pin in response to that exception just like other x87 FPU instruc-
tions, but then, unlike the other x87 FPU instructions, FERR# will be de-asserted.
This de-assertion was implemented to allow the no-wait class of instructions to
proceed without an interrupt due to any pending numeric exception. However, the
brief assertion of FERR# is sufficient to latch the x87 FPU exception request into most
hardware interface implementations (including Intel’s recommended circuit).
All the x87 FPU instructions are implemented such that during their execution, there
is a window in which the processor will sample and accept external interrupts. If
there is a pending interrupt, the processor services the interrupt first before
resuming the execution of the instruction. Consequently, it is possible that the no-
wait floating-point instruction may accept the external interrupt caused by it’s own
assertion of the FERR# pin in the event of a pending unmasked numeric exception,

D-8 Vol. 1
 GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

which is not an explicitly documented behavior of a no-wait instruction. This process

is illustrated in Figure D-3.

Exception Generating
Floating-Point
Instruction

Assertion of FERR#
by the Processor Start of the “No-Wait”
Floating-Point
Instruction
System
Dependent
Delay
Case 1 External Interrupt
Sampling Window
Assertion of INTR Pin
by the System
Case 2
Window Closed

Figure D-3. Timing of Receipt of External Interrupt

Figure D-3 assumes that a floating-point instruction that generates a “deferred”

error (as defined in the Section D.2.1.1, “Basic Rules: When FERR# Is Generated”),
which asserts the FERR# pin only on encountering the next floating-point instruction,
causes an unmasked numeric exception. Assume that the next floating-point instruc-
tion following this instruction is one of the no-wait floating-point instructions. The
FERR# pin is asserted by the processor to indicate the pending exception on encoun-
tering the no-wait floating-point instruction. After the assertion of the FERR# pin the
no-wait floating-point instruction opens a window where the pending external inter-
rupts are sampled.
Then there are two cases possible depending on the timing of the receipt of the inter-
rupt via the INTR pin (asserted by the system in response to the FERR# pin) by the
processor.
Case 1 If the system responds to the assertion of FERR# pin by the no-wait
floating-point instruction via the INTR pin during this window then
the interrupt is serviced first, before resuming the execution of the
no-wait floating-point instruction.
Case 2 If the system responds via the INTR pin after the window has closed
then the interrupt is recognized only at the next instruction boundary.

Vol. 1 D-9
GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

There are two other ways, in addition to Case 1 above, in which a no-wait floating-
point instruction can service a numeric exception inside its interrupt window. First,
the first floating-point error condition could be of the “immediate” category (as
defined in Section D.2.1.1, “Basic Rules: When FERR# Is Generated”) that asserts
FERR# immediately. If the system delay before asserting INTR is long enough, rela-
tive to the time elapsed before the no-wait floating-point instruction, INTR can be
asserted inside the interrupt window for the latter. Second, consider two no-wait x87
FPU instructions in close sequence, and assume that a previous x87 FPU instruction
has caused an unmasked numeric exception. Then if the INTR timing is too long for
an FERR# signal triggered by the first no-wait instruction to hit the first instruction’s
interrupt window, it could catch the interrupt window of the second.
The possible malfunction of a no-wait x87 FPU instruction explained above cannot
happen if the instruction is being used in the manner for which Intel originally
designed it. The no-wait instructions were intended to be used inside the x87 FPU
exception handler, to allow manipulation of the x87 FPU before the error condition is
cleared, without hanging the processor because of the x87 FPU error condition, and
without the need to assert IGNNE#. They will perform this function correctly, since
before the error condition is cleared, the assertion of FERR# that caused the x87 FPU
error handler to be invoked is still active. Thus the logic that would assert FERR#
briefly at a no-wait instruction causes no change since FERR# is already asserted.
The no-wait instructions may also be used without problem in the handler after the
error condition is cleared, since now they will not cause FERR# to be asserted at all.
If a no-wait instruction is used outside of the x87 FPU exception handler, it may
malfunction as explained above, depending on the details of the hardware interface
implementation and which particular processor is involved. The actual interrupt
inside the window in the no-wait instruction may be blocked by surrounding it with
the instructions: PUSHFD, CLI, no-wait, then POPFD. (CLI blocks interrupts, and the
push and pop of flags preserves and restores the original value of the interrupt flag.)
However, if FERR# was triggered by the no-wait, its latched value and the PIC
response will still be in effect. Further code can be used to check for and correct such
a condition, if needed. Section D.3.6, “Considerations When x87 FPU Shared
Between Tasks,” discusses an important example of this type of problem and gives a
solution.

D.2.2 MS-DOS* Compatibility Sub-mode in the P6 Family

and Pentium® 4 Processors
When bit NE = 0 in CR0, the MS-DOS compatibility mode of the P6 family and
Pentium 4 processors provides FERR# and IGNNE# functionality that is almost iden-
tical to the Intel486 and Pentium processors. The same external hardware described
in Section D.2.1.2, “Recommended External Hardware to Support the MS-DOS*
Compatibility Sub-mode,” is recommended for the P6 family and Pentium 4 proces-
sors as well as the two previous generations. The only change to MS-DOS compati-
bility x87 FPU exception handling with the P6 family and Pentium 4 processors is that
all exceptions for all x87 FPU instructions cause immediate error reporting. That is,

D-10 Vol. 1
 GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

FERR# is asserted as soon as the x87 FPU detects an unmasked exception; there are
no cases in which error reporting is deferred to the next x87 FPU or WAIT instruction.
(As is discussed in Section D.2.1.1, “Basic Rules: When FERR# Is Generated,” most
exception cases in the Intel486 and Pentium processors are of the deferred type.)
Although FERR# is asserted immediately upon detection of an unmasked x87 FPU
error, this certainly does not mean that the requested interrupt will always be
serviced before the next instruction in the code sequence is executed. To begin with,
the P6 family and Pentium 4 processors execute several instructions simultaneously.
There also will be a delay, which depends on the external hardware implementation,
between the FERR# assertion from the processor and the responding INTR assertion
to the processor. Further, the interrupt request to the PICs (IRQ13) may be tempo-
rarily blocked by the operating system, or delayed by higher priority interrupts, and
processor response to INTR itself is blocked if the operating system has cleared the
IF bit in EFLAGS. Note that Streaming SIMD Extensions numeric exceptions will not
cause assertion of FERR# (independent of the value of CR0.NE). In addition, they
ignore the assertion/deassertion of IGNNE#).
However, just as with the Intel486 and Pentium processors, if the IGNNE# input is
inactive, a floating-point exception which occurred in the previous x87 FPU instruc-
tion and is unmasked causes the processor to freeze immediately when encountering
the next WAIT or x87 FPU instruction (except for no-wait instructions). This means
that if the x87 FPU exception handler has not already been invoked due to the earlier
exception (and therefore, the handler not has cleared that exception state from the
x87 FPU), the processor is forced to wait for the handler to be invoked and handle the
exception, before the processor can execute another WAIT or x87 FPU instruction.
As explained in Section D.2.1.3, “No-Wait x87 FPU Instructions Can Get x87 FPU
Interrupt in Window,” if a no-wait instruction is used outside of the x87 FPU exception
handler, in the Intel486 and Pentium processors, it may accept an unmasked excep-
tion from a previous x87 FPU instruction which happens to fall within the external
interrupt sampling window that is opened near the beginning of execution of all x87
FPU instructions. This will not happen in the P6 family and Pentium 4 processors,
because this sampling window has been removed from the no-wait group of x87 FPU
instructions.

D.3 RECOMMENDED PROTOCOL FOR MS-DOS*

COMPATIBILITY HANDLERS
The activities of numeric programs can be split into two major areas: program control
and arithmetic. The program control part performs activities such as deciding what
functions to perform, calculating addresses of numeric operands, and loop control.
The arithmetic part simply adds, subtracts, multiplies, and performs other operations
on the numeric operands. The processor is designed to handle these two parts sepa-
rately and efficiently. An x87 FPU exception handler, if a system chooses to imple-
ment one, is often one of the most complicated parts of the program control code.

Vol. 1 D-11
GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

D.3.1 Floating-Point Exceptions and Their Defaults

The x87 FPU can recognize six classes of floating-point exception conditions while
executing floating-point instructions:
1. #I — Invalid operation
#IS — Stack fault
#IA — IEEE standard invalid operation
2. #Z — Divide-by-zero
3. #D — Denormalized operand
4. #O — Numeric overflow
5. #U — Numeric underflow
6. #P — Inexact result (precision)
For complete details on these exceptions and their defaults, see Section 8.4, “x87
FPU Floating-Point Exception Handling,” and Section 8.5, “x87 FPU Floating-Point
Exception Conditions.”

D.3.2 Two Options for Handling Numeric Exceptions

Depending on options determined by the software system designer, the processor
takes one of two possible courses of action when a numeric exception occurs:
1. The x87 FPU can handle selected exceptions itself, producing a default fix-up that
is reasonable in most situations. This allows the numeric program execution to
continue undisturbed. Programs can mask individual exception types to indicate
that the x87 FPU should generate this safe, reasonable result whenever the
exception occurs. The default exception fix-up activity is treated by the x87 FPU
as part of the instruction causing the exception; no external indication of the
exception is given (except that the instruction takes longer to execute when it
handles a masked exception.) When masked exceptions are detected, a flag is
set in the numeric status register, but no information is preserved regarding
where or when it was set.
2. A software exception handler can be invoked to handle the exception. When a
numeric exception is unmasked and the exception occurs, the x87 FPU stops
further execution of the numeric instruction and causes a branch to a software
exception handler. The exception handler can then implement any sort of
recovery procedures desired for any numeric exception detectable by the x87
FPU.

D.3.2.1 Automatic Exception Handling: Using Masked Exceptions

Each of the six exception conditions described above has a corresponding flag bit in
the x87 FPU status word and a mask bit in the x87 FPU control word. If an exception

D-12 Vol. 1
 GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

is masked (the corresponding mask bit in the control word = 1), the processor takes
an appropriate default action and continues with the computation.
The processor has a default fix-up activity for every possible exception condition it
may encounter. These masked-exception responses are designed to be safe and are
generally acceptable for most numeric applications.
For example, if the Inexact result (Precision) exception is masked, the system can
specify whether the x87 FPU should handle a result that cannot be represented
exactly by one of four modes of rounding: rounding it normally, chopping it toward
zero, always rounding it up, or always down. If the Underflow exception is masked,
the x87 FPU will store a number that is too small to be represented in normalized
form as a denormal (or zero if it’s smaller than the smallest denormal). Note that
when exceptions are masked, the x87 FPU may detect multiple exceptions in a single
instruction, because it continues executing the instruction after performing its
masked response. For example, the x87 FPU could detect a denormalized operand,
perform its masked response to this exception, and then detect an underflow.
As an example of how even severe exceptions can be handled safely and automati-
cally using the default exception responses, consider a calculation of the parallel
resistance of several values using only the standard formula (see Figure D-4). If R1
becomes zero, the circuit resistance becomes zero. With the divide-by-zero and
precision exceptions masked, the processor will produce the correct result. FDIV of
R1 into 1 gives infinity, and then FDIV of (infinity +R2 +R3) into 1 gives zero.

R1 R2 R3

1
Equivalent Resistance =
1 1 1
+ +
R1 R2 R3

Figure D-4. Arithmetic Example Using Infinity

By masking or unmasking specific numeric exceptions in the x87 FPU control word,
programmers can delegate responsibility for most exceptions to the processor,
reserving the most severe exceptions for programmed exception handlers. Excep-
tion-handling software is often difficult to write, and the masked responses have
been tailored to deliver the most reasonable result for each condition. For the
majority of applications, masking all exceptions yields satisfactory results with the

Vol. 1 D-13
GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

least programming effort. Certain exceptions can usefully be left unmasked during
the debugging phase of software development, and then masked when the clean
software is actually run. An invalid-operation exception for example, typically indi-
cates a program error that must be corrected.
The exception flags in the x87 FPU status word provide a cumulative record of excep-
tions that have occurred since these flags were last cleared. Once set, these flags can
be cleared only by executing the FCLEX/FNCLEX (clear exceptions) instruction, by
reinitializing the x87 FPU with FINIT/FNINIT or FSAVE/FNSAVE, or by overwriting the
flags with an FRSTOR or FLDENV instruction. This allows a programmer to mask all
exceptions, run a calculation, and then inspect the status word to see if any excep-
tions were detected at any point in the calculation.

D.3.2.2 Software Exception Handling

If the x87 FPU in or with an IA-32 processor (Intel 286 and onwards) encounters an
unmasked exception condition, with the system operated in the MS-DOS compati-
bility mode and with IGNNE# not asserted, a software exception handler is invoked
through a PIC and the processor’s INTR pin. The FERR# (or ERROR#) output from
the x87 FPU that begins the process of invoking the exception handler may occur
when the error condition is first detected, or when the processor encounters the next
WAIT or x87 FPU instruction. Which of these two cases occurs depends on the
processor generation and also on which exception and which x87 FPU instruction trig-
gered it, as discussed earlier in Section D.1, “MS-DOS Compatibility Sub-mode for
Handling x87 FPU Exceptions,” and Section D.2, “Implementation of the MS-DOS*
Compatibility Sub-mode in the Intel486™, Pentium®, and P6 Processor Family, and
Pentium® 4 Processors.” The elapsed time between the initial error signal and the
invocation of the x87 FPU exception handler depends of course on the external hard-
ware interface, and also on whether the external interrupt for x87 FPU errors is
enabled. But the architecture ensures that the handler will be invoked before execu-
tion of the next WAIT or floating-point instruction since an unmasked floating-point
exception causes the processor to freeze just before executing such an instruction
(unless the IGNNE# input is active, or it is a no-wait x87 FPU instruction).
The frozen processor waits for an external interrupt, which must be supplied by
external hardware in response to the FERR# (or ERROR#) output of the processor
(or coprocessor), usually through IRQ13 on the “slave” PIC, and then through INTR.
Then the external interrupt invokes the exception handling routine. Note that if the
external interrupt for x87 FPU errors is disabled when the processor executes an x87
FPU instruction, the processor will freeze until some other (enabled) interrupt occurs
if an unmasked x87 FPU exception condition is in effect. If NE = 0 but the IGNNE#
input is active, the processor disregards the exception and continues. Error reporting
via an external interrupt is supported for MS-DOS compatibility. Chapter 17, “IA-32
Architecture Compatibility,” of the Intel® 64 and IA-32 Architectures Software Devel-
oper’s Manual, Volume 3A, contains further discussion of compatibility issues.
The references above to the ERROR# output from the x87 FPU apply to the Intel 387
and Intel 287 math coprocessors (NPX chips). If one of these coprocessors encoun-
ters an unmasked exception condition, it signals the exception to the Intel 286 or

D-14 Vol. 1
 GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

Intel386 processor using the ERROR# status line between the processor and the
coprocessor. See Section D.1, “MS-DOS Compatibility Sub-mode for Handling x87
FPU Exceptions,” in this appendix, and Chapter 17, “IA-32 Architecture Compati-
bility,” in the Intel® 64 and IA-32 Architectures Software Developer’s Manual,
Volume 3A, for differences in x87 FPU exception handling.
The exception-handling routine is normally a part of the systems software. The
routine must clear (or disable) the active exception flags in the x87 FPU status word
before executing any floating-point instructions that cannot complete execution
when there is a pending floating-point exception. Otherwise, the floating-point
instruction will trigger the x87 FPU interrupt again, and the system will be caught in
an endless loop of nested floating-point exceptions, and hang. In any event, the
routine must clear (or disable) the active exception flags in the x87 FPU status word
after handling them, and before IRET(D). Typical exception responses may include:
• Incrementing an exception counter for later display or printing.
• Printing or displaying diagnostic information (e.g., the x87 FPU environment and
registers).
• Aborting further execution, or using the exception pointers to build an instruction
that will run without exception and executing it.
Applications programmers should consult their operating system's reference
manuals for the appropriate system response to numerical exceptions. For systems
programmers, some details on writing software exception handlers are provided in
Chapter 5, “Interrupt and Exception Handling,” in the Intel® 64 and IA-32 Architec-
tures Software Developer’s Manual, Volume 3A, as well as in Section D.3.4, “x87 FPU
Exception Handling Examples,” in this appendix.
As discussed in Section D.2.1.2, “Recommended External Hardware to Support the
MS-DOS* Compatibility Sub-mode,” some early FERR# to INTR hardware interface
implementations are less robust than the recommended circuit. This is because they
depended on the exception handler to clear the x87 FPU exception interrupt request
to the PIC (by accessing port 0F0H) before the handler causes FERR# to be de-
asserted by clearing the exception from the x87 FPU itself. To eliminate the chance of
a problem with this early hardware, Intel recommends that x87 FPU exception
handlers always access port 0F0H before clearing the error condition from the x87
FPU.

D.3.3 Synchronization Required for Use of x87 FPU Exception

Handlers
Concurrency or synchronization management requires a check for exceptions before
letting the processor change a value just used by the x87 FPU. It is important to
remember that almost any numeric instruction can, under the wrong circumstances,
produce a numeric exception.

Vol. 1 D-15
GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

D.3.3.1 Exception Synchronization: What, Why, and When

Exception synchronization means that the exception handler inspects and deals with
the exception in the context in which it occurred. If concurrent execution is allowed,
the state of the processor when it recognizes the exception is often not in the context
in which it occurred. The processor may have changed many of its internal registers
and be executing a totally different program by the time the exception occurs. If the
exception handler cannot recapture the original context, it cannot reliably determine
the cause of the exception or recover successfully from the exception. To handle this
situation, the x87 FPU has special registers updated at the start of each numeric
instruction to describe the state of the numeric program when the failed instruction
was attempted.
This provides tools to help the exception handler recapture the original context, but
the application code must also be written with synchronization in mind. Overall,
exception synchronization must ensure that the x87 FPU and other relevant parts of
the context are in a well defined state when the handler is invoked after an unmasked
numeric exception occurs.
When the x87 FPU signals an unmasked exception condition, it is requesting help.
The fact that the exception was unmasked indicates that further numeric program
execution under the arithmetic and programming rules of the x87 FPU will probably
yield invalid results. Thus the exception must be handled, and with proper synchro-
nization, or the program will not operate reliably.
For programmers using higher-level languages, all required synchronization is auto-
matically provided by the appropriate compiler. However, for assembly language
programmers exception synchronization remains the responsibility of the
programmer. It is not uncommon for a programmer to expect that their numeric
program will not cause numeric exceptions after it has been tested and debugged,
but in a different system or numeric environment, exceptions may occur regularly
nonetheless. An obvious example would be use of the program with some numbers
beyond the range for which it was designed and tested. Example D-1 and Example
D-2 in Section D.3.3.2, “Exception Synchronization Examples,” show a subtle way in
which unexpected exceptions can occur.
As described in Section D.3.1, “Floating-Point Exceptions and Their Defaults,”
depending on options determined by the software system designer, the processor
can perform one of two possible courses of action when a numeric exception occurs.
• The x87 FPU can provide a default fix-up for selected numeric exceptions. If the
x87 FPU performs its default action for all exceptions, then the need for exception
synchronization is not manifest. However, code is often ported to contexts and
operating systems for which it was not originally designed. Example D-1 and
Example D-2, below, illustrate that it is safest to always consider exception
synchronization when designing code that uses the x87 FPU.

D-16 Vol. 1
 GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

• Alternatively, a software exception handler can be invoked to handle the

exception. When a numeric exception is unmasked and the exception occurs, the
x87 FPU stops further execution of the numeric instruction and causes a branch
to a software exception handler. When an x87 FPU exception handler will be
invoked, synchronization must always be considered to assure reliable perfor-
mance.
Example D-1 and Example D-2, below, illustrate the need to always consider excep-
tion synchronization when writing numeric code, even when the code is initially
intended for execution with exceptions masked.

D.3.3.2 Exception Synchronization Examples

In the following examples, three instructions are shown to load an integer, calculate
its square root, then increment the integer. The synchronous execution of the x87
FPU will allow both of these programs to execute correctly, with INC COUNT being
executed in parallel in the processor, as long as no exceptions occur on the FILD
instruction. However, if the code is later moved to an environment where exceptions
are unmasked, the code in Example D-1 will not work correctly:

Example D-1. Incorrect Error Synchronization

FILD COUNT ;x87 FPU instruction

INC COUNT ;integer instruction alters operand
FSQRT ;subsequent x87 FPU instruction -- error
;from previous x87 FPU instruction detected here

Example D-2. Proper Error Synchronization

FILD COUNT ;x87 FPU instruction
FSQRT ;subsequent x87 FPU instruction -- error from
;previous x87 FPU instruction detected here
INC COUNT ;integer instruction alters operand

In some operating systems supporting the x87 FPU, the numeric register stack is
extended to memory. To extend the x87 FPU stack to memory, the invalid exception
is unmasked. A push to a full register or pop from an empty register sets SF (Stack
Fault flag) and causes an invalid operation exception. The recovery routine for the
exception must recognize this situation, fix up the stack, then perform the original
operation. The recovery routine will not work correctly in Example D-1. The problem
is that the value of COUNT increments before the exception handler is invoked, so
that the recovery routine will load an incorrect value of COUNT, causing the program
to fail or behave unreliably.

Vol. 1 D-17
GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

D.3.3.3 Proper Exception Synchronization

As explained in Section D.2.1.2, “Recommended External Hardware to Support the
MS-DOS* Compatibility Sub-mode,” if the x87 FPU encounters an unmasked excep-
tion condition a software exception handler is invoked before execution of the next
WAIT or floating-point instruction. This is because an unmasked floating-point
exception causes the processor to freeze immediately before executing such an
instruction (unless the IGNNE# input is active, or it is a no-wait x87 FPU instruction).
Exactly when the exception handler will be invoked (in the interval between when the
exception is detected and the next WAIT or x87 FPU instruction) is dependent on the
processor generation, the system, and which x87 FPU instruction and exception is
involved.
To be safe in exception synchronization, one should assume the handler will be
invoked at the end of the interval. Thus the program should not change any value
that might be needed by the handler (such as COUNT in Example D-1 and Example
D-2) until after the next x87 FPU instruction following an x87 FPU instruction that
could cause an error. If the program needs to modify such a value before the next
x87 FPU instruction (or if the next x87 FPU instruction could also cause an error),
then a WAIT instruction should be inserted before the value is modified. This will
force the handling of any exception before the value is modified. A WAIT instruction
should also be placed after the last floating-point instruction in an application so that
any unmasked exceptions will be serviced before the task completes.

D.3.4 x87 FPU Exception Handling Examples

There are many approaches to writing exception handlers. One useful technique is to
consider the exception handler procedure as consisting of “prologue,” “body,” and
“epilogue” sections of code.
In the transfer of control to the exception handler due to an INTR, NMI, or SMI,
external interrupts have been disabled by hardware. The prologue performs all func-
tions that must be protected from possible interruption by higher-priority sources.
Typically, this involves saving registers and transferring diagnostic information from
the x87 FPU to memory. When the critical processing has been completed, the
prologue may re-enable interrupts to allow higher-priority interrupt handlers to
preempt the exception handler. The standard “prologue” not only saves the registers
and transfers diagnostic information from the x87 FPU to memory but also clears the
floating-point exception flags in the status word. Alternatively, when it is not neces-
sary for the handler to be re-entrant, another technique may also be used. In this
technique, the exception flags are not cleared in the “prologue” and the body of the
handler must not contain any floating-point instructions that cannot complete execu-
tion when there is a pending floating-point exception. (The no-wait instructions are
discussed in Section 8.3.12, “Waiting vs. Non-waiting Instructions.”) Note that the
handler must still clear the exception flag(s) before executing the IRET. If the excep-
tion handler uses neither of these techniques, the system will be caught in an endless
loop of nested floating-point exceptions, and hang.

D-18 Vol. 1
 GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

The body of the exception handler examines the diagnostic information and makes a
response that is necessarily application-dependent. This response may range from
halting execution, to displaying a message, to attempting to repair the problem and
proceed with normal execution. The epilogue essentially reverses the actions of the
prologue, restoring the processor so that normal execution can be resumed. The
epilogue must not load an unmasked exception flag into the x87 FPU or another
exception will be requested immediately.
The following code examples show the ASM386/486 coding of three skeleton excep-
tion handlers, with the save spaces given as correct for 32-bit protected mode. They
show how prologues and epilogues can be written for various situations, but the
application-dependent exception handling body is just indicated by comments
showing where it should be placed.
The first two are very similar; their only substantial difference is their choice of
instructions to save and restore the x87 FPU. The trade-off here is between the
increased diagnostic information provided by FNSAVE and the faster execution of
FNSTENV. (Also, after saving the original contents, FNSAVE re-initializes the x87 FPU,
while FNSTENV only masks all x87 FPU exceptions.) For applications that are sensi-
tive to interrupt latency or that do not need to examine register contents, FNSTENV
reduces the duration of the “critical region,” during which the processor does not
recognize another interrupt request. (See the Section 8.1.10, “Saving the x87 FPU’s
State with FSTENV/FNSTENV and FSAVE/FNSAVE,” for a complete description of the
x87 FPU save image.) If the processor supports Streaming SIMD Extensions and the
operating system supports it, the FXSAVE instruction should be used instead of
FNSAVE. If the FXSAVE instruction is used, the save area should be increased to 512
bytes and aligned to 16 bytes to save the entire state. These steps will ensure that
the complete context is saved.
After the exception handler body, the epilogues prepare the processor to resume
execution from the point of interruption (for example, the instruction following the
one that generated the unmasked exception). Notice that the exception flags in the
memory image that is loaded into the x87 FPU are cleared to zero prior to reloading
(in fact, in these examples, the entire status word image is cleared).
Example D-3 and Example D-4 assume that the exception handler itself will not
cause an unmasked exception. Where this is a possibility, the general approach
shown in Example D-5 can be employed. The basic technique is to save the full x87
FPU state and then to load a new control word in the prologue. Note that considerable
care should be taken when designing an exception handler of this type to prevent the
handler from being reentered endlessly.

Example D-3. Full-State Exception Handler

SAVE_ALL PROC
;
;SAVE REGISTERS, ALLOCATE STACK SPACE FOR x87 FPU STATE IMAGE
PUSH EBP
.

Vol. 1 D-19
GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

.
MOV EBP, ESP
SUB ESP, 108 ; ALLOCATES 108 BYTES (32-bit PROTECTED MODE SIZE)
;SAVE FULL x87 FPU STATE, RESTORE INTERRUPT ENABLE FLAG (IF)
FNSAVE [EBP-108]
PUSH [EBP + OFFSET_TO_EFLAGS] ; COPY OLD EFLAGS TO STACK TOP
POPFD ;RESTORE IF TO VALUE BEFORE x87 FPU EXCEPTION
;
;APPLICATION-DEPENDENT EXCEPTION HANDLING CODE GOES HERE
;
;CLEAR EXCEPTION FLAGS IN STATUS WORD (WHICH IS IN MEMORY)
;RESTORE MODIFIED STATE IMAGE
MOV BYTE PTR [EBP-104], 0H
FRSTOR [EBP-108]
;DE-ALLOCATE STACK SPACE, RESTORE REGISTERS
MOV ESP, EBP
.
.
POP EBP
;
;RETURN TO INTERRUPTED CALCULATION
IRETD
SAVE_ALL ENDP

Example D-4. Reduced-Latency Exception Handler

SAVE_ENVIRONMENTPROC
;
;SAVE REGISTERS, ALLOCATE STACK SPACE FOR x87 FPU ENVIRONMENT
PUSH EBP
.
.
MOV EBP, ESP
SUB ESP, 28 ;ALLOCATES 28 BYTES (32-bit PROTECTED MODE SIZE)
;SAVE ENVIRONMENT, RESTORE INTERRUPT ENABLE FLAG (IF)
FNSTENV [EDP - 28]
PUSH [EBP + OFFSET_TO_EFLAGS] ; COPY OLD EFLAGS TO STACK TOP
POPFD ;RESTORE IF TO VALUE BEFORE x87 FPU EXCEPTION
;
;APPLICATION-DEPENDENT EXCEPTION HANDLING CODE GOES HERE
;
;CLEAR EXCEPTION FLAGS IN STATUS WORD (WHICH IS IN MEMORY)
;RESTORE MODIFIED ENVIRONMENT IMAGE

D-20 Vol. 1
 GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

MOV BYTE PTR [EBP-24], 0H

FLDENV [EBP-28]
;DE-ALLOCATE STACK SPACE, RESTORE REGISTERS
MOV ESP, EBP
.
.
POP EBP
;
;RETURN TO INTERRUPTED CALCULATION
IRETD
SAVE_ENVIRONMENT ENDP

Example D-5. Reentrant Exception Handler

.
.
LOCAL_CONTROL DW ?; ASSUME INITIALIZED
.
.
REENTRANTPROC
;
;SAVE REGISTERS, ALLOCATE STACK SPACE FOR x87 FPU STATE IMAGE
PUSH EBP
.
.
MOV EBP, ESP
SUB ESP, 108 ;ALLOCATES 108 BYTES (32-bit PROTECTED MODE SIZE)

;SAVE STATE, LOAD NEW CONTROL WORD, RESTORE INTERRUPT ENABLE FLAG (IF)
FNSAVE [EBP-108]
FLDCW LOCAL_CONTROL
PUSH [EBP + OFFSET_TO_EFLAGS] ;COPY OLD EFLAGS TO STACK TOP
POPFD ;RESTORE IF TO VALUE BEFORE x87 FPU EXCEPTION

.
.
;
;APPLICATION-DEPENDENT EXCEPTION HANDLING CODE
;GOES HERE - AN UNMASKED EXCEPTION
;GENERATED HERE WILL CAUSE THE EXCEPTION HANDLER TO BE REENTERED
;IF LOCAL STORAGE IS NEEDED, IT MUST BE ALLOCATED ON THE STACK
.
;CLEAR EXCEPTION FLAGS IN STATUS WORD (WHICH IS IN MEMORY)
;RESTORE MODIFIED STATE IMAGE

Vol. 1 D-21
GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

MOV BYTE PTR [EBP-104], 0H

FRSTOR [EBP-108]
;DE-ALLOCATE STACK SPACE, RESTORE REGISTERS
MOV ESP, EBP
.
.
POP EBP
;
;RETURN TO POINT OF INTERRUPTION
IRETD
REENTRANT ENDP

D.3.5 Need for Storing State of IGNNE# Circuit If Using x87 FPU
and SMM
The recommended circuit (see Figure D-1) for MS-DOS compatibility x87 FPU excep-
tion handling for Intel486 processors and beyond contains two flip flops. When the
x87 FPU exception handler accesses I/O port 0F0H it clears the IRQ13 interrupt
request output from Flip Flop #1 and also clocks out the IGNNE# signal (active) from
Flip Flop #2.
The assertion of IGNNE# may be used by the handler if needed to execute any x87
FPU instruction while ignoring the pending x87 FPU errors. The problem here is that
the state of Flip Flop #2 is effectively an additional (but hidden) status bit that can
affect processor behavior, and so ideally should be saved upon entering SMM, and
restored before resuming to normal operation. If this is not done, and also the SMM
code saves the x87 FPU state, AND an x87 FPU error handler is being used which
relies on IGNNE# assertion, then (very rarely) the x87 FPU handler will nest inside
itself and malfunction. The following example shows how this can happen.
Suppose that the x87 FPU exception handler includes the following sequence:

FNSTSW save_sw ; save the x87 FPU status word

; using a no-wait x87 FPU instruction
OUT0F0H, AL ; clears IRQ13 & activates IGNNE#
....
FLDCW new_cw ; loads new CW ignoring x87 FPU errors,
; since IGNNE# is assumed active; or any
; other x87 FPU instruction that is not a no-wait
; type will cause the same problem
....
FCLEX ; clear the x87 FPU error conditions & thus
; turn off FERR# & reset the IGNNE# FF

D-22 Vol. 1
 GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

The problem will only occur if the processor enters SMM between the OUT and the
FLDCW instructions. But if that happens, AND the SMM code saves the x87 FPU state
using FNSAVE, then the IGNNE# Flip Flop will be cleared (because FNSAVE clears the
x87 FPU errors and thus de-asserts FERR#). When the processor returns from SMM it
will restore the x87 FPU state with FRSTOR, which will re-assert FERR#, but the
IGNNE# Flip Flop will not get set. Then when the x87 FPU error handler executes the
FLDCW instruction, the active error condition will cause the processor to re-enter the
x87 FPU error handler from the beginning. This may cause the handler to malfunction.
To avoid this problem, Intel recommends two measures:
1. Do not use the x87 FPU for calculations inside SMM code. (The normal power
management, and sometimes security, functions provided by SMM have no need
for x87 FPU calculations; if they are needed for some special case, use scaling or
emulation instead.) This eliminates the need to do FNSAVE/FRSTOR inside SMM
code, except when going into a 0 V suspend state (in which, in order to save
power, the CPU is turned off completely, requiring its complete state to be saved).
2. The system should not call upon SMM code to put the processor into 0 V suspend
while the processor is running x87 FPU calculations, or just after an interrupt has
occurred. Normal power management protocol avoids this by going into power
down states only after timed intervals in which no system activity occurs.

D.3.6 Considerations When x87 FPU Shared Between Tasks

The IA-32 architecture allows speculative deferral of floating-point state swaps on
task switches. This feature allows postponing an x87 FPU state swap until an x87 FPU
instruction is actually encountered in another task. Since kernel tasks rarely use
floating-point, and some applications do not use floating-point or use it infrequently,
the amount of time saved by avoiding unnecessary stores of the floating-point state
is significant. Speculative deferral of x87 FPU saves does, however, place an extra
burden on the kernel in three key ways:
1. The kernel must keep track of which thread owns the x87 FPU, which may be
different from the currently executing thread.
2. The kernel must associate any floating-point exceptions with the generating task.
This requires special handling since floating-point exceptions are delivered
asynchronous with other system activity.
3. There are conditions under which spurious floating-point exception interrupts are
generated, which the kernel must recognize and discard.

D.3.6.1 Speculatively Deferring x87 FPU Saves, General Overview

In order to support multitasking, each thread in the system needs a save area for the
general-purpose registers, and each task that is allowed to use floating-point needs
an x87 FPU save area large enough to hold the entire x87 FPU stack and associated
x87 FPU state such as the control word and status word. (See Section 8.1.10,

Vol. 1 D-23
GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

“Saving the x87 FPU’s State with FSTENV/FNSTENV and FSAVE/FNSAVE,” for a
complete description of the x87 FPU save image.) If the processor and the operating
system support Streaming SIMD Extensions, the save area should be large enough
and aligned correctly to hold x87 FPU and Streaming SIMD Extensions state.
On a task switch, the general-purpose registers are swapped out to their save area
for the suspending thread, and the registers of the resuming thread are loaded. The
x87 FPU state does not need to be saved at this point. If the resuming thread does
not use the x87 FPU before it is itself suspended, then both a save and a load of the
x87 FPU state has been avoided. It is often the case that several threads may be
executed without any usage of the x87 FPU.
The processor supports speculative deferral of x87 FPU saves via interrupt 7 “Device
Not Available” (DNA), used in conjunction with CR0 bit 3, the “Task Switched” bit
(TS). (See “Control Registers” in Chapter 2 of the Intel® 64 and IA-32 Architectures
Software Developer’s Manual, Volume 3A.) Every task switch via the hardware
supported task switching mechanism (see “Task Switching” in Chapter 6 of the Intel®
64 and IA-32 Architectures Software Developer’s Manual, Volume 3A) sets TS. Multi-
threaded kernels that use software task switching1 can set the TS bit by reading CR0,
ORing a “1” into2 bit 3, and writing back CR0. Any subsequent floating-point instruc-
tions (now being executed in a new thread context) will fault via interrupt 7 before
execution.
This allows a DNA handler to save the old floating-point context and reload the x87
FPU state for the current thread. The handler should clear the TS bit before exit using
the CLTS instruction. On return from the handler the faulting thread will proceed with
its floating-point computation.
Some operating systems save the x87 FPU context on every task switch, typically
because they also change the linear address space between tasks. The problem and
solution discussed in the following sections apply to these operating systems also.

D.3.6.2 Tracking x87 FPU Ownership

Since the contents of the x87 FPU may not belong to the currently executing thread,
the thread identifier for the last x87 FPU user needs to be tracked separately. This is
not complicated; the kernel should simply provide a variable to store the thread iden-
tifier of the x87 FPU owner, separate from the variable that stores the identifier for
the currently executing thread. This variable is updated in the DNA exception

1 In a software task switch, the operating system uses a sequence of instructions to save the sus-
pending thread’s state and restore the resuming thread’s state, instead of the single long non-
interruptible task switch operation provided by the IA-32 architecture.
2 Although CR0, bit 2, the emulation flag (EM), also causes a DNA exception, do not use the EM bit as
a surrogate for TS. EM means that no x87 FPU is available and that floating-point instructions
must be emulated. Using EM to trap on task switches is not compatible with the MMX technology.
If the EM flag is set, MMX instructions raise the invalid opcode exception.

D-24 Vol. 1
 GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

handler, and is used by the DNA exception handler to find the x87 FPU save areas of
the old and new threads. A simplified flow for a DNA exception handler is then:
1. Use the “x87 FPU Owner” variable to find the x87 FPU save area of the last thread
to use the x87 FPU.
2. Save the x87 FPU contents to the old thread’s save area, typically using an
FNSAVE or FXSAVE instruction.
3. Set the x87 FPU Owner variable to the identify the currently executing thread.
4. Reload the x87 FPU contents from the new thread’s save area, typically using an
FRSTOR or FXSTOR instruction.
5. Clear TS using the CLTS instruction and exit the DNA exception handler.
While this flow covers the basic requirements for speculatively deferred x87 FPU
state swaps, there are some additional subtleties that need to be handled in a robust
implementation.

D.3.6.3 Interaction of x87 FPU State Saves and Floating-Point Exception

Association
Recall these key points from earlier in this document: When considering floating-
point exceptions across all implementations of the IA-32 architecture, and across all
floating-point instructions, a floating-point exception can be initiated from any time
during the excepting floating-point instruction, up to just before the next floating-
point instruction. The “next” floating-point instruction may be the FNSAVE used to
save the x87 FPU state for a task switch. In the case of “no-wait:” instructions such
as FNSAVE, the interrupt from a previously excepting instruction (NE = 0 case) may
arrive just before the no-wait instruction, during, or shortly thereafter with a system
dependent delay.
Note that this implies that an floating-point exception might be registered during the
state swap process itself, and the kernel and floating-point exception interrupt
handler must be prepared for this case.
A simple way to handle the case of exceptions arriving during x87 FPU state swaps is
to allow the kernel to be one of the x87 FPU owning threads. A reserved thread iden-
tifier is used to indicate kernel ownership of the x87 FPU. During an floating-point
state swap, the “x87 FPU owner” variable should be set to indicate the kernel as the
current owner. At the completion of the state swap, the variable should be set to indi-
cate the new owning thread. The numeric exception handler needs to check the x87
FPU owner and discard any numeric exceptions that occur while the kernel is the x87
FPU owner. A more general flow for a DNA exception handler that handles this case is
shown in Figure D-5.
Numeric exceptions received while the kernel owns the x87 FPU for a state swap
must be discarded in the kernel without being dispatched to a handler. A flow for a
numeric exception dispatch routine is shown in Figure D-6.

Vol. 1 D-25
GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

It may at first glance seem that there is a possibility of floating-point exceptions

being lost because of exceptions that are discarded during state swaps. This is not
the case, as the exception will be re-issued when the floating-point state is reloaded.
Walking through state swaps both with and without pending numeric exceptions will
clarify the operation of these two handlers.

DNA Handler Entry

<other handler set up code>

Current Thread
same as
FPU Owner? Yes

No
FPU Owner := Kernel

FNSAVE to Old Thread’s

FP Save Area
(may cause numeric exception) <handler final clean-up>

FRSTOR from Current Thread’s CLTS (clears CR0.TS)

FP Save Area

<other handler code> Exit DNA Handler

FPU Owner := Current Thread

Figure D-5. General Program Flow for DNA Exception Handler

D-26 Vol. 1
 GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

Numeric Exception Entry

Is Kernel
FPU Owner? Yes

No
Normal Dispatch to
Numeric Exception Handler Exit

Figure D-6. Program Flow for a Numeric Exception Dispatch Routine

Case #1: x87 FPU State Swap Without Numeric Exception

Assume two threads A and B, both using the floating-point unit. Let A be the thread
to have most recently executed a floating-point instruction, with no pending numeric
exceptions. Let B be the currently executing thread. CR0.TS was set when thread A
was suspended.
When B starts to execute a floating-point instruction the instruction will fault with the
DNA exception because TS is set.
At this point the handler is entered, and eventually it finds that the current x87 FPU
Owner is not the currently executing thread. To guard the x87 FPU state swap from
extraneous numeric exceptions, the x87 FPU Owner is set to be the kernel. The old
owner’s x87 FPU state is saved with FNSAVE, and the current thread’s x87 FPU state
is restored with FRSTOR. Before exiting, the x87 FPU owner is set to thread B, and
the TS bit is cleared.
On exit, thread B resumes execution of the faulting floating-point instruction and
continues.

Case #2: x87 FPU State Swap with Discarded Numeric Exception
Again, assume two threads A and B, both using the floating-point unit. Let A be the
thread to have most recently executed a floating-point instruction, but this time let
there be a pending numeric exception. Let B be the currently executing thread. When
B starts to execute a floating-point instruction the instruction will fault with the DNA
exception and enter the DNA handler. (If both numeric and DNA exceptions are
pending, the DNA exception takes precedence, in order to support handling the
numeric exception in its own context.)
When the FNSAVE starts, it will trigger an interrupt via FERR# because of the
pending numeric exception. After some system dependent delay, the numeric excep-
tion handler is entered. It may be entered before the FNSAVE starts to execute, or it
may be entered shortly after execution of the FNSAVE. Since the x87 FPU Owner is
the kernel, the numeric exception handler simply exits, discarding the exception. The

Vol. 1 D-27
GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

DNA handler resumes execution, completing the FNSAVE of the old floating-point
context of thread A and the FRSTOR of the floating-point context for thread B.
Thread A eventually gets an opportunity to handle the exception that was discarded
during the task switch. After some time, thread B is suspended, and thread A
resumes execution. When thread A starts to execute an floating-point instruction,
once again the DNA exception handler is entered. B’s x87 FPU state is Finessed, and
A’s x87 FPU state is Frustrate. Note that in restoring the x87 FPU state from A’s save
area, the pending numeric exception flags are reloaded into the floating-point status
word. Now when the DNA exception handler returns, thread A resumes execution of
the faulting floating-point instruction just long enough to immediately generate a
numeric exception, which now gets handled in the normal way. The net result is that
the task switch and resulting x87 FPU state swap via the DNA exception handler
causes an extra numeric exception which can be safely discarded.

D.3.6.4 Interrupt Routing From the Kernel

In MS-DOS, an application that wishes to handle numeric exceptions hooks interrupt
16 by placing its handler address in the interrupt vector table, and exiting via a jump
to the previous interrupt 16 handler. Protected mode systems that run MS-DOS
programs under a subsystem can emulate this exception delivery mechanism. For
example, assume a protected mode OS. that runs with CR0.NE[bit 5] = 1, and that
runs MS-DOS programs in a virtual machine subsystem. The MS-DOS program is
set up in a virtual machine that provides a virtualized interrupt table. The MS-DOS
application hooks interrupt 16 in the virtual machine in the normal way. A numeric
exception will trap to the kernel via the real INT 16 residing in the kernel at ring 0.
The INT 16 handler in the kernel then locates the correct MS-DOS virtual machine,
and reflects the interrupt to the virtual machine monitor. The virtual machine monitor
then emulates an interrupt by jumping through the address in the virtualized inter-
rupt table, eventually reaching the application’s numeric exception handler.

D.3.6.5 Special Considerations for Operating Systems that Support

Streaming SIMD Extensions
Operating systems that support Streaming SIMD Extensions instructions introduced
with the Pentium III processor should use the FXSAVE and FXRSTOR instructions to
save and restore the new SIMD floating-point instruction register state as well as the
floating-point state. Such operating systems must consider the following issues:
1. Enlarged state save area — FNSAVE/FRSTOR instructions operate on a
94-byte or 108-byte memory region, depending on whether they are executed in
16-bit or 32-bit mode. The FXSAVE/FXRSTOR instructions operate on a 512-byte
memory region.
2. Alignment requirements — FXSAVE/FXRSTOR instructions require the
memory region on which they operate to be 16-byte aligned (refer to the
individual instruction instructions descriptions in Chapter 3 of the Intel® 64 and

D-28 Vol. 1
 GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

IA-32 Architectures Software Developer’s Manual, Volume 2A, for information

about exceptions generated if the memory region is not aligned).
3. Maintaining compatibility with legacy applications/libraries — The
operating system changes to support Streaming SIMD Extensions must be
invisible to legacy applications or libraries that deal only with floating-point
instructions. The layout of the memory region operated on by the
FXSAVE/FXRSTOR instructions is different from the layout for the
FNSAVE/FRSTOR instructions. Specifically, the format of the x87 FPU tag word
and the length of the various fields in the memory region is different. Care must
be taken to return the x87 FPU state to a legacy application (e.g., when reporting
FP exceptions) in the format it expects.
4. Instruction semantic differences — There are some semantic differences
between the way the FXSAVE and FSAVE/FNSAVE instructions operate. The
FSAVE/FNSAVE instructions clear the x87 FPU after they save the state while the
FXSAVE instruction saves the x87 FPU/Streaming SIMD Extensions state but
does not clear it. Operating systems that use FXSAVE to save the x87 FPU state
before making it available for another thread (e.g., during thread switch time)
should take precautions not to pass a “dirty” x87 FPU to another application.

D.4 DIFFERENCES FOR HANDLERS USING NATIVE MODE

The 8087 has an INT pin which it asserts when an unmasked exception occurs. But
there is no interrupt input pin in the 8086 or 8088 dedicated to its attachment, nor an
interrupt vector number in the 8086 or 8088 specific for an x87 FPU error assertion.
Beginning with the Intel 286 and Intel 287 hardware, a connection was dedicated to
support the x87 FPU exception and interrupt vector 16 was assigned to it.

D.4.1 Origin with the Intel 286 and Intel 287, and Intel386
and Intel 387 Processors
The Intel 286 and Intel 287, and Intel386 and Intel 387 processor/coprocessor pairs
are each provided with ERROR# pins that are recommended to be connected
between the processor and x87 FPU. If this is done, when an unmasked x87 FPU
exception occurs, the x87 FPU records the exception, and asserts its ERROR# pin.
The processor recognizes this active condition of the ERROR# status line immediately
before execution of the next WAIT or x87 FPU instruction (except for the no-wait
type) in its instruction stream, and branches to the routine at interrupt vector 16.
Thus an x87 FPU exception will be handled before any other x87 FPU instruction
(after the one causing the error) is executed (except for no-wait instructions, which
will be executed without triggering the x87 FPU exception interrupt, but it will remain
pending).

Vol. 1 D-29
GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

Using the dedicated INT 16 for x87 FPU exception handling is referred to as the
native mode. It is the simplest approach, and the one recommended most highly by
Intel.

D.4.2 Changes with Intel486, Pentium and Pentium Pro

Processors with CR0.NE[bit 5] = 1
With these latest three generations of the IA-32 architecture, more enhancements
and speedup features have been added to the corresponding x87 FPUs. Also, the x87
FPU is now built into the same chip as the processor, which allows further increases
in the speed at which the x87 FPU can operate as part of the integrated system. This
also means that the native mode of x87 FPU exception handling, selected by setting
bit NE of register CR0 to 1, is now entirely internal.
If an unmasked exception occurs during an x87 FPU instruction, the x87 FPU records
the exception internally, and triggers the exception handler through interrupt 16
immediately before execution of the next WAIT or x87 FPU instruction (except for
no-wait instructions, which will be executed as described in Section D.4.1, “Origin
with the Intel 286 and Intel 287, and Intel386 and Intel 387 Processors”).
An unmasked numerical exception causes the FERR# output to be activated even
with NE = 1, and at exactly the same point in the program flow as it would have been
asserted if NE were zero. However, the system would not connect FERR# to a PIC to
generate INTR when operating in the native, internal mode. (If the hardware of a
system has FERR# connected to trigger IRQ13 in order to support MS-DOS, but an
operating system using the native mode is actually running the system, it is the oper-
ating system’s responsibility to make sure that IRQ13 is not enabled in the slave
PIC.) With this configuration a system is immune to the problem discussed in Section
D.2.1.3, “No-Wait x87 FPU Instructions Can Get x87 FPU Interrupt in Window,” where
for Intel486 and Pentium processors a no-wait x87 FPU instruction can get an x87
FPU exception.

D.4.3 Considerations When x87 FPU Shared Between Tasks Using

Native Mode
The protocols recommended in Section D.3.6, “Considerations When x87 FPU Shared
Between Tasks,” for MS-DOS compatibility x87 FPU exception handlers that are
shared between tasks may be used without change with the native mode. However,
the protocols for a handler written specifically for native mode can be simplified,
because the problem of a spurious floating-point exception interrupt occurring while
the kernel is executing cannot happen in native mode.
The problem as actually found in practical code in a MS-DOS compatibility system
happens when the DNA handler uses FNSAVE to switch x87 FPU contexts. If an x87
FPU exception is active, then FNSAVE triggers FERR# briefly, which usually will cause
the x87 FPU exception handler to be invoked inside the DNA handler. In native mode,
neither FNSAVE nor any other no-wait instructions can trigger interrupt 16. (As

D-30 Vol. 1
 GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

discussed above, FERR# gets asserted independent of the value of the NE bit, but
when NE = 1, the operating system should not enable its path through the PIC.)
Another possible (very rare) way a floating-point exception interrupt could occur
while the kernel is executing is by an x87 FPU immediate exception case having its
interrupt delayed by the external hardware until execution has switched to the
kernel. This also cannot happen in native mode because there is no delay through
external hardware.
Thus the native mode x87 FPU exception handler can omit the test to see if the kernel
is the x87 FPU owner, and the DNA handler for a native mode system can omit the
step of setting the kernel as the x87 FPU owner at the handler’s beginning. Since
however these simplifications are minor and save little code, it would be a reasonable
and conservative habit (as long as the MS-DOS compatibility mode is widely used) to
include these steps in all systems.
Note that the special DP (Dual Processing) mode for Pentium processors, and also
the more general Intel MultiProcessor Specification for systems with multiple
Pentium, P6 family, or Pentium 4 processors, support x87 FPU exception handling
only in the native mode. Intel does not recommend using the MS-DOS compatibility
mode for systems using more than one processor.

Vol. 1 D-31
GUIDELINES FOR WRITING X87 FPU EXCEPTION HANDLERS

D-32 Vol. 1
 APPENDIX E
GUIDELINES FOR WRITING SIMD FLOATING-POINT
EXCEPTION HANDLERS

See Section 11.5, “SSE, SSE2, and SSE3 Exceptions,” for a detailed discussion of
SIMD floating-point exceptions.
This appendix considers only SSE/SSE2/SSE3 instructions that can generate numeric
(SIMD floating-point) exceptions, and gives an overview of the necessary support for
handling such exceptions. This appendix does not address instructions that do not
generate floating-point exceptions (such as RSQRTSS, RSQRTPS, RCPSS, or RCPPS),
any x87 instructions, or any unlisted instruction.
For detailed information on which instructions generate numeric exceptions, and a
listing of those exceptions, refer to Appendix C, “Floating-Point Exceptions
Summary.” Non-numeric exceptions are handled in a way similar to that for the stan-
dard IA-32 instructions.

E.1 TWO OPTIONS FOR HANDLING FLOATING-POINT

EXCEPTIONS
Just as for x87 FPU floating-point exceptions, the processor takes one of two possible
courses of action when an SSE/SSE2/SSE3 instruction raises a floating-point excep-
tion:
• If the exception being raised is masked (by setting the corresponding mask bit in
the MXCSR to 1), then a default result is produced which is acceptable in most
situations. No external indication of the exception is given, but the corresponding
exception flags in the MXCSR are set and may be examined later. Note though
that for packed operations, an exception flag that is set in the MXCSR will not tell
which of the sub-operands caused the event to occur.
• If the exception being raised is not masked (by setting the corresponding mask
bit in the MXCSR to 0), a software exception handler previously registered by the
user with operating system support will be invoked through the SIMD floating-
point exception (#XM, vector 19). This case is discussed below in Section E.2,
“Software Exception Handling.”

E.2 SOFTWARE EXCEPTION HANDLING

The exception handling routine reached via interrupt vector 19 is usually part of the
system software (the operating system kernel). Note that an interrupt descriptor
table (IDT) entry must have been previously set up for this vector (refer to Chapter

Vol. 1 E-1
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

5, “Interrupt and Exception Handling,” in the Intel® 64 and IA-32 Architectures Soft-
ware Developer’s Manual, Volume 3A). Some compilers use specific run-time
libraries to assist in floating-point exception handling. If any x87 FPU floating-point
operations are going to be performed that might raise floating-point exceptions, then
the exception handling routine must either disable all floating-point exceptions (for
example, loading a local control word with FLDCW), or it must be implemented as re-
entrant (for the case of x87 FPU exceptions, refer to Example D-1 in Appendix D,
“Guidelines for Writing x87 FPU Exception Handlers”). If this is not the case, the
routine has to clear the status flags for x87 FPU exceptions or to mask all x87 FPU
floating-point exceptions. For SIMD floating-point exceptions though, the exception
flags in MXCSR do not have to be cleared, even if they remain unmasked (but they
may still be cleared). Exceptions are in this case precise and occur immediately, and
a SIMD floating-point exception status flag that is set when the corresponding excep-
tion is unmasked will not generate an exception.
Typical actions performed by this low-level exception handling routine are:
• Incrementing an exception counter for later display or printing
• Printing or displaying diagnostic information (e.g. the MXCSR and XMM registers)
• Aborting further execution, or using the exception pointers to build an instruction
that will run without exception and executing it
• Storing information about the exception in a data structure that will be passed to
a higher level user exception handler
In most cases (and this applies also to SSE/SSE2/SSE3 instructions), there will be
three main components of a low-level floating-point exception handler: a prologue, a
body, and an epilogue.
The prologue performs functions that must be protected from possible interruption
by higher-priority sources - typically saving registers and transferring diagnostic
information from the processor to memory. When the critical processing has been
completed, the prologue may re-enable interrupts to allow higher-priority interrupt
handlers to preempt the exception handler (assuming that the interrupt handler was
called through an interrupt gate, meaning that the processor cleared the interrupt
enable (IF) flag in the EFLAGS register - refer to Section 6.4.1, “Call and Return
Operation for Interrupt or Exception Handling Procedures”).
The body of the exception handler examines the diagnostic information and makes a
response that is application-dependent. It may range from halting execution, to
displaying a message, to attempting to fix the problem and then proceeding with
normal execution, to setting up a data structure, calling a higher-level user exception
handler and continuing execution upon return from it. This latter case will be
assumed in Section E.4, “SIMD Floating-Point Exceptions and the IEEE Standard
754” below.
Finally, the epilogue essentially reverses the actions of the prologue, restoring the
processor state so that normal execution can be resumed.
The following example represents a typical exception handler. To link it with Example
E-2 that will follow in Section E.4.3, “Example SIMD Floating-Point Emulation Imple-

E-2 Vol. 1
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

mentation,” assume that the body of the handler (not shown here in detail) passes
the saved state to a routine that will examine in turn all the sub-operands of the
excepting instruction, invoking a user floating-point exception handler if a particular
set of sub-operands raises an unmasked (enabled) exception, or emulating the
instruction otherwise.

Example E-1. SIMD Floating-Point Exception Handler

SIMD_FP_EXC_HANDLER PROC

;PROLOGUE
;SAVE REGISTERS THAT MIGHT BE USED BY THE EXCEPTION HANDLER
PUSH EBP ;SAVE EBP
PUSH EAX ;SAVE EAX
...
MOV EBP, ESP ;SAVE ESP in EBP
SUB ESP, 512 ;ALLOCATE 512 BYTES
AND ESP, 0fffffff0h ;MAKE THE ADDRESS 16-BYTE ALIGNED
FXSAVE [ESP] ;SAVE FP, MMX, AND SIMD FP STATE
PUSH [EBP+EFLAGS_OFFSET] ;COPY OLD EFLAGS TO STACK TOP
POPFD ;RESTORE THE INTERRUPT ENABLE FLAG IF
;TO VALUE BEFORE SIMD FP EXCEPTION

;BODY
;APPLICATION-DEPENDENT EXCEPTION HANDLING CODE GOES HERE
LDMXCSR LOCAL_MXCSR ;LOAD LOCAL MXCSR VALUE IF NEEDED
...
...
;EPILOGUE
FXRSTOR [ESP] ;RESTORE MODIFIED STATE IMAGE
MOV ESP, EBP ;DE-ALLOCATE STACK SPACE
...
POP EAX ;RESTORE EAX
POP EBP ;RESTORE EBP
IRET ;RETURN TO INTERRUPTED CALCULATION
SIMD_FP_EXC_HANDLER ENDP

E.3 EXCEPTION SYNCHRONIZATION

An SSE/SSE2/SSE3 instruction can execute in parallel with other similar instructions,
with integer instructions, and with floating-point or MMX instructions. Unlike for x87
instructions, special precaution for exception synchronization is not necessary in
this case. This is because floating-point exceptions for SSE/SSE2/SSE3 instructions

Vol. 1 E-3
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

occur immediately and are not delayed until a subsequent floating-point instruction
is executed. However, floating-point emulation may be necessary when unmasked
floating-point exceptions are generated.

E.4 SIMD FLOATING-POINT EXCEPTIONS AND THE IEEE

STANDARD 754
SSE/SSE2/SSE3 extensions are 100% compatible with the IEEE Standard 754 for
Binary Floating-Point Arithmetic, satisfying all of its mandatory requirements (when
the flush-to-zero or denormals-are-zeros modes are not enabled). But a program-
ming environment that includes SSE/SSE2/SSE3 instructions will comply with both
the obligatory and the strongly recommended requirements of the IEEE Standard
754 regarding floating-point exception handling, only as a combination of hardware
and software (which is acceptable). The standard states that a user should be able to
request a trap on any of the five floating-point exceptions (note that the denormal
exception is an IA-32 addition), and it also specifies the values (operands or result)
to be delivered to the exception handler.
The main issue is that for SSE/SSE2/SSE3 instructions that raise post-computation
exceptions (traps: overflow, underflow, or inexact), unlike for x87 FPU instructions,
the processor does not provide the result recommended by IEEE Standard 754 to the
user handler. If a user program needs the result of an instruction that generated a
post-computation exception, it is the responsibility of the software to produce this
result by emulating the faulting SSE/SSE2/SSE3 instruction. Another issue is that the
standard does not specify explicitly how to handle multiple floating-point exceptions
that occur simultaneously. For packed operations, a logical OR of the flags that would
be set by each sub-operation is used to set the exception flags in the MXCSR. The
following subsections present one possible way to solve these problems.

E.4.1 Floating-Point Emulation

Every operating system must provide a kernel level floating-point exception handler
(a template was presented in Section E.2, “Software Exception Handling” above). In
the following discussion, assume that a user mode floating-point exception filter is
supplied for SIMD floating-point exceptions (for example as part of a library of C
functions), that a user program can invoke in order to handle unmasked exceptions.
The user mode floating-point exception filter (not shown here) has to be able to
emulate the subset of SSE/SSE2/SSE3 instructions that can generate numeric
exceptions, and has to be able to invoke a user provided floating-point exception
handler for floating-point exceptions. When a floating-point exception that is not
masked is raised by an SSE/SSE2/SSE3 instruction, the low-level floating-point
exception handler will be called. This low-level handler may in turn call the user mode
floating-point exception filter. The filter function receives the original operands of the
excepting instruction as no results are provided by the hardware, whether a pre-
computation or a post-computation exception has occurred. The filter will unpack the

E-4 Vol. 1
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

operands into up to four sets of sub-operands, and will submit them one set at a time
to an emulation function (See Example E-2 in Section E.4.3, “Example SIMD
Floating-Point Emulation Implementation”). The emulation function will examine the
sub-operands, and will possibly redo the necessary calculation.
Two cases are possible:
• If an unmasked (enabled) exception would occur in this process, the emulation
function will return to its caller (the filter function) with the appropriate infor-
mation. The filter will invoke a (previously registered) user floating-point
exception handler for this set of sub-operands, and will record the result upon
return from the user handler (provided the user handler allows continuation of
the execution).
• If no unmasked (enabled) exception would occur, the emulation function will
determine and will return to its caller the result of the operation for the current
set of sub-operands (it has to be IEEE Standard 754 compliant). The filter
function will record the result (plus any new flag settings).
The user level filter function will then call the emulation function for the next set of
sub-operands (if any). When done with all the operand sets, the partial results will be
packed (if the excepting instruction has a packed floating-point result, which is true
for most SSE/SSE2/SSE3 numeric instructions) and the filter will return to the low-
level exception handler, which in turn will return from the interruption, allowing
execution to continue. Note that the instruction pointer (EIP) has to be altered to
point to the instruction following the excepting instruction, in order to continue
execution correctly.
If a user mode floating-point exception filter is not provided, then all the work for
decoding the excepting instruction, reading its operands, emulating the instruction
for the components of the result that do not correspond to unmasked floating-point
exceptions, and providing the compounded result will have to be performed by the
user-provided floating-point exception handler.
Actual emulation might have to take place for one operand or pair of operands for
scalar operations, and for all sub-operands or pairs of sub-operands for packed oper-
ations. The steps to perform are the following:
• The excepting instruction has to be decoded and the operands have to be read
from the saved context.
• The instruction has to be emulated for each (pair of) sub-operand(s); if no
floating-point exception occurs, the partial result has to be saved; if a masked
floating-point exception occurs, the masked result has to be produced through
emulation and saved, and the appropriate status flags have to be set; if an
unmasked floating-point exception occurs, the result has to be generated by the
user provided floating-point exception handler, and the appropriate status flags
have to be set.
• The partial results have to be combined and written to the context that will be
restored upon application program resumption.

Vol. 1 E-5
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

A diagram of the control flow in handling an unmasked floating-point exception is

presented below.

User Application

Low-Level Floating-Point Exception Handler

User Level Floating-Point Exception Filter

User Floating-Point Exception Handler

Figure E-1. Control Flow for Handling Unmasked Floating-Point Exceptions

From the user-level floating-point filter, Example E-2 in Section E.4.3, “Example
SIMD Floating-Point Emulation Implementation,” will present only the floating-point
emulation part. In order to understand the actions involved, the expected response
to exceptions has to be known for all SSE/SSE2/SSE3 numeric instructions in two
situations: with exceptions enabled (unmasked result), and with exceptions disabled
(masked result). The latter can be found in Section 6.4, “Interrupts and Exceptions.”
The response to NaN operands that do not raise an exception is specified in Section
4.8.3.4, “NaNs.” Operations on NaNs are explained in the same source. This response
is also discussed in more detail in the next subsection, along with the unmasked and
masked responses to floating-point exceptions.

E.4.2 SSE/SSE2/SSE3 Response To Floating-Point Exceptions

This subsection specifies the unmasked response expected from the SSE/SSE2/SSE3
instructions that raise floating-point exceptions. The masked response is given in
parallel, as it is necessary in the emulation process of the instructions that raise
unmasked floating-point exceptions. The response to NaN operands is also included
in more detail than in Section 4.8.3.4, “NaNs.” For floating-point exception priority,
refer to “Priority Among Simultaneous Exceptions and Interrupts” in Chapter 5,

E-6 Vol. 1
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

“Interrupt and Exception Handling,” of Intel® 64 and IA-32 Architectures Software

Developer’s Manual, Volume 3A.

E.4.2.1 Numeric Exceptions

There are six classes of numeric (floating-point) exception conditions that can occur:
Invalid operation (#I), Divide-by-Zero (#Z), Denormal Operand (#D), Numeric
Overflow (#O), Numeric Underflow (#U), and Inexact Result (precision) (#P). #I,
#Z, #D are pre-computation exceptions (floating-point faults), detected before the
arithmetic operation. #O, #U, #P are post-computation exceptions (floating-point
traps).
Users can control how the SSE/SSE2/SSE3 floating-point exceptions are handled by
setting the mask/unmask bits in MXCSR. Masked exceptions are handled by the
processor, or by software if they are combined with unmasked exceptions occurring
in the same instruction. Unmasked exceptions are usually handled by the low-level
exception handler, in conjunction with user-level software.

E.4.2.2 Results of Operations with NaN Operands or a NaN Result for

SSE/SSE2/SSE3 Numeric Instructions
The tables below (E-1 through E-10) specify the response of SSE/SSE2/SSE3
instructions to NaN inputs, or to other inputs that lead to NaN results.
These results will be referenced by subsequent tables (e.g., E-10). Most operations
do not raise an invalid exception for quiet NaN operands, but even so, they will have
higher precedence over raising floating-point exceptions other than invalid opera-
tion.
Note that the single precision QNaN Indefinite value is 0xffc00000, the double preci-
sion QNaN Indefinite value is 0xfff8000000000000, and the Integer Indefinite value
is 0x80000000 (not a floating-point number, but it can be the result of a conversion
instruction from floating-point to integer).
For an unmasked exception, no result will be provided by the hardware to the user
handler. If a user registered floating-point exception handler is invoked, it may
provide a result for the excepting instruction, that will be used if execution of the
application code is continued after returning from the interruption.
In Tables E-1 through Table E-12, the specified operands cause an invalid exception,
unless the unmasked result is marked with “not an exception”. In this latter case, the
unmasked and masked results are the same.

Vol. 1 E-7
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

Table E-1. ADDPS, ADDSS, SUBPS, SUBSS, MULPS, MULSS, DIVPS, DIVSS, ADDPD,
ADDSD, SUBPD, SUBSD, MULPD, MULSD, DIVPD, DIVSD, ADDSUBPS, ADDSUBPD,
HADDPS, HADDPD, HSUBPS, HSUBPD
Source Operands Masked Result Unmasked Result
1
SNaN1 op SNaN2 SNaN1 | 00400000H or None
SNaN1 |
0008000000000000H2
SNaN1 op QNaN2 SNaN1 | 00400000H or None
SNaN1 |
0008000000000000H2
QNaN1 op SNaN2 QNaN1 None
QNaN1 op QNaN2 QNaN1 QNaN1 (not an exception)
SNaN op real value SNaN | 00400000H or None
SNaN1 |
0008000000000000H2
Real value op SNaN SNaN | 00400000H or None
SNaN1 |
0008000000000000H2
QNaN op real value QNaN QNaN (not an exception)
Real value op QNaN QNaN QNaN (not an exception)
Neither source operand is Single precision or double None
SNaN, precision QNaN Indefinite
but #I is signaled (e.g. for Inf -
Inf,
Inf ∗ 0, Inf / Inf, 0/0)
NOTES:
1. For Tables E-1 to E-12: op denotes the operation to be performed.
2. SNaN | 0x00400000 is a quiet NaN in single precision format (if SNaN is in single precision) and
SNaN | 0008000000000000H is a quiet NaN in double precision format (if SNaN is in double
precision), obtained from the signaling NaN given as input.
3. Operations involving only quiet NaNs do not raise floating-point exceptions.

E-8 Vol. 1
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

Table E-2. CMPPS.EQ, CMPSS.EQ, CMPPS.ORD, CMPSS.ORD,

CMPPD.EQ, CMPSD.EQ, CMPPD.ORD, CMPSD.ORD
Source Operands Masked Result Unmasked Result
NaN op Opd2 (any Opd2) 00000000H or 00000000H or
0000000000000000H1 0000000000000000H1 (not
an exception)
Opd1 op NaN (any Opd1) 00000000H or 00000000H or
0000000000000000H1 0000000000000000H1 (not
an exception)
NOTE:
1. 32-bit results are for single, and 64-bit results for double precision operations.

Table E-3. CMPPS.NEQ, CMPSS.NEQ, CMPPS.UNORD, CMPSS.UNORD, CMPPD.NEQ,

CMPSD.NEQ, CMPPD.UNORD, CMPSD.UNORD
Source Operands Masked Result Unmasked Result
NaN op Opd2 (any Opd2) FFFFFFFFH or FFFFFFFFH or
FFFFFFFFFFFFFFFFH1 FFFFFFFFFFFFFFFFH1 (not an
exception)
Opd1 op NaN (any Opd1) FFFFFFFFH or FFFFFFFFH or
FFFFFFFFFFFFFFFFH1 FFFFFFFFFFFFFFFFH1 (not an
exception)
NOTE:
1. 32-bit results are for single, and 64-bit results for double precision operations.

Table E-4. CMPPS.LT, CMPSS.LT, CMPPS.LE, CMPSS.LE, CMPPD.LT, CMPSD.LT,

CMPPD.LE, CMPSD.LE
Source Operands Masked Result Unmasked Result
NaN op Opd2 (any Opd2) 00000000H or None
0000000000000000H1
Opd1 op NaN (any Opd1) 00000000H or None
0000000000000000H1
NOTE:
1. 32-bit results are for single, and 64-bit results for double precision operations.

Vol. 1 E-9
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

Table E-5. CMPPS.NLT, CMPSS.NLT, CMPPS.NLE, CMPSS.NLE, CMPPD.NLT, CMPSD.NLT,

CMPPD.NLE, CMPSD.NLE
Source Operands Masked Result Unmasked Result
NaN op Opd2 (any Opd2) FFFFFFFFH or None
FFFFFFFFFFFFFFFFH1
Opd1 op NaN (any Opd1) FFFFFFFFH or None
FFFFFFFFFFFFFFFFH1
NOTE:
1. 32-bit results are for single, and 64-bit results for double precision operations.

Table E-6. COMISS, COMISD

Source Operands Masked Result Unmasked Result
SNaN op Opd2 (any Opd2) OF, SF, AF = 000 None
ZF, PF, CF = 111
Opd1 op SNaN (any Opd1) OF, SF, AF = 000 None
ZF, PF, CF = 111
QNaN op Opd2 (any Opd2) OF, SF, AF = 000 None
ZF, PF, CF = 111
Opd1 op QNaN (any Opd1) OF, SF, AF = 000 None
ZF, PF, CF = 111

Table E-7. UCOMISS, UCOMISD

Source Operands Masked Result Unmasked Result
SNaN op Opd2 (any Opd2) OF, SF, AF = 000 None
ZF, PF, CF = 111
Opd1 op SNaN (any Opd1) OF, SF, AF = 000 None
ZF, PF, CF = 111
QNaN op Opd2 OF, SF, AF = 000 OF, SF, AF = 000
(any Opd2 ≠ SNaN) ZF, PF, CF = 111 ZF, PF, CF = 111 (not an
exception)
Opd1 op QNaN OF, SF, AF = 000 OF, SF, AF = 000
(any Opd1 ≠ SNaN) ZF, PF, CF = 111 ZF, PF, CF = 111 (not an
exception)

E-10 Vol. 1
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

Table E-8. CVTPS2PI, CVTSS2SI, CVTTPS2PI, CVTTSS2SI, CVTPD2PI, CVTSD2SI,

CVTTPD2PI, CVTTSD2SI, CVTPS2DQ, CVTTPS2DQ, CVTPD2DQ, CVTTPD2DQ
Source Operand Masked Result Unmasked Result
SNaN 80000000H or None
80000000000000001
(Integer Indefinite)
QNaN 80000000H or None
80000000000000001
(Integer Indefinite)
NOTE:
1. 32-bit results are for single, and 64-bit results for double precision operations.

Table E-9. MAXPS, MAXSS, MINPS, MINSS, MAXPD, MAXSD, MINPD, MINSD
Source Operands Masked Result Unmasked Result
Opd1 op NaN2 (any Opd1) NaN2 None
NaN1 op Opd2 (any Opd2) Opd2 None
NOTE:
1. SNaN and QNaN operands raise an Invalid Operation fault.

Table E-10. SQRTPS, SQRTSS, SQRTPD, SQRTSD

Source Operand Masked Result Unmasked Result
QNaN QNaN QNaN (not an exception)
SNaN SNaN | 00400000H or None
SNaN |
0008000000000000H1
Source operand is not SNaN; Single precision or None
but #I is signaled (e.g. for double precision QNaN
sqrt (-1.0)) Indefinite
NOTE:
1. SNaN | 00400000H is a quiet NaN in single precision format (if SNaN is in single precision) and
SNaN | 0008000000000000H is a quiet NaN in double precision format (if SNaN is in double
precision), obtained from the signaling NaN given as input.

Vol. 1 E-11
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

Table E-11. CVTPS2PD, CVTSS2SD

Source Operands Masked Result Unmasked Result
1
QNaN QNaN1 QNaN11 (not an exception)
SNaN QNaN12 None
NOTES:
1. The double precision output QNaN1 is created from the single precision input QNaN as follows:
the sign bit is preserved, the 8-bit exponent FFH is replaced by the 11-bit exponent 7FFH, and
the 24-bit significand is extended to a 53-bit significand by appending 29 bits equal to 0.
2. The double precision output QNaN1 is created from the single precision input SNaN as follows:
the sign bit is preserved, the 8-bit exponent FFH is replaced by the 11-bit exponent 7FFH, and
the 24-bit significand is extended to a 53-bit significand by pending 29 bits equal to 0. The sec-
ond most significant bit of the significand is changed from 0 to 1 to convert the signaling NaN
into a quiet NaN.

Table E-12. CVTPD2PS, CVTSD2SS

Source Operands Masked Result Unmasked Result
QNaN QNaN11 QNaN11 (not an exception)
SNaN QNaN12 None
NOTES:
1. The single precision output QNaN1 is created from the double precision input QNaN as follows:
the sign bit is preserved, the 11-bit exponent 7FFH is replaced by the 8-bit exponent FFH, and
the 53-bit significand is truncated to a 24-bit significand by removing its 29 least significant
bits.
2. The single precision output QNaN1 is created from the double precision input SNaN as follows:
the sign bit is preserved, the 11-bit exponent 7FFH is replaced by the 8-bit exponent FFH, and
the 53-bit significand is truncated to a 24-bit significand by removing its 29 least significant
bits. The second most significant bit of the significand is changed from 0 to 1 to convert the sig-
naling NaN into a quiet NaN.

E.4.2.3 Condition Codes, Exception Flags, and Response for Masked and
Unmasked Numeric Exceptions
In the following, the masked response is what the processor provides when a masked
exception is raised by an SSE/SSE2/SSE3 numeric instruction. The same response is
provided by the floating-point emulator for SSE/SSE2/SSE3 numeric instructions,
when certain components of the quadruple input operands generate exceptions that
are masked (the emulator also generates the correct answer, as specified by IEEE
Standard 754 wherever applicable, in the case when no floating-point exception
occurs). The unmasked response is what the emulator provides to the user handler
for those components of the packed operands of SSE/SSE2/SSE3 instructions that
raise unmasked exceptions. Note that for pre-computation exceptions (floating-point

E-12 Vol. 1
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

faults), no result is provided to the user handler. For post-computation exceptions

(floating-point traps), a result is provided to the user handler, as specified below.
In the following tables, the result is denoted by 'res', with the understanding that for
the actual instruction, the destination coincides with the first source operand (except
for COMISS, UCOMISS, COMISD, and UCOMISD, whose destination is the EFLAGS
register).

Table E-13. #I - Invalid Operations

Unmasked
Response and
Instruction Condition Masked Response Exception Code
ADDPS src1 or src2 = SNaN
1
Refer to Table E-1 for src1, src2
ADDPD NaN operands, #IA = 1 unchanged; #IA =
ADDSS 1
ADDSD
HADDPS
HADDPD
ADDSUBPS (the src1 = +Inf, src2 = -Inf or res1 = QNaN Indefinite,
addition src1 = -Inf, src2 = +Inf #IA = 1
component)
ADDSUBPD (the
addition
component)
SUBPS src1 or src2 = SNaN Refer to Table E-1 for NaN src1, src2
SUBPD operands, #IA = 1 unchanged; #IA =
SUBSS 1
SUBSD
HSUBPS
HSUBPD
ADDSUBPS (the src1 = +Inf, src2 = +Inf or res = QNaN Indefinite,
subtraction src1 = -Inf, src2 = -Inf #IA = 1
component)
ADDSUBPD (the
subtraction
component)
MULPS src1 or src2 = SNaN Refer to Table E-1 for src1, src2
MULPD NaN operands, #IA = 1 unchanged;
#IA = 1
MULSS src1 = ±Inf, src2 = ±0 or res = QNaN Indefinite,
MULSD src1 = ±0, src2 = ±Inf #IA = 1
DIVPS src1 or src2 = SNaN Refer to Table E-1 for src1, src2
DIVPD NaN operands, #IA = 1 unchanged;
#IA = 1
DIVSS src1 = ±Inf, src2 = ±Inf or res = QNaN Indefinite,
DIVSD src1 = ±0, src2 = ±0 #IA = 1

Vol. 1 E-13
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

Table E-13. #I - Invalid Operations (Contd.)

Unmasked
Response and
Instruction Condition Masked Response Exception Code
SQRTPS src = SNaN Refer to Table E-10 for src unchanged,
SQRTPD NaN operands, #IA = 1 #IA = 1
SQRTSS
src < 0 res = QNaN Indefinite,
SQRTSD
(note that -0 < 0 is false) #IA = 1
MAXPS src1 = NaN or src2 = NaN res = src2, #IA = 1 src1, src2
MAXSS unchanged; #IA =
MAXPD 1
MAXSD
MINPS src1 = NaN or src2 = NaN res = src2, #IA = 1 src1, src2
MINSS unchanged; #IA =
MINPD 1
MINSD

E-14 Vol. 1
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

Table E-13. #I - Invalid Operations (Contd.)

Unmasked
Response and
Instruction Condition Masked Response Exception Code
CMPPS.LT src1 = NaN or src2 = NaN Refer to Table E-4 and src1, src2
CMPPS.LE Table E-5 for NaN unchanged; #IA =
CMPPS.NLT operands; #IA = 1 1
CMPPS.NLE
CMPSS.LT
CMPSS.LE
CMPSS.NLT
CMPSS.NLE
CMPPD.LT
CMPPD.LE
CMPPD.NLT
CMPPD.NLE
CMPSD.LT
CMPSD.LE
CMPSD.NLT
CMPSD.NLE
COMISS src1 = NaN or src2 = NaN Refer to Table E-6 for NaN src1, src2, EFLAGS
COMISD operands unchanged; #IA =
1
UCOMISS src1 = SNaN or src2 = SNaN Refer to Table E-7 for NaN src1, src2, EFLAGS
UCOMISD operands unchanged; #IA =
1
CVTPS2PI src = NaN, ±Inf, or res = Integer Indefinite, src unchanged,
CVTSS2SI |(src)rnd | > 7FFFFFFFH and #IA = 1 #IA = 1
CVTPD2PI (src)rnd ≠ 80000000H
CVTSD2SI
CVTPS2DQ See Note2 for information
CVTPD2DQ on rnd.
CVTTPS2PI src = NaN, ±Inf, or res = Integer Indefinite, src unchanged,
CVTTSS2SI |(src)rz | > 7FFFFFFFH and #IA = 1 #IA = 1
CVTTPD2PI (src)rz ≠ 80000000H
CVTTSD2SI
CVTTPS2DQ See Note2 for information
CVTTPD2DQ on rz.
CVTPS2PD src = NAN Refer to Table E-11 for src unchanged,
CVTSS2SD NaN operands #IA = 1
CVTPD2PS src = NAN Refer to Table E-12 for src unchanged,
CVTSD2SS NaN operands #IA = 1

Vol. 1 E-15
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

Table E-13. #I - Invalid Operations (Contd.)

Unmasked
Response and
Instruction Condition Masked Response Exception Code
NOTES:
1. For Tables E-13 to E-18:
- src denotes the single source operand of a unary operation.
- src1, src2 denote the first and second source operand of a binary operation.
- res denotes the numerical result of an operation.
2. rnd signifies the user rounding mode from MXCSR, and rz signifies the rounding mode toward
zero. (truncate), when rounding a floating-point value to an integer. For more information, refer
to Table 4-8.
3. For NAN encodings, see Table 4-3.

Table E-14. #Z - Divide-by-Zero

Unmasked
Response and
Instruction Condition Masked Response Exception Code
DIVPS src1 = finite non-zero (normal, res = ±Inf, src1, src2
DIVSS or denormal) #ZE = 1 unchanged;
DIVPD src2 = ±0 #ZE = 1
DIVPS

E-16 Vol. 1
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

Table E-15. #D - Denormal Operand

Unmasked Response
Instruction Condition Masked Response and Exception Code
ADDPS src1 = denormal1 or res = Result rounded to src1, src2 unchanged;
ADDPD src2 = denormal (and the destination precision #DE = 1
ADDSUBPS the DAZ bit in MXCSR and using the bounded
ADDSUBPD is 0) exponent, but only if no Note that SQRT,
HADDPS unmasked post- CVTPS2PD, CVTSS2SD,
HADDPD computation exception CVTPD2PS, CVTSD2SS
SUBPS occurs. have only 1 src.
SUBPD
HSUBPS
HSUBPD
MULPS
MULPD
DIVPS
DIVPD
SQRTPS
SQRTPD
MAXPS
MAXPD
MINPS
MINPD
CMPPS
CMPPD
ADDSS
ADDSD
SUBSS
SUBSD
MULSS
MULSD
DIVSS
DIVSD
SQRTSS
SQRTSD
MAXSS
MAXSD
MINSS
MINSD
CMPSS
CMPSD
COMISS
COMISD
UCOMISS
UCOMISD
CVTPS2PD

Vol. 1 E-17
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

Table E-15. #D - Denormal Operand (Contd.)

Unmasked Response
Instruction Condition Masked Response and Exception Code
CVTSS2SD
CVTPD2PS
CVTSD2SS
NOTE:
1. For denormal encodings, see Section 4.8.3.2, “Normalized and Denormalized Finite Numbers.”

Table E-16. #O - Numeric Overflow

Unmasked Response
Instruction Condition Masked Response and Exception Code
ADDPS Rounded result Roundi Sign Result & Status res = (result calculated
ADDSUBPS > largest single ng Flags with unbounded
HADDPS precision finite exponent and rounded to
To #OE = 1, #PE = 1
SUBPS normal value the destination precision)
nearest + res = + ∞
HSUBPS / 2192
- res = – ∞
MULPS #OE = 1
DIVPS Toward #OE = 1, #PE = 1 #PE = 1 if the result is
ADDSS –∞ + res = 1.11…1 * 2127 inexact
SUBSS - res = – ∞
MULSS Toward #OE = 1, #PE = 1
DIVSS +∞ + res = + ∞
CVTPD2PS - res = -1.11…1 * 2127
CVTSD2SS
Toward #OE = 1, #PE = 1
0 + res = 1.11…1 * 2127
- res = -1.11…1 * 2127

E-18 Vol. 1
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

Table E-16. #O - Numeric Overflow (Contd.)

Unmasked Response
Instruction Condition Masked Response and Exception Code
ADDPD Rounded result Roundi Sign Result & Status res = (result calculated
ADDSUBPD > largest double ng Flags with unbounded
HADDPD precision finite exponent and rounded to
To #OE = 1, #PE = 1
SUBPD normal value the destination precision)
nearest + res = + ∞
HSUBPD / 21536
- res = – ∞
MULPD • #OE = 1
DIVPD Toward #OE = 1, #PE = 1 • #PE = 1 if the result is
ADDSD –∞ + res = 1.11…1 * inexact
SUBSD - 21023
MULSD res = – ∞
DIVSD Toward #OE = 1, #PE = 1
+∞ + res = + ∞
- res = -1.11…1 *
21023
Toward #OE = 1, #PE = 1
0 + res = 1.11…1 *
- 21023
res = -1.11…1 *
21023

Vol. 1 E-19
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

Table E-17. #U - Numeric Underflow

Unmasked Response
Instruction Condition Masked Response and Exception Code
ADDPS Result calculated with res = ±0, denormal, or res = (result calculated
ADDSUBPS unbounded exponent and normal with unbounded
HADDPS rounded to the exponent and rounded to
SUBPS destination precision < #UE = 1 and #PE = 1, the destination precision)
HSUBPS smallest single precision but only if the result is * 2192
MULPS finite normal value. inexact • #UE = 1
DIVPS • #PE = 1 if the result is
ADDSS inexact
SUBSS
MULSS
DIVSS
CVTPD2PS
CVTSD2SS
ADDPD Result calculated with res = ±0, denormal or res = (result calculated
ADDSUBPD unbounded exponent and normal with unbounded
HADDPD rounded to the exponent and rounded to
SUBPD destination precision < #UE = 1 and #PE = 1, the destination precision)
HSUBPD smallest double precision but only if the result is * 21536
MULPD finite normal value. inexact • #UE = 1
DIVPD • #PE = 1 if the result is
ADDSD inexact
SUBSD
MULSD
DIVSD

E-20 Vol. 1
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

Table E-18. #P - Inexact Result (Precision)

Unmasked Response and Exception
Instruction Condition Masked Response Code
ADDPS The result is not res = Result rounded Only if no underflow/overflow
ADDPD exactly to the destination condition occurred, or if the
ADDSUBPS representable in precision and using corresponding exceptions are masked:
ADDSUBPD the destination the bounded • Set #OE if masked overflow and set
HADDPS format. exponent, but only if result as described above for
HADDPD no unmasked masked overflow.
SUBPS underflow or • Set #UE if masked underflow and
SUBPD overflow conditions set result as described above for
masked underflow.
HSUBPS occur (this exception
If neither underflow nor overflow, res
HSUBPD can occur in the
equals the result rounded to the
MULPS presence of a
destination precision and using the
MULPD masked underflow
bounded exponent set #PE = 1.
DIVPS or overflow); #PE =
DIVPD 1.
SQRTPS
SQRTPD
CVTDQ2PS
CVTPI2PS
CVTPS2PI
CVTPS2DQ
CVTPD2PI
CVTPD2DQ
CVTPD2PS
CVTTPS2PI
CVTTPD2PI
CVTTPD2DQ
CVTTPS2DQ
ADDSS
ADDSD
SUBSS
SUBSD
MULSS
MULSD
DIVSS
DIVSD
SQRTSS
SQRTSD
CVTSI2SS
CVTSS2SI
CVTSD2SI
CVTSD2SS
CVTTSS2SI
CVTTSD2SI

Vol. 1 E-21
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

E.4.3 Example SIMD Floating-Point Emulation Implementation

The sample code listed below may be considered as being part of a user-level
floating-point exception filter for the SSE/SSE2/SSE3 numeric instructions. It is
assumed that the filter function is invoked by a low-level exception handler (reached
via interrupt vector 19 when an unmasked floating-point exception occurs), and that
it operates as explained in Section E.4.1, “Floating-Point Emulation.” The sample
code does the emulation only for the SSE instructions for addition, subtraction, multi-
plication, and division. For this, it uses C code and x87 FPU operations. Operations
corresponding to other SSE/SSE2/SSE3 numeric instructions can be emulated simi-
larly. The example assumes that the emulation function receives a pointer to a data
structure specifying a number of input parameters: the operation that caused the
exception, a set of sub-operands (unpacked, of type float), the rounding mode (the
precision is always single), exception masks (having the same relative bit positions
as in the MXCSR but starting from bit 0 in an unsigned integer), and flush-to-zero
and denormals-are-zeros indicators.
The output parameters are a floating-point result (of type float), the cause of the
exception (identified by constants not explicitly defined below), and the exception
status flags. The corresponding C definition is:

typedef struct {
unsigned int operation; //SSE or SSE2 operation: ADDPS, ADDSS, ...
unsigned int operand1_uint32; //first operand value
unsigned int operand2_uint32; //second operand value (if any)
float result_fval; // result value (if any)
unsigned int rounding_mode; //rounding mode
unsigned int exc_masks; //exception masks, in the order P,U,O,Z,D,I
unsigned int exception_cause; //exception cause
unsigned int status_flag_inexact; //inexact status flag
unsigned int status_flag_underflow; //underflow status flag
unsigned int status_flag_overflow; //overflow status flag
unsigned int status_flag_divide_by_zero;
//divide by zero status flag
unsigned int status_flag_denormal_operand;
//denormal operand status flag
unsigned int status_flag_invalid_operation;
//invalid operation status flag
unsigned int ftz; // flush-to-zero flag
unsigned int daz; // denormals-are-zeros flag
} EXC_ENV;

The arithmetic operations exemplified are emulated as follows:

E-22 Vol. 1
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

1. If the denormals-are-zeros mode is enabled (the DAZ bit in MXCSR is set to 1),
replace all the denormal inputs with zeroes of the same sign (the denormal flag is
not affected by this change).
2. Perform the operation using x87 FPU instructions, with exceptions disabled, the
original user rounding mode, and single precision. This reveals invalid, denormal,
or divide-by-zero exceptions (if there are any) and stores the result in memory as
a double precision value (whose exponent range is large enough to look like
“unbounded” to the result of the single precision computation).
3. If no unmasked exceptions were detected, determine if the result is less than the
smallest normal number (tiny) that can be represented in single precision
format, or greater than the largest normal number that can be represented in
single precision format (huge). If an unmasked overflow or underflow occurs,
calculate the scaled result that will be handed to the user exception handler, as
specified by IEEE Standard 754.
4. If no exception was raised, calculate the result with a “bounded” exponent. If the
result is tiny, it requires denormalization (shifting the significand right while
incrementing the exponent to bring it into the admissible range of [-126,+127]
for single precision floating-point numbers).

The result obtained in step 2 cannot be used because it might incur a double
rounding error (it was rounded to 24 bits in step 2, and might have to be rounded
again in the denormalization process). To overcome this is, calculate the result as
a double precision value, and store it to memory in single precision format.

Rounding first to 53 bits in the significand, and then to 24 never causes a double
rounding error (exact properties exist that state when double-rounding error
occurs, but for the elementary arithmetic operations, the rule of thumb is that if
an infinitely precise result is rounded to 2p+1 bits and then again to p bits, the
result is the same as when rounding directly to p bits, which means that no
double-rounding error occurs).
5. If the result is inexact and the inexact exceptions are unmasked, the calculated
result will be delivered to the user floating-point exception handler.
6. The flush-to-zero case is dealt with if the result is tiny.
7. The emulation function returns RAISE_EXCEPTION to the filter function if an
exception has to be raised (the exception_cause field indicates the cause).
Otherwise, the emulation function returns DO_NOT_ RAISE_EXCEPTION. In the
first case, the result is provided by the user exception handler called by the filter
function. In the second case, it is provided by the emulation function. The filter
function has to collect all the partial results, and to assemble the scalar or packed
result that is used if execution is to continue.

Vol. 1 E-23
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

Example E-2. SIMD Floating-Point Emulation

// masks for individual status word bits
#define PRECISION_MASK 0x20
#define UNDERFLOW_MASK 0x10
#define OVERFLOW_MASK 0x08
#define ZERODIVIDE_MASK 0x04
#define DENORMAL_MASK 0x02
#define INVALID_MASK 0x01

// 32-bit constants
static unsigned ZEROF_ARRAY[] = {0x00000000};
#define ZEROF *(float *) ZEROF_ARRAY
// +0.0
static unsigned NZEROF_ARRAY[] = {0x80000000};
#define NZEROF *(float *) NZEROF_ARRAY
// -0.0
static unsigned POSINFF_ARRAY[] = {0x7f800000};
#define POSINFF *(float *)POSINFF_ARRAY
// +Inf
static unsigned NEGINFF_ARRAY[] = {0xff800000};
#define NEGINFF *(float *)NEGINFF_ARRAY
// -Inf

// 64-bit constants
static unsigned MIN_SINGLE_NORMAL_ARRAY [] = {0x00000000, 0x38100000};
#define MIN_SINGLE_NORMAL *(double *)MIN_SINGLE_NORMAL_ARRAY
// +1.0 * 2^-126
static unsigned MAX_SINGLE_NORMAL_ARRAY [] = {0x70000000, 0x47efffff};
#define MAX_SINGLE_NORMAL *(double *)MAX_SINGLE_NORMAL_ARRAY
// +1.1...1*2^127
static unsigned TWO_TO_192_ARRAY[] = {0x00000000, 0x4bf00000};
#define TWO_TO_192 *(double *)TWO_TO_192_ARRAY
// +1.0 * 2^192
static unsigned TWO_TO_M192_ARRAY[] = {0x00000000, 0x33f00000};
#define TWO_TO_M192 *(double *)TWO_TO_M192_ARRAY
// +1.0 * 2^-192

// auxiliary functions
static int isnanf (unsigned int ); // returns 1 if f is a NaN, and 0 otherwise
static float quietf (unsigned int ); // converts a signaling NaN to a quiet
// NaN, and leaves a quiet NaN unchanged
static unsigned int check_for_daz (unsigned int ); // converts denormals
// to zeros of the same sign;
// does not affect any status flags

// emulation of SSE and SSE2 instructions using

// C code and x87 FPU instructions

unsigned int
simd_fp_emulate (EXC_ENV *exc_env)

E-24 Vol. 1
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

int uiopd1; // first operand of the add, subtract, multiply, or divide

int uiopd2; // second operand of the add, subtract, multiply, or divide
float res; // result of the add, subtract, multiply, or divide
double dbl_res24; // result with 24-bit significand, but "unbounded" exponent
// (needed to check tininess, to provide a scaled result to
// an underflow/overflow trap handler, and in flush-to-zero mode)
double dbl_res; // result in double precision format (needed to avoid a
// double rounding error when denormalizing)
unsigned int result_tiny;
unsigned int result_huge;
unsigned short int sw; // 16 bits
unsigned short int cw; // 16 bits

// have to check first for faults (V, D, Z), and then for traps (O, U, I)

// initialize x87 FPU (floating-point exceptions are masked)

_asm {
fninit;
}

result_tiny = 0;
result_huge = 0;

switch (exc_env->operation) {

case ADDPS:
case ADDSS:
case SUBPS:
case SUBSS:
case MULPS:
case MULSS:
case DIVPS:
case DIVSS:

uiopd1 = exc_env->operand1_uint32; // copy as unsigned int

// do not copy as float to avoid conversion
// of SNaN to QNaN by compiled code
uiopd2 = exc_env->operand2_uint32;
// do not copy as float to avoid conversion of SNaN
// to QNaN by compiled code
uiopd1 = check_for_daz (uiopd1); // operand1 = +0.0 * operand1 if it is
// denormal and DAZ=1
uiopd2 = check_for_daz (uiopd2); // operand2 = +0.0 * operand2 if it is
// denormal and DAZ=1

// execute the operation and check whether the invalid, denormal, or

// divide by zero flags are set and the respective exceptions enabled

// set control word with rounding mode set to exc_env->rounding_mode,

// single precision, and all exceptions disabled

Vol. 1 E-25
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

switch (exc_env->rounding_mode) {
case ROUND_TO_NEAREST:
cw = 0x003f; // round to nearest, single precision, exceptions masked
break;
case ROUND_DOWN:
cw = 0x043f; // round down, single precision, exceptions masked
break;
case ROUND_UP:
cw = 0x083f; // round up, single precision, exceptions masked
break;
case ROUND_TO_ZERO:
cw = 0x0c3f; // round to zero, single precision, exceptions masked
break;
default:
;
}
__asm {
fldcw WORD PTR cw;
}

// compute result and round to the destination precision, with

// "unbounded" exponent (first IEEE rounding)
switch (exc_env->operation) {

case ADDPS:
case ADDSS:
// perform the addition
__asm {
fnclex;
// load input operands
fld DWORD PTR uiopd1; // may set denormal or invalid status flags
fld DWORD PTR uiopd2; // may set denormal or invalid status flags
faddp st(1), st(0); // may set inexact or invalid status flags
// store result
fstp QWORD PTR dbl_res24; // exact
}
break;

case SUBPS:
case SUBSS:
// perform the subtraction
__asm {
fnclex;
// load input operands
fld DWORD PTR uiopd1; // may set denormal or invalid status flags
fld DWORD PTR uiopd2; // may set denormal or invalid status flags
fsubp st(1), st(0); // may set the inexact or invalid status flags

// store result
fstp QWORD PTR dbl_res24; // exact
}
break;

E-26 Vol. 1
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

case MULPS:
case MULSS:
// perform the multiplication
__asm {
fnclex;
// load input operands
fld DWORD PTR uiopd1; // may set denormal or invalid status flags
fld DWORD PTR uiopd2; // may set denormal or invalid status flags
fmulp st(1), st(0); // may set inexact or invalid status flags

// store result
fstp QWORD PTR dbl_res24; // exact
}
break;

case DIVPS:
case DIVSS:
// perform the division
__asm {
fnclex;
// load input operands
fld DWORD PTR uiopd1; // may set denormal or invalid status flags
fld DWORD PTR uiopd2; // may set denormal or invalid status flags
fdivp st(1), st(0); // may set the inexact, divide by zero, or
// invalid status flags
// store result
fstp QWORD PTR dbl_res24; // exact
}
break;

default:
; // will never occur

// read status word

__asm {
fstsw WORD PTR sw;
}

if (sw & ZERODIVIDE_MASK)

sw = sw & ~DENORMAL_MASK; // clear D flag for (denormal / 0)

// if invalid flag is set, and invalid exceptions are enabled, take trap
if (!(exc_env->exc_masks & INVALID_MASK) && (sw & INVALID_MASK)) {
exc_env->status_flag_invalid_operation = 1;
exc_env->exception_cause = INVALID_OPERATION;
return (RAISE_EXCEPTION);
}

// checking for NaN operands has priority over denormal exceptions;

Vol. 1 E-27
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

// also fix for the SSE and SSE2

// differences in treating two NaN inputs between the
// instructions and other IA-32 instructions
if (isnanf (uiopd1) || isnanf (uiopd2)) {

if (isnanf (uiopd1) && isnanf (uiopd2))

exc_env->result_fval = quietf (uiopd1);
else
exc_env->result_fval = (float)dbl_res24; // exact

if (sw & INVALID_MASK) exc_env->status_flag_invalid_operation = 1;

return (DO_NOT_RAISE_EXCEPTION);
}

// if denormal flag set, and denormal exceptions are enabled, take trap
if (!(exc_env->exc_masks & DENORMAL_MASK) && (sw & DENORMAL_MASK)) {
exc_env->status_flag_denormal_operand = 1;
exc_env->exception_cause = DENORMAL_OPERAND;
return (RAISE_EXCEPTION);
}

// if divide by zero flag set, and divide by zero exceptions are

// enabled, take trap (for divide only)
if (!(exc_env->exc_masks & ZERODIVIDE_MASK) && (sw & ZERODIVIDE_MASK)) {
exc_env->status_flag_divide_by_zero = 1;
exc_env->exception_cause = DIVIDE_BY_ZERO;
return (RAISE_EXCEPTION);
}

// done if the result is a NaN (QNaN Indefinite)

res = (float)dbl_res24;
if (isnanf (*(unsigned int *)&res)) {
exc_env->result_fval = res; // exact
exc_env->status_flag_invalid_operation = 1;
return (DO_NOT_RAISE_EXCEPTION);
}

// dbl_res24 is not a NaN at this point

if (sw & DENORMAL_MASK) exc_env->status_flag_denormal_operand = 1;

// Note: (dbl_res24 == 0.0 && sw & PRECISION_MASK) cannot occur

if (-MIN_SINGLE_NORMAL < dbl_res24 && dbl_res24 < 0.0 ||
0.0 < dbl_res24 && dbl_res24 < MIN_SINGLE_NORMAL) {
result_tiny = 1;
}

// check if the result is huge

if (NEGINFF < dbl_res24 && dbl_res24 < -MAX_SINGLE_NORMAL ||
MAX_SINGLE_NORMAL < dbl_res24 && dbl_res24 < POSINFF) {
result_huge = 1;
}

E-28 Vol. 1
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

// at this point, there are no enabled I,D, or Z exceptions

// to take; the instr.
// might lead to an enabled underflow, enabled underflow and inexact,
// enabled overflow, enabled overflow and inexact, enabled inexact, or
// none of these; if there are no U or O enabled exceptions, re-execute
// the instruction using IA-32 double precision format, and the
// user's rounding mode; exceptions must have
// been disabled before calling
// this function; an inexact exception may be reported on the 53-bit
// fsubp, fmulp, or on both the 53-bit and 24-bit conversions, while an
// overflow or underflow (with traps disabled) may be reported on the
// conversion from dbl_res to res

// check whether there is an underflow, overflow,

// or inexact trap to be taken
// if the underflow traps are enabled and the result is
// tiny, take underflow trap

if (!(exc_env->exc_masks & UNDERFLOW_MASK) && result_tiny) {

dbl_res24 = TWO_TO_192 * dbl_res24; // exact
exc_env->status_flag_underflow = 1;
exc_env->exception_cause = UNDERFLOW;
exc_env->result_fval = (float)dbl_res24; // exact
if (sw & PRECISION_MASK) exc_env->status_flag_inexact = 1;
return (RAISE_EXCEPTION);
}

// if overflow traps are enabled and the result is huge, take

// overflow trap
if (!(exc_env->exc_masks & OVERFLOW_MASK) && result_huge) {
dbl_res24 = TWO_TO_M192 * dbl_res24; // exact
exc_env->status_flag_overflow = 1;
exc_env->exception_cause = OVERFLOW;
exc_env->result_fval = (float)dbl_res24; // exact
if (sw & PRECISION_MASK) exc_env->status_flag_inexact = 1;
return (RAISE_EXCEPTION);
}

// set control word with rounding mode set to exc_env->rounding_mode,

// double precision, and all exceptions disabled
cw = cw | 0x0200; // set precision to double
__asm {
fldcw WORD PTR cw;
}

switch (exc_env->operation) {

case ADDPS:
case ADDSS:
// perform the addition
__asm {

Vol. 1 E-29
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

// load input operands

fld DWORD PTR uiopd1; // may set the denormal status flag
fld DWORD PTR uiopd2; // may set the denormal status flag
faddp st(1), st(0); // rounded to 53 bits, may set the inexact
// status flag
// store result
fstp QWORD PTR dbl_res; // exact, will not set any flag
}
break;

case SUBPS:
case SUBSS:
// perform the subtraction
__asm {
// load input operands
fld DWORD PTR uiopd1; // may set the denormal status flag
fld DWORD PTR uiopd2; // may set the denormal status flag
fsubp st(1), st(0); // rounded to 53 bits, may set the inexact
// status flag
// store result
fstp QWORD PTR dbl_res; // exact, will not set any flag
}
break;

case MULPS:
case MULSS:
// perform the multiplication
__asm {
// load input operands
fld DWORD PTR uiopd1; // may set the denormal status flag
fld DWORD PTR uiopd2; // may set the denormal status flag
fmulp st(1), st(0); // rounded to 53 bits, exact
// store result
fstp QWORD PTR dbl_res; // exact, will not set any flag
}
break;

case DIVPS:
case DIVSS:
// perform the division
__asm {
// load input operands
fld DWORD PTR uiopd1; // may set the denormal status flag
fld DWORD PTR uiopd2; // may set the denormal status flag
fdivp st(1), st(0); // rounded to 53 bits, may set the inexact
// status flag
// store result
fstp QWORD PTR dbl_res; // exact, will not set any flag
}
break;

default:

E-30 Vol. 1
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

; // will never occur

// calculate result for the case an inexact trap has to be taken, or

// when no trap occurs (second IEEE rounding)
res = (float)dbl_res;
// may set P, U or O; may also involve denormalizing the result

// read status word

__asm {
fstsw WORD PTR sw;
}

// if inexact traps are enabled and result is inexact, take inexact trap
if (!(exc_env->exc_masks & PRECISION_MASK) &&
((sw & PRECISION_MASK) || (exc_env->ftz && result_tiny))) {
exc_env->status_flag_inexact = 1;
exc_env->exception_cause = INEXACT;
if (result_tiny) {
exc_env->status_flag_underflow = 1;

// if ftz = 1 and result is tiny, result = 0.0

// (no need to check for underflow traps disabled: result tiny and
// underflow traps enabled would have caused taking an underflow
// trap above)
if (exc_env->ftz) {
if (res > 0.0)
res = ZEROF;
else if (res < 0.0)
res = NZEROF;
// else leave res unchanged
}
}
if (result_huge) exc_env->status_flag_overflow = 1;
exc_env->result_fval = res;
return (RAISE_EXCEPTION);
}

// if it got here, then there is no trap to be taken; the following must

// hold: ((the MXCSR U exceptions are disabled or
//
// the MXCSR underflow exceptions are enabled and the underflow flag is
// clear and (the inexact flag is set or the inexact flag is clear and
// the 24-bit result with unbounded exponent is not tiny)))
// and (the MXCSR overflow traps are disabled or the overflow flag is
// clear) and (the MXCSR inexact traps are disabled or the inexact flag
// is clear)
//
// in this case, the result has to be delivered (the status flags are
// sticky, so they are all set correctly already)

Vol. 1 E-31
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

// read status word to see if result is inexact

__asm {
fstsw WORD PTR sw;
}

if (sw & UNDERFLOW_MASK) exc_env->status_flag_underflow = 1;

if (sw & OVERFLOW_MASK) exc_env->status_flag_overflow = 1;
if (sw & PRECISION_MASK) exc_env->status_flag_inexact = 1;

// if ftz = 1, and result is tiny (underflow traps must be disabled),

// result = 0.0
if (exc_env->ftz && result_tiny) {
if (res > 0.0)
res = ZEROF;
else if (res < 0.0)
res = NZEROF;
// else leave res unchanged

exc_env->status_flag_inexact = 1;
exc_env->status_flag_underflow = 1;
}

exc_env->result_fval = res;
if (sw & ZERODIVIDE_MASK) exc_env->status_flag_divide_by_zero = 1;
if (sw & DENORMAL_MASK) exc_env->status_flag_denormal= 1;
if (sw & INVALID_MASK) exc_env->status_flag_invalid_operation = 1;
return (DO_NOT_RAISE_EXCEPTION);

break;

case CMPPS:
case CMPSS:

...

break;

case COMISS:
case UCOMISS:

...

break;

case CVTPI2PS:
case CVTSI2SS:

...

break;

case CVTPS2PI:

E-32 Vol. 1
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

case CVTSS2SI:
case CVTTPS2PI:
case CVTTSS2SI:

...

break;

case MAXPS:
case MAXSS:
case MINPS:
case MINSS:

...

break;

case SQRTPS:
case SQRTSS:

...

break;

...

case UNSPEC:

...

break;

default:
...

Vol. 1 E-33
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS

E-34 Vol. 1
 INDEX

Numerics JMP instruction, 6-12, 7-25

128-bit LAHF instruction, 7-31
packed byte integers data type, 4-13, 11-5 LDTR register, 3-6
packed double-precision floating-point legacy modes, 2-27
data type, 4-13, 11-5 LODS instruction, 7-28
packed doubleword integers data type, 4-13 LOOP instruction, 6-12, 7-25
packed quadword integers data type, 4-13 memory models, 3-11
packed SIMD data types, 4-12 memory operands, 3-29
packed single-precision floating-point MMX technology, 9-2
data type, 4-13, 10-8 MOVS instruction, 7-28
packed word integers data type, 4-13, 11-5 MOVSXD instruction, 7-11
16-bit near pointer, 4-9
address size, 3-11 operand addressing, 3-32
operand size, 3-11 operand size, 3-25
286 processor, 2-1 operands, 3-28, 3-29
32-bit POPF instruction, 7-31
address size, 3-11 promoted instructions, 3-2
operand size, 3-11 PUSHA, PUSHAD, POPA, POPAD, 7-10
64-bit PUSHF instruction, 7-31
packed byte integers data type, 4-12, 9-3 PUSHFD instruction, 7-31
packed doubleword integers data type, 4-12 real address mode, 3-11
packed doubleword integers data types, 9-4 register operands, 3-28
packed word integers data type, 4-12, 9-4 REP prefix, 7-28
64-bit mode RET instruction, 6-12, 7-25
sub-mode of IA-32e, 3-2 REX prefix, 3-2, 3-16, 3-25
address calculation, 3-12 RFLAGS register, 7-31
address size, 3-25 RIP register, 3-12
address space, 3-6 RIP-relative addressing, 3-24, 3-32
BOUND instruction, 7-26 SAHF instruction, 7-31
branch behavior, 6-11 SCAS instruction, 7-28
byte register limitation, 3-17 segment registers, 3-20
CALL instruction, 6-12, 7-25 segmentation, 3-11, 3-30
canonical address, 3-13 SSE extensions, 10-4
CMPS instruction, 7-28 SSE2 extensions, 11-4
CMPXCHG16B instruction, 7-7 SSE3 extensions, 12-1
data types, 7-2 SSSE3 extensions, 12-1
DEC instruction, 7-12 stack behavior, 6-5
decimal arithmetic instructions, 7-15 STOS instruction, 7-28
default operand and address sizes, 3-2 TR register, 3-6
exceptions, 6-19 x87 FPU, 8-2
far pointer, 4-9 See also: IA-32e mode, compatibility mode
feature list, 2-26 8086 processor, 2-1
GDTR register, 3-6 8088 processor, 2-1
IDTR register, 3-6
INC instruction, 7-12 A
instruction pointer, 3-12, 3-24
AAA instruction, 7-14
instructions introduced, 5-36
AAD instruction, 7-14
interrupts, 6-19
AAM instruction, 7-14
introduction, 2-26, 3-2, 7-2
AAS instruction, 7-14
IRET instruction, 7-26
AC (alignment check) flag, EFLAGS register, 3-23
I/O instructions, 7-29
Access rights, segment descriptor, 6-9, 6-14
JCC instruction, 6-12, 7-25
ADC instruction, 7-12
JCXZ instruction, 6-12, 7-25
ADD instruction, 7-12

Vol. 1 INDEX-1
INDEX

ADDPD instruction, 11-8 Basic programming environment, 7-1, 7-2

ADDPS instruction, 10-12 B-bit, x87 FPU status word, 8-7
Address size attribute BCD integers
code segment, 3-24 packed, 4-14
description of, 3-24 relationship to status flags, 3-22
of stack, 6-3 unpacked, 4-13, 7-14
Address sizes, 3-11 x87 FPU encoding, 4-14, 4-15
Address space BH register, 3-16
64-bit mode, 3-2, 3-6 Bias value
compatibility mode, 3-2 numeric overflow, 8-41
overview of, 3-3 numeric underflow, 8-42
physical, 3-8 Biased exponent, 4-18
Addressing modes Biasing constant, for floating-point numbers, 4-8
assembler, 3-32 Binary numbers, 1-7
base, 3-30, 3-31, 3-32 Binary-coded decimal (see BCD)
base plus displacement, 3-31 Bit field, 4-10
base plus index plus displacement, 3-32 Bit order, 1-5
base plus index time scale plus displacement, 3-32 BL register, 3-16
canonical address, 3-13 BOUND instruction, 6-18, 7-26, 7-32
displacement, 3-30, 3-31, 3-32 BOUND range exceeded exception (#BR), 6-18
effective address, 3-30 BP register, 3-16
immediate operands, 3-27 Branch
index, 3-30, 3-32 control transfer instructions, 7-21
index times scale plus displacement, 3-32 hints, 11-18
memory operands, 3-28, 3-29 on EFLAGS register status flags, 7-23, 8-9
register operands, 3-27, 3-28 on x87 FPU condition codes, 8-9, 8-28
RIP-relative addressing, 3-24, 3-32 prediction, 2-9
scale factor, 3-30, 3-32 BSF instruction, 7-20
specifying a segment selector, 3-29 BSR instruction, 7-20
specifying an offset, 3-30 BSWAP instruction, 7-5
specifying offsets in 64-bit mode, 3-32 BT instruction, 3-20, 3-22, 7-20
ADDSD instruction, 11-8 BTC instruction, 3-20, 3-22, 7-20
ADDSS instruction, 10-12 BTR instruction, 3-20, 3-22, 7-20
ADDSUBPD instruction, 5-26, 12-5 BTS instruction, 3-20, 3-22, 7-20
ADDSUBPS instruction, 5-25, 12-5 BX register, 3-16
Advanced media boost, 2-14 Byte, 4-1
advanced smart cache, 2-14 Byte order, 1-5
AF (adjust) flag, EFLAGS register, 3-21, A-1
AH register, 3-16
AL register, 3-16 C
Alignment C1 flag, x87 FPU status word, 8-7, 8-37, 8-41, 8-43
words, doublewords, quadwords, 4-2 C2 flag, x87 FPU status word, 8-7
AND instruction, 7-15 cache, smart, 2-6
ANDNPD instruction, 11-9 Call gate, 6-9
ANDNPS instruction, 10-13 CALL instruction, 3-24, 6-4, 6-5, 6-9, 7-22, 7-32
ANDPD instruction, 11-9 Calls (see Procedure calls)
ANDPS instruction, 10-13 Canonical address, 3-13
Arctangent, x87 FPU operation, 8-29 CBW instruction, 7-11
Arithmetic instructions, x87 FPU, 8-35 CDQ instruction, 7-11
Assembler, addressing modes, 3-32 Celeron processor
Asymmetric processing model, 12-2 description of, 2-3
AX register, 3-16 CF (carry) flag, EFLAGS register, 3-21, A-1
CH register, 3-16
CL register, 3-16
B CLC instruction, 3-22, 7-29
B (default size) flag, segment descriptor, 3-24 CLD instruction, 3-22, 7-30
Base (operand addressing), 3-30, 3-31, 3-32 CLFLUSH instruction, 11-17
Basic execution environment, 3-2 CLI instruction, 13-5

INDEX-2 Vol. 1
 INDEX

CMC instruction, 3-22, 7-29 MMX feature flag, 9-11

CMOVcc instructions, 7-4, 7-6 processor identification, 14-1
CMP instruction, 7-12 serializing use, 13-7
CMPPD instruction, 11-10 SSE feature flag, 10-1, 10-9
CMPPS instruction, 10-13 SSE2 feature flag, 11-1, 12-7, 12-8
CMPS instruction, 3-22, 7-27 SSE3 feature flag, 12-8
CMPSD instruction, 11-10 SSSE2 feature flag, 12-13, 12-28, 12-29
CMPSS instruction, 10-14 summary of, 7-33
CMPXCHG instruction, 7-6 CS register, 3-17, 3-19
CMPXCHG16B instruction, 7-7 CTI instruction, 7-31
CMPXCHG8B instruction, 7-6 Current privilege level (see CPL)
Code segment, 3-19 Current stack, 6-2, 6-4
COMISD instruction, 11-10 CVTDQ2PD instruction, 11-14
COMISS instruction, 10-14 CVTDQ2PS instruction, 11-14
Compare CVTPD2DQ instruction, 11-14
compare and exchange, 7-6 CVTPD2PI instruction, 11-13
integers, 7-12 CVTPD2PS instruction, 11-12
real numbers, x87 FPU, 8-27 CVTPI2PD instruction, 11-13
strings, 7-27 CVTPI2PS instruction, 10-16
Compatibility mode CVTPS2DQ instruction, 11-14
address space, 3-2 CVTPS2PD instruction, 11-12
branch functions, 6-12 CVTPS2PI instruction, 10-16
call gate descriptors, 6-12 CVTSD2SI instruction, 11-14
introduction, 2-26, 3-2 CVTSD2SS instruction, 11-12
memory models, 3-11 CVTSI2SD instruction, 11-14
MMX technology, 9-2 CVTSI2SS instruction, 10-16
segmentation, 3-30 CVTSS2SD instruction, 11-12
SSE extensions, 10-4 CVTSS2SI instruction, 10-16
SSE2 extensions, 11-4 CVTTPD2DQ instruction, 11-14
SSE3 extensions, 12-1 CVTTPD2PI instruction, 11-13
SSSE3 extensions, 12-1 CVTTPS2DQ instruction, 11-14
x87 FPU, 8-2 CVTTPS2PI instruction, 10-16
See also: IA-32e mode, 64-bit mode CVTTSD2SI instruction, 11-14
Compatibility, software, 1-5 CVTTSS2SI instruction, 10-16
compilers CWD instruction, 7-11
documentation, 1-9 CWDE instruction, 7-11
Condition code flags, x87 FPU status word CX register, 3-16
branching on, 8-9
conditional moves on, 8-9
description of, 8-6 D
interpretation of, 8-8 D (default size) flag, segment descriptor, 6-3
use of, 8-27 DAA instruction, 7-14
Conditional moves, x87 FPU condition codes, 8-9 DAS instruction, 7-14
Constants (floating point), 8-24 Data movement instructions, 7-3
Control registers Data pointer, x87 FPU, 8-13
64-bit mode, 3-6 Data registers, x87 FPU, 8-2
overview of, 3-5 Data segment, 3-19
Core microarchitecture, 2-13, 2-16, 2-17 Data types
core microarchitecture, 2-13, 2-16, 2-17 128-bit packed SIMD, 4-12
Core Solo and Core Duo, 2-5 64-bit mode, 7-2
Cosine, x87 FPU operation, 8-29 64-bit packed SIMD, 4-11
CPUID instruction alignment, 4-2
AP-485, 1-9 BCD integers, 4-13, 7-14
CLFLUSH flag, 11-17 bit field, 4-10
CMOVcc feature flag, 7-5 byte, 4-1
determine support for, 3-23 doubleword, 4-1
earlier processors, 14-2 floating-point, 4-6
FXSAVE-FXRSTOR flag, 10-21 fundamental, 4-1

Vol. 1 INDEX-3
INDEX

integers, 4-4 x87 FPU control word, 8-11

numeric, 4-3 Double-extended-precision FP format, 4-6
operated on by GP instructions, 7-1, 7-2 Double-precision floating-point format, 4-6
operated on by MMX technology, 9-3 Doubleword, 4-1
operated on by SSE extensions, 10-8 DS register, 3-17, 3-19
operated on by SSE2 extensions, 11-5 Dual-core technology
operated on by x87 FPU, 8-17 introduction, 2-22
operated on in 64-bit mode, 4-9 DX register, 3-16
packed bytes, 9-3 Dynamic data flow analysis, 2-9
packed doublewords, 9-3 Dynamic execution, 2-9, 2-14, 2-16, 2-17
packed SIMD, 4-11
packed words, 9-3
pointers, 4-9 E
quadword, 4-1, 9-3 EAX register, 3-14, 3-16
signed integers, 4-5 EBP register, 3-14, 3-16, 6-4, 6-8
strings, 4-11 EBX register, 3-14, 3-16
unsigned integers, 4-5 ECX register, 3-14, 3-16
word, 4-1 EDI register, 3-14, 3-16
DAZ (denormals-are-zeros) flag EDX register, 3-14, 3-16
MXCSR register, 10-7 Effective address, 3-30
DE (denormal operand exception) flag EFLAGS register
MXCSR register, 11-21 64-bit mode, 7-2
x87 FPU status word, 8-7, 8-39 condition codes, B-1
Debug registers cross-reference with instructions, A-1
64-bit mode, 3-6 description of, 3-20
legacy modes, 3-5 instructions that operate on, 7-29
DEC instruction, 7-12 overview, 3-14
Decimal integers, x87 FPU, 4-15 part of basic programming environment, 7-1
Deeper sleep, 2-6 restoring from stack, 6-8
Denormal number (see Denormalized finite number) saving on a procedure call, 6-8
Denormal operand exception (#D) status flags, 8-9, 8-10, 8-28
overview of, 4-28 use with CMOVcc instructions, 7-4
SSE and SSE2 extensions, 11-21 EIP register
x87 FPU, 8-38 description of, 3-24
Denormalization process, 4-21 overview, 3-14
Denormalized finite number, 4-7, 4-20 part of basic programming environment, 7-1
Denormals-are-zero relationship to CS register, 3-19
DAZ flag, MXCSR register, 10-7, 11-3, 11-4, EMMS instruction, 9-10, 9-12
11-28 Enhanced Intel Deeper Sleep, 2-6
mode, 10-7, 11-28 ENTER instruction, 6-19, 6-20, 7-29
DF (direction) flag, EFLAGS register, 3-22, A-1 GETSEC, 5-37
DH register, 3-16 ES register, 3-17, 3-19
DI register, 3-16 ES (exception summary) flag
Digital media boost, 2-6 x87 FPU status word, 8-44
Displacement (operand addressing), 3-30, 3-31, 3-32 ESC instructions, x87 FPU, 8-22
DIV instruction, 7-13 ESI register, 3-14, 3-16
Divide, 4-29 ESP register, 3-16
Divide by zero exception (#Z) ESP register (stack pointer), 3-14, 6-3, 6-4
SSE and SSE2 extensions, 11-22 Exception flags, x87 FPU status word, 8-7
x87 FPU, 8-40 Exception handlers
DIVPD instruction, 11-8 overview of, 6-13
DIVPS instruction, 10-12 SIMD floating-point exceptions, E-1
DIVSD instruction, 11-8 SSE and SSE2 extensions, 11-25, 11-26
DIVSS instruction, 10-12 typical actions of a FP exception handler, 4-33
DL register, 3-16 x87 FPU, 8-45
DM (denormal operand exception) mask bit Exception priority, floating-point exceptions, 4-32
MXCSR register, 11-21 Exception-flag masks, x87 FPU control word, 8-11
x87 FPU, 8-39 Exceptions

INDEX-4 Vol. 1
 INDEX

64-bit mode, 6-19 FLD1 instruction, 8-24

description of, 6-13 FLDCW instruction, 8-10, 8-32
handler, 6-13 FLDENV instruction, 8-7, 8-13, 8-15, 8-33
implicit call to handler, 6-1 FLDL2E instruction, 8-24
in real-address mode, 6-17 FLDL2T instruction, 8-24
notation, 1-8 FLDLG2 instruction, 8-24
vector, 6-13 FLDLN2 instruction, 8-24
Exponent, floating-point number, 4-16 FLDPI instruction, 8-24
FLDSW instruction, 8-32
FLDZ instruction, 8-24
F Floating-point data types
F2XM1 instruction, 8-31 biasing constant, 4-8
FABS instruction, 8-25 denormalized finite number, 4-7
FADD instruction, 8-24 description of, 4-6
FADDP instruction, 8-24 double extended precision format, 4-6, 4-7
Far call double precision format, 4-6, 4-7
description of, 6-5 infinites, 4-7
operation, 6-6 normalized finite number, 4-7
Far pointer single precision format, 4-6, 4-7
16-bit addressing, 3-11 SSE extensions, 10-8
32-bit addressing, 3-11 SSE2 extensions, 11-5
64-bit mode, 4-9 storing in memory, 4-9
description of, 3-8, 4-9 x87 FPU, 8-17
legacy modes, 4-9 zeros, 4-7
Far return operation, 6-6 Floating-point exception handlers
FBLD instruction, 8-23 SSE and SSE2 extensions, 11-25, 11-26
FBSTP instruction, 8-23 typical actions, 4-33
FCHS instruction, 8-25 x87 FPU, 8-45
FCLEX/FNCLEX instructions, 8-7 Floating-point exceptions
FCMOVcc instructions, 8-10, 8-23 denormal operand exception (#D), 4-28, 8-39,
FCOM instruction, 8-9, 8-26 11-21, C-1
FCOMI instruction, 8-10, 8-26 divide by zero exception (#Z), 4-29, 8-40, 11-22,
FCOMIP instruction, 8-10, 8-26 C-1
FCOMP instruction, 8-9, 8-26 exception conditions, 4-28
FCOMPP instruction, 8-9, 8-26 exception priority, 4-32
FCOS instruction, 8-7, 8-29 inexact result (precision) exception (#P), 4-31,
FDIV instruction, 8-25 8-42, 11-22, C-1
FDIVP instruction, 8-25 invalid operation exception (#I), 4-28, 8-36, 11-20
FDIVR instruction, 8-25 invalid-operation exception (#IA), C-1
FDIVRP instruction, 8-25 invalid-operation exception (#IS), C-1
Feature determination, of processor, 14-1 invalid-operation exception (#I), C-1
FIADD instruction, 8-25 numeric overflow exception (#O), 4-29, 8-40,
FICOM instruction, 8-9, 8-26 11-22, C-1
FICOMP instruction, 8-9, 8-26 numeric underflow exception (#U), 4-30, 8-41,
FIDIV instruction, 8-25 11-22, C-1
FIDIVR instruction, 8-25 summary of, 4-26, C-1
FILD instruction, 8-23 typical handler actions, 4-33
FIMUL instruction, 8-25 Floating-point format
FINIT/FNINIT instructions, 8-7, 8-11, 8-12, 8-32 biased exponent, 4-18
FIST instruction, 8-23 description of, 8-17
FISTP instruction, 8-23 exponent, 4-16
FISTTP instruction, 5-25, 12-4 fraction, 4-16
FISUB instruction, 8-25 indefinite, 4-8
FISUBR instruction, 8-25 QNaN floating-point indefinite, 4-24
Flags real number system, 4-15
cross-reference with instructions, A-1 sign, 4-16
Flat memory model, 3-8, 3-18 significand, 4-16
FLD instruction, 8-22 Floating-point numbers

Vol. 1 INDEX-5
INDEX

defined, 4-16 basic programming environment, 7-1

encoding, 4-8 data types operated on, 7-1, 7-2
Flush-to-zero description of, 7-1
FZ flag, MXCSR register, 10-7, 11-3 origin of, 7-1
mode, 10-7 programming with, 7-1
FMUL instruction, 8-25 summary of, 5-3, 7-3
FMULP instruction, 8-25 GS register, 3-17, 3-19
FNOP instruction, 8-32
Fopcode compatibility mode, 8-14
FPATAN instruction, 8-29 H
FPREM instruction, 8-7, 8-25, 8-30 HADDPD instruction, 5-26, 12-6
FPREM1 instruction, 8-7, 8-25, 8-30 HADDPS instruction, 5-26, 12-5
FPTAN instruction, 8-7 Hexadecimal numbers, 1-7
Fraction, floating-point number, 4-16 Horizontal processing model, 12-2
FRNDINT instruction, 8-25 HSUBPD instruction, 5-26, 12-6
FRSTOR instruction, 8-7, 8-13, 8-15, 8-33 HSUBPS instruction, 5-26, 12-6
FS register, 3-17, 3-19 HT Technology
FSAVE/FNSAVE instructions, 8-6, 8-7, 8-13, 8-15, first processor, 2-4
8-33 implementing, 2-22
FSCALE instruction, 8-31 introduction, 2-21
FSIN instruction, 8-7, 8-29
FSINCOS instruction, 8-7, 8-29 I
FSQRT instruction, 8-25
IA-32 architecture
FST instruction, 8-23
history of, 2-1
FSTCW/FNSTCW instructions, 8-10, 8-32
introduction to, 2-1
FSTENV/FNSTENV instructions, 8-6, 8-13, 8-15, 8-33
IA-32e mode
FSTP instruction, 8-23
introduction, 2-26
FSTSW/FNSTSW instructions, 8-6, 8-32
segmentation, 3-30
FSUB instruction, 8-25
See also: 64-bit mode, compatibility mode
FSUBP instruction, 8-25
IA32_MISC_ENABLE MSR, 8-14
FSUBR instruction, 8-25
ID (identification) flag, EFLAGS register, 3-23
FSUBRP instruction, 8-25
IDIV instruction, 7-13
FTST instruction, 8-9, 8-26
IDTR register, 3-5, 3-6
FUCOM instruction, 8-26
IE (invalid operation exception) flag
FUCOMI instruction, 8-10, 8-26
MXCSR register, 11-20
FUCOMIP instruction, 8-10, 8-26
x87 FPU status word, 8-7, 8-37, 8-38
FUCOMP instruction, 8-26
IEEE Standard 754, 4-6, 4-15, 8-1
FUCOMPP instruction, 8-9, 8-26
IF (interrupt enable) flag
FXAM instruction, 8-7, 8-26
EFLAGS register, 3-23, 6-14, 13-5, A-1
FXCH instruction, 8-23
IM (invalid operation exception) mask bit
FXRSTOR instruction, 5-13, 8-17, 10-20, 11-34
MXCSR register, 11-20
FXSAVE instruction, 5-13, 8-17, 10-20, 11-34
x87 FPU control word, 8-11
FXTRACT instruction, 8-25
Immediate operands, 3-27
FYL2X instruction, 8-31
IMUL instruction, 7-13
FYL2XP1 instruction, 8-31
IN instruction, 5-8, 7-28, 13-4
INC instruction, 7-12
G Indefinite
GDTR register, 3-5, 3-6 description of, 4-24
General purpose registers floating-point format, 4-8, 4-19
64-bit mode, 3-6, 3-17 integer, 4-6, 8-20
description of, 3-13, 3-14 packed BCD integer, 4-15
overview of, 3-3, 3-6 QNaN floating-point, 4-22, 4-24
parameter passing, 6-7 Index (operand addressing), 3-30, 3-32
part of basic programming environment, 7-1, 7-2 Inexact result (precision)
using REX prefix, 3-17 exception (#P), overview, 4-31
General-purpose instructions exception (#P), SSE-SSE2 extensions, 11-23
64-bit mode, 7-2 exception (#P), x87 FPU, 8-42

INDEX-6 Vol. 1
 INDEX

on floating-point operations, 4-25 description of, 4-4

Infinity control flag, x87 FPU control word, 8-12 indefinite, 4-6, 8-20
Infinity, floating-point format, 4-7, 4-21 signed integer encodings, 4-6
INIT pin, 3-20 signed, description of, 4-5
Input/output (see I/O) unsigned integer encodings, 4-5
INS instruction, 5-8, 7-28, 13-4 unsigned, description of, 4-5
Instruction operands, 1-6 Intel 64 architecture
Instruction pointer 64-bit mode, 3-2
64-bit mode, 7-2 64-bit mode instructions, 5-36
EIP register, 3-14, 3-24 address space, 3-8
RIP register, 3-24 compatibility mode, 3-2
RIP, EIP, IP compared, 3-12 data types, 4-1
x87 FPU, 8-13 definition of, 1-3
Instruction prefixes executing calls, 6-1
effect on SSE and SSE2 instructions, 11-37 general purpose instructions, 7-1
REX prefix, 3-2, 3-16 generations, 2-27
Instruction set history of, 2-1
binary arithmetic instructions, 7-12 IA32e mode, 3-2
bit scan instructions, 7-20 introduction, 2-26
bit test and modify instructions, 7-20 memory organization, 3-8, 3-10
byte-set-on-condition instructions, 7-20 relation to IA-32, 1-3
cacheability control instructions, 5-20, 5-24 See also: IA-32e mode
comparison and sign change instruction, 7-12 Intel Advanced Digital Media Boost, 2-6, 2-14
control transfer instructions, 7-21 Intel Advanced Smart Cache, 2-14
data movement instructions, 7-3 Intel Advanced Thermal Manager, 2-6
decimal arithmetic instructions, 7-13 Intel Core 2 Extreme processor family, 2-6, 2-24
EFLAGS cross-reference, A-1 Intel Core Duo processor, 2-5, 2-23
EFLAGS instructions, 7-29 Intel Core microarchitecture, 2-6, 2-13, 2-16, 2-17,
exchange instructions, 7-5 2-24
FXSAVE and FXRSTOR instructions, 5-13 Intel Core Solo processor, 2-5
general-purpose instructions, 5-3 Intel developer link, 1-10
grouped by processor, 5-1 Intel Dynamic Power Coordination, 2-6
increment and decrement instructions, 7-12 Intel NetBurst microarchitecture, 1-2
instruction ordering instructions, 5-20, 5-24 description of, 2-10
I/O instructions, 5-8, 7-28 introduction, 2-10
logical instructions, 7-15 Intel Pentium D processor, 2-23
MMX instructions, 5-13, 9-6 Intel Pentium processor Extreme Edition, 2-22
multiply and divide instructions, 7-13 Intel Smart Cache, 2-6
processor identification instruction, 7-33 Intel Smart Memory Access, 2-6, 2-14
repeating string operations, 7-27 Intel software network link, 1-10
rotate instructions, 7-18 Intel VTune Performance Analyzer
segment register instructions, 7-31 related information, 1-9
shift instructions, 7-15 Intel Wide Dynamic Execution, 2-6, 2-14, 2-16, 2-17
SIMD instructions, introduction to, 2-18 Intel Xeon processor, 1-1
software interrupt instructions, 7-25 description of, 2-4
SSE instructions, 5-16 Intel Xeon processor 5100 series, 2-6, 2-24
SSE2 instructions, 5-20 Intel386 processor, 2-2
stack manipulation instructions, 7-7 Intel486 processor
string operation instructions, 7-26 history of, 2-2
summary, 5-1 Inter-privilege level call
system instructions, 5-35 description of, 6-8
test instruction, 7-21 operation, 6-10
type conversion instructions, 7-10 Inter-privilege level return
x87 FPU and SIMD state management instructions description of, 6-8
, 5-13 operation, 6-10
x87 FPU instructions, 5-9 Interrupt gate, 6-14
INT instruction, 6-18, 7-32 Interrupt handler, 6-13
Integers Interrupt vector, 6-13

Vol. 1 INDEX-7
INDEX

Interrupts defined, 3-8

64-bit mode, 6-19 maximum size, 3-8
description of, 6-13 LOCK signal, 7-5
handler, 6-13 LODS instruction, 3-22, 7-27
implicit call to an interrupt handler Log epsilon, x87 FPU operation, 8-31
procedure, 6-14 Logical address, 3-8
implicit call to an interrupt handler task, 6-17 LOOP instructions, 7-24
implicit call to interrupt handler procedure, 6-14 LOOPcc instructions, 3-22, 7-24
implicit call to interrupt handler task, 6-17 LSS instruction, 7-32
in real-address mode, 6-17
maskable, 6-13
user-defined, 6-13 M
vector, 6-13 Machine check registers, 3-5
INTn instruction, 7-26 Machine specific registers (see MSRs)
INTO instruction, 6-18, 7-26, 7-32 Maskable interrupts, 6-13
Invalid arithmetic operand exception (#IA) Masked responses
description of, 8-38 denormal operand exception (#D), 4-28, 8-39
masked response to, 8-38 divide by zero exception (#Z), 4-29, 8-40
Invalid operation exception (#I) inexact result (precision) exception (#P), 4-32,
overview, 4-28 8-43
SSE and SSE2 extensions, 11-20 invalid arithmetic operation (#IA), 8-38
x87 FPU, 8-36 invalid operation exception (#I), 4-28
IOPL (I/O privilege level) field numeric overflow exception (#O), 4-30, 8-40
EFLAGS register, 3-23, 13-4 numeric underflow exception (#U), 4-31, 8-42
IRET instruction, 3-24, 6-17, 6-18, 7-22, 7-32, 13-5 stack overflow or underflow
I/O exception (#IS), 8-37
address space, 13-2 MASKMOVDQU instruction, 11-17, 11-36
instruction serialization, 13-7 MASKMOVQ instruction, 10-18, 11-36
instructions, 5-8, 7-28, 13-3 Masks, exception-flags
I/O privilege level (see IOPL) MXCSR register, 10-6
map base, 13-5 x87 FPU control word, 8-11
permission bit map, 13-5 MAXPD instruction, 11-9
ports, 3-5, 13-1, 13-2, 13-4, 13-7 MAXPS instruction, 10-12
sensitive instructions, 13-4 MAXSD instruction, 11-9
MAXSS instruction, 10-13
Memory
J flat memory model, 3-8
J-bit, 4-16 management registers, 3-5
Jcc instructions, 3-22, 3-24, 7-23 memory type range registers (MTRRs), 3-5
JMP instruction, 3-24, 7-21, 7-32 modes of operation, 3-10
organization, 3-8
physical, 3-8
L real address mode memory model, 3-9, 3-10
L1 (level 1) cache, 2-9, 2-12 segmented memory model, 3-8
L2 (level 2) cache, 2-9, 2-12 virtual-8086 mode memory model, 3-9, 3-10
LAHF instruction, 3-20, 7-30 Memory operands
Last instruction opcode, x87 FPU, 8-14 64-bit mode, 3-29
LDDQU instruction, 5-25, 12-4 legacy modes, 3-28
LDMXCSR instruction, 10-17, 11-34 Memory-mapped I/O, 13-2
LDS instruction, 7-32 MFENCE instruction, 11-17, 11-37
LDTR register, 3-5, 3-6 Microarchitecture
LEA instruction, 7-33 (see Intel NetBurst microarchitecture)
LEAVE instruction, 6-19, 6-25, 7-29 (see P6 family microarchitecture)
LES instruction, 7-32 MINPD instruction, 11-9
LFENCE instruction, 11-17 MINPS instruction, 10-13
LGS instruction, 7-32 MINSD instruction, 11-9
Linear address, 3-8 MINSS instruction, 10-13
Linear address space MMX instruction set

INDEX-8 Vol. 1
 INDEX

arithmetic instructions, 9-8 MOVHLPS instruction, 10-11

comparison instructions, 9-9 MOVHPD instruction, 11-8
conversion instructions, 9-9 MOVHPS instruction, 10-11
data transfer instructions, 9-8 MOVLHPS instruction, 10-11
EMMS instruction, 9-10 MOVLPD instruction, 11-8
logical instructions, 9-10 MOVLPS instruction, 10-11
overview, 9-6 MOVMSKPD instruction, 11-8
shift instructions, 9-10 MOVMSKPS instruction, 10-11
MMX registers MOVNTDQ instruction, 11-17, 11-36
description of, 9-3 MOVNTI instruction, 11-17, 11-36
overview of, 3-3 MOVNTPD instruction, 11-17, 11-36
MMX technology MOVNTPS instruction, 10-18, 11-36
64-bit mode, 9-2 MOVNTQ instruction, 10-18, 11-36
64-bit packed SIMD data types, 4-11 MOVQ instruction, 9-8
compatibility mode, 9-2 MOVQ2DQ instruction, 11-16
compatibility with FPU architecture, 9-10 MOVS instruction, 3-22, 7-27
data types, 9-3 MOVSD instruction, 11-7, 11-34
detecting MMX technology with CPUID instruction MOVSHDUP instruction, 5-26, 12-4
, 9-11 MOVSLDUP instruction, 5-26, 12-4
effect of instruction prefixes on MMX instructions MOVSS instruction, 10-11, 11-34
, 9-14 MOVSX instruction, 7-11
exception handling in MMX code, 9-14 MOVSXD instruction, 7-11
IA-32e mode, 9-2 MOVUPD instruction, 11-7, 11-34
instruction set, 5-13, 9-6 MOVUPS instruction, 10-9, 10-11, 11-34
interfacing with MMX code, 9-13 MOVZX instruction, 7-11
introduction to, 9-1 MS-DOS compatibility mode, 8-45, D-1
memory data formats, 9-4 MSRs, 3-5
mixing MMX and floating-point instructions, 9-13 MTRRs, 3-5
MMX registers, 9-3 MUL instruction, 7-13
programming environment (overview), 9-2 MULPD instruction, 11-8
register mapping, 9-14 MULPS instruction, 10-12
saturation arithmetic, 9-5 MULSD instruction, 11-8
SIMD execution environment, 9-4 MULSS instruction, 10-12
transitions between x87 FPU - MMX code, 9-12 Multi-core technology, 2-22
updating MMX technology routines using 128-bit Multi-threading capability, 2-22
SIMD integer instructions, 11-35 MWAIT instruction, 5-27, 12-7
using MMX code in a multitasking operating MXCSR register, 11-23
system environment, 9-14 denormals-are-zero (DAZ) flag, 10-7, 11-3, 11-4
using the EMMS instruction, 9-12 description, 10-5
wraparound mode, 9-5 flush-to-zero flag (FZ), 10-7
Modes of operation FXSAVE and FXRSTOR instructions, 11-34
64-bit mode, 3-2 LDMXCSR instruction, 11-34
compatibility mode, 3-2 load and store instructions, 10-17
memory models used with, 3-10 RC field, 4-25
overview, 3-1, 3-6 saving on a procedure or function call, 11-34
protected mode, 3-1 SIMD floating-point mask and flag bits, 10-6
real address mode, 3-1 SIMD floating-point rounding control field, 10-7
system management mode (SMM), 3-1 state management instructions, 5-19, 10-17
MONITOR instruction, 5-27, 12-7 STMXCSR instruction, 11-34
Moore’s law, 2-27 writing to while preventing general-protection
MOV instruction, 7-4, 7-31 exceptions (#GP), 11-30
MOVAPD instruction, 11-7, 11-34
MOVAPS instruction, 10-11, 11-34
MOVD instruction, 9-8 N
MOVDDUP instruction, 5-26, 12-5 NaNs
MOVDQ2Q instruction, 11-16 description of, 4-19, 4-21
MOVDQA instruction, 11-15, 11-34 encoding of, 4-7, 4-8, 4-19
MOVDQU instruction, 11-15, 11-34 SNaNs vs. QNaNs, 4-21

Vol. 1 INDEX-9
INDEX

Near call Ordering I/O, 13-7

description of, 6-5 ORPD instruction, 11-9
operation, 6-5 ORPS instruction, 10-13
Near pointer OSXMMEXCPT flag
64-bit mode, 4-9 control register CR4, 11-25
legacy modes, 4-9 OUT instruction, 5-8, 7-28, 13-4
Near return operation, 6-5 OUTS instruction, 5-8, 7-28, 13-4
NEG instruction, 7-12 Overflow exception (#OF), 6-18
NetBurst microarchitecture (see Intel NetBurst Overflow, x87 FPU stack, 8-36, 8-37
microarchitecture)
Non-arithmetic instructions, x87 FPU, 8-35
Non-number encodings, floating-point format, 4-19 P
Non-temporal data P6 family microarchitecture
caching of, 10-18 description of, 2-8
description, 10-18 history of, 2-3
temporal vs. non-temporal data, 10-18 P6 family processors
Non-waiting instructions, x87 FPU, 8-33, 8-45 description of, 1-1
NOP instruction, 7-33 history of, 2-3
Normalized finite number, 4-7, 4-18, 4-20 P6 family microarchitecture, 2-8
NOT instruction, 7-15 PABSB instruction, 5-28, 12-11
Notation PABSD instruction, 12-11
bit and byte order, 1-5 PABSW instruction, 5-28, 12-11
exceptions, 1-8 Packed
hexadecimal and binary numbers, 1-7 BCD integer indefinite, 4-15
instruction operands, 1-6 BCD integers, 4-14
notational conventions, 1-5 bytes, 9-3
reserved bits, 1-5 doublewords, 9-3
segmented addressing, 1-7 SIMD data types, 4-11
NT (nested task) flag, EFLAGS register, 3-23, A-1 SIMD floating-point values, 4-12
Numeric overflow exception (#O) SIMD integers, 4-11, 4-12
overview, 4-29 words, 9-3
SSE and SSE2 extensions, 11-22 PACKSSWB instruction, 9-9
x87 FPU, 8-7, 8-40 PACKUSWB instruction, 9-9
Numeric underflow exception (#U) PADDB instruction, 9-8
overview, 4-30 PADDD instruction, 9-8
SSE and SSE2 extensions, 11-22 PADDQ instruction, 11-15
x87 FPU, 8-7, 8-41 PADDSB instruction, 9-8
PADDSW instruction, 9-8
PADDUSB instruction, 9-8
O PADDUSW instruction, 9-8
OE (numeric overflow exception) flag PADDW instruction, 9-8
MXCSR register, 11-22 PALIGNR instruction, 5-29, 12-12
x87 FPU status word, 8-7, 8-40 PAND instruction, 9-10
OF (overflow) flag PANDN instruction, 9-10
EFLAGS register, 3-21, 6-18 Parameter passing
OF (overflow) flag, EFLAGS register, A-1 argument list, 6-8
Offset (operand addressing), 3-30 on stack, 6-7
Offset (operand addressing, 64-bit mode), 3-32 on the stack, 6-7
OM (numeric overflow exception) mask bit through general-purpose registers, 6-7
MXCSR register, 11-22 x87 FPU register stack, 8-5
x87 FPU control word, 8-11, 8-40 XMM registers, 11-34
Operand PAUSE instruction, 11-18
addressing, modes, 3-26 PAVGB instruction, 10-16
instruction, 1-6 PC (precision) field, x87 FPU control word, 8-11
size attribute, 3-24 PCMPEQB instruction, 9-9
sizes, 3-11, 3-25 PCMPEQD instruction, 9-9
x87 FPU instructions, 8-22 PCMPEQW instruction, 9-9
OR instruction, 7-15 PCMPGTB instruction, 9-9

INDEX-10 Vol. 1
 INDEX

PCMPGTD instruction, 9-9 Pointer data types, 4-9

PCMPGTW instruction, 9-9 Pointers
PE (inexact result exception) flag, 11-23 64-bit mode, 4-9
MXCSR register, 4-25 far pointer, 4-9
x87 FPU status word, 4-25, 8-7, 8-43 near pointer, 4-9
Pentium 4 processor, 1-1 POP instruction, 6-1, 6-3, 7-8, 7-31
description of, 2-4, 2-5 POPA instruction, 6-8, 7-9
Pentium 4 processor supporting Hyper-Threading POPF instruction, 3-20, 6-8, 7-30, 13-5
Technology POPFD instruction, 3-20, 6-8, 7-30
description of, 2-4, 2-5 POR instruction, 9-10
Pentium II processor, 1-2 Power coordination, 2-6
description of, 2-3 PREFETCHh instructions, 10-19, 11-36
P6 family microarchitecture, 2-8 Privilege levels
Pentium II Xeon processor description of, 6-9
description of, 2-3 inter-privilege level calls, 6-8
Pentium III processor, 1-2 protection rings, 6-9
description of, 2-4 stack switching, 6-15
P6 family microarchitecture, 2-8 Procedure calls
Pentium III Xeon processor description of, 6-5
description of, 2-4 far call, 6-5
Pentium M processor for block-structured languages, 6-19
description of, 2-5 inter-privilege level call, 6-10
instructions supported, 2-5 linking, 6-4
Pentium Pro processor, 1-2 near call, 6-5
description of, 2-3 overview, 6-1
P6 family microarchitecture, 2-8 return instruction pointer (EIP register), 6-4
Pentium processor, 1-1 saving procedure state information, 6-8
history of, 2-2 stack, 6-1
Pentium processor Extreme Edition stack switching, 6-10
introduction, 2-5 to exception handler procedure, 6-14
Pentium processor with MMX technology, 2-3 to exception task, 6-17
Performance monitoring counters, 3-5 to interrupt handler procedure, 6-14
PEXTRW instruction, 10-17 to interrupt task, 6-17
PF (parity) flag, EFLAGS register, 3-21, A-1 to other privilege levels, 6-8
PHADDD instruction, 5-28, 12-10 types of, 6-1
PHADDSW instruction, 5-28, 12-10 Processor identification
PHADDW instruction, 5-27, 12-10 earlier Intel architecture processors, 14-2
PHSUBD instruction, 5-28, 12-10 early processors, 14-2
PHSUBSW instruction, 5-28, 12-10 notes on where to start, 14-1
PHSUBW instruction, 5-28, 12-10 using CPUID, 14-1
Physical using CPUID instruction, 14-1
address space, 3-8 Processor state information, saving, 6-8
memory, 3-8 Protected mode
PINSRW instruction, 10-17 I/O, 13-4
Pi, x87 FPU constant, 8-30 memory models used, 3-10
PM (inexact result exception) mask bit overview, 3-1
MXCSR register, 11-23 Protection rings, 6-9
x87 FPU control word, 8-11, 8-43 PSADBW instruction, 10-17
PMADDUBSW instruction, 5-28, 12-11 PSHUFB instruction, 5-29, 12-12
PMADDWD instruction, 9-9 PSHUFD instruction, 11-16
PMAXSW instruction, 10-17 PSHUFHW instruction, 11-15
PMAXUB instruction, 10-17 PSHUFLW instruction, 11-15
PMINSW instruction, 10-17 PSHUFW instruction, 10-17, 11-16
PMINUB instruction, 10-17 PSIGNB/W/D instruction, 5-29, 12-12
PMOVMSKB instruction, 10-17 PSLLD instruction, 9-10
PMULHRSW instruction, 5-28, 12-11 PSLLDQ instruction, 11-16
PMULHUW instruction, 10-17 PSLLQ instruction, 9-10
PMULUDQ instruction, 11-15 PSLLW instruction, 9-10

Vol. 1 INDEX-11
INDEX

PSRLDQ instruction, 11-16 not in 64-bit mode, 3-11

PSUBB instruction, 9-8 overview, 3-1
PSUBD instruction, 9-8 Real numbers
PSUBQ instruction, 11-15 continuum, 4-16
PSUBSB instruction, 9-8 encoding, 4-19
PSUBSW instruction, 9-8 notation, 4-18
PSUBUSB instruction, 9-8 system, 4-15
PSUBUSW instruction, 9-8 Register operands
PSUBW instruction, 9-8 64-bit mode, 3-28
PUNPCKHBW instruction, 9-9 legacy modes, 3-27
PUNPCKHDQ instruction, 9-9 Register stack, x87 FPU, 8-2
PUNPCKHQDQ instruction, 11-16 Registers
PUNPCKHWD instruction, 9-9 64-bit mode, 3-16, 3-20
PUNPCKLBW instruction, 9-9 control registers, 3-5
PUNPCKLDQ instruction, 9-9 CR in 64-bit mode, 3-6
PUNPCKLQDQ instruction, 11-16 debug registers, 3-5
PUNPCKLWD instruction, 9-9 EFLAGS register, 3-14, 3-20
PUSH instruction, 6-1, 6-3, 7-7, 7-31 EIP register, 3-14, 3-24
PUSHA instruction, 6-8, 7-8 general purpose registers, 3-13, 3-14
PUSHF instruction, 3-20, 6-8, 7-30 instruction pointer, 3-14
PUSHFD instruction, 3-20, 6-8, 7-30 machine check registers, 3-5
PXOR instruction, 9-10 memory management registers, 3-5
MMX registers, 3-3, 9-3
MSRs, 3-5
Q MTRRs, 3-5
QNaN floating-point indefinite, 4-7, 4-22, 4-24, 8-20 MXCSR register, 10-6
QNaNs performance monitoring counters, 3-5
description of, 4-22 REX prefix, 3-16
effect on COMISD and UCOMISD, 11-10 segment registers, 3-13, 3-17
encodings, 4-7 x87 FPU registers, 8-1
operating on, 4-22 XMM registers, 3-3, 10-4
rules for generating, 4-23 Related literature, 1-9
using in applications, 4-23 REP/REPE/REPZ/REPNE/REPNZ
Quadword, 4-1, 9-3 prefixes, 7-27, 13-4
Quiet NaN (see QNaN) Reserved bits, 1-5
RESET pin, 3-20
R RET instruction, 3-24, 6-4, 6-5, 7-22, 7-32
Return instruction pointer, 6-4
R8D-R15D registers, 3-16
Returns, from procedure calls
R8-R15 registers, 3-16
exception handler, return from, 6-14
RAX register, 3-16
far return, 6-6
RBP register, 3-16, 6-5
inter-privilege level return, 6-10
RBX register, 3-16
interrupt handler, return from, 6-14
RC (rounding control) field
near return, 6-5
MXCSR register, 4-25, 10-7
REX prefixes, 3-2, 3-16, 3-25
x87 FPU control word, 4-25, 8-12
RF (resume) flag, EFLAGS register, 3-23, A-1
RCL instruction, 7-19
RFLAGS, 3-24
RCPPS instruction, 10-12
RFLAGS register, 7-31
RCPSS instruction, 10-12
See EFLAGS register
RCR instruction, 7-19
RIP register, 6-5
RCX register, 3-16
64-bit mode, 7-2
RDI register, 3-16
description of, 3-24
RDX register, 3-16
relation to EIP, 7-2
Real address mode
ROL instruction, 7-19
handling exceptions in, 6-17
ROR instruction, 7-19
handling interrupts in, 6-17
Rounding
memory model, 3-9, 3-10
modes, floating-point operations, 4-25
memory model used, 3-11
modes, x87 FPU, 8-12

INDEX-12 Vol. 1
 INDEX

toward zero (truncation), 4-26 zero, 4-20

Rounding control (RC) field Significand, of floating-point number, 4-16
MXCSR register, 4-25, 10-7 Sign, floating-point number, 4-16
x87 FPU control word, 4-25, 8-12 SIMD floating-point exception (#XM), 11-25
RSI register, 3-16 SIMD floating-point exceptions
RSP register, 3-16, 6-5 denormal operand exception (#D), 11-21
RSQRTPS instruction, 10-12 divide-by-zero (#Z), 11-22
RSQRTSS instruction, 10-12 exception conditions, 11-19
exception handlers, E-1
inexact result exception (#P), 11-23
S invalid operation exception (#I), 11-20
SAHF instruction, 3-20, 7-30 list of, 11-19
SAL instruction, 7-15 numeric overflow exception (#O), 11-22
SAR instruction, 7-17 numeric underflow exception (#U), 11-22
Saturation arithmetic (MMX instructions), 9-5 precision exception (#P), 11-23
SBB instruction, 7-12 software handling, 11-26
Scalar operations summary of, C-1
defined, 10-10, 11-7 writing exception handlers for, E-1
scalar double-precision FP operands, 11-7 SIMD floating-point flag bits, 10-6
scalar single-precision FP operands, 10-10 SIMD floating-point mask bits, 10-6
Scale (operand addressing), 3-30, 3-32 SIMD floating-point rounding control field, 10-7
Scale, x87 FPU operation, 8-31 SIMD (single-instruction, multiple-data)
Scaling bias value, 8-41, 8-42 execution model, 2-3, 2-4, 9-4
SCAS instruction, 3-22, 7-27 instructions, 2-18, 5-20, 10-10
Segment MMX instructions, 5-13
defined, 3-8 operations, on packed double-precision
maximum number, 3-8 floating-point operands, 11-6
Segment override prefixes, 3-29 operations, on packed single-precision
Segment registers floating-point operands, 10-9
64-bit mode, 3-20, 3-30, 7-2 packed data types, 4-11
default usage rules, 3-29 SSE instructions, 5-16
description of, 3-13, 3-17 SSE2 instructions, 11-6, 12-3, 12-9
part of basic programming environment, 7-1 Sine, x87 FPU operation, 8-29
Segment selector Single-precision floating-point format, 4-6
description of, 3-8, 3-17 Sleep, 2-6
segment override prefixes, 3-29 Smart cache, 2-6
specifying, 3-29 Smart memory access, 2-14
Segmented memory model, 1-7, 3-8, 3-18 smart memory access, 2-6
Serialization of I/O instructions, 13-7 SMM
Serializing instructions, 13-7 memory model used, 3-11
SETcc instructions, 3-22, 7-20 overview, 3-1
SF (sign) flag, EFLAGS register, 3-21, A-1 SNaNs
SF (stack fault) flag, x87 FPU status word, 8-9, 8-37 description of, 4-22
SFENCE instruction, 10-20, 11-17, 11-37 effect on COMISD and UCOMISD, 11-10
SHL instruction, 7-15 encodings, 4-7
SHLD instruction, 7-18 operating on, 4-22
SHR instruction, 7-16 typical uses of, 4-22
SHRD instruction, 7-18 using in applications, 4-23
Shuffle instructions Software compatibility, 1-5
SSE extensions, 10-14 SP register, 3-16
SSE2 extensions, 11-10 Speculative execution, 2-9, 2-12
SHUFPD instruction, 11-10 Spin-wait loops
SI register, 3-16 programming with PAUSE instruction, 11-18
Signaling NaN (see SNaN) SQRTPD instruction, 11-8
Signed SQRTPS instruction, 10-12
infinity, 4-21 SQRTSD instruction, 11-9
integers, description of, 4-5 SQRTSS instruction, 10-12
integers, encodings, 4-6 SS register, 3-17, 3-19, 6-1

Vol. 1 INDEX-13
INDEX

SSE extensions packed and scalar floating-point instructions, 10-9

128-bit packed single-precision data type, 10-8 programming environment, 10-3
64-bit mode, 10-4 QNaN floating-point indefinite, 4-24
64-bit SIMD integer instructions, 10-16 restoring SSE and SSE2 state, 11-30
branching on arithmetic operations, 11-36 REX prefixes, 10-4
cacheability control instructions, 10-18 saving SSE and SSE2 state, 11-30
cacheability hint instructions, 11-36 saving XMM register state on a procedure or
caller-save requirement for procedure and function call, 11-34
function calls, 11-35 shuffle instructions, 10-14
checking for SSE and SSE2 support, 11-28 SIMD floating-point exception conditions, 11-19
comparison instructions, 10-13 SIMD floating-point exception cross reference,
compatibility mode, 10-4 C-4
compatibility of SIMD and x87 FPU floating-point SIMD floating-point exception (#XM), 11-25,
data types, 11-32 11-26
conversion instructions, 10-15 SIMD floating-point exceptions, 11-19
data movement instructions, 10-11 SIMD floating-point mask and flag bits, 10-6
data types, 10-8, 12-1 SIMD floating-point rounding control field, 10-7
denormal operand exception (#D), 11-21 SSE and SSE2 conversion instruction chart, 11-13
denormals-are-zeros mode, 10-7 SSE feature flag, CPUID instruction, 11-28
divide by zero exception (#Z), 11-22 SSE2 compatibility, 10-8
exceptions, 11-18 unpack instructions, 10-14
floating-point format, 4-15, 4-16 updating MMX technology routines
flush-to-zero mode, 10-7 using128-bit SIMD integer instructions, 11-35
generating SIMD FP exceptions, 11-23 x87 FPU compatibility, 10-8
guidelines for using, 11-27 XMM registers, 10-4
handling combinations of masked and unmasked SSE feature flag, CPUID instruction, 11-28, 12-7
exceptions, 11-26 SSE instructions
handling masked exceptions, 11-23 descriptions of, 10-9
handling SIMD floating-point exceptions in SIMD floating-point exception cross-reference,
software, 11-26 C-4
handling unmasked exceptions, 11-25, 11-26 summary of, 5-16
inexact result exception (#P), 11-23 SSE2 extensions
instruction prefixes, effect on SSE and SSE2 128-bit packed single-precision
instructions, 11-37 data type, 11-4
instruction set, 5-16, 10-9 128-bit packed single-precision data type, 12-2
interaction of SIMD and x87 FPU floating-point 128-bit SIMD integer instruction
exceptions, 11-26 extensions, 11-16
interaction of SSE and SSE2 instructions with x87 64-bit and 128-bit SIMD integer instructions,
FPU and MMX instructions, 11-31 11-15
interfacing with SSE and SSE2 procedures and 64-bit mode, 11-4
functions, 11-34 arithmetic instructions, 11-8
intermixing packed and scalar floating-point branch hints, 11-18
and 128-bit SIMD integer instructions branching on arithmetic operations, 11-36
and data, 11-32 cacheability control instructions, 11-17
introduction, 2-4 cacheability hint instructions, 11-36
invalid operation exception (#I), 11-20 caller-save requirement for procedure and
logical instructions, 10-13 function calls, 11-35
masked responses to invalid arithmetic operations checking for SSE and SSE2 support, 11-28
, 11-20 comparison instructions, 11-9
memory ordering instruction, 10-20 compatibility mode, 11-4
MMX technology compatibility, 10-8 compatibility of SIMD and x87 FPU floating-point
MXCSR register, 10-5 data types, 11-32
MXCSR state management instructions, 10-17 conversion instructions, 11-12
non-temporal data, operating on, 10-18 data movement instructions, 11-7
numeric overflow exception (#O), 11-22 data types, 11-4, 11-5, 12-2
numeric underflow exception (#U), 11-22 denormal operand exception (#D), 11-21
overview, 10-1 denormals-are-zero mode, 11-4
packed 128-Bit SIMD data types, 4-12 divide by zero exception (#Z), 11-22

INDEX-14 Vol. 1
 INDEX

exceptions, 11-18 SSE2 feature flag, CPUID instruction, 11-28, 12-7

floating-point format, 4-15, 4-16 SSE2 instructions
generating SIMD floating-point exceptions, 11-23 descriptions of, 11-6, 12-3, 12-9
guidelines for using, 11-27 SIMD floating-point exception cross-reference,
handling combinations of masked and unmasked C-7
exceptions, 11-26 summary of, 5-20
handling masked exceptions, 11-23 SSE3 extensions
handling SIMD floating-point exceptions in 64-bit mode, 12-1
software, 11-26 asymmetric processing, 12-2
handling unmasked exceptions, 11-25, 11-26 compatibility mode, 12-1
inexact result exception (#P), 11-23 DNA exceptions, 12-13
initialization of, 11-29 emulation, 12-14
instruction prefixes, effect on SSE and SSE2 enabling support in a system executive, 12-7,
instructions, 11-37 12-28
instruction set, 5-20 exceptions, 12-13
instructions, 11-6, 12-3, 12-9 guideline for packed addition/subtraction
interaction of SIMD and x87 FPU floating-point instructions, 12-8
exceptions, 11-26 horizontal addition/subtraction instructions, 12-5
interaction of SSE and SSE2 instructions with x87 horizontal processing, 12-2
FPU and MMX instructions, 11-31 instruction that addresses cache line splits, 5-25
interfacing with SSE and SSE2 procedures and instruction that improves X87-FP integer
functions, 11-34 conversion, 5-25
intermixing packed and scalar floating-point and instructions for horizontal addition/subtraction,
128-bit SIMD integer instructions and data, 5-26
11-32 instructions for packed addition/subtraction, 5-26
invalid operation exception (#I), 11-20 instructions that enhance
logical instructions, 11-9 LOAD/MOVE/DUPLICATE, 5-26
masked responses to invalid arithmetic operations instructions that improve synchronization
, 11-20 between agents, 5-27
memory ordering instructions, 11-17 LOAD/MOVE/DUPLICATE enhancement
MMX technology compatibility, 11-4 instructions, 12-4
numeric overflow exception (#O), 11-22 MMX technology compatibility, 12-2
numeric underflow exception (#U), 11-22 numeric error flag and IGNNE#, 12-14
overview of, 11-1 packed addition/subtraction instructions, 12-5
packed 128-Bit SIMD data types, 4-12 programming environment, 12-1
packed and scalar floating-point instructions, 11-6 REX prefixes, 12-1
programming environment, 11-3 SIMD floating-point exception cross reference,
QNaN floating-point indefinite, 4-24 C-11, C-13
restoring SSE and SSE2 state, 11-30 specialized 120-bit load instruction, 12-4
REX prefixes, 11-4 SSE compatibility, 12-2
saving SSE and SSE2 state, 11-30 SSE2 compatibility, 12-2
saving XMM register state on a procedure or x87 FPU compatibility, 12-2
function call, 11-34 SSE3 instructions
shuffle instructions, 11-10 descriptions of, 12-3
SIMD floating-point exception conditions, 11-19 SIMD floating-point exception
SIMD floating-point exception cross reference, cross-reference, C-11, C-13
C-7 summary of, 5-25
SIMD floating-point exception (#XM), 11-25, SSSE3 extensions
11-26 64-bit mode, 12-1
SIMD floating-point exceptions, 11-19 asymmetric processing, 12-2
SSE and SSE2 conversion instruction chart, 11-13 checking for support, 12-13
SSE compatibility, 11-4 compatibility, 12-2
SSE2 feature flag, CPUID instruction, 11-28 compatibility mode, 12-1
unpack instructions, 11-10 data types, 12-1
updating MMX technology routines using 128-bit DNA exceptions, 12-13
SIMD integer instructions, 11-35 emulation, 12-14
writing applications with, 11-27 enabling support in a system executive, 12-12
x87 FPU compatibility, 11-4 exceptions, 12-13

Vol. 1 INDEX-15
INDEX

horizontal add/subtract instructions, 12-9 Superscalar microarchitecture

horizontal processing, 12-2 P6 family microarchitecture, 2-3
MMX technology compatibility, 12-2 P6 family processors, 2-8
multiply and add packed instructions, 12-11 Pentium 4 processor, 2-11
numeric error flag and IGNNE#, 12-14 Pentium Pro processor, 2-3
packed absolute value instructions, 12-11 Pentium processor, 2-2
packed align instruction, 12-12 System management mode (see SMM)
packed multiply high instructions, 12-11
packed shuffle instruction, 12-12
programming environment, 12-1 T
SSSE2 compatibility, 12-2 Tangent, x87 FPU operation, 8-29
x87 FPU compatibility, 12-2 Task gate, 6-17
SSSE3 instructions Task register, 3-5
descriptions of, 12-8 Task state segment (see TSS)
summary of, 5-27 Tasks
Stack exception handler, 6-17
64-bit mode, 3-6, 6-5 interrupt handler, 6-17
64-bit mode behavior, 6-19 Temporal data, 10-18
address-size attribute, 6-3 TEST instruction, 7-21
alignment, 6-3 TF (trap) flag, EFLAGS register, 3-23, A-1
alignment of stack pointer, 6-3 Thermal Monitor, 2-6
current stack, 6-2, 6-4 Tiny number, 4-20
description of, 6-1 TOP (stack TOP) field
EIP register (return instruction pointer), 6-4 x87 FPU status word, 8-3, 9-12
maximum size, 6-1 TR register, 3-6
number allowed, 6-1 Trace cache, 2-12
overview of, 3-5 Transcendental instruction accuracy, 8-31
passing parameters on, 6-7 Trap gate, 6-14
popping values from, 6-1 Truncation
procedure linking information, 6-4 description of, 4-26
pushing values on, 6-1 with SSE-SSE2 conversion instructions, 4-26
return instruction pointer, 6-4 TSS
SS register, 6-1 I/O map base, 13-5
stack segment, 3-19, 6-1 I/O permission bit map, 13-5
stack-frame base pointer, EBP register, 6-4 saving state of EFLAGS register, 3-20
switching
on calls to interrupt and exception handlers, U
6-15
UCOMISD instruction, 11-10
on inter-privilege level calls, 6-11, 6-16
UCOMISS instruction, 10-14
privilege levels, 6-10
UD2 instruction, 7-33
width, 6-3
UE (numeric underflow exception) flag
Stack, x87 FPU
MXCSR register, 11-22
stack fault, 8-9
x87 FPU status word, 8-7, 8-42
stack overflow and underflow exception (#IS),
UM (numeric underflow exception) mask bit
8-7, 8-36, 8-37
MXCSR register, 11-22
Status flags
x87 FPU control word, 8-11, 8-42
EFLAGS register, 3-21, 8-9, 8-10, 8-28
Underflow
STC instruction, 3-22, 7-29
FPU exception
STD instruction, 3-22, 7-30
(see Numeric underflow exception)
STI instruction, 7-31, 13-5
numeric, floating-point, 4-20
Sticky bits, 8-7
x87 FPU stack, 8-36, 8-37
STMXCSR instruction, 10-17, 11-34
Underflow, x87 FPU stack, 8-37
STOS instruction, 3-22, 7-27
Unpack instructions
Streaming SIMD extensions 2 (see SSE2 extensions)
SSE extensions, 10-14
Streaming SIMD extensions (see SSE extensions)
SSE2 extensions, 11-10
String data type, 4-11
UNPCKHPD instruction, 11-11
ST(0), top-of-stack register, 8-4
UNPCKHPS instruction, 10-15
SUB instruction, 7-12

INDEX-16 Vol. 1
 INDEX

UNPCKLPD instruction, 11-11 instruction pointer, 8-13

UNPCKLPS instruction, 10-15 instruction set, 8-21
Unsigned integers last instruction opcode, 8-14
description of, 4-5 overview of registers, 3-3
range of, 4-5 programming, 8-1
types, 4-5 QNaN floating-point indefinite, 4-24
Unsupported, 8-20 register stack, 8-2
floating-point formats, x87 FPU, 8-20 register stack, parameter passing, 8-5
x87 FPU instructions, 8-34 registers, 8-1
save and restore state instructions, 5-13
saving registers, 11-34
V state, 8-15
Vector (see Interrupt vector) state, image, 8-16, 8-17
VIF (virtual interrupt) flag, EFLAGS register, 3-23 state, saving, 8-15, 8-17
VIP (virtual interrupt pending) flag status register, 8-6
EFLAGS register, 3-23 tag word, 8-12
Virtual 8086 mode transcendental instruction accuracy, 8-31
description of, 3-23 x87 FPU control word
memory model, 3-9, 3-10 description of, 8-10
VM (virtual 8086 mode) flag, EFLAGS register, 3-23 exception-flag mask bits, 8-11
VMCALL instruction, 5-37 infinity control flag, 8-12
VMCLEAR instruction, 5-36, 5-37 precision control (PC) field, 8-11
VMLAUNCH instruction, 5-37 rounding control (RC) field, 4-25, 8-12
VMPTRLD instruction, 5-36, 5-37 x87 FPU exception handling
VMPTRST instruction, 5-36, 5-37 description of, 8-45
VMREAD instruction, 5-37 floating-point exception summary, C-2
VMRESUME instruction, 5-37 MS-DOS compatibility mode, 8-45
VMWRITE instruction, 5-37 native mode, 8-45
VMX x87 FPU floating-point exceptions
instruction set, 5-36, 5-37 denormal operand exception, 8-39
introduction, 2-27 division-by-zero, 8-40
Virtual machine monitor (VMM), 2-27 exception conditions, 8-36
virtualization, 2-27 exception summary, C-2
VMXOFF instruction, 5-37 guidelines for writing exception handlers, D-1
VMXON instruction, 5-37 inexact-result (precision), 8-42
interaction of SIMD and x87 FPU floating-point
W exceptions, 11-26
invalid arithmetic operand, 8-36, 8-38
Waiting instructions, x87 FPU, 8-33
MS-DOS compatibility mode, D-1
WAIT/FWAIT instructions, 8-33, 8-44
numeric overflow, 8-40
WC memory type, 10-18
numeric underflow, 8-41
wide dynamic execution, 2-6
software handling, 8-45
Word, 4-1
stack overflow, 8-7, 8-37
Wraparound mode (MMX instructions), 9-5
stack underflow, 8-7, 8-36, 8-37
summary of, 8-34
X synchronization, 8-43
x87 FPU x87 FPU instructions
64-bit mode, 8-2 arithmetic vs. non-arithmetic instructions, 8-35
compatibility mode, 8-2 basic arithmetic, 8-24
control word, 8-10 comparison and classification, 8-26
data pointer, 8-13 control, 8-32
data registers, 8-2 data transfer, 8-22
execution environment, 8-1 exponential, 8-31
floating-point data types, 8-17 instruction set, 8-21
floating-point format, 4-15, 4-16 load constant, 8-24
fopcode compatibility mode, 8-14 logarithmic, 8-31
FXSAVE and FXRSTOR instructions, 11-34 operands, 8-22
IEEE Standard 754, 8-1 overview, 8-21

Vol. 1 INDEX-17
INDEX

save and restore state, 8-32

scale, 8-31
transcendental, 8-31
transitions between x87 FPU and MMX code, 9-12
trigonometric, 8-29
unsupported, 8-34
x87 FPU status word
condition code flags, 8-6
DE flag, 8-39
description of, 8-6
exception flags, 8-7
OE flag, 8-40
PE flag, 8-7
stack fault flag, 8-9
TOP field, 8-3
top of stack (TOP) pointer, 8-6
x87 FPU tag word, 8-12, 9-12
XADD instruction, 7-6
XCHG instruction, 7-5
XLAT/XLATB instruction, 7-33
XMM registers
64-bit mode, 3-6
description, 10-4
FXSAVE and FXRSTOR instructions, 11-34
overview of, 3-3
parameters passing in, 11-34
saving on a procedure or function call, 11-34
XOR instruction, 7-15
XORPD instruction, 11-9
XORPS instruction, 10-13

Z
ZE (divide by zero exception) flag
x87 FPU status word, 8-7, 8-40
ZE (divide by zero exception) flag bit
MXCSR register, 11-22
Zero, floating-point format, 4-7, 4-20
ZF (zero) flag, EFLAGS register, 3-21, A-1
ZM (divide by zero exception) mask bit
MXCSR register, 11-22
x87 FPU control word, 8-11, 8-40

INDEX-18 Vol. 1