LX5280 Data Sheet

Lexra, Inc.

Release 1.9

April 30, 2001

Lexra Proprietary and Confidential

 April 30, 2001 LX5280

LX5280 Data Sheet Revision 1.1, for RTL Release 1.9.

This document is proprietary and confidential to Lexra, Inc.

Copyright  2001 Lexra, Inc.
ALL RIGHTS RESERVED

MIPS, MIPS16, MIPS ABI, MIPSII, MIPSIV, MIPSV, MIPS32, R3000, R4000, and other MIPS
common law marks are trademarks and/or registered trademarks of MIPS Technologies, Inc. Lexra,
Inc. is not associated with MIPS Technologies, Inc. in any way.

SmoothCore, Radiax, and NetVortex are trademarks of Lexra, Inc.

Lexra Proprietary & Confidential Release 1.9

 April 30, 2001 LX5280

Table of Contents
1. LX5280 Product Overview ........................................................................................1
1.1. Introduction ..................................................................................................................... 1
1.2. LX5280 Processor Overview .......................................................................................... 3
1.3. System Level Building Blocks ....................................................................................... 4
1.3.1. SMMU .............................................................................................................. 4
1.3.2. Local Memory Interface ................................................................................... 4
1.3.3. Coprocessor Interface ....................................................................................... 5
1.3.4. Custom Engine Interface .................................................................................. 5
1.3.5. Lexra Bus Controller ........................................................................................ 5
1.3.6. Building Block Integration ............................................................................... 5
1.4. RTL Core & SmoothCore ............................................................................................... 5
1.5. EDA Tool Support .......................................................................................................... 6
2. LX5280 Architecture .................................................................................................7
2.1. Motivation ....................................................................................................................... 7
2.2. Hardware Architecture .................................................................................................... 7
2.2.1. Module Partitioning .......................................................................................... 7
2.2.2. Six Stage Pipeline ............................................................................................. 8
2.3. Dual Issue ....................................................................................................................... 9
2.3.1. Instruction Fetch ............................................................................................... 9
2.3.2. Instruction Analysis and Select Logic .............................................................. 9
2.3.3. MIPS16 ............................................................................................................. 9
2.4. RALU Data Path ........................................................................................................... 10
2.4.1. Overview ........................................................................................................ 10
2.4.2. Assignment of Instructions to Pipe A, Pipe B. ............................................... 11
2.5. System Control Coprocessor (CP0) .............................................................................. 12
2.6. Dual Multiliply-Accumulate (MAC) ............................................................................ 13
2.6.1. Dual MAC Operations ................................................................................... 13
2.6.2. MAC MODE (MMD) register ....................................................................... 14
2.6.3. Architecture .................................................................................................... 15
2.7. Data Addressing ............................................................................................................ 18
2.7.1. Twinword Data Movement ............................................................................ 18
2.7.2. Vector Addressing .......................................................................................... 18
2.7.3. Circular Buffers .............................................................................................. 20
2.8. Radiax ALU Operations ............................................................................................... 21
2.8.1. Extensions to MIPS ALU Operations ............................................................ 21
2.8.2. New ALU Instructions ................................................................................... 21
2.8.3. Conditional Move Operations ........................................................................ 21
2.9. Zero Overhead Loop Facility ........................................................................................ 22
2.10. Low-Overhead Prioritized Interrupts ............................................................................ 24
3. LX5280 RISC Programming Model .......................................................................27
3.1. Summary of MIPS-I Instructions .................................................................................. 27
3.1.1. ALU Instructions ............................................................................................ 28
3.1.2. Load and Store Instructions ............................................................................ 29
3.1.3. Conditional Move Instructions ....................................................................... 29
3.1.4. Branch and Jump Instructions ........................................................................ 30
3.1.5. Control Instructions ........................................................................................ 31
3.1.6. Coprocessor Instructions ................................................................................ 31

Lexra Proprietary & Confidential -i- Release 1.9

 April 30, 2001 LX5280

3.2. Opcode Extension Using the Custom Engine Interface (CEI) ..................................... 32
3.2.1. CEI Operations ............................................................................................... 32
3.2.2. Interface Signals ............................................................................................. 32
3.3. Memory Management ................................................................................................... 33
3.4. Exception Processing .................................................................................................... 33
3.4.1. Exception Processing Registers ..................................................................... 35
3.4.2. Exception Processing: Entry and Exit ............................................................ 36
3.5. The Coprocessor Interface (CI) .................................................................................... 36
3.6. Power Savings Mode .................................................................................................... 36
4. MIPS16 ......................................................................................................................39
4.1. MIPS16 Instructions ..................................................................................................... 39
4.2. Mode switching ............................................................................................................ 42
4.3. Exceptions ..................................................................................................................... 42
4.4. No Delay Slots .............................................................................................................. 42
5. LX5280 DSP Programming Model .........................................................................43
5.1. Radiax Instructions ....................................................................................................... 43
5.1.1. Radiax Dual-MAC Instructions ..................................................................... 43
5.1.2. Cycle-by-Cycle Usage for Dual MAC Instructions ...................................... 46
5.1.3. Vector Addressing Instructions ...................................................................... 48
5.1.4. Radiax ALU Operations ................................................................................. 50
5.1.5. Conditional Operations .................................................................................. 52
5.2. Instruction Encoding ..................................................................................................... 53
5.2.1. Lexra Formats ................................................................................................ 53
5.3. Load/Store Formats ...................................................................................................... 54
5.3.1. Arithmetic Format .......................................................................................... 56
5.3.2. MAC Format A .............................................................................................. 57
5.3.3. MAC Format B ............................................................................................... 58
5.3.4. MAC Format C ............................................................................................... 59
5.3.5. RADIAX MOVE Format and Lexra-Cop0 MTLXC0/MFLXC0 .................. 59
5.3.6. CMOVE Format ............................................................................................. 61
5.3.7. Lexra SUBOP Bit Encodings ......................................................................... 61
6. Integer Multiply-Divide-Accumulate .....................................................................63
6.1. Summary of Instructions ............................................................................................... 63
6.2. MAC-DIV Instruction Overview .................................................................................. 64
6.3. Op-codes for Standard Mode (32-Bit) MAC Instructions ............................................ 65
6.4. Op-codes for MIPS-16 (16-Bit) Mode MAC Instructions .......................................... 66
6.5. Non-Standard Instruction Descriptions ........................................................................ 67
6.6. Accessing HI and LO after multiply instructions ......................................................... 69
6.7. Divider Overview and Register Usage ......................................................................... 70
7. LX5280 Local Memory ............................................................................................71
7.1. Local Memory Overview .............................................................................................. 71
7.2. Cache Control Register: CCTL .................................................................................... 72
7.3. Instruction Cache (ICACHE) LMI ............................................................................... 73
7.4. Instruction Memory (IMEM) LMI ............................................................................... 75
7.5. Instruction ROM (IROM) LMI .................................................................................... 76
7.6. Direct Mapped Write Through Data Cache (DCACHE) LMI ..................................... 77
7.7. Scratch Pad Data Memory (DMEM) LMI .................................................................... 78
8. LX5280 System Bus .................................................................................................81
8.1. Connecting the LX5280 to internal devices ................................................................. 81

Lexra Proprietary & Confidential - ii - Release 1.9

 April 30, 2001 LX5280

8.2. Terminology ................................................................................................................. 81

8.3. Bus Operations .............................................................................................................. 82
8.3.1. Single-Cycle Read .......................................................................................... 82
8.3.2. Read Line ....................................................................................................... 82
8.3.3. Burst Read ...................................................................................................... 83
8.3.4. Single-Cycle Write ......................................................................................... 83
8.3.5. Line Write ...................................................................................................... 83
8.3.6. Burst Write ..................................................................................................... 83
8.4. Signal Descriptions ....................................................................................................... 84
8.5. LBus Commands .......................................................................................................... 84
8.6. Byte Alignment ............................................................................................................. 85
8.7. Lexra Bus Controller .................................................................................................... 85
8.7.1. LBC Commands ............................................................................................. 85
8.7.2. LBC Write Buffer and Out-of-Order Processing ........................................... 86
8.7.3. LBC Read Buffer ........................................................................................... 86
8.7.4. Transfer Descriptions ..................................................................................... 87
8.7.5. Single Cycle Read with No Waits .................................................................. 88
8.7.6. Single Cycle Read with Target Wait .............................................................. 89
8.7.7. Line Read with No Waits ............................................................................... 89
8.7.8. Line Read with Target Waits ......................................................................... 90
8.7.9. Line Read with Initiator Waits ....................................................................... 90
8.7.10. Burst Read ...................................................................................................... 91
8.7.11. Single-Cycle Write with No Waits ................................................................. 91
8.7.12. Single-Cycle Write with Waits ...................................................................... 92
8.7.13. Burst Write with No Waits ............................................................................. 92
8.7.14. Burst Write with Target Waits ....................................................................... 93
8.7.15. Burst Write with Initiator Waits ..................................................................... 93
8.8. LBC Signals .................................................................................................................. 94
8.9. Arbitration ..................................................................................................................... 95
8.9.1. Rules ............................................................................................................... 95
8.9.2. LBC behavior ................................................................................................. 95
8.10. Connecting Devices to the Bus ..................................................................................... 95
9. LX5280 Coprocessor Interface ...............................................................................97
9.1. Attaching a Coprocessor Using the Coprocessor Interface (CI) .................................. 97
9.2. Coprocessor Interface (CI) Signals ............................................................................... 97
9.3. Coprocessor Write Operations ...................................................................................... 98
9.4. Coprocessor Read Operations ....................................................................................... 98
9.5. Coprocessor Interface and Pipeline Stages ................................................................... 99
9.5.1. Pipeline Holds ................................................................................................ 99
9.5.2. Pipeline Invalidation ...................................................................................... 99
10. LX5280 EJTAG ......................................................................................................101
10.1. Introduction ................................................................................................................. 101
10.2. Overview ..................................................................................................................... 101
10.2.1. IEEE JTAG-specific Pinout ......................................................................... 102
10.3. Single Processor PC Trace .......................................................................................... 102
10.3.1. PC Trace DCLK - Debug Clock .................................................................. 103
10.3.2. PC Trace PCST - Program Counter Status Trace ........................................ 103
10.3.3. PC Trace TPC - Target Program Counter .................................................... 103
10.3.4. Dual Pipe PC Trace ...................................................................................... 103
10.3.5. Single-Processor PC Trace Pinout ............................................................... 104

Lexra Proprietary & Confidential - iii - Release 1.9

 April 30, 2001 LX5280

10.3.6. Vectored Interrupts and PC Trace ................................................................ 104

10.3.7. Demultiplexing of TDO and TDI During PC Trace .................................... 105
Appendix A.LX5280 Lconfig Forms ...........................................................................107
A.1. Configuration Options for the LX5280 Processor ...................................................... 107
Appendix B.LX5280 Port Descriptions ......................................................................109
Appendix C. LX5280 Pipeline Stalls ...........................................................................117
C.1. Stall Definitions .......................................................................................................... 117
C.2. Instruction Groupings ................................................................................................. 117
C.3. Dual Pipe Issue Rules ................................................................................................. 118
C.4. M16 32-bit Instructions .............................................................................................. 119
C.5. Non-Sequential Program Flow Issue Stalls ................................................................ 119
C.6. Load/Store Rules ........................................................................................................ 119
C.7. Load/Store Ops Stall Matrix ....................................................................................... 121
C.8. Mac Ops Interlock Matrix .......................................................................................... 122
C.9. MVCz Stall ................................................................................................................. 124
C.10. ZovLoop Rules ........................................................................................................... 124
C.11. IMMU Stalls ............................................................................................................... 125
C.12. Cache Miss Stalls ........................................................................................................ 125
C.13. Non-Sequential Program Flow Issue Stall Pipeline Diagrams ................................... 126
C.14. Load/Store Stall Pipeline Diagrams ............................................................................ 127
C.15. Mac Ops Interlock Pipeline Diagram ......................................................................... 129
C.16. MVCz Stall Pipeline Diagrams .................................................................................. 129
C.17. ZovLoop Pipeline Diagrams ....................................................................................... 129
C.18. Cache Miss Pipeline Diagrams ................................................................................... 130

Lexra Proprietary & Confidential - iv - Release 1.9

 April 30, 2001 LX5280

List of Tables
Table 1: EDA Tool Support .................................................................................................... 6
Table 2: Assignment of Instructions of Pipe A, Pipe B ........................................................ 11
Table 3: CP0 Registers.......................................................................................................... 13
Table 4: MMD Fields (Radiax User Register 24)................................................................. 15
Table 5: Prioritized Interrupt Exception Vectors.................................................................. 25
Table 6: ALU Instructions .................................................................................................... 28
Table 7: Load and Store Instructions .................................................................................... 29
Table 8: Conditional Move Instructions ............................................................................... 29
Table 9: Branch and Jump Instructions................................................................................. 30
Table 10: Control Instructions ................................................................................................ 31
Table 11: Coprocessor Instructions......................................................................................... 31
Table 12: Custom Engine Interface Operations...................................................................... 32
Table 13: Custom Engine Interface Signals............................................................................ 32
Table 14: SMMU Address Mapping....................................................................................... 33
Table 15: List of Exceptions ................................................................................................... 34
Table 16: MIPS I Instructions Not Supported by MIPS16 ..................................................... 40
Table 17: MIPS16 Instructions that Support MIPS I.............................................................. 40
Table 18: New MIPS16 Instructions....................................................................................... 41
Table 19: PC-Relative Addressing.......................................................................................... 41
Table 20: Radiax Dual-MAC Instructions .............................................................................. 43
Table 21: Vector Addressing Instructions .............................................................................. 48
Table 22: Radiax ALU Operations ......................................................................................... 50
Table 23: Conditional Operations ........................................................................................... 52
Table 24: Summary of MAC-DIV Instructions. ..................................................................... 63
Table 25: 16-bit Multiply and Multiply-Accumulate Instructions.......................................... 67
Table 26: 32-Bit Multiply-Accumulate Instructions............................................................... 68
Table 27: Local Memory Interface Modules .......................................................................... 71
Table 28: ICACHE Configurations......................................................................................... 73
Table 29: ICACHE RAM Interfaces....................................................................................... 74
Table 30: IMEM Configurations............................................................................................. 75
Table 31: IMEM RAM Interfaces........................................................................................... 75
Table 32: IROM Configurations ............................................................................................. 76
Table 33: IROM ROM Interfaces ........................................................................................... 77
Table 34: DCACHE Configurations ....................................................................................... 78
Table 35: DCACHE RAM Interfaces ..................................................................................... 78
Table 36: DMEM Configurations ........................................................................................... 79
Table 37: DMEM RAM Interfaces ......................................................................................... 79
Table 38: Line Read Interleave Order..................................................................................... 83
Table 39: LBus Signal Description......................................................................................... 84
Table 40: LBus Byte Lane Assignment .................................................................................. 85
Table 41: LBus Commands Issued by the LBC...................................................................... 86
Table 42: LBC Interface Signals............................................................................................. 94
Table 43: Coprocessor Interface Signals ................................................................................ 97
Table 44: EJTAG Pinout....................................................................................................... 102
Table 45: EJTAG AC Characteristics................................................................................... 102
Table 46: EJTAG Synthesis Constraints............................................................................... 102
Table 47: Single-Processor PC Trace Pinout........................................................................ 104
Table 48: Single-Processor PC Trace AC Characteristics .................................................... 104
Table 49: LX5280 Processor Port Summary ........................................................................ 109
Table 50: Instruction Groupings For Stall Definition........................................................... 117
Table 51: Load/Store Ops Stall Matrix ................................................................................. 122
Table 52: Cycles Required Between Dual MAC Instructions .............................................. 123

Lexra Proprietary & Confidential -v- Release 1.9

 April 30, 2001 LX5280

Lexra Proprietary & Confidential - vi - Release 1.9

 April 30, 2001 LX5280

List of Figures
Figure 1: LX5280 Processor Overview.................................................................................... 3
Figure 2: Superscalar Processor Core Module Partitioning ..................................................... 8
Figure 3: Superscalar Instruction Issue .................................................................................. 10
Figure 4: Dual MAC Data Path.............................................................................................. 16
Figure 5: Post-modified Pointers with Circular Buffer Support ............................................ 20
Figure 6: Lexra System Bus Diagram .................................................................................... 81

Lexra Proprietary & Confidential - vii - Release 1.9

 April 30, 2001 LX5280

Lexra Proprietary & Confidential - viii - Release 1.9

 April 30, 2001 LX5280

1. LX5280 Product Overview

1.1. Introduction

This data sheet describes the LX5280, a high-performance RISC-DSP, developed for Intellectual Property
(IP) licensing. However, the LX5280 is not just a highly specialized DSP architecture, but also a carefully
engineered extension to the MIPS ISA. As a result, system functions and computationally intensive DSP
algorithms can be integrated on a single, low-cost subsystem. Key applications include data communication
products such as network protocol processors, cable and xDSL modems, Voice over Packetized Network
gateways, and set-top boxes, as well as disk controllers.

As DSP-intensive applications have gained commercial importance, there has been an increasing recognition
of the benefit of implementing DSP functions on the CPU. A CPU is usually required for memory
management, user interface and control software; CPUs also have excellent third-party software tool support.
However, software implementations of DSP algorithms such as the FIR Filter or Discrete Cosine Transform
(DCT) typically suffer by an order of magnitude in performance compared to specialized DSPs. The problem
is compounded by the difficulty of deterministically allocating real-time in sophisticated CPUs. Vendors have
addressed these problems by offering DSP Coprocessors which have separate instruction sets, separate
instruction stores and execution units; and DSP Accelerators, which share the same I-stream with the CPU
but have separate DSP execution units. Each of these approaches imposes a substantial burden on the CPU in
managing the DSP functions. The LX5280, on the other hand, tightly integrates its DSP extensions into the
MIPS ISA. As a result, a wide variety of third-party tools are available and the LX5280 programmer can
switch seamlessly from RISC code to DSP code.

The LX5280 adds to the MIPS-I instruction set a collection of DSP-oriented instructions called the RadiaxTM
instruction set. The Radiax instruction set adds dual 16-bit multiply and multiply-accumulate operations
including DSP modes such as saturation, rounding and fractional arithmetic. It includes DSP addressing
modes such as post-modified address pointers, circular buffers and zero-overhead loops. It also includes dual
16-bit SIMD ALU operations and data alignment operations for applications where 16 bits of data is
sufficient.

The LX5280 pipeline is a dual-issue, six-stage architecture. Pipe A is the Load/Store Pipe and includes data
memory access and all MIPS instructions except multiply and divide operations, while Pipe B is the
Multiply-Accumulate Pipe and includes a MAC and ALU, each with dual 16-bit operations. DSP algorithms
will typically use Pipe A to load a pair of operands into a general register while executing Dual MAC
operations in Pipe B on earlier data. Decoupling register loads from the MAC allows loop unrolling and takes
effective advantage of the 32 general registers for temporary storage. As compared to memory-based
operands which are common to specialized DSP instruction sets, dual-issue allows the LX5280 to achieve the
memory bandwidth required by DSP within a RISC architecture. In addition, the LX5280 introduces
twinword load and store functions, allowing 64-bits of data to be moved between the local cache and register
file in a single cycle. This provides sufficient data-movement for data-hungry DSP inner loops.

Features introduced in Lexra’s RISC product line to support System-on-Chip (SoC) design, including
customer-defined Coprocessors and customer extensions to the MIPS ISA, are standard in the LX5280.
Configuration options include Extended-JTAG (EJTAG) support for debug and In-Circuit Emulation (ICE).
Lexra’s products include the same memory management stub (SMMU) as the LX4189.

Because the LX5280 executes the MIPS instruction set, a wide variety of third-party software tools are
available including compilers, operating systems, debuggers and in-circuit emulators. The assembler
extensions and a cycle accurate Instruction Set Simulator (ISS) are developed by Lexra. Programmers can
use “off-the-shelf” C Compilers for initial coding; then replace performance-critical loops with optimized
assembler code.

Third parties provide C compiler support for the new DSP instructions and will supply DSP macro libraries

Lexra Proprietary & Confidential -1- Release 1.9

 April 30, 2001 LX5280

and application packages. Compiler support is provided by the GreenHills MULTI IDE package. A DSP
library of functions such as filters and transforms will be available from Lexra.

Memory sizes and peripherals can be tailored by the licensee to system requirements, avoiding excess silicon
cost and power dissipation. It is expected that the LX5280 will deliver exceptional price/performance for
numerous consumer products and that multiple LX5280 subsystems on a single die can cost-effectively
implement high-performance next generation telecommunications systems.

Key Features

• Complete Processor Subsystem

• Executes MIPS I ISA (except unaligned loads, stores).

• Extensive third-party tool support.
• Dual instruction issue.
• High-performance 6-stage pipeline.
• Local instruction memory and/or cache, configurable sizes.
• Local data memory and/or cache, configurable sizes.
• Memory interface logic included.
• System bus controller.
• Optional customer-defined coprocessors.
• Optional customer-defined instruction extensions.
• Supports EJTAG Draft 2.0 for debugging.

• Portable RTL Model

• Available as a synthesizable RTL.

• Portable to any 0.25µm, 0.18µm or 0.15µm.
logic and SRAM process.
• Foundry partners include IBM, TSMC, and UMC.

• Easy ASIC Design

• Single phase clocking.

• Fully synchronous design.
• Easy to interface system bus protocol.
• Supports popular EDA tools.

• Executes Lexra’s RadiaxTM Instruction Set

• SIMD operations.
• Zero-overhead loop.
• Multiply-accumulate instructions.
• Vector and circular buffer addressing modes.

• Easy RTL Customization

• User-configurable local memory, reset method, clock distribution.

• User-configurable EJTAG breakpoints.
• Over 30 other configuration options.
• Interfaces for adding application-specific instructions.

Lexra Proprietary & Confidential -2- Release 1.9

 April 30, 2001 LX5280

1.2. LX5280 Processor Overview

The LX5280 is a RISC-DSP processor that executes the MIPS-I instruction set1 along with Lexra’s
RadiaxTM DSP extensions. However, the clocking, pipeline structure, pin-out, and memory interfaces have
all been designed by Lexra to reflect system-on-silicon design needs, deep submicron process technology, as
well as design methodology advances.

The figure below shows the structure of the LX5280 processor.

Icache Inst Inst

RAM RAM ROM

Instruction Bus (Addr, Data, Ctl)

LX5280 Inst LMI
Custom CPU System
Engine
CEI LBC
Core Data Bus (Addr, Data, Ctl) Bus
Data LMI

CI (1-3)
Dcache Data
RAM RAM

denotes customer logic Coprocessor(s)

Figure 1: LX5280 Processor Overview

MIPS ISA Execution. The LX5280 supports the MIPS I programming model. Two source operands can be
supplied and one destination update performed per cycle. The second operand is either a register or 16-bit
immediate. The instruction set includes a wide selection of ALU operations executed by the RALU, Lexra’s
proprietary register based ALU. The RALU also generates memory addresses for 8-bit, 16-bit, and 32-bit
register loads from (stores to) memory by adding a register base to an immediate offset. An extension to the
MIPS ISA allows a pair of 32-bit registers to be loaded from (stored to) memory. Branches are based on
comparisons between registers, rather than flags, and are therefore easy to relocate. Optional links following
jump or branch instructions assist with subroutine programming.

The MIPS unaligned load and store instructions are not supported, because they represent poor
price/performance trade-off for embedded applications.

Pipeline. LX5280 instructions are executed by a six-stage pipeline that has been designed so that all
transactions internal to the LX5280, as well as at the interfaces, occur on the positive edge of the processor
clock. Two-phase clocks are not used.

Exception Handling. The MIPS R3000 exception handling model is supported. Exceptions include both
instruction-synchronous traps as well as hardware and software interrupts. The STATUS register controls the
interrupt mask and operating mode. Exceptions are prioritized. When an exception is taken, control is
transferred to the exception vector, the current instruction address is saved in the EPC register, and the
exception source is identified in the CAUSE register. A user program located at the exception vector identifies
the cause of the exception, and transfers control to the application-specific handler. In the event of an address
error exception, the BADVADDR holds the failing address.

Coprocessor Operations. The LX5280 supports 32-bit Coprocessor operations. These include moves to and
from the Coprocessor general registers and control registers (MTCz, MFCz, CTCz, CFCz), Coprocessor
loads and stores (LWCz, SWCz) and branches based on Coprocessor condition flags (BCzT, BCzF). The
Lexra-supplied Coprocessor Interface can support Coprocessor operations in a single cycle, without pipeline

1. The MIPS unaligned load and store instructions (LWL, LWR, SWL, SWR) are not supported.

Lexra Proprietary & Confidential -3- Release 1.9

 April 30, 2001 LX5280

stalls.

LX5280 provides excellent price/performance and time-to-market. There are two main approaches which
Lexra has taken to achieve this:

• Deliver simple building blocks outside the processor core to enable system level
customizations such as coprocessors, application specific instructions, memories, and
busses.

• Deliver either a fully synthesizable Verilog source model or fully implemented hardcore
(called SmoothCore) for popular pure-play foundries.

Section 1.3 describes the building blocks, and Section 1.4 describes the deliverable models.

1.3. System Level Building Blocks

The LX5280 processor is designed to easily fit into different target applications. It provides the following
building blocks.

• A simple memory management unit (SMMU).

• An optimized Custom Engine Interface (CEI).

• Up to three Coprocessor Interfaces (CI).

• A flexible Local Memory Interface (LMI) that supports instruction cache, instruction
RAM, instruction ROM, data cache and data RAM.

• A Lexra Bus Controller (LBC) to connect peripheral devices and secondary memories to
the processor’s own local buses.

The following sections discuss each of these system building block interfaces.

1.3.1. SMMU

The LX5280 SMMU is designed for embedded applications using a single address space. Its primary
function is to provide memory protection between user space and kernel space. The SMMU is consistent
with the MIPS address space scheme for User/Kernel modes, mapping, and cached/uncached regions.

1.3.2. Local Memory Interface

The LX5280’s Harvard Architecture provides Local Memory Interfaces (LMIs) that support instruction
memory and data memory. Synchronous memory interfaces are employed for all memory blocks. The LMI
block is designed to easily interface with standard memory blocks provided by ASIC vendors or by third-
party library vendors.

The LMIs provide a two-way set associative instruction cache interface, and a direct-mapped write-through
data cache interface. The tag compare logic as well as a cache replacement algorithm are provided as part of
the LMI. One of the instruction cache sets may be locked down as un-swappable local memory. Local
instruction and data memories can also be mapped to fixed regions of the physical address space, and include
non-volatile memory (such as ROM, flash, or EPROM).

Lexra Proprietary & Confidential -4- Release 1.9

 April 30, 2001 LX5280

1.3.3. Coprocessor Interface

Lexra supplies an optional Coprocessor Interface (CI) for applications requiring this functionality. Up to three
CIs may be implemented in one design. The Coprocessor Interface “eavesdrops” on the Instruction bus. If a
Coprocessor load (LWCz) or “move to” (MTCz, CTCz) is decoded, data is passed over the Data Bus into a
CI register, then supplied to the designer-defined Coprocessor. Similarly, if a Coprocessor store (SWCz) or
“move from” (MFCz, CFCz) is decoded, data is obtained from the Coprocessor and loaded into a CI register,
then transferred onto the Data Bus in the following cycle. The design interface includes a data bus, five-bit
address, and independent read and write selects for Coprocessor registers and control registers. The LX5280
pipeline and Harvard Architecture permit single cycle Coprocessor access and transfer. An application-
defined Coprocessor condition flag is synchronized by the CI then passed to the Sequencer for testing in
branch instructions.

1.3.4. Custom Engine Interface

The LX5280 includes a Custom Engine Interface (CEI) that the application may use to extend the MIPS I
ALU opcodes with application-specific or proprietary operations. Similar to the standard ALU, the CEI
supplies the Custom Engine two input 32-bit operands, SRC1 and SRC2. One operand is selected from the
Register File. Depending on the most significant 6 bits of the opcode, the second operand is either selected
from the Register File or is a 16-bit sign-extended immediate. The opcode is locally decoded by the custom
engine, and following execution by the custom engine, the result is returned on the 32-bit result bus to the
LX5280. To support multi-cycle operations, a stall input is included in the interface.

1.3.5. Lexra Bus Controller

The Lexra Bus Controller (LBC) is the interface between the LX5280 and the outside world, which includes
DRAM and various peripherals. It is a non-multiplexed, non-pipelined, and non-parity checked bus to
provide the easiest bus protocol for design integration. On the processor side, the LBC provides a write-buffer
of configurable depth to support the write-through cache, as well as the control for byte and half-word
transfers. On the peripheral side, the LBC is designed to easily interface to industry standard bus protocols,
such as PCI, USB, and FireWire.

The LBC can run at any speed from 33 MHz, up to the speed of the LX5280 processor core in both the RTL
core and SmoothCore.

1.3.6. Building Block Integration

The LX5280 configuration script, lconfig, provides a menu of selections for designers to specify building
blocks needed, number of different memory blocks, target speed, and target standard cell library. Next, the
configuration software automatically generates a top level Verilog model, makefiles, and scripts for all steps
of the design flow.

For testability purposes, all building blocks contain scan control signals. The Lexra synthesis scripts include
scan insertion, which allows ATPG testing of the entire LX5280 core.

1.4. RTL Core & SmoothCore

Lexra delivers LX5280 as RTL Core and SmoothCore.

RTL Core: For full ASIC designs, the RTL is fully synthesizable and scan-testable Verilog source code, and
may be targeted to any ASIC vendor’s standard cell libraries. In this case, the designer may simply follow the
ASIC vendor’s design flow to ensure proper sign-off. In addition to the Verilog source code and system level
test bench, Lexra provides synthesis scripts as well as floor plan guidelines to maximize the performance of
the LX5280.

Lexra Proprietary & Confidential -5- Release 1.9

 April 30, 2001 LX5280

SmoothCore: For COT designs that are manufactured at popular foundries such as IBM, TSMC, and UMC,
a SmoothCore port is the quickest, lowest cost, and best performance choice. In this case, the LX5280 has
been fully implemented and verified as a hard macro. All data path, register file, and interface optimizations
have been performed to ensure the smallest die size and fastest performance possible. Furthermore, there is a
scan based test pattern that provides excellent fault coverage during manufacturing tests.

1.5. EDA Tool Support

Lexra supports mainstream EDA software, so designers do not have to alter their design methodology. The
following is a snapshot of EDA tools currently supported:

Table 1: EDA Tool Support

Design Flow Tools Supported

Simulation Synopsys VCS

Cadence Verilog XL
Cadence NC-Verilog

Synthesis Synopsys Design Compiler

Static Timing Synopsys PrimeTime

DFT Synopsys TetraMax

P&R Avant! Apollo II

Lexra Proprietary & Confidential -6- Release 1.9

 April 30, 2001 LX5280

2. LX5280 Architecture

2.1. Motivation

The LX5280 issues dual 32-bit instructions to two distinct 6-stage execution pipelines. Superscalar
architectures have been widely deployed in RISC CPUs where increased performance is obtained at the cost
of significantly increased area. Although superscalar issue significantly increases the area of the LX5280
processor Core, performance analysis at Lexra demonstrates benefits on key DSP algorithms well beyond
that which is obtained in typical CPU benchmarks.

Sustaining peak computational performance in DSP algorithms typically requires at least one operand from
memory per instruction cycle. DSPs have traditionally implemented specialized instruction sets that support
memory-based operands. Single-issue RISC architectures operate on register-based operands and thus
degrade performance by a factor of two in order to pre-load the operand into the register file. Grafting
memory-based operands onto a RISC architecture is inconsistent with both the RISC pipeline and ISA. Dual-
issue superscalar design, on the other hand, allows operands to be loaded from memory by one instruction
while the Multiple Accumulate data path (MAC) operates on register-based data loaded earlier from memory.
The RISC data path is 32-bits, but few DSP algorithms require more than 16-bits of precision. Thus two
values can be fetched simultaneously from memory. Simultaneous dual 16-bit ALU and MAC operations
further improve the LX5280 DSP performance.

Compared to specialized 32-bit (or even 16-bit) DSP instructions which allow memory reference, the
superscalar approach will have lesser code density, but only within the DSP loop, or kernel. These kernels are
typically small sections of code which are executed many times. Thus the overall degradation of code density
is minimal and can be offset by use of MIPS16 code compression in “outer loop” code which is not
performance critical.

2.2. Hardware Architecture

2.2.1. Module Partitioning

The LX5280 processor core includes two major blocks: the RALU (register file and ALU) and the CP0
(Control Processor). The RALU performs ALU operations and generates data addresses while CP0 includes
instruction address sequencing, exception processing, and product specific mode control. The RALU and
CP0 are loosely-coupled and include their own independent instruction decoders.

Lexra Proprietary & Confidential -7- Release 1.9

 April 30, 2001 LX5280

Instruction Address and Control

Instructions

RALU

addr_wA2 addr_wB2
addr_wA1 Register File addr_wB1

CP0 addr_rA1 32 x 32b (r0 = 0) addr_rB1

addr_rA2
8p (4r/4w) addr_rB2
PC and
Sequencer
wA1 wA2 wB1 wB2

rA1 rA2 rB1 rB2

Exception
Processing Link address
Logic Jump address ALU A (Bypass ALU B
Data)
N, Z
PRiD Ov_A

Dbus1 cbs0-cbs2 CEI-1

cbe0-cbe2
CEI-0
Data addr
and control

Res0 OpA OpB Res1 OpA OpB

Dual MAC Custom

(DMAC) Engine
m0-m3 (Optional)

Figure 2: Superscalar Processor Core Module Partitioning

2.2.2. Six Stage Pipeline

The LX5280 has a six stage pipeline:

Stage 1 I Instruction fetch

Stage 2 D Decode
Stage 3 S Source fetch (register file read)
Stage 4 E Execution and address generation
Stage 5 M Memory data select (read data cache store and tags)
Stage 6 W Write back to register file

The LX5280 I-Cache and IRAM can fetch two 32-bit instructions I0_I, I1_I simultaneously. Following the
superscalar instruction buffer and issue logic, described below, the instructions are issued to Pipe B and Pipe
A as appropriate. To avoid degrading operating frequency, the superscalar issue logic operates during the
Decode stage (D-stage) of the pipeline. Support for fully synchronous memories in the LX5280 has the
added benefit of isolating the processor logic from the customer-supplied memories in the instruction cache,
thus facilitating integration of the LX5280 into SoC designs.

As a result of the D-Stage, a two cycle penalty is incurred on branch prediction failure vs. the one-cycle
penalty in the LX4180 five stage pipeline. However, the LX5280’s zero-overhead loop hardware and

Lexra Proprietary & Confidential -8- Release 1.9

 April 30, 2001 LX5280

conditional move instructions can be used to avoid any wasted cycles in the control of real-time critical loops.

2.3. Dual Issue

2.3.1. Instruction Fetch

Two instructions are fetched during each instruction cache access. In the event of a cache miss, the processor
will be stalled until the cache line containing the requested instructions is retrieved. In the event that only one
instruction of a fetched pair can issue, the fetch will be stalled until the second instruction is issued to the
pipeline.

Instruction fetches always occur on an aligned 64-bit address boundary. In the event of a branch to an odd 32-
bit address in the 64-bit boundary, both instructions in the 64-bit window will be fetched, but only the second
(odd) instruction will issue to the pipeline. The first, or even, instruction will be ignored.

2.3.2. Instruction Analysis and Select Logic

The Instruction Analysis and Select Logic is located in the D-stage of the pipeline. During this stage, the
processor analyzes both instructions in a fetched pair and determines which pipeline can execute the
instructions. For example, if the first instruction in the pair, I0, is an ADD, and the second instruction I1is a
MAC, the processor will determine that I0 can be executed by either Pipe A or Pipe B while I1 can be
executed by Pipe B. The Instruction Select Logic will then issue I0 to pipe A and I1 to pipe B, since only pipe
B can execute the MAC instruction.

If both instructions of the fetched pair can only be issued to one pipeline (for example, a pair of MAC
instructions, which can only issue to Pipe B), the two instructions will be issued serially. The instruction fetch
will be stalled by one cycle until the second instruction has been issued to the pipeline.

If the result of the first instruction, I0, is used by the second instruction, I1, only one of the two instructions
will issue. The second instruction, I1, will issue in the next cycle, and the instruction fetch will be stalled for
one cycle until I1 has been issued.

2.3.3. MIPS16

The MIPS16_N signal indicates whether or not MIPS16 code compression has been enabled. If so, each 32-
bit fetch is interpreted as a pair of 16-bit instructions encoded according to the MIPS16 Specification.
MIPS16 instructions are not dual-issued, but always issued to Pipe A. It is expected that MIPS16 code
compression is enabled for “outer loop” code where code density is more important than performance. The
critical Register File read addresses for MIPS16 are resolved during the D-stage so that register file access for
MIPS16 instructions, as for 32-bit MIPS instructions, can begin on the rising edge of the S-Stage clock.

Lexra Proprietary & Confidential -9- Release 1.9

 April 30, 2001 LX5280

I$ and IRAM
1 V_I 32 I0_I 32 I1_I

Mux Mux Mux

1
MIPS16_N
V0/V1 I0 I1

Instructions
Analysis
Logic

Inst Issue
State

Instructions NOP NOP

1 Select
Stall Logic Mux Mux

Reg Reg Reg

1 32 32

ORD_S_R IB_S_R to Pipe B IA_S_R to Pipe A

Figure 3: Superscalar Instruction Issue

2.4. RALU Data Path

2.4.1. Overview

The Superscalar RALU Datapath is illustrated in the Figure. Operations are divided between Pipe A and Pipe
B in such a way that the RALU is the only major section of the processor which requires both Pipe A and B
instructions. Coprocessor 0, as well as the optional customer-defined Coprocessors 1-3, only require the Pipe
A instruction.

To “first approximation” the superscalar RALU is a “doubling” of the LX4180 RALU: it includes an 8-port
(4r/4w) general register file with 4-ports (2r/2w) assigned to Pipe A, and 4-ports (2r/2w) assigned to Pipe B.
In each Pipe, one write port is dedicated to register file updates from the Data Bus (Loads, MFCz, CFCz -
moves from Coprocessor). The remaining three ports (2r/1w) are available for the other operations assigned
to that Pipe. As a result, loads, including “twinword” loads of register pairs can dual-issue with any MAC or
ALU instruction without register port access restriction.

Each Pipe has an ALU and a nearly-independent control section. Differences occur in the assignment of
operations to Pipe A and Pipe B, and in the pipeline features to support superscalar. The pipeline differences
in the RALU to support superscalar issue are:

• Data must be forwarded from Pipe A (Pipe B) to Pipe B (Pipe A) when the input to a Pipe
B (Pipe A) execution unit requires a result computed earlier in Pipe A (Pipe B). The

Lexra Proprietary & Confidential - 10 - Release 1.9

 April 30, 2001 LX5280

forwarding paths are illustrated in Figure 2.

• If both Pipe A and Pipe B operations write the same register, the RALU control examines
the instruction order and suppresses the write for the earlier instruction based on program
order.

In addition, the RALU interfaces with the dual-MAC as a custom engine; this interface can supply two, 32-
bit operands per cycle, and return a single 32-bit operand per cycle. In the case of the dual-MAC, each of the
32-bit operands can be interpreted as two independent 16-bit operands.

2.4.2. Assignment of Instructions to Pipe A, Pipe B.

Table 2 lists the detailed assignment of instructions to Pipe A and Pipe B. Pipe B is called the “MAC Pipe”
because it uniquely supports multiply-accumulate, as well as multiply and divide operations. The Dual MAC
unit, which is attached to Pipe B as Custom Engine 0 (CE0), includes the accumulator registers (including HI
and LO) and therefore also supports the move to and move from operations which transfer data between these
registers and the general register file.

Pipe A is called the “Load/Store Pipe” because it uniquely supports the Load and Store operations. DSP
extensions to memory addressing are therefore also unique to Pipe A. These extensions include pointer post-
modification and circular buffer addressing. The Figure illustrates the circular buffer start registers (cbs0-
cbs2) and circular buffer end registers (cbe0-cbe2) located in ALU A.

The Coprocessor operations, and all “sequencing control instructions” (branches, jumps) are unique to Pipe
A. As a result, Pipe B instructions are not routed to Coprocessors.

The opcodes reserved for a customer defined Custom Engine 1 (CE1) are routed to Pipe B, since CE1 is
attached to Pipe B.

All ALU operations are available in both Pipe A and Pipe B. As a result performance is improved,
particularly in computation-intensive programs, and, the design is simplified because major sub-blocks in
ALU A and ALU B are replicated.

The Custom Engine Interface (CEI) is available for customer proprietary operations in Pipe B. This allows
the customer extensions to maintain high throughput since they can dual-issue with Load and Store
instructions which issue to Pipe A.

Table 2: Assignment of Instructions of Pipe A, Pipe B

Pipe A Pipe B

The Load/Store Pipe The MAC Pipe

MIPS 32-bit MIPS 32-bit General Instructions MULT(U), DIV(U), MFHI, MFLO,
General Instructions except: MTHI, MTLO,MAD(U),MSUB(U)
CE1 Custom Engine Opcodes, CE1 Custom Engine Opcodes,
MULT(U), DIV(U), MFHI, MFLO, MIPS 32-bit ALU Instructions
MTHI, MTLO,MAD(U),MSUB(U) Note: No Load or Store
Instructions

MIPS 32-bit J, JAL, JR, JALR, JALX

Control Instructions SYSCALL, BREAK,
All Branch Instructions,
All MFCz, MTCz, SWCz, LWCz

Lexra Proprietary & Confidential - 11 - Release 1.9

 April 30, 2001 LX5280

Pipe A Pipe B

The Load/Store Pipe The MAC Pipe

MIPS16 Instructions All MIPS16 Instructions except: MULT(U), DIV(U), MFHI, MFLO
(No Doubleword
Instructions) MULT(U), DIV(U), MFHI, MFLO

EJTAG Instructions DERET, SDBBP

(including MIPS16 SDBBP)

Lexra Control MTRU, MFRU, MTRK, MFRK,

Instructions MTLXC0,MFLXC0

Lexra Vector LT, ST,

Addressing LTP, LWP,
LHP(U), LBP(U),
STP, SWP,
SHP, SBP

Lexra MAC MTA2, MFA, MFA2, MULTA,

Instructions MULTA2, MULNA2, CMULTA,
MADDA, MSUBA, ADDMA,
SUBMA, DIVA, RNDA2

Lexra Extensions to SLLV2, SRLV2, SRAV2, SLLV2, SRLV2, SRAV2,

MIPS ALU ADDR, ADDR2, SUBR, SUBR2, ADDR, ADDR2, SUBR, SUBR2,
Instructions SLTR2 SLTR2

New Lexra ALU MIN, MIN2, MAX, MAX2, ABSR, MIN, MIN2, MAX, MAX2, ABSR,
Operations ABS2, CLS, MUX2, BITREV, ABS2, CLS, MUX2, BITREV,
CMVEQZ, CMVNEZ CMVEQZ, CMVNEZ

2.5. System Control Coprocessor (CP0)

The System Control Coprocessor (CP0) is responsible for instruction address sequencing and exception
processing.

For normal execution, the next instruction address has several potential sources: the increment of the previous
address, a branch address computed using a pc-relative offset, or a jump target address. For jump addresses,
the absolute target can be included in the instruction, or it can be the contents of a general-purpose register
transferred from the RALU.

Branches are assumed (or predicted) to be taken. In the event of prediction failure, two stall cycles are
incurred and the correct address is selected from a special “backup” register. Statistics from several large
programs suggest that these stalls will degrade average LX5280 throughput by several percent. However, the
net effect of the LX5280’s branch prediction on performance is positive because this technique eliminates
certain critical paths and therefore, permits a higher speed system clock.

If an exception occurs, CP0 selects one of several hardwired vectors for the next instruction address. The
exception vector depends on the mode and specific trap which occurred. This is described further in
Section 3.4, Exception Processing.

The following registers, which are visible to the programming model, are located in CP0:

Lexra Proprietary & Confidential - 12 - Release 1.9

 April 30, 2001 LX5280

Table 3: CP0 Registers

CP0 register Number Function

BADVADDR 8 Holds bad virtual address if address exception error occurs

STATUS 12 Interrupt masks, mode selects

CAUSE 13 Exception cause

EPC 14 Holds address for return after exception handler

PRID 15 Processor ID (read-only) 0x0000c601 for LX5280

CCTL 20 Instruction and data memory control

EPC, STATUS, CAUSE, and BADVADDR are described further in the Section 3.4. PRID is a read-only
register that allows the customer’s software to identify the specific version of the LX5280 that has been
implemented in their product. The CCTL register is a Lexra defined CP0 register used to control the
instruction and data memories, as described in Section 7.2, Cache Control Register: CCTL.

The contents of the above registers can be transferred to and from the RALU’s general-purpose register file
using CP0 operations. (Unlike registers located in Coprocessors 1-3, they cannot be loaded or stored directly
to data memory.)

2.6. Dual Multiliply-Accumulate (MAC)

2.6.1. Dual MAC Operations

The Dual MAC data path is illustrated in Figure 4 on page 16. The major subsystems are:

• Two 16-bit Multiply-Accumulate data paths each with:

• 16-bit x 16-bit Multiplier

• 32-bit Product Register
• Four 40-bit Accumulator Registers with optional saturate
• Output Scalers

• 40-bit Add/Subtract/Dual Round Unit with optional saturate

• One or two 16-bit x 16-bit multiply or multiply-accumulate operations can be initiated

every cycle, with a three cycle latency.

• 32-bit x 32-bit multiply executes on a single Multiply-Accumulate data path with five
cycle latency. By using both data paths, a 32-bit x 32-bit multiply-accumulate can be
initiated every other cycle.

• 32-bit x 32-bit multiply-accumulate executes using a combination of a single Multiply-

Accumulate data path, followed by the Add-Round. The total latency is six cycles. By
using both data paths, a 32-bit x 32-bit multiply-accumulate can be initiated every other
cycle.

• Complex Multiply (16-bit Real, 16-bit Imaginary per-product) uses both Multiply-
Accumulate data paths with three cycle latency. A new Complex Multiply can be initiated
every two cycles.

Lexra Proprietary & Confidential - 13 - Release 1.9

 April 30, 2001 LX5280

• One Divide unit.

2.6.2. MAC MODE (MMD) register

Several new DSP features are controlled using the MMD (“MAC Mode”) register. MMD is a new Radiax
User register (24) which is accessed using Radiax User Move instructions MTRU and MFRU. If MMD is
updated between a MAC instruction and the MFA instruction that retrieves the result of that instruction, the
resulting operation is undefined.

The fields in MMD are as follows. Note that MMD is reset to all zeroes.

MF selects arithmetic mode for multiplies in the Dual MAC:

0: use integer arithmetic mode

1: use fractional arithmetic mode

MS selects saturation boundary in the Dual MAC accumulators:

0: saturate at 40 bits
1: saturate at 32 bits

MT selects truncation of 32x32 multiplies in the Dual MAC:

0: perform full 32x32 multiply (sum all four partial products)

1: omit partial product rS[15:00] x rT[15:00] when performing 32x32
multiply.

RND selects the rounding mode used in the RNDA2 instruction.

00: Convergent Rounding. (Sometimes called “round-to-nearest-even”)

Round to nearest number; when the number to be rounded is midway
between two numbers representable in the smaller format, round to the
even number. The rounded result will always have 0 in the lsb.
Assuming that the lsb left of the roundoff point is random, convergent
rounding is unbiased.

01: Round-to-Nearest
Round to nearest number; when the number to be rounded is midway
between two numbers representable in the smaller format, round to the
more positive number. (This rounding mode is common because it is
easily implemented by always adding 0...0^10...0 to the number to be
rounded. Digits to the right of “^” are dropped after rounding.)

1x: reserved

Lexra Proprietary & Confidential - 14 - Release 1.9

 April 30, 2001 LX5280

MMD (Radiax User Register 24)

• A new Radiax User register (24)

• Accessed with MTRU, MFRU operations

• Reset to 0

31 5 4 3 2 1 0

0 RND MF MS MT

27 2 1 1 1

Table 4: MMD Fields (Radiax User Register 24)

Width
Field Description
(Bits)

RND 2 Rounding mode

00 = convergent
01 = round-to-nearest
10 = reserved
11 = reserved

MF 1 MAC fractional mode

MS 1 MAC 32-bit saturate mode

MT 1 MAC 32x32 truncate mode

2.6.3. Architecture

The Multiply-Accumulate data paths can operate on 16-bit input data, either individually or in parallel. The
same Assembler mnemonic is used for individual or parallel operation. The output register specified
determines whether MAC0 or MAC1 or both, operate. For example,

MADDA2 m0l, r2, r3 // MAC0: m0l ← m0l + r2[15:00] * r3[15:00]

// MAC1: IDLE

MADDA2 m0h, r2, r3 // MAC0: IDLE

// MAC1: m0h ← m0h + r2[31:16] * r3[31:16]

MADDA2 m0, r2, r3 // MAC0: m0l ← m0l + r2[15:00] * r3[15:00]

// MAC1: m0h ← m0h + r2[31:16] * r3[31:16]

Each Multiplier can initiate a new 16-bit x 16-bit product every cycle (single cycle throughput). Each 16-bit x
16-bit multiply-accumulate completes in three cycles. (Figure 4 illustrates the intermediate pipeline registers
TEMP0, TEMP1, Product 0, Product 1 to help the reader remember that the Multipliers require two cycles
but have single cycle throughput. TEMP0, TEMP1, Product 0, Product 1 are not accessible by the
programmer.). Thus, there are two delay slots for Multiplication or Multiply-Accumulate. For example,

Cycle 1: MADDA2 m1h, r2, r3

Cycle 2: delay slot 1 // new m1h is not available
Cycle 3: delay slot 2 // new m1h is not available
Cycle 4: MFA r3, m1h // new m1h is available

Lexra Proprietary & Confidential - 15 - Release 1.9

 April 30, 2001 LX5280

x y

32 32

Multiplier 1 Multiplier 0
16b x 16b 16b x 16b Divider
DIV, DIVU
TEMP1 (32) TEMP0 (32)

32 32

Product 1 (32) Product 0 (32)

40 40

Accumulator 1 Accumulator 0
mh ← p ml ← p
mh ← [sat]mh + {p,ax} ml ← [sat]ml + {p,ax}
mh ← [sat]mh - {p,ax} ml ← [sat]ml - {p,ax}

40 40

ACC m0h(40) ACC m0l(40)

Add /
ACC m1h(40) ACC m1l(40) Subtract/
Dual Round
ACC m2h(40) ACC m2l(40) (w/ saturate)
ACC m3h(40) ACC m3l(40)

40 40

Scaler Scaler
>> [0-8] >> [0-8]
16, 32 16, 32

32

Figure 4: Dual MAC Data Path

The accumulator m1h can be referenced by MFA in Inst2(Inst3), however two (one) stall cycles will be
incurred. It is expected that the number of stall cycles in DSP algorithms will be minimal because, typically
many products are accumulated before the accumulator must be stored. In a 64-tap FIR, for example, 64
terms are accumulated before the filter sample is updated in memory. Also, the four accumulator pairs allow
loops to be “unrolled” so that up to three additional independent MAC operations can be initiated before the
result of the first is available.

Compared to a typical RISC multiply-accumulate unit the LX5280 MAC includes a number of features
critical to high-fidelity DSP arithmetic. These features are optionally selected by opcodes and/or mode bits in
the MMD register, and are compatible with conventional integer arithmetic, also supported by the LX5280:

Lexra Proprietary & Confidential - 16 - Release 1.9

 April 30, 2001 LX5280

• Accumulator guard bits,

• Fractional arithmetic,

• Saturation,

• Rounding,

• Output Scaling.

Accumulation is performed at 40-bit precision, using eight guard bits for overflow protection. The alternative
is to require the programmer to right-shift (scale) products prior to accumulation, which complicates
programming and causes loss of precision. Prior to accumulation, the product is sign-extended to 40-bits.
With guard bits, typically the only loss of precision will occur at the end of a lengthy calculation when the 40-
bit result must be stored to the general register file or to memory in 32-bit or 16-bit format.

Fractional arithmetic is implemented by the program’s interpretation of the 16-, 32- or 40-bit quantities and
is controlled by a bit in the MMD register. When fractional mode is selected, the Dual MAC shifts the results
of any Radiax multiply operation left by one bit to maintain the alignment of the implied radix point.
Furthermore, since -1 can be represented in fractional format but +1 cannot be represented in fractional
mode, the Dual MAC detects when both operands of a multiply are equal to -1. If so, it generates the
approximate product consisting of 0 for the sign bit (representing a positive result) and all ones for the
remaining bits. This is true for both 16x16 bit and 32x32 bit Radiax multiplications. The least significant bit
of a product is always zero in fractional mode (due to the left shift).

The Accumulation Units can add the product to, or subtract it from, one of the four accumulator registers.
This operation can be performed with optional saturation; that is, if a result overflows (underflows), the
accumulator is updated with the largest (smallest) positive (negative) number rather than the “wraparound”
result with incorrect sign. The LX5280 instructions include a Multiply-add and Multiply-sub, each with and
without saturation. There are also instructions for adding or subtracting any pair of 40-bit accumulator
registers together, with and without saturation. A bit in the MMD register determines whether the saturation
is performed on the full 40 bits or whether saturation is performed at 32 bits. The latter capability is useful for
emulating the results of other architectures that do not have guard bits. In 32-bit saturation mode, a full 40- bit
compare is used to determine if the result is greater (less) than the maximum (minimum) value which can be
stored in a 32-bit quantity; this provides the most robust solution.

In the case that the instruction requires multiplication, but no accumulation, the product is passed through the
accumulation unit unchanged. (Thus, both 16-bit multiplication and multiply-accumulate require three MAC
cycles.)

A Round instruction can also be executed on one (or a pair) of the accumulator registers to reduce precision
prior to storage. The rounding mode is selectable in the MMD register.

The output Scaler is used to right shift, (scale), the accumulator register when it is transferred to the general
register file.

The Dual MAC is also used to execute the 32-bit MULT(U) and DIV(U) instructions specified in the MIPS
ISA. In the case of MULT(U), one of the 16-bit Multiply-Accumulate data paths works iteratively to produce
the 64-bit product in five cycles. (The least significant 32 bits are available one cycle earlier than the most
significant 32 bits.)

Note: The MMD mode bits have no effect on the operation of the standard MIPS ISA instructions. By
contrast, the LX5280 MULTA instruction is subject to the MMD mode bits for fractional arithmetic and
truncated 32x32 multiplication.

32-bit x 32-bit Multiply-Accumulate instructions (MADDA, MSUBA) are implemented using one of the 16-

Lexra Proprietary & Confidential - 17 - Release 1.9

 April 30, 2001 LX5280

bit Multiply-Accumulate data paths, and the Add-Round unit. It provides a 64-bit Multiply result which is
sign-extended and accumulated at 72-bits. The result is available in six cycles. (The least significant 32 bits
are available one cycle earlier than the most significant 40 bits.) The MAD(U) and MSUB(U) instructions of
the MIPS32 ISA are also supported.

For the LX5280 MULTA, an accumulator pair M0h[39:0]/M0l[31:0], M1h[39:0]/M1l[31:0] etc. is the target.
M0h[39:0] is aliased to HI; M0l[31:0] is aliased to LO. The most significant 8-bits of the 40-bit HI
accumulator are used as the guard bits, while the LO accumulator is simply zero-extended to 40 bits. Unlike
the (dual) 16-bit operations, single-cycle throughput is not available for 32-bit data. However since there are
two available data paths, two 32-bit x 32-bit multiply operations can be initiated every four cycles. The Dual
MAC hardware automatically allocates the second operation to the available data path. If a third 32-bit
multiplication is programmed too soon, stall cycles are inserted until one of the data paths is free.

The Dual MAC also supports a complex multiply instruction, CMULTA. For this instruction, each of the 32-
bit general register operands is considered to represent a 16-bit real part (in bits 31:16) and a 16-bit imaginary
part (in bits 15:00). One of the multiply-accumulate engines calculates the real part (33 bits) of the complex
product (namely XrYr - XiYi) and stores it in the “h” half of the target accumulator pair. The other MAC
engine calculates the imaginary part (32 bits) of the complex product (namely XrYi + XiYr) and stores it in
the “l” half of the target accumulator pair. This instruction can be initiated every two cycles (2-cycle
throughput) and takes four cycles to complete. As in the other Dual MAC operations, programming
CMULTA instructions too close together causes stall cycles but the correct results are always obtained.

The Dual MAC includes a separate Divide Unit for executing the 32-bit DIV(U) operations specified by the
MIPS ISA. The Divide requires 19 cycles to complete. The quotient is loaded into M0l[31:0], M1l[31:0],
M2l[31:0] or M3l[31:0] and the remainder is loaded into the lower 32-bits of the other accumulator in the
target pair. There is no special support for fractional arithmetic for the divide operations.

2.7. Data Addressing

2.7.1. Twinword Data Movement

Since the Dual MAC is capable of consuming four 16-bit operands every cycle (in Pipe B) by performing two
16x16 multiply-accumulates, it is desirable to be able to fetch four 16-bit operands from memory every cycle
(in Pipe A). Therefore, the LX5280 extends the MIPS load and store instructions to include twinword
accesses and implements a 64-bit data path from memory. A twinword memory operation accesses an (even-
odd) pair of 32-bit general registers with a single instruction and executes in a single pipeline cycle. The
nomenclature “twinword” is used to distinguish these operations from “doubleword” operations which (in
other extensions to the MIPS ISA) access a single 64-bit general register.

Like the standard byte, halfword, and word load/store instructions, the twinword load/store instructions use a
register and an immediate field to specify the memory address. However, in order to obtain the maximum
range from the LEXOP instruction format, the available signed 11-bit immediate field (called the
displacement) is considered a twinword quantity, so is left-shifted by 3 bits before being added to the base
register. This is equivalent to a 14-bit byte offset, in comparison to the full 16-bit immediate byte offset used
in the byte, halfword and word instructions. Also, the target register pair for the twinword load/store must be
an even-odd pair, so that only 4 bits are used to specify it.

2.7.2. Vector Addressing

DSP algorithms usually operate on vectors or matrices of data; for example Discrete Cosine Transforms
operate on 8x8 pixel blocks. As a result data memory pointers are incremented from one operand to the next.
The extra instruction cycle required to increment RISC memory pointers is eliminated in DSPs with auto-
increment. This capability is provided in the LX5280. Memory pointers are used unmodified to create the

Lexra Proprietary & Confidential - 18 - Release 1.9

 April 30, 2001 LX5280

address, then updated in the general register file before the next use:

address ← pointer
pointer ← pointer + stride

In the LX5280 the 8-bit immediate field containing the stride is sign-extended to 32-bits before being added
to the pointer for the latter’s update. The nomenclature “pointer” is used to distinguish the update performed
after memory addressing from the “offset”, in which the “base” register (in the MIPS ISA) which is
augmented by the offset before addressing memory in the standard instructions. The nomenclature “stride”,
which is dependent on the granularity of the access, is used to distinguish it from the invariant byte offset
used in the standard load and store instructions. For twinword/word/halfword addressing the 8-bit field is first
left-shifted by three/two/one places and zero-filled, before sign extension to 32-bits. This use of left shifts for
the twinword, word, and halfword word and halfword strides is similar to MIPS16 and is used to extend the
effective address range. Thus, increments of between -128 and +127 twinwords1, words, halfwords or bytes
are available for each data type.

In the case of Loads (but not Stores) pointer update requires a second general register file write port. The
LX5280 includes an 8p(4r/4w) register file with two of the four write ports dedicated to register Loads. As a
result, twinword loads can execute in parallel with any Pipe B operation.

For some DSP algorithms - notably Filters - DSP data is organized into “circular buffers”. In this case, at the
end of the buffer the next reference is to the beginning of the buffer. Implementing this structure in RISC
requires:

Inst 1: LW reg, AddressReg

Inst 2: BNEL AddressReg, BufferEnd, Continue
Inst 3: ADDIU AddressReg, AddressREG + 4
Inst 4: MOVE AddressReg, BufferStart
Continue:

Note that the above example is written so that a branch prediction failure will only be incurred at the end of
the buffer. Nevertheless, the combination of post-modified pointers together with hardware support for
circular buffers in the LX5280 allows this typical DSP addressing operation to be reduced from four cycles to
one.

1. Twinwords are supported only on the LX5280, and not the LX5180

Lexra Proprietary & Confidential - 19 - Release 1.9

 April 30, 2001 LX5280

for pointer update

ALU REG A REG BI REG BR REG

TEMP
ALU A
DBUS_M

A + BI A REG

DADDR_E
M
U
X

select
compare

cbs0 cbe0
3 3
cbs1 cbe1
enable cbs2 enable cbe2
CB[2:0] CB[2:0]

Figure 5: Post-modified Pointers with Circular Buffer Support

2.7.3. Circular Buffers

The LX5280 supports three circular buffers. To initialize the circular buffers, the MTRU instructions are used
to set the twinword start addresses CBS0-CBS2[31:3] and twinword end addresses CBE0-CBE2[31:3].
Circular buffers are only used when memory pointers are post-modified, and consist of an integral number of
twinwords.

When a circular buffer pointer is used in a post-modified address calculation, the pointer is compared to the
associated CBE address; if they match (and the stride is non-negative), the CBS address (rather than the post-
modified address) is restored to the register file. Similarly, to allow for traversing the circular buffer in the
reverse direction, the pointer is compared to the CBS address; if they match (and the stride is negative) the
CBE address (rather than the post-modified address) is restored to the register file.

It is worth noting that circular buffers can also be accessed with byte, halfword, or word Load/Store with
Pointer Increment instructions. In those cases, the several least significant bits of the pointer register are
examined to determine if the start or end of the buffer has been reached, taking into account the granularity of
the access, before replacing the pointer with the CBS or CBE as appropriate.

Any general register memory pointer can be used with circular buffers using the “.Cn” option. To use general
register rP as a circular buffer pointer. For example, the instruction

LWP.C2 r3, (r4)stride

Lexra Proprietary & Confidential - 20 - Release 1.9

 April 30, 2001 LX5280

associates the r4 memory pointer with circular buffer C2 which is defined by the start address CBS2 and end
address CBE2.

2.8. Radiax ALU Operations

The LX5280 introduces extensions to the MIPS instructions to support dual 16-bit operations. The LX5280
also introduces a number of new ALU instructions which improve performance on DSP algorithms. These
instructions will also be described in this section.

2.8.1. Extensions to MIPS ALU Operations

To support high-performance dual 16-bit operations in the RISC-DSP it is necessary to support not only Dual
MAC instructions but also dual 16-bit versions of other arithmetic operations that the programmer may
require. To maintain a simple, orthogonal instruction set, the following criteria were used to determine the
MIPS ALU extensions:

• Dual 16-bit versions of all MIPS ALU operations without immediate data,

• Optional saturation for every ALU instruction (without immediate data) that can produce
signed overflow or underflow.

It is expected that the above organizing principles will simplify the LX5280 ISA for both programmers and
tool developers. Obviously, dual 16-bit versions of logical operations such as AND are not required.
However, dual 16-bit versions have been provided for all 3-register operand shifts and add/subtracts included
in the MIPS R-Format. The character “2” in the assembler mnemonic indicates an operation on dual 16-bit
data.

DSP algorithms are often somewhat tolerant of data errors. For example, a bad audio sample may cause a
brief distortion, but no lasting effect as new audio samples arrive and the bad sample is cleared out of the
buffer. Accordingly, the saturated result of signed arithmetic is a closer, more desirable, approximation than
the wraparound result. Therefore, all LX5280 arithmetic operations which may, potentially, produce
arithmetic overflow or underflow, and do not have immediate operands, support optional saturation. For
example, not only the dual 16-bit add (ADDR2) but also the 32-bit add (ADDR) have optional saturate in the
LX5280. Saturation options are not provided for MIPS I-Format 32-bit instructions; for example ADDIU.
However, in this case the programmer selects the immediate operand and, as a result, saturation is less likely,
or at least more predictable.

Neither the dual 16-bit instructions nor the new 32-bit saturating adds and subtracts cause exceptions.

2.8.2. New ALU Instructions

The LX5280 adds several new ALU instructions which have proven useful in DSP performance analysis.
Consistent with the approach described above, each new instruction has both a 32-bit and a dual 16-bit
version. If signed overflow/underflow is possible, a saturation option is provided.

2.8.3. Conditional Move Operations

The LX5280 includes new instructions (MOVZ and MOVN) to support conditional operations. These
instructions are described in this section.

A number of DSPs and RISC processors have deployed extensive “conditional execution.” In these
processors the branch prediction penalty is three cycles or more. Conditional execution can mitigate the effect
of the branch prediction penalty by allowing the branch to be avoided in some cases. However, conditional
execution is a costly alternative: it uses instruction opcode bits and consequently limits the size of immediate

Lexra Proprietary & Confidential - 21 - Release 1.9

 April 30, 2001 LX5280

data and/or limits the number of general purpose registers visible to the program. The LX5280 branch
prediction penalty is only two cycles; therefore the need for conditional execution is minimized and only a
restricted set of “conditional move” instructions is needed. It is notable, however, that the effect of any
conditional execution can be “emulated” in the LX5280 with a sequence of two instructions by using the
conditional move. For example:

Processor with conditional execution:

Inst 1: ALU operation sets condition flags

Inst 2: COND: ALU operation

LX5280:

Inst 1: ALU operation updates register rB (condition setting operation)

Inst 2: ALU operation with result directed to temp register rA
Inst 3: MOV<COND> rD, rA, rB

If rB satisfies the COND, rD is updated with rA; i.e. the 2nd ALU operation is executed to “completion”.
Note that this sequence is interruptible.

Another use of the conditional move instructions is to code “if-then-else” constructs as follows:

if (rB COND)
rD = rA
else
rD = rC

can be coded if the previous example is prefaced with:

MOVE rD, rC // move rC to rD

One reason Lexra has provided conditional move is to facilitate initial porting of Assembler code from
processors with conditional execution to the LX5280.

2.9. Zero Overhead Loop Facility

Because DSP algorithms spend much of their time in short real-time critical code loops, DSPs often include
hardware support for “zero-overhead looping.” The goal of zero-overhead looping is that branching from the
end-to-beginning of the loop can be accomplished without explicit program overhead if the loop is to be
executed a fixed number of times, known at compile time.

The LX5280 supplies such a facility but allows the loop count to be determined at run time as well. The
facility consists of three new Radiax registers, which are accessible by a program running in User mode using
the Radiax instructions MFRU and MTRU. The operating system should consider these registers as part of
the context of the executing process and must save and restore them in the case of an interrupt.

LPE0[31:2] —virtual address of the ending instruction of the loop

LPS0[28:2] —low order bits of the virtual address of the “starting” instruction of the loop.

LPC0[15:0] —the loop count.

Although the facility is intended for use in loops, the algorithm executed by the hardware can be described
more simply. In particular it should be noted that there is no “knowledge” of being “inside” the loop. All that
matters is the contents of the three registers when an attempt is made to execute the instruction at the address

Lexra Proprietary & Confidential - 22 - Release 1.9

 April 30, 2001 LX5280

specified by LPE0:

if (M32-mode, AND current-instruct-addr[31:2] = LPE0, AND LPC0 ≠ 0) then

execute current instruction (at LPE0[31:2] || 00),
decrement LPC0[15:0] by one,
execute instruction at LPE0[31:29] || LPS0[28:2] || 00
continue (LPS0 could be a jump/branch)
else
execute current instruction,
continue (current instruction could be a jump/branch)

The following restrictions apply to the usage of the Zero Overhead Loop Facility:

• It is only active in 32-bit ISA mode. It is disabled in MIPS16 mode.

• LPS may not be exactly equal to LPE if LPC is non-zero. Therefore, the loop must contain
at least two instructions. Otherwise, operation is undefined.

• LPE may not be in the delay slot of a branch, nor may it be a branch or jump instruction
itself if LPC is non-zero. Otherwise, operation is undefined.

• For correct operation, the order of loading the registers must be: first LPS, then LPE, then
LPC with a non-zero value.

• For correct operation, there must be at least two (2) instructions between the instruction
which loads LPC with a non-zero value, and the instruction at the LPE address. To
guarantee that no stall cycles are incurred, there must be at least three (3) cycles between
the instruction which loads LPC with a non-zero value, and the instruction at the LPE
address.1

• If the instruction at LPE is a load type instruction, then the immediately executed
instruction at LPS is considered to be in the load delay slot and cannot rely on seeing the
result of the load.

The following items are not restrictions that apply to the usage of the Zero Overhead Loop Facility but are
features to be aware of:

• The loop count LPC may be reloaded multiple times after LPS and LPE are loaded.
Typically this would be done in an outer loop.

• The instruction at LPE may be the target of a jump or branch, including a change in mode
from 16-bit to 32-bit ISA.

• Any of the instructions before or at LPE may be subject to exceptions or interrupts and
processing will conform to the normal exception handling rules. Note that the BD bit will

1. The following discussion is only relevant if LPC will be updated in an instruction that is “close” to LPE. That case can have a per-
formance impact although correct operation will still be obtained. The programming guideline is: keep the LPC update in an outer
loop as far as possible from the (end of) the inner loop:

The updates of LPS, LPE, and LPC use the MTRU instruction. Therefore the new LPS, LPE, and LPC values are only known after
the E-stage of the pipeline. But in order to perform the pseudo-branch they must be used in the I-stage of the pipeline. Because of
the restriction on the order of setting these registers, the hardware introduces a minimum number of stalls after setting LPS and LPE
to test for an LPE match against the current instruction address. However, if the LPC update is still in the pipeline when the LPE
match is detected, the hardware stalls to check and update the new value of LPC. To avoid these stalls, LPC should not be updated
within 3 cycles (which could be as many as 6 instruction issue slots) of an expected LPE-matching instruction. As noted above, for
correct operation there must be at least 2 instructions between the LPC update and any expected LPE-matching instruction.

Lexra Proprietary & Confidential - 23 - Release 1.9

 April 30, 2001 LX5280

always be off since LPE must not be in the delay slot of a branch. The return from the
exception handler to LPE will also be handled normally, since it is just a special case of
LPE being the target of a jump.

• If the instruction at LPE causes a Reserved Instruction Trap, it is necessary for the
Exception Handler to decrement LPC prior to return, after emulating the instruction at
LPE and before returning to the instruction at LPS. Similar restrictions apply if the
instruction at LPE is not to be re-executed for any other reason, such as BREAK or
SYSCALL execution.

2.10. Low-Overhead Prioritized Interrupts

The LX5280 includes eight new low-overhead hardware interrupt signals. These signals are compatible with
the R3000 Exception Processing model and are useful for real-time applications.

These interrupts are supported with three new Lexra CP0 registers, ESTATUS, ECAUSE, and INTVEC,
accessed with the new MTLXC0 and MFLXC0 variants of the MTC0 and MFC0 instructions. As with any
COP0 instruction, a Coprocessor Unusable Exception is taken if these instructions are executed while in User
Mode and the Cu0 bit is 0 in the CP0 STATUS register.

The three new Lexra CP0 registers are ESTATUS (0), ECAUSE (1), and INTVEC (2), and are defined as
follows:

ESTATUS (LX COP0 Reg 0) Read/Write

31 - 24 23 - 16 15 - 0

0 IM[15:8] 0

ECAUSE (LX COP0 Reg 1) Read-only

31 - 24 23 - 16 15 - 0

0 IP[15:8] 0

INTVEC (LX COP0 Reg 2) Read/Write

31 - 6 5-0

BASE 0

ESTATUS contains the new interrupt mask bits IM[15:8], which are reset to 0 so that none of the new
interrupts will be activated, regardless of the global interrupt signal IEc. IP[15:8] for the new interrupt signals
is located in ECAUSE and is read-only. These fields are similar to the IM and IP fields defined in the R3000
Exception Processing Model, except that the new interrupts are prioritized in hardware, and each have a
dedicated exception vector.

IP[15] has the highest priority, while IP[8] has the lowest priority, however, all new interrupts are higher
priority than IP[7:0]. The processor concatenates the program defined BASE address for the exception
vectors with the interrupt number for form the interrupt vector, as shown in the table below. Two instructions
can be executed in each vector; typically these will consist of a jump instruction and its delay slot, with the
target of the jump being either a shared interrupt handler or one that is unique to that particular interrupt.

Lexra Proprietary & Confidential - 24 - Release 1.9

 April 30, 2001 LX5280

Table 5: Prioritized Interrupt Exception Vectors

Interrupt Number Exception Vector

15 { BASE, 6’b111000 }

14 { BASE, 6’b110000 }

13 { BASE, 6’b101000 }

12 { BASE, 6’b100000 }

11 { BASE, 6’b011000 }

10 { BASE, 6’b010000 }

9 { BASE, 6’b001000 }

8 { BASE, 6’b000000 }

When a vectored interrupt causes an exception, all of the standard actions for an exception occur. These
include updating the EPC register and certain subfields of the standard STATUS and CAUSE registers. In
particular, the Exception Code of the CAUSE register indicates “Interrupt”, and the “current” and “previous”
mode bits of the STATUS register are updated in the usual manner.

Lexra Proprietary & Confidential - 25 - Release 1.9

 April 30, 2001 LX5280

Lexra Proprietary & Confidential - 26 - Release 1.9

 April 30, 2001 LX5280

3. LX5280 RISC Programming Model

This section describes the LX5280 Programming Model. Section 3.1, Summary of MIPS-I Instructions,
contains a list summarizing all MIPS-I operations supported by the LX5280. These opcodes may be extended
by the customer using Lexra’s Custom Engine Interface (CEI). This capability is described in Section 3.2,
Opcode Extension Using the Custom Engine Interface (CEI).

Section 3.3, Memory Management, describes the Simplified Memory Management Unit (SMMU) which is
physically incorporated in the LX5280 LMI. The SMMU provides sufficient memory management
capabilities for most embedded applications while ensuring execution of third-party MIPS software
development tools.

The LX5280 supports the MIPS R3000 Exception Processing model, as described in Section 3.4, Exception
Processing.

The LX5280 supports all MIPS-I Coprocessor operations. The customer can include one to three application-
specific Coprocessors. Lexra provides a functional block called the Coprocessor Interface (CI) which allows
the customer a simplified connection between their Coprocessor and the internal signals of the LX5280. The
CI is described in Section 3.5, The Coprocessor Interface (CI).

3.1. Summary of MIPS-I Instructions

The LX5280 executes MIPS-I instructions as detailed in the tables below. To summarize, the LX5280
executes MIPS-I instructions with the following exclusions: the unaligned loads and stores (LWL, SWL,
LWR, SWR) are not supported because they add significant silicon area for little benefit in most applications.

The following conventions are employed in the instruction descriptions.

«» Encloses a list of syntax choices, from which one must be chosen.

{ } Encloses a list of values that are concatented to form a larger value.
n { value } Replicates (concatenates) a value n times.
value[3] Bits selected from a value.
[rA + offset] Memory address computation and corresponding memory contents.
4’b0000 A sized constant binary value.
32’h1234_5678 A sized constant hexadecimal value.
expr ? A : B Select A if expr is true, otherwise select B.

Lexra Proprietary & Confidential - 27 - Release 1.9

 April 30, 2001 LX5280

3.1.1. ALU Instructions

Table 6: ALU Instructions

Instruction Description

ADD rD, rA, rB rD <- rA + «rB, immediate»

ADDU rD, rA, rB Add reg rA to either reg rB or a 16-bit immediate sign-
ADDI rD, rA, immediate extended to 32 bits. Result is stored in reg rD. ADD and ADDI
ADDIU rD, rA, immediate can generate overflow trap; ADDU and ADDIU do not.

SUB rD, rA, rB rD <- rA - rB

SUBU rD, rA, rB Subtract reg rB from reg rA. Result is stored in register rD.
SUB can generate overflow trap. SUBU does not.

AND rD, rA, rB rD <- rA & «rB, immediate»

ANDI rD, rA, immediate Logical and of reg rA with either reg rB or a 16-bit immediate
zero-extended to 32 bits. Result is stored in reg rD.

OR rD, rA, rB rD <- rA | «rB, immediate»

ORI rD, rA, immediate Logical or of reg rA with either reg rB or a 16-bit immediate
zero-extended to 32 bits. Result is stored in reg rD.

XOR rD, rA, rB rD <- rA ^ «rB, immediate»

XORI rD, rA, immediate Logical xor of reg rA with either reg rB or a 16-bit immediate
zero-extended to 32 bits. Result is stored in reg rD.

NOR rD, rA, rB rD <- ~(rA | rB)

Logical nor of reg rA with either reg rB or a zero-extended 16-
bit immediate. Result is stored in reg rD.

LUI rD, immediate rD <- {immediate, 16’b0}

The 16-bit immediate is stored into the upper half of reg rD.
The lower half is loaded with zeroes.

SLL rD, rB, immediate rD <- rB << «rA, immediate»

SLLV rD, rB, rA Reg rB is left-shifted by 0-31. The shift amount is either the 5b
immediate of the 5 lsb of rA. Result is store in reg rD.

SRL rD, rB, immediate rD <- rB >> «rA, immediate»

SRLV rD, rB, rA Reg rB is right-shifted by 0-31. The unsigned shift amount is
either the 5b immediate or the 5 lsb of rA. Result is stored in
reg rD.

SRA rD, rB, immediate rD <- rB >>(a) «rA, immediate»

SRAV rD, rB, rA Reg rB is arithmetic right-shifted by 0-31. The unsigned shift
amount is either the 5b immediate or the 5 lsb of rA. Result is
stored in reg rD.

SLT rD, rA, rB rD <- (rA < «rB, immediate») ? 1 : 0

SLTU rD, rA, rB If reg rA is less than «rB, immediate» set rD to 1, else 0. The
SLTI rD, rA, immediate 16-bit immediate is sign extended. For SLT, SLTI, the compari-
SLTIU rD, rA, immediate son is signed; for SLU, SLTIU, the comparison is unsigned.

Lexra Proprietary & Confidential - 28 - Release 1.9

 April 30, 2001 LX5280

3.1.2. Load and Store Instructions

Table 7: Load and Store Instructions

Instruction Description

LB rD, offset(rA) rD <- Memory[rA + offset]

LBU rD, offset(rA) Reg rD is loaded from data memory. The memory address is
LH rD, offset(rA) computed as base + offset, where the base is reg rA and the
LHU rD, offset(rA) offset is the 16-bit offset sign-extended to 32 bits.
LW rD, offset(rA) LB, LBU addresses are interpreted as byte addresses to data
memory; LH, LHU as halfword (16-bit) addresses; LW as word
(32-bit) addresses.
The data fetched in LB, LH (LBU, LHU) is sign-extended (zero-
extended) to 32-bits for storage to reg rD.
rD cannot be referenced in the instruction following a load
instruction.

SB rB, offset(rA) rB -> Memory[rA + offset]

SH rB, offset(rA) Reg rB is stored to data memory. The memory address is
SW rB, offset(rA) computed as base + offset, where the base is reg rA and the
offset is the 16-bit offset sign-extended to 32 bits.
SB addresses are interpreted as byte addresses to data mem-
ory; the 8 lsb of rB are stored. SH addresses are interpreted
as halfword addresses to data memory; the 16 lsb of rB are
stored.

3.1.3. Conditional Move Instructions

Table 8: Conditional Move Instructions

Instruction Description

MOVZ rD, rS, rT rD <- (rT== 0) ? rS : rD

If the contents of general register rT are equal to 0, the general
register rD is updated with rS; otherwise rD is unchanged.

MOVN rD, rS, rT rD <- (rT != 0) ? rS : rD

If the contents of general register rT are not equal to 0, the gen-
eral register rD is updated with rS; otherwise rD is unchanged.

Lexra Proprietary & Confidential - 29 - Release 1.9

 April 30, 2001 LX5280

3.1.4. Branch and Jump Instructions

Table 9: Branch and Jump Instructions

Instruction Description

BEQ rA, rB, destination if COND

BNE rA, rB, destination pc <- pc + 4 + { 14 { (destination[15] }, destination, 2’b00 }
else
pc <- pc + 8
where COND = (rA = rB) for EQ, (rA ne rB) for NE, and desti-
nation is a 16-bit value.
For BEQ, BNE the instruction after the branch (delay slot) is
always executed.

BLEZ rA, destination if COND

BGTZ rA, destination pc <- pc + 4 + { 14 {destination[15] }, destination, 2’b00 }
else
pc <- pc + 8
where COND = (rA <= 0) for LE, (rA > 0) for GT, and destina-
tion is a 16-bit value
For BLEZ, BGTZ the instruction after the branch (delay slot) is
always executed.

BLTZ rA, destination if COND

BGEZ rA, destination pc <- pc + 4 + { 14 { destination[15] }, destination, 2’b00 }
else
pc <- pc + 8
where COND = (rA < 0) for LT, (rA >= 0) for GE, and destina-
tion is a 16-bit value
For BLTZ, BGEZ the instruction after the branch (delay slot) is
always executed.

BLTZAL rA, destination Similar to the BLTZ and BGEZ except that the address of the
BGEZAL rA, destination instruction following the delay slot is saved in r31 (regardless
of whether the branch is taken.)

J target pc <- { pc[31:28], target, 2’b00 }

target is a 26-bit absolute. The instruction following J (delay
slot) is always executed.

JAL target Same as above except that the address of the instruction fol-
lowing the delay slot is saved in r31.

JR rA pc <- (rA)
The instruction following JR (delay slot) is always executed.

JALR rA, rD Same as above except that the address of the instruction fol-
lowing the delay slot is saved in rD.

Lexra Proprietary & Confidential - 30 - Release 1.9

 April 30, 2001 LX5280

3.1.5. Control Instructions

Table 10: Control Instructions

Instruction Description

SYSCALL The Sys Trap occurs when SYSCALL is executed.

BREAK The Bp Trap occurs when BREAK is executed.

RFE Causes the KU/IE stack to be popped. Used when returning

from the exception handler. See “Exception Processing”
below.

SLEEP Initiates low-power standby mode. This is a Lexra specific

operation (LEXOP). See Section 3.6, Power Savings Mode.

3.1.6. Coprocessor Instructions

Table 11: Coprocessor Instructions

Instruction Description

LWCz rCGEN, offset(rA) rCGEN <- Memory[rA + offset]

Coprocessor z general reg rCGEN is loaded from data mem-
ory. The memory address is computed as base + offset,
where the base is reg rA and the offset is the 16-bit offset
sign-extended to 32 bits.
rCGEN cannot be referenced in the following instruction (one
cycle delay).

SWCz rCGEN, offset(rA) rCGEN <- Memory[rA + offset]

Coprocessor z general reg rCGEN is stored to data memory.
The memory address is computed as base + offset, where
the base is reg rA and the offset is the16-bit offset sign-
extended to 32 bits.

MTCz rB, rCGEN In MTCz(CTCz), the general register rB is moved to Copro-

CTCz rB, rCCON cessor z general (control) reg rCGEN(rCCON).
rCGEN and rCCON cannot be referenced in the following
instruction.

MFCz rB, rCGEN In MFCz(CFCz), the Coprocessor z general (control) reg

CFCz rB, rCCON rCGEN(rCCON) is moved to the general register rB.
rB cannot be referenced in the following instruction.

BCzT destination if COND

BCzF destination pc <- pc + 4 + { 14’ { destination[15] } , destination, 2’b00 }
else
pc <- pc + 8
where COND = (CpCondz = True) for BCzT, (CpCondz =
False) for BCzF.
For BCzT, BCzF the instruction after the branch (delay slot) is
always executed.

Lexra Proprietary & Confidential - 31 - Release 1.9

 April 30, 2001 LX5280

3.2. Opcode Extension Using the Custom Engine Interface (CEI)

3.2.1. CEI Operations

Customers may add proprietary or application-specific opcodes to their LX5280 based products using the
Custom Engine Interface (CEI). The new instructions take one of the following forms illustrated below and
use reserved opcodes.

Table 12: Custom Engine Interface Operations

New Instruction Description Available Opcodes

NEWOPI rD, rA, immed rD <- rA NEWOPI immed INST[31:26] = 24 - 27

Reg rA is supplied to the SRC1 port of
CEI and the 16-bit immediate, sign-
extended to 32-bits is supplied to
SRC2.
The result of the customer’s NEWOPI
is placed on the CEI input port RES
and stored in reg rD.

NEWOPR rD, rA, rB rD <- rA NEWOPR rB INST[31:26] = 0 and

Reg rA is supplied to the SRC1 port of INST[5:0] = 56,58-
CEI and reg rB is supplied to SRC2. 60,62-63
The result of the customer’s NEWOPI
is placed on the CEI input port RES
and stored in reg rD.

Lexra permits customer operations to be added using the four (4) I-Format opcodes and six (6) R-Format
opcodes listed in the table above. Other opcode extensions in future Lexra products will not utilize the
opcodes reserved above.

When the CEI decodes NEWOPI or NEWOPR, it must signal the Core that a custom operation has been
executed so that the Reserved Instruction trap will not be taken. Multi-cycle custom operations may be
executed by asserting CESEL.

Note: The custom operation may choose to ignore the SRC1 and SRC2 operands supplied by the CEI and
reference customer registers instead. Results can also be written to an implicit customer register; however,
unless D = 0 is coded, a register in the Core will also be written.

3.2.2. Interface Signals

Table 13: Custom Engine Interface Signals

Signal I/O Description

SRC1[31:0] output Operand supplied to customer logic.

SRC2[31:0] output Operand supplied to customer logic.

RES[31:0] input Result of customer logic. Supplied to Core.

CEIOP[11:0] output Instruction OP and SUBOP fields – to be decoded by

customer logic.

Lexra Proprietary & Confidential - 32 - Release 1.9

 April 30, 2001 LX5280

Signal I/O Description

CEHALT input Indicates that a multi-cycle custom operation is in

progress.

CESEL input Indicates that a CEI operation has been decoded.

3.3. Memory Management

The LX5280 includes a Simplified Memory Management Unit (SMMU) for the instruction memory address
and the data memory address. These units are physically located in the Local Memory Interface (LMI)
modules. The hardwired virtual-to-physical address mapping performed by the SMMU is sufficient to ensure
execution of third-party software development tools.

Table 14: SMMU Address Mapping

Virtual Address Space Description Mapped to Physical Address

0xFF00_0000 to EJTAG address space. 0xFF00_0000 to 0xFFFF_FFFF

0xFFFF_FFFF 16 Mbyte. Uncached.
This address range is
reserved for EJTAG use
only.

0xC000_0000 to KSEG2. 1Gbyte (minus 0xC000_0000 to 0xFEFF_FFFF

0xFEFF_FFFF 16 Mbyte). Addressable
only in kernel mode.
Cached.

0xA000_0000 to KSEG1. 0.5 Gbyte. 0x0000_0000 to 0x1FFF_FFFF

0xBFFF_FFFF Addressable only in ker-
nel mode. Uncached.
Used for I/O devices.

0x8000_0000 to KSEG0. 0.5 Gbyte. 0x0000_0000 to 0x1FFF_FFFF

0x9FFF_FFFF Addressable only in ker- (differentiated from KSEG1
nel mode. Cached. addresses with an internal signal)

0x0000_0000 to KUSEG. 2Gbyte. 0x4000_0000 to 0xBFFF_FFFF

0x7FFF_FFFF Addressable in kernel or
user mode. Cached.

3.4. Exception Processing

The LX5280 implements the MIPS R3000 exception processing model as described below. Features specific
to on-chip TLB support are not included. In the discussion below, the term exception refers to both traps,
which are non-maskable program synchronous events, and interrupts, which result from unmasked
asynchronous events.

The list below is numbered from highest to lowest priority. ExcCode is stored in CAUSE when an exception
is taken. Note that Sys, Bp, RI, CpU can share the same priority level because only one can occur in a
particular time slot.

Lexra Proprietary & Confidential - 33 - Release 1.9

 April 30, 2001 LX5280

Table 15: List of Exceptions

Exception Priority ExcCode Description

Reset 1 -- Reset trap.

AdEL – 2 4 Address exception trap. Instruction

instruction fetch. Occurs if the instruction address
is not word-aligned or if a kernel
address is referenced in user mode.

Ov 3 12 Arithmetic overflow trap. Can occur as a

result of signed add or subtract opera-
tions.

Sys 4 8 SYSCALL instruction trap. Occurs

when SYSCALL instruction is executed.

Bp 4 9 BREAK instruction trap. Occurs when

BREAK instruction is executed.

RI 4 10 Reserved instruction trap. Occurs when

a reserved opcode is fetched. Reserved
opcodes are listed below.

CpU 4 11 Coprocessor Usability trap. Occurs

when an attempt is made to execute a
Coprocessor n operation and Copro-
cessor n is not enabled.

AdEL – data 5 4 Address exception trap. Data fetch.

Occurs if the data address is not prop-
erly aligned or if a kernel address is
generated in user mode.

AdES 6 5 Address exception trap. Data store.

Occurs if the data address is not prop-
erly aligned or if a kernel address is
generated in user mode.

Int 7 0 Unmasked interrupt. There are six (6)

level-sensitive hardware interrupt
request signals into the LX5280 Core.
Each is synchronized by the Core to the
LX5280 system clock. In addition, pro-
gram writes to CAUSE[9:8] are soft-
ware-initiated interrupt requests. Each
of the eight (8) requests has an associ-
ated mask bit in STATUS. Int is gener-
ated by any unmasked request (when
Interrupts are globally enabled).

Lexra Proprietary & Confidential - 34 - Release 1.9

 April 30, 2001 LX5280

3.4.1. Exception Processing Registers

STATUS: Coprocessor 0 General Register Address = 12

31-28 27-23 22 21-16 15-8 7-6 5 4 3 2 1 0

CU(3:0) 0 BEV 0 IM(7:0) 0 KUo IEo KUp IEp KUc IEc

CU CU[n] = 1(0) indicates that Coprocessor n is usable(unusable) in Coprocessor instructions.

BEV Bootstrap Exception Vector. Selects between two trap vectors. (see below)

IM Interrupt masks for the six hardware interrupts and two software interrupts.

KU/IE KU = 0(1) indicates kernel (user) mode. In the LX5280, user mode virtual addresses must have
msb = 0. In kernel mode, the full address space is addressable. IE = 1(0) indicates that
interrupts are enabled (disabled).
The KUo, IEo, KUp, IEp, KUc and IEc fields form a three-level stack hardware stack KU/IE
signals. The current values are KUc/IEc, the previous values are KUp/IEp, and the old values
(those before previous) are KUo/IEo. (See Section 3.4.2.)

STATUS is read or written using MTC0 and MTF0 operations. On reset, BEV = 1, KUc = IEc = 0. The other
bits in STATUS are undefined. The 0 fields are ignored on write and are 0 on read. It is recommended that the
user explicitly write them to 0 to insure compatibility with future versions of the LX5280.

CAUSE: Coprocessor 0 General Register Address = 13

31 30 29-28 27-16 15-8 7 6-2 1-0

BD 0 CE(1:0) 0 IP(7:0) 0 ExcCode(4:0) 0

BD Branch Delay. Indicates that the exception was taken in a branch or jump delay slot.

CE Coprocessor Exception. In the case of a Coprocessor Usability exception, indicates the number
of the responsible Coprocessor.

IP Interrupt Pending. Each bit in IP(7:0) indicated an associated unmasked interrupt request.

ExcCode The ExcCode listed above for the different exceptions are stored here when as exception
occurs.

CAUSE is read or written using MTC0 and MTF0 operations. The only program writable bits in CAUSE are
IP(1:0), which are called software interrupts. CAUSE is undefined at reset. The 0 fields are ignored on write
and are 0 on read.

EPC: Coprocessor 0 General Register Address = 14

EPC is a 32-bit read-only register which contains the virtual address of the next instruction to be executed
following return from the exception handler. If the exception occurs in the delay slot of a branch, EPC will
hold the address of the branch instruction and BD will be set in CAUSE. The branch will typically be re-
executed following the exception handler.

BADVADDR: Coprocessor 0 General Register Address = 8

BADVADDR is a 32-bit read-only register containing the virtual address (instruction or data) which

Lexra Proprietary & Confidential - 35 - Release 1.9

 April 30, 2001 LX5280

generated an AdEL or AdES exception error.

3.4.2. Exception Processing: Entry and Exit

When an exception occurs, the instruction address changes to one of the following locations:

RESET 0xbfc0_0000

Other exceptions, BEV = 0 0x8000_0080

Other exceptions, BEV = 1 0xbfc0_0180

The KU/IE stack is pushed:

{ KUo, IEo, KUp, IEp, KUc, IEc } (before push)

{ KUp, IEp, KUc, IEc, 0, 0 } (after push)

which disables interrupts and puts the program in kernel mode. The code (ExcCode) for the exception source
is loaded into CAUSE so that the application-specific exception handler can determine the appropriate action.
The exception handler should not re-enable Interrupts until necessary context has been saved.

To return from the exception, the exception handler first moves EPC to a general register using MFC0,
followed by a JR operation. RFE only pops the KU/IE stack:

{ KUp, IEp, KUc, IEc, 0, 0 } (before pop)

{ KUp, IEp, KUp, IEp, KUc, IEc } (after pop)

(This example assumes that KU/IE were not modified by the exception handler). Therefore, a typical
sequence of operations to return from the exception handler would be:

MFC0 EPC, r26 // r26 is a temporary storage register in the RALU

. . .
JR r26
RFE

3.5. The Coprocessor Interface (CI)

Designers may implement up to three Coprocessors to interface with the LX5280. The contents of these
Coprocessors may include up to thirty-two (32) 32-bit general registers and up to thirty-two (32) 32-bit
control registers. The general registers may be moved to and from the RALU’s registers using MTCz, MFCz
operations, or be loaded and stored from data memory using LWCz, SWCz operations. The control registers
may only be moved to and from the RALU’s registers using CTCz, CFCz operations.

Lexra supplies a simple Coprocessor Interface (CI) model allowing the customer to easily interface a
Coprocessor to the LX5280. The CI supplies a set of control, address, and data busses that may be tied
directly to the Coprocessor general and special registers.

The CI is described in more detail in Section 9, LX5280 Coprocessor Interface.

3.6. Power Savings Mode

The operating system kernel can initiate a power savings standby mode using the Lexra specific SLEEP

Lexra Proprietary & Confidential - 36 - Release 1.9

 April 30, 2001 LX5280

instruction. This holds the LX5280's internal clocks in the high state until an external hardware interrupt is
received.

Before executing the SLEEP instruction, the kernel must ensure that the interrupt condition that will
ultimately terminate standby mode has been enabled via the IM field of the coprocessor 0 Status register.
When the SLEEP instruction enters the W stage, the standby logic stalls the processor and waits for the LBC
to complete any outstanding processor initiated system bus operations. After these are completed, the
standby logic holds the system and bus clocks high. These are held high until an enabled interrupt is
received.

When standby mode is terminated by an interrupt, the standby logic allows the clocks to toggle. The
processor honors the interrupt by branching to the exception handler as is normally done for interrupt
servicing. Because several instructions are held in the pipeline while the clocks are frozen prior to the
interrupt, the exception PC will not point to the SLEEP instruction, but rather some later instruction.
Typically, a kernel would enter an idle loop just after executing the SLEEP instruction, so the interrupt will be
serviced from the kernel's normal idle interrupt service level.

The LX5280 takes a minimum of 6 cycles after the SLEEP instruction enters the W stage to safely
synchronize the initiation of standby mode, i.e. hold the clocks in the high state. Two cycles are required
terminate standby mode. The processor is stalled during these periods.

The standby logic receives the free running system and bus clocks, and generates gated clocks for distribution
to the LX5280. The standby logic must use flip-flops tied to free running clocks, which results in about a
dozen loads on the free running clocks.

Two pins, SL_SLEEPING_R and SL_SLEEPING_BR, are available from the standby logic and are asserted
high when the processor is in standby mode. The _R pin is for use in the system clock domain, and the _BR
pin is for use in the bus clock domain.

Lexra Proprietary & Confidential - 37 - Release 1.9

 April 30, 2001 LX5280

Lexra Proprietary & Confidential - 38 - Release 1.9

 April 30, 2001 LX5280

4. MIPS16

MIPS16 is an extension to the MIPS Instruction Set Architecture (ISA) that was developed to improve code
density, especially for System-on-Chip (SoC) designs. In these designs, on-chip instruction storage is often a
significant, even dominant, portion of the silicon component cost. This is especially true for real-time
applications because, in order to meet real-time requirements, instruction cache miss penalties cannot be
tolerated and thus a large portion of the instruction storage must be resident on-chip.

MIPS16 provides a set of 16-bit instruction formats to encode the most common operations. The key
compromises required to achieve 16-bit encoding include: (i) some MIPS I instructions are not available, (ii)
immediate widths are reduced, (iii) only 8 of the 32 general registers may be directly addressed. As a result
some operations cannot be executed in MIPS16 or require multiple MIPS16 instructions. Thus realistic
programs need to include both MIPS16 and MIPS I instructions, using MIPS16 where possible to save
storage, at some cost to performance.1 Mode switching between MIPS16 and MIPS I is discussed below. To
permit occasional access to all 32 general registers without the overhead of mode switching, MIPS16
provides MOVE instructions to move data between the MIPS16-visible registers and the full general register
set. Also, to permit occasional use of 16-bit immediates without mode switching, MIPS16 provides the
EXTEND instruction to allow a full width immediate in two MIPS16 instruction cycles. (Programs requiring
a large register set or frequent full-width immediates should be compiled in MIPS I.)

MIPS16 is difficult to program effectively at the assembler level. This is because of the limited register set
and the restricted size immediates. In fact, according to Sweetman2, "MIPS16 is not a suitable language for
assembly coding". Rather, MIPS16 is viewed as a compiler option which can be effectively applied to
achieve significant code size reduction where performance is not critical.

4.1. MIPS16 Instructions

This section describes the MIPS16 instructions, with emphasis on the differences between MIPS16 and the
32-bit MIPS ISA. The first table lists MIPS I Instructions that are not supported in MIPS16.

The second table lists MIPS I instructions which are supported in MIPS16. In most cases these are
specialized versions of the MIPS I instruction. MIPS16 is compatible with MIPS I, II and III, IV or V. The
LX5280 implements all MIPS16 for 32-bit data operations.3 The table lists all MIPS16 instructions together
with the corresponding MIPS I instruction and the specialization required to produce the MIPS16 instruction
(other than smaller register set and smaller immediates).

The third table lists the several new instructions introduced by MIPS16.

It is notable that MULT(U), DIV(U) are supported in MIPS16. MFHI and MFLO are also supported and are
necessary to access the result of MULT(U) or DIV(U). However, MTHI and MTLO are not supported. These
are used primarily to restore the state after exception handling and are used within the kernel, typically in
MIPS I.

1. The MIPS16 performance penalty results from occasionally using two instructions where one MIPS I instruction would suffice.
Some of this penalty is recovered in applications where a larger number of instructions per cache line reduces cache miss rate.
2. “See MIPS Run”, Dominic Sweetman, Appendix D, p. 425.
3. MIPS16 includes 16-bit formats for a number of MIPS III 64-bit doubleword operations which are not supported in the MIPS I
ISA.They are also not supported in Radiax.

Lexra Proprietary & Confidential - 39 - Release 1.9

 April 30, 2001 LX5280

Table 16: MIPS I Instructions Not Supported by MIPS16

MIPS I Not Supported by MIPS16 Assembler Mnemonics

Coprocessor operations CTCz, CFCz, MTCz, MFCz, LWCz, SWCz,

BCzT, BCzF, COPz

Unaligned loads, stores LWL, LWR, SWL, SWR

Arithmetic operations ADD, ADDI, SUB

Conditional branches BLEZ, BGTZ, BLTZ, BGEZ, BLTZAL, BGEZAL

Logical operations with immediates ANDI, ORI, XORI, LUI

Jump J

Miscellaneous SYSCALL, RFE, MTHI, MTLO

Table 17: MIPS16 Instructions that Support MIPS I

MIPS16 Instruction MIPS I Equivalent Instructiona

LB(U) ry, offset(rx)

LH(U) ry, offset(rx)
LW ry, offset(rx)
LW rx, offset(sp) (r29 base) LW rx, offset(base); base = r29
SB ry, offset(rx)
SH ry, offset(rx)
SW ry, offset(rx)
SW rx, offset(sp) (r29 base) SW rx, offset(base); base = r29

ADDIU ry, rx, immediate

ADDIU rx, immediate ADDIU rt, rs, immediate; rt=rs
ADDIU sp, immediate (1-operand) ADDIU rt, rs, immediate; rt=rs=r29
ADDIU rx, sp, immediate (2-operand) ADDIU rt, rs, immediate; rs=r29
ADDU rz, rx, ry
SUBU rz, rx, ry
NEG rx, ry (2-operand) SUBU rd, rs, rt; rs=r0

SLT(U) rx, ry (r24 dest. implied) SLT(U) rd, rs, rt; rd=r24
SLTI(U) rx, immediate (2-op., r24 dest) SLTI(U) rt, rs, immediate; rt=rs

CMPI rx, immediate (r24 dest. implied) XORI rt, rs, immediate; rt=r24
CMP rx, ry (r24 dest. implied) XOR rd, rs, rt; rd=r24

AND rx, ry (2-operand) AND rd, rs, rt; rd=rs

OR rx, ry (2-operand) OR rd, rs, rt; rd=rs
XOR rx, ry (2-operand) XOR rd, rs, rt; rd=rs
NOT rx, ry (2-operand) NOR rt, rs, rt; rs=r0
MOVE ry, r32 (2-operand) OR rd, rs, rt; rs=r0
MOVE r32, ry (2-operand) OR rd, rs, rt; rs=r0
LI rx, immediate ORI rd, rs, immediate; rs=r0

SLL rx, ry, immediate

SRL rx, ry, immediate
SRA rx, ry, immediate
SLLV ry, rx (2-operand) SLLV rd, rt, rs; rd=rs
SRLV ry, rx (2-operand) SRLV rd, rt, rs; rd=rs
SRAV ry, rx (2-operand) SRAV rd, rt, rs; rd=rs

MULT(U) rx, ry
DIV(U) rx, ry
MFHI rx
MFLO rx

Lexra Proprietary & Confidential - 40 - Release 1.9

 April 30, 2001 LX5280

MIPS16 Instruction MIPS I Equivalent Instructiona

JAL target
JR rx
JR ra JR rs; rs=r31
JALR ra, rx (2-operand; link = r31) JALR rs, rd; rs=r31

BEQZ rx, offset (1-operand) BEQ rs, rt, offset; rt=r0

BNEZ rx, offset (1-operand) BNE rs, rt, offset; rt=r0
BTEQ offset (implied operands) BEQ rs, rt, offset; rs=r24, rt=r0
BTNE offset (implied operands) BNE rs, rt, offset; rs=r24, rt=r0
B offset (implied operands) BEQ rs, rt, offset; rs=r0, rt=r0

BREAK

a. If no 32-bit MIPS instruction is listed, no specialization beyond limited size register set and
limited size immediates is required.

As noted earlier, MIPS16 restricts the MIPS I directly addressable register set and immediate field. Another
common MIPS16 restriction is that two, rather than three, register operands, are permitted. MIPS16 provides
a number of instructions that are not found MIPS I, as shown in Table 18.

Table 18: New MIPS16 Instructions

New MIPS16 Instruction Comment

LW rx, offset(pc) Load word with pc-relative address

ADDIU rx, pc, immediate ADDIU with pc operand

EXTEND immediate Supplies 11-bit immediate for use in the

following MIPS16 instruction

JALX target Jump to target, store return in r31 and toggle

the ISA mode between MIPS16 and MIPS I.

The pc-relative load LW is important to overcoming the drawback of smaller immediates in MIPS16. It
allows full 32-bit immediates to be embedded in the program and loaded into registers in a single instruction.
The ADDIU with pc operand is useful to support immediates embedded in the program. The pc value
referenced in LW or ADDIU depends on the context of the pc-relative instruction as shown in Table 19.

Table 19: PC-Relative Addressing

Context for PC-Relative Instruction pc Value

Normal case. (Non-extended pc-relative pc of the pc-relative instruction.

instruction, not in jump delay slot.)

pc-relative instruction with extended immediate pc of the EXTEND instruction

Non-extended pc-relative in the delay slot of pc of the jump instruction

jump, JR, JALR, JAL(X) (extended instructions
are not permitted in the delay slot of the jump.)

EXTEND is used to supply an extra 11-bits of immediate. It is used together with the restricted size
immediate field of the next instruction to supply a full width immediate. EXTEND cannot occur in the delay
slot of a Jump. It is not necessary for the assembly programmer to code EXTEND instructions. It will
automatically be assembled by MIPS16 assemblers wherever the immediate is too large to be encoded in a
single MIPS16 instruction.

Lexra Proprietary & Confidential - 41 - Release 1.9

 April 30, 2001 LX5280

Another new instruction JALX, is available in both MIPS16 and also in MIPS I on machines implementing
MIPS16 and is discussed below. [in MIPS I machines not implementing MIPS16, the JALX opcode 000111
causes an RI trap.]

4.2. Mode switching

Mode is switched between MIPS16 and MIPS I in one of two ways:

1. The instruction,

JALX target

toggles the mode.

2. The lsb of the general register rx in

JR rx
JALR rs, rx (in MIPS16 rs=ra)

causes the mode to be set to MIPS16 if rx[0] = 1; to MIPS I if rx[0] = 0. However, the lsb of the instruction
memory address from JR/JALR is forced to 0. As a consequence, machines that implement MIPS16 never
take AdEL exceptions on the lsb of the instruction address (this is true regardless of whether the machine is
operating in MIPS16 or MIPS I mode.).

The mode bit is saved in the lsb of the link register in JAL, JALX, JALR.

4.3. Exceptions

Upon Exception, the mode is automatically switched to MIPS I. The mode is saved in the lsb of the
Exception PC (EPC). EPC[0] = 0 indicates that the Exception occurred while executing code in MIPS I
mode; EPC[0] = 1 indicates that the Exception occurred in MIPS16 mode. The typical program will save the
EPC to a general register and later return to the main program with a JR instruction, causing the proper ISA
mode to be restored.

4.4. No Delay Slots

Consistent with the MIP16 emphasis on code density, there are no load delay or branch delay slots. In other
words, the instruction following the branch is executed only if the branch is not taken. [MIPS16 jumps (JAL,
JALX, JR, JALR) have a single delay slot, the same as in MIPS I. For jumps, the target address is always
taken. Thus, there is no risk that the delay slot cannot be used to do useful work: the instruction from the
target can be moved to the delay slot, if necessary.]

For MIPS16 loads, the instruction following the load can reference the loaded register (as in MIPS II). This
feature is present because the MIPS I compiler is not always successful in scheduling a useful instruction in
the delay slot and must occasionally resort to a NOP, reducing code density. This possibility is eliminated in
MIPS16.

Lexra Proprietary & Confidential - 42 - Release 1.9

 April 30, 2001 LX5280

5. LX5280 DSP Programming Model

The LX5280 supports Lexra’s Radiax DSP extensions to the MIPS-1 instruction set. This chapter describes
the Radiax extensions in detail. Section 5.1 describes each of the Radiax instructions. Section 5.2 describes
the instruction encoding.

The following conventions are employed in the instruction descriptions.

«» Encloses a list of syntax choices, from which one must be chosen.

value[3] Bits selected from a value.
MNEM[.OPT] Indicates an optional form of instruction an mnemonic.

5.1. Radiax Instructions

The Radiax instruction extensions include MAC operations, vector-addressing, and enhanced extensions to
the MIPS-1 ALU instructions.

5.1.1. Radiax Dual-MAC Instructions

Table 20: Radiax Dual-MAC Instructions

Instruction Syntax and Description

Dual Move to MTA2[.G] rS, «mD, mDh, mDl»

Accumulator If MTA2, and mDh(mDl) is selected, sign-extend the contents of general register rS to
40-bits and move to accumulator register mDh(mDl). If MTA2, and mD is selected,
update both mDh and mDl with the 40-bit, sign-extended contents of the same rS. If
MTA2.G is selected, the accumulator register bits [39:32] are updated with rS[31:24];
bits [31:00] of the accumulator are unchanged. (The .G option is used to restore the
upper-bits of the accumulator from the general register file; typically, following an
exception.)

Move From MFA rD, «mTh, mTl» [,n]

Accumulator Move the contents of accumulator register mTh or accumulator register mTl to
register rD with optional right shift. Bits [31+n : n] from the accumulator register are
transferred to rD[31:00]. The range n = 0 - 8 is permitted for the output alignment shift
amount. In the case of n = 0, the field may be omitted.

Dual Move From MFA2 rD, mT [,n]

Accumulator Move the contents of the upper halves of accumulator register pair mT to register rD
with optional right shift. The rD[31:16] are taken from mTh and rD[15:00] from the
corresponding mTl. mTh[31+n: 16+n] || mTl[31+n : 16+n] from the accumulator
register pair are transferred to rD[31:00]. The range n = 0 - 8 is permitted for the
output alignment shift amount. In the case of n = 0, the field may be omitted.

Divide DIVA mD, rS, rT

The contents of register rS is divided by rT, treating the operands as signed 2’s
complement values. The remainder is sign-extended to 40-bits and stored in mDh
and the quotient is sign-extended to 40-bits and stored in mDl. m0h[31:00] is also
called HI. m0l[31:00] is also called LO.

Divide Unsigned DIVAU mD, rS, rT

The contents of register rS is divided by rT, treating the operands as unsigned values.
The remainder is zero-extended to 40-bits and stored in mDh and the quotient is
zero-extended to 40-bits and stored in mDl. m0h[31:00] is also called HI. m0l[31:00]
is also called LO.

Lexra Proprietary & Confidential - 43 - Release 1.9

 April 30, 2001 LX5280

Instruction Syntax and Description

Multiply MULTA mD, rS, rT

(32-bit) The contents of register rS is multiplied by rT, treating the operands as signed 2’s
complement values. The upper 32-bits of the 64-bit product is sign-extended to 40-
bits and stored in mDh and the lower 32-bits is zero-extended to 40-bits and stored in
the corresponding mDl. m0h[31:00] is also called HI. m0l[31:00] is also called LO. If
MMD[MT] is 1, then the partial product rS[15:00] x rT[15:00] is not included in the
total product. If MMD[MF] is 1, then the product is left shifted by one bit, and
furthermore, if both operands are -1 then the product is set to positive signed, all
ones fraction, prior to the shift. If both MMD[MT] and MMD[MF] are 1, the result is
undefined.

Multiply MULTAU mD, rS, rT

Unsigned The contents of register rS is multiplied by rT, treating the operands as unsigned
(32-bit) values. The upper 32-bits of the 64-bit product is zero-extended to 40-bits and stored
in mDh and the lower 32-bits is zero-extended to 40-bits and stored in the
corresponding mDl. m0h[31:00] is also called HI. m0l[31:00] is also called LO. If
MMD[MT] is 1, then the partial product rS[15:00] x rT[15:00] is not included in the
total product. If MMD[MF] is 1, then the result is undefined.

Dual Multiply MULTA2 «mD, mDh, mDl», rS, rT

(16-bit) The contents of register rS is multiplied by rT, treating the operands as signed 2’s
complement values. If the destination register is mDh, rS[31:16] is multiplied by
rT[31:16] and the product is sign-extended to 40-bits and stored in mDh. If the
destination register is mDl, rS[15:00] is multiplied by rT[15:00] and the product is
sign-extended to 40-bits and stored in mDl. If the destination is mD, both operations
are performed and the two products are stored in the accumulator register pair mD. If
MMD[MF] is 1, then each product is left shifted by one bit, and furthermore, for each
multiply, if both operands are -1 then the product is set to positive signed, all ones
fraction.

Dual Multiply and Negate MULNA2 «mD, mDh, mDl», rS, rT

(16-bit) The contents of register rS is multiplied by rT, treating the operands as signed 2’s
complement values. If the destination register is mDh, rS[31:16] is multiplied by
rT[31:16] and the product is sign-extended to 40-bits, negated (i.e. subtracted from
zero) and stored in mDh. If the destination register is mDl, rS[15:00] is multiplied by
rT[15:00] and the product is sign-extended to 40-bits, negated (i.e. subtracted from
zero) and stored in mDl. If the destination is mD, both operations are performed and
the two products are stored in the accumulator register pair mD. If MMD[MF] is 1,
then each product is left shifted by one bit prior to sign-extension and negation, and
furthermore, for each multiply, if both operands are -1 then the product is set to
positive signed, all ones fraction prior to sign-extension and negation.

Complex CMULTA mD, rS, rT

Multiply, rS[31:16] is interpreted as the real part of a complex number. rS[15:00] is interpreted
as the imaginary part of the same complex number. Similarly for the contents of
general register rT. As the result of CMULTA, mDh is updated with the real part of the
product, sign-extended to 40-bits and mDl is updated with the imaginary part of the
product, sign-extended to 40-bits. If MMD[MF] is 1, then each product is left shifted
by one bit, and furthermore, for each multiply, if both operands are -1 then the
product is set to positive signed, all ones fraction, prior to the addition of terms.

32-bit Multiply-Add with MADDA mD, rS, rT

72-bit accumulate The contents of register rS is multiplied by rT treating the operands as signed 2’s
complement values. If MMD[MT] is 1, then the partial product rS[15:00] x rT[15:00] is
not included in the total product. If MMD[MF] is 1, then the product is left shifted by
one bit, and furthermore, if both operands are -1 then the product is set to a positive
signed, all ones fraction. If both MMD[MT] and MMD[MF] are 1, then the result of the
multiply is undefined.
The 64-bit product is sign-extended to 72-bits and added to the concatenation
mDh[39:0] || mDl[31:0], ignoring mDl[39:32]. The lower 32 bits of the result are zero-
extended to 40-bits and stored into mDl. The upper 40-bits of the result are stored
into mDh.

Lexra Proprietary & Confidential - 44 - Release 1.9

 April 30, 2001 LX5280

Instruction Syntax and Description

32-bit unsigned Multiply- MADDAU mD, rS, rT

Add with 72-bit The contents of register rS is multiplied by rT treating the operands as unsigned
accumulate values. If MMD[MT] is 1, then the partial product rS[15:00] x rT[15:00] is not included
in the total product. If MMD[MF] is 1, then the result of the multiply is undefined.
The 64-bit product is zero-extended to 72-bits and added to the concatenation
mDh[39:0] || mDl[31:0], ignoring mDl[39:32]. The lower 32 bits of the result are zero-
extended to 40-bits and stored into mDl. The upper 40-bits of the result are stored
into mDh.

Dual Multiply-Add, MADDA2[.S]«mD, mDh, mDl», rS, rT

optional The contents of register rS is multiplied by rT and added to an accumulator register,
saturation treating the operands as signed 2’s complement values. If the destination register is
mDh, rS[31:16] is multiplied by rT[31:16] then sign-extended and added to
mDh[39:00]. If the destination register is mDl, rS[15:00] is multiplied by rT[15:00] then
sign-extended and added to mDl[39:00]. If the destination is mD, both operations
are performed and the two results are stored in the accumulator register pair mD. If
MADDA2.S the result of each addition is saturated before storage in the accumulator
register. The multiplies are subject to MMD[MF] as in MULTA2. The saturation point
is selected as either 40 or 32 bits by MMD[MS].

32-bit Multiply-Subtract MSUBA mD, rS, rT

with 72-bit accumulate The contents of register rS is multiplied by rT treating the operands as signed 2’s
complement values. If MMD[MT] is 1, then the partial product rS[15:00] x rT[15:00] is
not included in the total product. If MMD[MF] is 1, then the product is left shifted by
one bit, and furthermore, if both operands are -1 then the product is set to a positive
signed, all ones fraction. If both MMD[MT] and MMD[MF] are 1, then the result of the
multiply is undefined.
The 64-bit product is sign-extended to 72-bits and subtracted from the concatenation
mDh[39:0] || mDl[31:0], ignoring mDl[39:32]. The lower 32 bits of the result are zero-
extended to 40-bits and stored into mDl. The upper 40-bits of the result are stored
into mDh.

32-bit unsigned Multiply- MSUBAU mD, rS, rT

Subtract with 72-bit The contents of register rS is multiplied by rT treating the operands as unsigned
accumulate values. If MMD[MT] is 1, then the partial product rS[15:00] x rT[15:00] is not included
in the total product. If MMD[MF] is 1, then the result of the multiply is undefined.
The 64-bit product is zero-extended to 72-bits and subtracted from the concatenation
mDh[39:0] || mDl[31:0], ignoring mDl[39:32]. The lower 32 bits of the result are zero-
extended to 40-bits and stored into mDl. The upper 40-bits of the result are stored
into mDh.

Dual Multiply-Sub, MSUBA2[.S] «mD, mDh, mDl», rS, rT

optional The contents of register rS is multiplied by rT and subtracted from an accumulator
saturation register, treating the operands as signed 2’s complement values. If the destination
register is mDh, rS[31:16] is multiplied by rT[31:16] then sign-extended and
subtracted from mDh[39:00]. If the destination register is mDl, rS[15:00] is multiplied
by rT[15:00] then sign-extended and subtracted from mDl[39:00]. If the destination
is mD, both operations are performed and both results are stored in the accumulator
register pair mD. If MSUBA2.S the result of each subtraction is saturated before
storage in the accumulator register.

Add ADDMA[.S]mD«h,l», mS«h,l», mT«h,l»

Accumulators The contents of accumulator mTh or mTl is added to the contents of accumulator
mSh or mSl, treating both registers as signed 40-bit values. mDh or mDl is updated
with the result. If ADDMA.S, the result is saturated before storage. The saturation
point is selected as either 40 or 32 bits by MMD[MS].

Subtract SUBMA[.S] mD«h,l», mS«h,l», mT«h,l»

Accumulators The contents of accumulator mTh or mTl is subtracted from the contents of
accumulator mSh or mSl, treating both registers as signed 40-bit values. mDh or mDl
is updated with the result. If SUBMA.S, the result is saturated before storage.The
saturation point is selected as either 40 or 32 bits by MMD[MS].

Lexra Proprietary & Confidential - 45 - Release 1.9

 April 30, 2001 LX5280

Instruction Syntax and Description

Dual Round RNDA2 «mT, mTh, mTl» [,n]

The accumulator register mTh or mTl is rounded, then updated. If mT, the
accumulator register pair mTh/mTl are each rounded, then updated. The rounding
mode is selected in MMD field “RND”. The least significant bit of precision in the
accumulator register after rounding is: 16+n. Bits [15+n : 00] are zeroed. The range n
= 0 - 8 is permitted for the output alignment shift amount. In the case of n = 0, the
field may be omitted.

Nomenclature:

rS, rT = r0 - r31
mD = mDh || mDl; also for mT
mDh = m0h - m3h; also for mSh, mTh
mDl = m0l - m3l; also for mSh, mTh
HI = m0h[31:00]
LO = m0l[31:00]

5.1.2. Cycle-by-Cycle Usage for Dual MAC Instructions

The Dual MAC eliminates all programming hazards for its instructions by stalling the pipeline when
necessary. It does this both to avoid resource conflicts and to wait for results of a first instruction to be ready
before attempting to use those results in a second instruction. This means that there are no programming
restrictions in order to obtain correct results from a sequence of Dual MAC instructions.

However, the most efficient use of the Dual MAC hardware is obtained when the program avoids these stalls.
This can be done by scheduling the instructions properly. Table 52 on page 123 indicates the number of
cycles that must be present between MAC instructions to avoid stalls. In addition several instruction
sequences are presented that represent the most efficient use of the Dual MAC for the “inner loop” of some
common DSP algorithms. Typically, these make use of the multiple accumulators in the Dual MAC.

The following code sequences indicate the most efficient use of the Dual MAC for coding the inner loop of
some common DSP algorithms. The algorithms are presented for 16-bit operands with 16-bit results, as well
as 32-bit operands with 32-bit results. The algorithms assume that fractional arithmetic is used. Therefore, for
the 32-bit results of a 32x32 multiply, only the HI half of the target accumulator pair is retrieved or used.

In these examples, only the Dual MAC instructions are shown. The other pipe is used to fetch and store
operands and take care of loop housekeeping functions. The loops may need to be unrolled to take full
advantage of the multiple Dual MAC accumulators.

CASE 1: 16-bit inner product. SUM = SUM + Ai*Bi

Assuming packed operands, two multiply-adds per cycle:

MADDA2 m0,r1,r2
MADDA2 m0,r3,r4
MADDA2 m0,r5,r6
MADDA2 m0,r7,r8
...

CASE 2: 16-bit vector product loop. Ci = Ai*Bi

Assuming packed fractional operands, two multiplies per two cycles using two accumulator pairs.

Lexra Proprietary & Confidential - 46 - Release 1.9

 April 30, 2001 LX5280

MULTA2 m0,r1,r2
MFA2 m1,r8
MULTA2 m1,r3,r4
MFA2 m0,r7
...

CASE 3: 16-bit complex vector product. Ci = Ai *complex Bi

Assuming fractional operands packed as 16-bit real, 16-bit imaginary. One complex multiply every two
cycles using two accumulator pairs.

CMULTA m0,r1,r2
MFA2 m1,r8
CMULTA m1,r3,r4
MFA2 m0,r7
...

CASE 4: 32-bit inner product loop. SUM = SUM + Ai*Bi

Achieves a multiply-accumulate every other cycle using one accumulator.

MADDA m0, r1, r2

non-DualMAC op
MADDA m0, r3, r4
non-DualMAC op
...

CASE 5: 32-bit vector product loop. Ci = Ai*Bi

Assuming fractional 32-bit operands so that the MFA waits for the HI result of the MULTA. Achieves one
multiply per two cycles using all the accumulators.

MULTA m0, r1, r2

MFA r9, m1h
MULTA m1, r3, r4
MFA r10,m2h
MULTA m2, r5, r6
MFA r11,m3h
MULTA m3, r7, r8
MFA r12,m0h
...

CASE 6: 32-bit complex vector product. Ci = Ai *complex Bi

Assuming fractional 32-bit operands so that the ADDMA/SUBMA waits for the HI result of the second
MULTA. Achieves one complex multiply per ten cycles using all the accumulators, with two inserted
instructions. This is a good example of the cycles needed from MULTA to SUBMA/ADDMA (5 cycles for
HI) and from SUBMA/ADDMA to MFA (2 cycles).

MULTA m0, r1, r4 ; a[2i] * b[2i+1]

MFA rimag, m1h
MULTA m1, r2, r3 ; a[2i+1] * b[2i]
SUBMA m3h,m2h,m3h ; c[2i-2] = a[2i-2]*b[2i-2] - a[2i-1]*b[2i-1]
non-DualMAC op
MULTA m2, r1, r3 ; a[2i] * b[2i]
MFA rreal, m3h

Lexra Proprietary & Confidential - 47 - Release 1.9

 April 30, 2001 LX5280

MULTA m3, r2, r4 ; a[2i+1] * b[2i+1]

ADDMA m1h,m0h,m1h ; c[2i+1] = a[2i+1]*b[2i] + a[2i]*b[2i+1]
non-DualMAC op
...

5.1.3. Vector Addressing Instructions

Table 21: Vector Addressing Instructions

Instruction Syntax and Description

Load Twinword LT rT, displacement(base)

The displacement, in bytes, is a signed 14-bit quantity that must be divisible by 8
(since it occupies only 11 bits of the instruction word). Sign-extend the displacement
to 32-bits and add to the contents of register base to form the address temp. Load
contents of word addressed by temp into register rT (which must be an even
register). Load contents of word addressed by temp+4 into register rT+1.

Store Twinword ST rT, displacement(base)

The displacement, in bytes, is a signed 14-bit quantity that must be divisible by 8
(since it occupies only 11 bits of the instruction word). Sign-extend the displacement
to 32-bits and add to the contents of register base to form the address temp. Store
contents of register rT (which must be an even register) into word addressed by
temp. Store contents of register rT+1 into word addressed by temp+4.

Load Twinword, LTP[.Cn] rT, (pointer)stride

Pointer Increment, Let temp = contents of register pointer. Load contents of word addressed by temp
optional circular buffer into register rT (which must be an even register). Load contents of word addressed
by temp+4 into register rT+1. The stride, in bytes, is a signed 11-bit quantity that must
be divisible by 8 (since it occupies only 8 bits of the instruction word). Sign-extend the
stride to 32-bits and add to contents of register pointer to form next address. Update
pointer with the calculated next address. “.Cn” selects circular buffer n = 0 - 2. See
Note 2.

Load Word, LWP[.Cn] rT, (pointer)stride

Pointer Increment, Load contents of word addressed by register pointer into register rT. The stride, in
optional circular buffer bytes, is a signed 10-bit quantity that must be divisible by 4 (since it occupies only 8
bits of the instruction word). Sign-extend the stride to 32-bits and add to contents of
register pointer to form next address.Update pointer with the calculated next address.
“.Cn” selects circular buffer n = 0 - 2. See Note 2.

Load Halfword, LHP[.Cn] rT, (pointer)stride

Pointer Increment, Load contents of sign-extended halfword addressed by register pointer into register
optional circular buffer rT. The stride, in bytes, is a signed 9-bit quantity that must be divisible by 2 (since it
occupies only 8 bits of the instruction word). Sign-extend the stride to 32-bits and add
to contents of register pointer to form next address. Update pointer with the
calculated next address. “.Cn” selects circular buffer n = 0 - 2. See Note 2.

Load Halfword Unsigned, LHPU[.Cn] rT, (pointer)stride

Pointer Increment, Load contents of zero-extended halfword addressed by register pointer into register
optional circular buffer rT. The stride, in bytes, is a signed 9-bit quantity that must be divisible by 2 (since it
occupies only 8 bits of the instruction word). Sign-extend the stride to 32-bits and add
to contents of register pointer to form next address. Update pointer with the
calculated next address. “.Cn” selects circular buffer n = 0 - 2. See Note 2.

Load Byte, LBP[.Cn] rT, (pointer)stride

Pointer Increment, Load contents of sign-extended byte addressed by register pointer into register rT.
optional circular buffer The stride, in bytes, is a signed 8-bit quantity. Sign-extend the stride to 32-bits and
add to contents of register pointer to form next address. Update pointer with the
calculated next address. “.Cn” selects circular buffer n = 0 - 2. See Note 2.

Lexra Proprietary & Confidential - 48 - Release 1.9

 April 30, 2001 LX5280

Instruction Syntax and Description

Load Byte LBPU[.Cn] rT, (pointer)stride

Unsigned, Load contents of zero-extended byte addressed by register pointer into register rT.
Pointer Increment, The stride, in bytes, is a signed 8-bit quantity. Sign-extend the stride to 32-bits and
optional circular buffer add to contents of register pointer to form next address. Update pointer with the
calculated next address. “.Cn” selects circular buffer n = 0 - 2. See Note 2.

Store Twinword, STP[.Cn] rT, (pointer)stride

Pointer Increment, Let temp = contents of register pointer. Store contents of register rT (which must be
optional circular buffer an even register) into word addressed by temp. Store contents of register rT+1 into
word addressed by temp+4. The stride, in bytes, is a signed 11-bit quantity that must
be divisible by 8 (since it occupies only 8 bits of the instruction word). Sign-extend the
stride to 32-bits and add to contents of register pointer to form next address.Update
pointer with the calculated next address. “.Cn” selects circular buffer n = 0 - 2. See
Note 2.

Store Word, SWP[.Cn] rT, (pointer)stride

Pointer Increment, Store contents of register rT into word addressed by register pointer. The stride, in
optional circular buffer bytes, is a signed 10-bit quantity that must be divisible by 4 (since it occupies only 8
bits of the instruction word). Sign-extend the stride to 32-bits and add to contents of
register pointer to form next address. Update pointer with the calculated next
address. “.Cn” selects circular buffer n = 0 - 2. See Note 2.

Store Halfword, SHP[.Cn] rT, (pointer)stride

Pointer Increment, Store contents of register rT[15:00] into 16-bit halfword addressed by register pointer.
optional circular buffer The stride, in bytes, is a signed 9-bit quantity that must be divisible by 2 (since it
occupies only 8 bits of the instruction word). Sign-extend the stride to 32-bits and add
to contents of register pointer to form next address. Update pointer with the
calculated next address. “.Cn” selects circular buffer n = 0 - 2. See Note 2.

Store Byte, SBP[.Cn] rT, (pointer)stride

Pointer Increment, Store contents of register rT[07:00] into byte addressed by register pointer. The
optional circular buffer stride, in bytes, is a signed 8-bit quantity. Sign-extend the stride to 32-bits and add to
contents of register pointer to form next address. Update pointer with the calculated
next address. “.Cn” selects circular buffer n = 0 - 2. See Note 2.

Move To Radiax, User MTRU rT, RADREG

Move the contents of register rT to one of the User Radiax registers: cbs0 - cbs2,
cbe0 - cbe2, mmd, lpc0, lpe0, lps0. This instruction has a single delay slot before the
MMD register takes effect (all other registers have no delay slot).

Move From Radiax, User MFRU rT, RADREG

Move the contents of the designated User Radiax register (cbs0 - cbs2, cbe0 - cbe2,
mmd, lpc0, lps0, lpe0) to register rT.

Nomenclature:

rT = r0 - r31, and must be even for LT, ST, LTP[.Cn], STP[.Cn]

base, pointer = r0 - r31
stride = 8/9/10/11-bit signed value (in bytes) for byte/halfword/
word/twinword ops.
displacement = 14-bit signed value, in bytes
RADREG = cbs0 - cbs2, cbe0 - cbe2, mmd, lpc0, lps0, lpe0

Notes:
1. For LTP[.Cn], LWP[.Cn], LHP(U)[.Cn], LBP(U)[.Cn], rT = pointer is unsupported.
2. When a circular buffer is selected, the update of the pointer register is performed according to the
following algorithm, which depends on the sign of the stride and the granularity of the access. A stride
exactly equal to 0 is not supported:

Lexra Proprietary & Confidential - 49 - Release 1.9

 April 30, 2001 LX5280

LBP(U).Cn and SBP.Cn

if (stride > 0 && pointer[2:0] == 111 && pointer[31:3] == CBEn)

tpointer <= { CBSn[31:3], 3’b000 }
else if (stride < 0 && pointer[2:0] == 000 && pointer[31:3] == CBSn)
pointer <= { CBEn[31:3], 3’b111 }
else
pointer <= pointer + stride.

LHP(U).Cn and SHP.Cn

if (stride > 0 && pointer[2:0] == 11x && pointer[31:3] == CBEn)

pointer <= { CBSn[31:3], 3’b000 }
else if (stride < 0 && pointer[2:0] == 00x && pointer[31:3] == CBSn)
pointer <= { CBEn[31:3] 3’b110 }
else
pointer <= pointer + stride.

LWP.Cn and SWP.Cn

if (stride > 0 && pointer[2:0] == 1xx && pointer[31:3] == CBEn)

pointer <= { CBSn[31:3] 3’b 000 }
else if (stride < 0 && pointer[2:0] == 0xx && pointer[31:3] == CBSn)
pointer <= { CBEn[31:3], 3’b100 }
else
pointer <= pointer + stride.

LTP.Cn and STP.Cn

if (stride > 0 && pointer[31:3] == CBEn)

pointer <= { CBSn[31:3], 3’b000 }
else if (stride < 0 && pointer[31:3] == CBSn)
pointer <= { CBEn[31:3], 3’b000 }
else
pointer <= pointer + stride.

5.1.4. Radiax ALU Operations

The Radiax ALU operations include both dual 16-bit and saturating versions of the MIPS-1 ALU operations
and several new ALU operations which are useful for common DSP algorithms.

Table 22: Radiax ALU Operations

Instruction Syntax and Description

Dual Shift Left SLLV2 rD, rT, rS

Logical Variable The contents of rT[31:16] and the contents of rT[15:00] are independently shifted left
by the number of bits specified by the low order four bits of the contents of general
register rS, inserting zeros into the low order bits of rT[31:16] and rT[15:00]. For
SLLV2, the high and low results are concatenated and placed in register rD. (Note
that a [.S] option is not provided because this is a logical rather than arithmetic shift
and thus the concept of arithmetic overflow is not relevant.)

Lexra Proprietary & Confidential - 50 - Release 1.9

 April 30, 2001 LX5280

Instruction Syntax and Description

Dual Shift Right SRLV2 rD, rT, rS

Logical Variable The contents of rT[31:16] and the contents of rT[15:00] are independently shifted
right by the number of bits specified by the low order four bits of the contents of
general register rS, inserting zeros into the high order bits of rT[31:16] and rT[15:00].
The high and low results are concatenated and placed in register rD. (Note that a [.S]
option is not provided because this is a logical rather than arithmetic shift and thus
the concept of arithmetic overflow is not relevant.)

Dual Shift Right SRAV2 rD, rT, rS

Arithmetic Variable The contents of rT[31:16] and the contents of rT[15:00] are independently shifted
right by the number of bits specified by the low order four bits of the contents of
general register rS, sign-extending the high order bits of rT[31:16] and rT[15:00]. The
high and low results are concatenated and placed in register rD. (Note that a [.S]
option is not provided because arithmetic overflow/underflow is not possible.)

Add, ADDR[.S] rD, rS, rT

optional saturation 32-bit addition. Considering both quantities as signed 32-bit integers, add the
contents of register rS to rT. For ADDR, the result is placed in register rD, ignoring
any overflow or underflow. For ADDR.S, the result is saturated to 0 || 131 (if overflow)
or 1 || 031 (if underflow) then placed in rD. ADDR[.S] will not cause an Overflow Trap.

Dual Add, ADDR2[.S] rD, rS, rT

optional saturation Dual 16-bit addition. Considering all quantities as signed 16-bit integers, add the
contents of register rS[15:00] to rT[15:00] and, independently add the contents of
register rS[31:16] to rT[31:16]. For ADDR2, the high and low results are
concatenated and placed in register rD ignoring any overflow or underflow. For
ADDR2.S, the two results are independently saturated to 0 || 115 (if overflow) or 1 ||
015 (if underflow) then placed in rD. ADDR2[.S] will not cause an Overflow Trap.

Subtract, SUBR[.S] rD, rS, rT

optional saturation 32-bit subtraction. Considering both quantities as signed 32-bit integers, subtract the
contents of register rT from the contents of register rS. For SUBR, the result is placed
in register rD ignoring any overflow or underflow. For SUBR.S, the result is saturated
to 0 || 131 (if overflow) or 1 || 031 (if underflow) then placed in rD. SUBR[.S] will not
cause an Overflow Trap.

Dual Subtract, SUBR2[.S] rD, rS, rT

optional saturation Dual 16-bit subtraction. Considering all quantities as signed 16-bit integers, subtract
the contents of register rT[15:00] from rS[15:00] and, independently subtract the
contents of register rT[31:16] from rS[31:16]. For SUBR2, the high and low results
are concatenated and placed in register rD ignoring any overflow or underflow. For
SUBR2.S, the two results are independently saturated to 0 || 115 (if overflow) or 1 ||
015 (if underflow) then placed in rD. SUBR2[.S] will not cause an Overflow Trap.

Dual Set On Less Than SLTR2 rD, rS, rT

Dual 16-bit comparison. Considering both quantities as signed 16-bit integers, if
rS[15:00] is less than rT[15:00] then set rD[15:00] to 015 || 1, else to zero.
Independently, considering both quantities as signed 16-bit integers, if rS[31:16] is
less than rT[31:16] then set rD[31:16] to 015 || 1, else to zero.

Minimum MIN rD, rS, rT

The contents of the general register rT are compared with rS considering both
quantities as signed 32-bit integers. If rS < rT or rS = rT, rS is placed into rD. If, rS >
rT, rT is placed into rD.

Dual Minimum MIN2 rD, rS, rT

The contents of rT[31:16] are compared with rS[31:16] considering both quantities as
signed 16-bit integers. If rS[31:16] < rT[31:16] or rS[31:16] = rT[31:16], rS[31:16] is
placed into rD[31:16]. If, rS[31:16] > rT[31:16], rT[31:16] is placed into rD[31:16]. A
similar, independent operation is performed on rT[15:00] and rS[15:00] to determine
rD[15:00].

Lexra Proprietary & Confidential - 51 - Release 1.9

 April 30, 2001 LX5280

Instruction Syntax and Description

Maximum MAX rD, rS, rT

The contents of the general register rT are compared with rS considering both
quantities as signed 32-bit integers. If rS > rT or rS = rT, rS is placed into rD. If, rS <
rT, rT is placed into rD.

Dual Maximum MAX2 rD, rS, rT

The contents of rT[31:16] are compared with rS[31:16] considering both quantities as
signed 16-bit integers. If rS[31:16] > rT[31:16] or rS[31:16] = rT[31:16], rS[31:16] is
placed into rD[31:16]. If, rS[31:16] < rT[31:16], rT[31:16] is placed into rD[31:16]. A
similar, independent operation is performed on rT[15:00] and rS[15:00] to determine
rD[15:00].

Absolute, ABSR[.S] rD, rT

optional saturation Considering rT as a signed 32-bit integer, if rT > 0, rT is placed into rD. If rT < 0, -rT is
placed into rD. If ABSR.S and rT = 1 || 031 (the smallest negative number) then 0 ||
131 (the largest positive number) is placed into rD; otherwise, if ABSR and rT = 1 ||
031, rT is placed into rD.

Dual Absolute, ABSR2[.S] rD, rT

optional saturation ABS[.S] operations are performed independently on rT[31:16] and rT[15:00],
considering each to be 16-bit signed integers. rD is updated with the absolute value
of rT[31:16] concatenated with the absolute value of rT[15:00].

Dual Mux MUX2«[.HH], [.HL], [.LH], [.LL]» rD, rS, rT

rD[31:16] is updated with rS[31:16] for MUX2.HH or MUX2.HL.
rD[31:16] is updated with rS[15:00] for MUX2.LH or MUX2.LL.
rD[15:00] is updated with rT[31:16] for MUX2.HH or MUX2.LH
rD[15:00] is updated with rT[15:00] for MUX2.HL or MUX2.LL

Count Leading Sign bits CLS rD, rT

The binary-encoded number of redundant sign bits of general register rT is placed
into rD. If rT[31:30] = 10 or 01, rD is updated with 0. If rT = 0, or if rT = 132 , rD is
updated with 027 || 15 (decimal 31).

Bit Reverse BITREV rD, rT, rS

A bit-reversal of the contents of general register rT is performed. The result is then
shifted right logically by the amount specified in the lower 5-bits of the contents of
general register rS, then stored in rD.

5.1.5. Conditional Operations

The LX5280 provides conditional move instructions that reduce the need for program branches, resulting in
greater program efficiency.

Table 23: Conditional Operations

Instruction Syntax and Description

Conditional Move on CMVEQZ[.H] [.L] rD, rS, rT

Equal Zero If the general register rT is equal to 0, the general register rD is updated with rS;
otherwise rD is unchanged. For [.H] if rT[31:16] is equal to 0, the full 32-bit general
register rD[31:00] is updated with rS; otherwise rD is unchanged. For [.L] if rT[15:00]
is equal to 0, the full 32-bit general register rD[31:00] is updated with rS; otherwise rD
is unchanged.

Lexra Proprietary & Confidential - 52 - Release 1.9

 April 30, 2001 LX5280

Instruction Syntax and Description

Conditional Move on Not CMVNEZ[.H] [.L] rD, rS, rT

Equal Zero If the general register rT is not equal to 0, the general register rD is updated with rS;
otherwise rD is unchanged. For [.H] if rT[31:16] is not equal to 0, the full 32-bit
general register rD[31:00] is updated with rS; otherwise rD is unchanged. For [.L] if
rT[15:00] is not equal to 0, the full 32-bit general register rD[31:00] is updated with rS;
otherwise rD is unchanged.

Usage Note:
When combined with the SLT or SLTR2 instructions, the conditional move instructions can be used to
construct a complete set of conditional move macro-operations. For example:

if ( r3 < r4 ) r1 <-- r2:

CMVLT r1,r2,r3,r4 => SLT AT,r3,r4
CMVNEZ r1,r2,AT

if ( r3 >= r4 ) r1 <-- r2:

CMVGE r1,r2,r3,r4 => SLT AT,r3,r4
CMVEQZ r1,r2,AT

if ( r3 <= r4 ) r1 <-- r2:

CMVLE r1,r2,r3,r4 => SLT AT,r4,r3
CMVEQZ r1,r2,AT

if ( r3 > r4 ) r1 <-- r2:

CMVGT r1,r2,r3,r4 => SLT AT,r4,r3
CMVNEZ r1,r2,AT

5.2. Instruction Encoding

5.2.1. Lexra Formats

The Lexra Formats are introduced into the MIPS instruction set by designating a single I-Format as
“LEXOP”, then using the INST[5:0] “subop” field to permit up to 64 new Lexra opcodes. Thus the new DSP
opcodes model the MIPS “special” opcodes encoded in R-Format. The diagrams below illustrate the LEXOP
codes using I-Format 011_111 which is unused in the MIPS I-IV ISA.

The following principles are used to resolve potential ambiguity of encoding between the new LX5280 DSP
extensions and MIPS instructions:

a. LX5280 instructions with similar operations to existing MIPS instructions, but with additional operands
permitted, are programmed with new Assembler mnemonics and encoded as a LEXOP. For instance:

multa m1, r1, r2 is encoded as a LEXOP instruction

mult r1, r2 is encoded as a MIPS instruction
multa m0, r1, r2 is encoded as a LEXOP instruction (m0 is an alias for HI/LO)

b. If a MIPS instruction is “extended” with new functionality, it is programmed with new Assembler
mnemonics and encoded as a LEXOP. Lexra mnemonics which end in “r” indicate general register file
targets; mnemonics which end in “a” indicate accumulator register targets. This convention removes

Lexra Proprietary & Confidential - 53 - Release 1.9

 April 30, 2001 LX5280

ambiguity between the Lexra op and a similar MIPS op. For example,

addr r3, r1, r2 is encoded as a LEXOP instruction

add r3, r1, r2 is encoded as a MIPS instruction

The MIPS add and the LEXOP addr are both signed 32-bit additions. However, on overflow the MIPS
instruction triggers the Overflow Exception, while the LEXOP does not. Alternatively the result of the
LEXOP will saturate if the “.s” option is selected (addr.s).

5.3. Load/Store Formats

31 26 25 21 20 16 15 6 5 0

Assembler LEXOP Lexra

Mnemonic 011 111 base rt immediate SUBOP

LT LEXOP base rt-even displacement/8 LT

ST LEXOP base rt-even displacement/8 ST

6 5 5 10 6

31 26 25 21 20 16 15 8 7 6 5 0

Assembler LEXOP Lexra

Mnemonic 011 111 base rt immediate cc SUBOP

LBP[.Cn] LEXOP pointer rt stride cc LBP

LBPU[.Cn] LEXOP pointer rt stride cc LBPU

LHP[.Cn] LEXOP pointer rt stride/2 cc LHP

LHPU[.Cn] LEXOP pointer rt stride/2 cc LHPU

LWP[.Cn] LEXOP pointer rt stride/4 cc LWP

LTP[.Cn] LEXOP pointer rt stride/8 cc LTP

SBP[.Cn] LEXOP pointer rt stride cc SBP

SHP[.Cn] LEXOP pointer rt stride/2 cc SHP

SWP[.Cn] LEXOP pointer rt stride/4 cc SWP

STP[.Cn] LEXOP pointer rt stride/8 cc STP

6 5 5 8 2 6

Lexra Proprietary & Confidential - 54 - Release 1.9

 April 30, 2001 LX5280

base, pointer, rt Selects general register r0 - r31.

rt-even Selects general register even-odd pair r0/r1, r2/r3, ... r30/r31
stride Signed 2s-complement number in bytes. Must be an integral
number of halfwords/words/twinwords for the corresponding
instructions.
displacement Signed 2s-complement number in bytes. Must be an integral
number of twinwords.
cc 00 = select circular buffer 0 (cbs0, cbe0)
01 = select circular buffer 1 (cbs1, cbe1)
10 = select circular buffer 2 (cbs2, cbe2)
11 = no circular buffer selected

Lexra Proprietary & Confidential - 55 - Release 1.9

 April 30, 2001 LX5280

5.3.1. Arithmetic Format

31 26 25 21 20 16 15 11 10 9 8 7 6 50

Assembler LEXOP Lexra

Mnemonic 011 111 rs rt rd hl 0 s d SUBOP

ADDR[.S],ADDR2[.S] LEXOP rs rt rd 0 0 s d ADDR

SUBR.S, SUBR2[.S] LEXOP rs rt rd 0 0 s d SUBR

SLTR2 LEXOP rs rt rd 0 0 0 1 SLTR

SLLV2 LEXOP rs rt rd 0 0 0 1 SLLV

SRLV2 LEXOP rs rt rd 0 0 0 1 SRLV

SRAV2 LEXOP rs rt rd 0 0 0 1 SRAV

MIN, MIN2 LEXOP rs rt rd 0 0 0 d MIN

MAX, MAX2 LEXOP rs rt rd 0 0 0 d MAX

ABSR[.S], ABSR2[.S] LEXOP 0 rt rd 0 0 s d ABSR

MUX2.[LL,LH,HL,HH] LEXOP rs rt rd hl 0 0 1 MUX

CLS LEXOP 0 rt rd 0 0 0 0 CLS

BITREV LEXOP rs rt rd 0 0 0 0 BITREV

6 5 5 5 2 1 1 1 6

rs, rt, rd Selects general register r0 - r31.

s Selects saturation of result. s=1 indicates that saturation is
performed.
d d=1 indicates that dual operations on 16-bit data are
performed.
hl (for MUX2) 00 = LL: rD = rs[15:00] || rt[15:00]
01 = LH: rD = rs[15:00] || rt[31:16]
10 = HL: rD = rs[31:16] || rt[15:00]
11 = HH: rD = rs[31:16] || rt[31:16]

Lexra Proprietary & Confidential - 56 - Release 1.9

 April 30, 2001 LX5280

5.3.2. MAC Format A

31 26 25 21 20 16 15 11 10 9 8 5 6 50

Assembler LEXOP Lexra

Mnemonic 011 111 rs rt md 0 u gz s d SUBOP

CMULTA LEXOP rs rt md 0 0 0 0 0 CMULTA

DIVA(U) LEXOP rs rt md 0 u 0 0 0 DIVA

MULTA(U) LEXOP rs rt md 0 u 1 0 0 MADDA

MULTA2 LEXOP rs rt md 0 0 1 0 1 MADDA

MADDA(U) LEXOP rs rt md 0 u 0 0 0 MADDA

MADDA2[.S] LEXOP rs rt md 0 0 0 s 1 MADDA

MSUBA(U) LEXOP rs rt md 0 u 0 0 0 MSUBA

MSUBA2[.S] LEXOP rs rt md 0 0 0 s 1 MSUBA

MULNA2 LEXOP rs rt md 0 0 1 0 1 MSUBA

MTA2[.G] LEXOP rs rt md 0 0 g 0 1 MTA

6 5 5 5 1 1 1 1 1 6

base, pointer, Selects general register r0 - r31.

rt
rs, rt Selects general register r0 - r31.
md Selects accumulator, 0NNHL where, NN = m0 - m3
HL 00 = reserved
01 = mNl
10 = mNh
11 = mN
s Selects saturation of result. s=1 indicates that saturation is
performed.
d d=1 indicates that dual operations on 16-bit data are performed.
gz For MTA2, used as “guard” bit. If g=1, bits [39:32] of the
accumulator (pair) are loaded and bits [31:00] are unchanged. If
g=0, all 40 bits [39:00] of the accumulator (or pair) are updated.
For MADDA, MSUBA, used as a “zero” bit. If z = 1, the result is
added to (subtracted from) zero rather than the previous
accumulator value; this performs a MULTA, MULTA2 or
MULNA2. If z = 0, performs a MADDA, MSUBA, MADDA2 or
MSUBA2.
u Treat operands as unsigned values (0 = signed, 1 = unsigned)

Lexra Proprietary & Confidential - 57 - Release 1.9

 April 30, 2001 LX5280

5.3.3. MAC Format B

31 26 25 21 20 16 15 11 10 7 6 5 0

Assembler LEXOP Lexra

Mnemonic 011 111 00000 mt rd so d SUBOP

MFA, MFA2 LEXOP 000000 mt rd so d MFA

RNDA2 LEXOP 000000 mt 0 so 1 RNDA

6 5 5 5 4 1 6

rd Selects general register r0 - r31.

mt Selects accumulator, 0NNHL where, NN = m0 - m3
HL 00 = reserved
01 = mNl
10 = mNh
11 = mN
d d=1 indicates that dual operations on 16-bit data are performed.
so Encoded (“output”) shift amount n = 0 - 8 for RNDA2, MFA,
MFA2 instructions.

Lexra Proprietary & Confidential - 58 - Release 1.9

 April 30, 2001 LX5280

5.3.4. MAC Format C

31 26 25 21 20 16 15 11 10 8 7 6 5 0

Assembler LEXOP Lexra

Mnemonic 011 111 ms mt md 000 s 0 SUBOP

ADDMA[.S] LEXOP ms mt md 000 s 0 ADDMA

SUBMA[.S] LEXOP ms mt md 000 s 0 SUBMA

6 5 5 5 3 1 1 6

mt, ms, md Selects accumulator, 0NNHL where, NN = m0 - m3

HL 00 = reserved
01 = mNl
10 = mNh
11 = reserved
s Selects saturation of result. s=1 indicates that saturation is
performed.

5.3.5. RADIAX MOVE Format and Lexra-Cop0 MTLXC0/MFLXC0

31 26 25 21 20 16 15 11 10 8 7 6 5 0

Assembler LEXOP Lexra

Mnemonic 011 111 00000 rt ru/rk 000 k 0 SUBOP

MFRU LEXOP 00000 rt ru 000 0 0 MFRAD

MTRU LEXOP 00000 rt ru 000 0 0 MTRAD

MFRK LEXOP 00000 rt rk 000 1 0 MFRAD

MTRK LEXOP 00000 rt rk 000 1 0 MTRAD

6 5 5 5 3 1 1 6

rt Selects general register r0 - r31.

rk Selects Radiax Kernel register in MFRK, MTRK instructions —
currently all reserved. However, a Coprocessor Unusable
Exception is taken in User mode if the Cu0 bit is 0 in the CP0
Status register when MFRK or MTRK is executed.
ru Selects Radiax User register in MFRU, MTRU instructions.
00000 cbs0
00001 cbs1
00010 cbs2
00011 reserved
00100 cbe0
00101 cbe1
00110 cbe2
00111 reserved
01xxx reserved
10000 lps0
10001 lpe0
10010 lpc0

Lexra Proprietary & Confidential - 59 - Release 1.9

 April 30, 2001 LX5280

1001 reserved
101xx reserved
11000 mmd
11001 reserved
111xx reserved

31 26 25 21 20 16 15 11 10 0

Assembler COP0
Mnemonic 010 000 Copz rs rt rd 000 0000 0000

MFLX COP0 00011 rt rd 000 0000 0000

MTLX COP0 00111 rt rd 000 0000 0000

6 5 5 5 11

These are not LEXOP instructions. They are variants of the standard MTC0 and MFC0 instructions
that allow access to the Lexra Coprocessor0 Registers listed below. As with any COP0 instruction, a
Coprocessor Unusable Exception is taken in User mode if the Cu0 bit is 0 in the CP0 Status register
when these instructions are executed.

rt Selects general register r0 - r31.

rd Selects Lexra Coprocessor0 register:
00000 ESTATUS
00001 ECAUSE
00010 INTVEC
00011 reserved
001xx reserved
01xxx reserved
1xxxx reserved

Lexra Proprietary & Confidential - 60 - Release 1.9

 April 30, 2001 LX5280

5.3.6. CMOVE Format

31 26 25 21 20 16 15 11 10 9 8 6 5 0

Assembler LEXOP Lexra

Mnemonic 011 111 rs rt rd 00 cond SUBOP

CMVEQZ[.H][.L] LEXOP rs rt rd 00 cond CMOVE

CMVNEZ[.H][.L] LEXOP rs rt rd 00 cond CMOVE

6 5 5 5 2 3 6

rs, rt, rd Selects general register r0 - r31.

cond Condition code for rT operand referenced by the
conditional move.
000 EQZ
001 NEZ
010 EQZ.H
011 NEZ.H
100 EQZ.L
101 NEZ.L
11x reserved

5.3.7. Lexra SUBOP Bit Encodings

Inst[2:0]

Inst[5:3] 0 1 2 3 4 5 6 7

0 CMOVE SLLV SRLV SRAV

1 BITREV MUX CLS ABSR

2 MADDA MSUBA RNDA

3 DIVA CMULTA MFA MTA ADDMA SUBMA

4 ADDR SUBR MFRAD MTRAD

5 MIN MAX SLTR

6 LBP LHP LTP* LWP LBPU LHPU LT*

7 SBP SHP STP* SWP ST*

* Indicates instructions which are implemented only in the LX5280, and not the LX5180 product.

Lexra Proprietary & Confidential - 61 - Release 1.9

 April 30, 2001 LX5280

Lexra Proprietary & Confidential - 62 - Release 1.9

 April 30, 2001 LX5280

6. Integer Multiply-Divide-Accumulate

The integer Multiply-Divide-Accumulate instructions, which are optional on other Lexra processors, are a
standard feature of the LX5280 processor.

6.1. Summary of Instructions

Table 24 provides a summary of the integer Multiply-Divide-Accumulate instructions.

Table 24: Summary of MAC-DIV Instructions.

Mnemonic Operation Description

MTHI HI <- Rs pre-load accumulator, or restore saved HI

MTLO LO <- Rs pre-load accumulator, or restore saved LO

MFHI Rd <- HI read accumulator, or part of 64 bit result

MFLO Rd <- LO read accumulator, or part of 64 bit result

MULT {HI,LO} <- Rs * Rt 32x32 signed multiply 64bit result

MULTU {HI,LO} <- Rs * Rt 32x32 unsigned multiply, 64bit result

MAD {HI,LO}<- {HI,LO} + (Rs * Rt) 32x32 signed multiply, with 64bit signed add
to accum

MADU {HI,LO}<- {HI,LO} + (Rs * Rt) 32x32 unsigned multiply, with 64bit
unsigned add to accum

MSUB {HI,LO}<- {HI,LO} - (Rs * Rt) 32x32 signed multiply, with 64bit signed add
to accum

MSUBU {HI,LO}<- {HI,LO} - (Rs * Rt) 32x32 unsigned multiply, with 64bit
unsigned add to accum

MADH HI <- HI + (Rs[15:0] * Rt[15:0]) 16x16 signed multiply, with 32 bit signed
add to accum

MADL LO <- LO + (Rs[15:0] * Rt[15:0]) 16x16 signed multiply, with 32 bit signed
add to accum

MAZH HI <- 0 + (Rs[15:0] * Rt[15:0]) 16x16 signed multiply, add to pre-zeroed

32bit accum

MAZL LO <- 0 + (Rs[15:0] * Rt[15:0]) 16x16 signed multiply, add to pre-zeroed

32bit accum

MSBH HI <- HI - (Rs[15:0] * Rt[15:0]) 16x16 signed multiply, with 32 bit signed
sub from accum

MSBL LO <- LO - (Rs[15:0] * Rt[15:0]) 16x16 signed multiply, with 32 bit signed
sub from accum

MSZH HI <- 0 - (Rs[15:0] * Rt[15:0]) 16x16 signed multiply, sub from pre-zeroed
32bit accum

MSZL LO <- 0 - (Rs[15:0] * Rt[15:0]) 16x16 signed multiply, sub from pre-zeroed
32bit accum

DIV HI <- Rs%Rt; LO <- Rs/Rt 32 by 32 signed divide with remainder

DIVU HI <- Rs%Rt; LO <- Rs/Rt 32 by 32 unsigned divide with remainder

Lexra Proprietary & Confidential - 63 - Release 1.9

 April 30, 2001 LX5280

The processor may stall if a new MAC instruction is executed while a prior MAC operation is pending.
Table 52 on page 123 indicates the number of cycles that must be present between MAC instructions to avoid
stalls.

6.2. MAC-DIV Instruction Overview

• All ops except Move-to-accumulator and 32-bit multiply-accumulate functions are

supported in M16 mode as well as M32 for best code compression.

• Independent 32-bit HI and LO accumulators for 16-bit Multiply-accumulate allow

optimal performance in the FIR filter, or other applications which allow generation of a
new result while the previous result is pending.

• Multiply-subtract instructions eliminate the need to negate coefficients.

• In case of resource conflicts, hardware manages all hazards simplifying software debug.

• There are no coding restrictions.

Lexra Proprietary & Confidential - 64 - Release 1.9

 April 30, 2001 LX5280

6.3. Op-codes for Standard Mode (32-Bit) MAC Instructions

31 26 25 21 20 16 15 6 5 0

Mnemonic Major Op Base Rt Immediate Subop

MFHI 000000 Rs Rt 0000000000 010000

MTHI 000000 Rs Rt 0000000000 010001

MFLO 000000 Rs Rt 0000000000 010010

MTLO 000000 Rs Rt 0000000000 010011

MULT 000000 Rs Rt 0000000000 011000

MULTU 000000 Rs Rt 0000000000 011001

MAD 011100 Rs Rt 0000000000 00000

MADU 011100 Rs Rt 0000000000 000001

MSUB 011100 Rs Rt 0000000000 000100

MSUBU 011100 Rs Rt 0000000000 000101

DIV 000000 Rs Rt 0000000000 011010

DIVU 000000 Rs Rt 0000000000 011011

MADH 111100 Rs Rt 0000000000 000000

MADL 111100 Rs Rt 0000000000 000010

MAZH 111100 Rs Rt 0000000000 000100

MAZL 111100 Rs Rt 0000000000 000110

MSBH 111100 Rs Rt 0000000000 010000

MSBL 111100 Rs Rt 0000000000 010010

MSZH 111100 Rs Rt 0000000000 010100

MSZL 111100 Rs Rt 0000000000 010110

6 5 5 10 6

Lexra Proprietary & Confidential - 65 - Release 1.9

 April 30, 2001 LX5280

6.4. Op-codes for MIPS-16 (16-Bit) Mode MAC Instructions

15 11 10 8 7 5 4 0

Mnemonic major op base rt subop

MFHI 11101 rx ry 10000

MTHI not supported by MIPS-16 architecture

MFLO 11101 rx ry 10010

MTLO not supported by MIPS-16 architecture

MULT 11101 rx ry 11000

MULTU 11101 rx ry 11001

MAD not supported by MIPS-16 architecture

MADU not supported by MIPS-16 architecture

MSUB not supported by MIPS-16 architecture

MSUBU not supported by MIPS-16 architecture

DIV 11101 rx ry 11010

DIVU 11101 rx ry 11011

MADH 11111 rx ry 00000

MADL 11111 rx ry 00010

MAZH 11111 rx ry 00100

MAZL 11111 rx ry 00110

MSBH 11111 rx ry 10000

MSBL 11111 rx ry 10010

MSZH 11111 rx ry 10100

MSZL 11111 rx ry 10110

5 3 3 5

Lexra Proprietary & Confidential - 66 - Release 1.9

 April 30, 2001 LX5280

6.5. Non-Standard Instruction Descriptions

Table 25: 16-bit Multiply and Multiply-Accumulate Instructions

Signed 16-bit Multiply MAZH rS, rT

to {HI,LO} MAZL rS, rT
The contents of rS[15:0] is multiplied by rT[15:0], treating the operands as signed
2's complement values. The 32-bit product is stored in the {HI,LO} register.
{HI,LO} <- 0 + Rs * Rt

Signed 16-bit Multiply- MADH rS, rT

Accumulate to {HI,LO} MADL rS, rT
The contents of rS[15:0] is multiplied by rT[15:0], treating the operands as signed
2's complement values. The 32-bit product is added to {HI,LO}, ignoring any
overflow. The result is stored in the {HI,LO} register.
{HI,LO} <- {HI,LO} + Rs * Rt

Signed 16-bit Multiply- MSZH rS, rT

Negate to {HI,LO} MSZL rS, rT
The contents of rS[15:0] is multiplied by rT[15:0], treating the operands as signed
2's complement values. The 32-bit product is negated (subtracted from zero) and
stored in the {HI,LO} register.
{HI,LO} <- 0 - Rs * Rt

Signed 16-bit Multiply- MSBH rS, rT

Subtract from {HI,LO} MSBL rS, rT
The contents of rS[15:0] is multiplied by rT[15:0], treating the operands as signed
2's complement values. The 32-bit product is subtracted from {HI,LO}, ignoring
any overflow. The result is stored in the {HI,LO} register.
{HI,LO} <- {HI,LO} - Rs * Rt

Lexra Proprietary & Confidential - 67 - Release 1.9

 April 30, 2001 LX5280

Table 26: 32-Bit Multiply-Accumulate Instructions

Signed 32-bit Multiply- MAD rS, rT

Accumulate The contents of rS is multiplied by rT, treating the operands as signed 2's
complement values. The 64-bit product is added to the concatenation HI and LO
to form a 64-bit result ignoring any overflow. The upper 32-bits of the 64-bit result
are stored in the HI register. The lower 32-bits are stored in the LO register.
t <- {HI,LO} + Rs * Rt
LO <- t<31:0>
HI <- t<63:32>

32-bit Multiply- MADU rS, rT

Accumulate The contents of rS is multiplied by rT, treating the operands as unsigned values.
The 64-bit product is added to the concatenation HI and LO to form a 64-bit result
ignoring any overflow. The upper 32-bits of the 64-bit result are stored in the HI
register. The lower 32-bits are stored in the LO register.
t <- {HI,LO} + Rs * Rt
LO <- t<31:0>
HI <- t<63:32>

Signed 32-bit Multiply- MSUB rS, rT

Subtract The contents of rS is multiplied by rT, treating the operands as signed 2's
complement values. The 64-bit product is subtracted from the concatenation HI
and LO to form a 64-bit result ignoring any overflow. The upper 32-bits of the 64-
bit result are stored in the HI register. The lower 32-bits are stored in the LO
register.
t <- {HI,LO} - Rs * Rt
LO <- t<31:0>
HI <- t<63:32>

32-bit Multiply-Subtract MSUBU rS, rT

The contents of rS is multiplied by rT, treating the operands as unsigned values.
The 64-bit product is subtracted from the concatenation HI and LO to form a 64-
bit result ignoring any overflow. The upper 32-bits of the 64-bit result are stored in
the HI register. The lower 32-bits are stored in the LO register.
t <- {HI,LO} - Rs * Rt
LO <- t<31:0>
HI <- t<63:32>

Notes:
The 32-bit op-codes are unchanged (from the MIPS-I standard) for the existing MULT, DIV, MF, and MT
instructions. The MAD, MADU, MSUB, and MSUBU are new Special2 opcodes, also standard to several
processors. In M32 mode, the new instructions are all R-format with bits 31:26 = 6'b111100. Bits 5:0
determine the specific operation, as shown. In M16 mode, the new instructions are all RR-format with bits
15:11 = 5'b11111. Bits 4:0 determine the specific operation, as shown in Section 6.4.

The upper 16 bits of both operand registers are ignored by 16-bit instructions.

The MxxH and MxxL instructions can be freely interleaved. That is, adds and subtracts from either
accumulator can be combined in a sequence with the two accumulators functioning "in parallel."

The MxZx instructions can be used as stand-alone 16-bit signed multiply. This removes the need for a
"MTHI, zero" instruction at the beginning of a multiply-accumulate sequence, for example:

Lexra Proprietary & Confidential - 68 - Release 1.9

 April 30, 2001 LX5280

MAZH r1,r2
MADH r3,r4
MADH r5,r6
MADH r7,r8
any op that doesn't write HI
any op that doesn't write HI
MFHI r9
In the above sequence, the two non-HI ops are not necessary for correct operation but the pipeline will stall if
they are not used, so it is more efficient to perform useful work in those slots.

For the MULTx, MADx or MSUBx instructions, the most efficient use is:

MULTx r1,r2
MADx r3,r4
MSUBx r5,r6
any op that doesn't write HI or LO
any op that doesn't write HI or LO
any op that doesn't write HI or LO
MFLO r7 /* LO or HI is available this cycle*/
MFHI r8

6.6. Accessing HI and LO after multiply instructions

The MFLO (MFHI) instruction reads the contents of the LO (HI) register during the E cycle of the pipeline.
The following descriptions indicate how the latency of the multiply instructions affects the usage of the MF
instructions. The most efficient sequence is shown. If the MF instruction is coded earlier, the correct result
will still be obtained because the hardware will stall the MF instruction in the E-cycle until the result is valid.

During the E cycle of any multiply operation, the initial operands are re-coded and loaded into the
MANDHW and MIERHW (MBOOTH) registers. For the MULTx operations, the multiply cycles can be
labeled M1 through M3. Then the following timing diagram is valid:

MULTx I S E M1 M2 M3
LO/HI valid X
any op I S E M W
any op I S E M W
MFLO I S E M W
or MFHI I S E M W
For the MADx operations, the pipeline cycles after E can be labeled as C (carry save), and A (accumulate).
Then the following timing diagram is valid:

MAZH0 I S E C A
MADH1 I S E C A
MADH2 I S E C A
MADH3 I S E C A
any op I S E M W
any op I S E M W
MFHI I S E M W
HI contains A0 A1 A2 A3

Lexra Proprietary & Confidential - 69 - Release 1.9

 April 30, 2001 LX5280

6.7. Divider Overview and Register Usage

Given a dividend DEND, and divisor DVSR, the divider generates a quotient QUOT and remainder REM
that satisfy the following conditions, regardless of the signs of DEND and DVSR:

DEND = DVSR * QUOT + REM,

0 <= abs(REM) < abs(DVSR)
where REM and DEND have the same sign.

It is worth noting that the requirement that REM and DEND have the same sign is not universally accepted if
DEND and DVSR are not both positive. (For example the Modula-3 language expects: -5DIV3=-2, -
5MOD3=+1, whereas the divider generates QUOT=-1, REM=-2 in agreement with FORTRAN and others.)
These examples show the possible combinations of signs:

DEND DVSR QUOT REM

---- ---- ---- ---
+19 +5 +3 +4
-19 +5 -3 -4
+19 -5 -3 +4
-19 -5 +3 -4
The divider is an iterative circuit that generates 2 quotient bit per cycle, with an additional 3 cycles required
due to pipelining considerations.

Thus the pipeline flow of a division instruction and the most efficient subsequent read of the quotient (using
MFLO) is as shown in the following diagram, assuming that all the intervening instructions complete in one
cycle. If the MFLO is issued earlier it will stall until the divide completes. Less than 19 instructions may be
issued if some of them take more than one cycle to complete (due to cache misses or data dependent stalls, for
example).

DIV I S E D0 D1 D2 ... D17 D18

...
18 cycles
...
MFLO I S E M W

Lexra Proprietary & Confidential - 70 - Release 1.9

 April 30, 2001 LX5280

7. LX5280 Local Memory

7.1. Local Memory Overview

This chapter describes how memories are configured and connected to the LX5280 using the Local Memory
Interfaces (LMIs). This section provides a brief summary of the conventions and supported memories.
Section 7.2 describes the control register that allows software control over certain aspects of the LMIs. The
subsequent sections cover each of the LMIs in detail.

This chapter also discusses configuration options and the ports that customers must access to connect
application specific RAM and ROM devices that are used by the LX5280 LMIs. All of the signals between
the processor core, the LMIs, RAMs and the system bus controller are automatically configured by lconfig,
the LX5280 configuration tool. Lconfig also produces documentation of the exact RAMs required for the
chosen configuration settings, and writes RAM models used for RTL simulation.

The LMIs connect to RAMs that service the LX5280 processor’s local instruction and data busses. The LMIs
also provide the pathways from the processor to the system bus. The LX5280 includes an LMI for each of the
local memory types. The sizes of the RAMs and ROMs are customer selectable. The LX5280 LMIs directly
support synchronous RAMs that register the address, write data, and control signals at the RAM inputs. The
LMIs also supply redundant read enable and chip select lines for each RAM, which may be required for some
RAM types. ROMs may also be connected, but may require a customer supplied address register at the
address inputs.

Lexra supplies an integration layer for the LMIs and the memory devices connected to them. In this layer,
memory devices are instanced as generic modules satisfying the depth and width requirements for each
specific memory instance. The lconfig utility supplies a summary of the memory devices required for the
chosen configuration. In most cases, customers simply need to write a wrapper that connects the generic
module port list to a technology specific RAM instance inside the RAM wrapper.

The LX5280 is configurable for a 16, 32, 64, or 128-byte cache line size. The tag store RAM sizes shown in
the tables of this chapter assume a 16-byte line size. The documentation produced by lconfig indicates the
required tag RAMs for the selected configuration options, including the line size. As a general rule, a
doubling of the line size results in halving the tag store depth.

The valid bits within tag stores are automatically cleared by the LMIs upon reset. The data cache implements
a write-through protocol. Caches do not snoop the system bus. The LX5280 is configurable to work with
RAMs with a write granularity of 8 bits (byte) or 32 bits (word). Byte write granularity results in more
efficient operation of store byte and store half-word instructions.

Table 27 summarizes the LMIs that can be integrated on the local busses.

Table 27: Local Memory Interface Modules

Name Description

ICACHE Direct mapped or two-way set associative instruction cache.

IMEM Instruction RAM.

IROM Instruction ROM.

DCACHE Direct mapped data cache.

DMEM Data RAM or ROM.

Lexra Proprietary & Confidential - 71 - Release 1.9

 April 30, 2001 LX5280

7.2. Cache Control Register: CCTL

CCTL. CP0 General Register Address = 20

31-8 7 6 5 4 3-2 1 0

Reserved IROMOff IROMOn IMEMOff IMEMFill ILock IInval DInval

When reading this register, the contents of the Reserved bits are undefined. When writing this register, the
contents of the Reserved bits should be preserved.

Changes in the contents of the CCTL register are observed in the W stage. However, these changes affect
instruction fetches currently in progress in the I stage, and data load or store operations in progress in the M
stage.

The IROMOn and IROMOff bits of the CCTL register control the and use of the optional local IROM
memory configured into the LX5280. When IROM is present and the LX5280 is reset, the LMI enables
access to the IROM. A transition from 0 to 1 on IROMOff disables the IROM, allowing instruction references
to be serviced IMEM, ICACHE or the system bus. A transition from 0 to 1 on IROMOn enables the IROM.

The IMEMFill and IMEMOff bits of the CCTL register control the contents and use of any local IMEM
memory configured into the LX5280. When the LX5280 is reset, the LMI clears an internal register to
indicate that the entire IMEM LMI contents are invalid. When IMEM is invalid, all cacheable fetches from
the IMEM region will be serviced by the instruction cache, if an instruction cache is present.

A transition from 0 to 1 on IMEMFill causes the LMI to initiate a series of line read operations to fill the
IMEM contents. The addresses used for these reads are defined by the configured BASE and TOP addresses
of the IMEM, described in Section 7.4. The processor stalls while the entire IMEM contents are filled by the
LMI. Thereafter, the LMI sets its internal IMEM valid bit and will service any access to the IMEM range
from the local IMEM memory. The time that an IMEM fill takes to complete is the number of line reads
needed to fill the IMEM range, multiplied by the latency of one line read, assuming there is no other system
bus traffic.

A transition from 0 to 1 on IMEMOff causes the LMI to clear its internal IMEM valid bit. Subsequent
cacheable fetches from the IMEM region will be serviced by the instruction cache. To use the IMEM again,
an application must re-initialize the IMEM contents through the IMEMFill bit of the CCTL register.

The ILock field controls set locking in the two set associative instruction cache. When ILock is 00 or 01, the
instruction cache operates normally. When ILock is 10, all cached instruction references are forced to occupy
set 1. The hardware will invalidate lines in set 0 if necessary to accomplish this. When ILock is 11, lines in set
1 are never displaced – i.e. they are locked in the cache. Set 0 is used to hold other lines as needed.

To utilize the cache locking feature, software should execute at least one pass of critical subroutines or loops
with ILock set to 10. After this has been done, ILock should be set to 11 to lock the critical code into set 1,
and use set 0 for other code.

The IInval and DInval fields control hardware invalidation of the instruction cache and data cache. A
transition from 0 to 1 on IInval will initiate a hardware invalidation sequence of the entire instruction cache.
Likewise, a 0 to 1 transition on DInval will initiate a hardware invalidation sequence of the entire data cache.
The DMEM, if present, is unaffected by this operation.

The hardware invalidation sequence for the instruction and data caches requires one cycle per cache line to
complete.

Depending on the circumstances, software may be able to employ an alternative to a full invalidation of the
data cache. If a small number of lines must be invalidated, software may perform cached reads from aliases of

Lexra Proprietary & Confidential - 72 - Release 1.9

 April 30, 2001 LX5280

the memory locations of concern. This displaces data in the addressed locations of the data cache, even if they
do not encache the affected memory location.

Another alternative, if the affected memory location has an alias in uncacheable (KSEG1) space, is to simply
perform an uncached read of the affected memory locations. If the location is resident in the data cache it will
be invalidated. This method has the advantage of not displacing data in the cache unless it is absolutely
necessary to maintain coherency. Note that a write to a KSEG1 address has no affect on the contents of the
data cache.

With either of these two alternatives, it is only necessary to reference one word of each affected cache line.

7.3. Instruction Cache (ICACHE) LMI

The ICACHE LMI supplies the interface for a direct mapped or two-way set associative instruction cache
attached to the LX5280 local bus. The degree of associativity is specified through lconfig. The ICACHE LMI
participates in cacheable instruction fetches, but only if the address is not claimed by the IMEM module. The
configurations supported by ICACHE, and the synchronous RAMs required for each, are summarized in
Table 28.

The instruction store for the two-way ICACHE consists of two 64-bit wide banks, with separate write-enable
controls. The tag store consists of one RAM bank with tag and valid bits for set 0, and a second RAM for set
1 that holds the tag, valid, LRU (Least Recently Used), and lock bits. When a miss occurs in the two-way
ICACHE, the LRU bit is examined to determine which element of the set to replace, with element 0 being
replaced if LRU is 0, and element 1 being replaced if LRU is 1. The state of the LRU bit is then inverted. To
optimize the timing of cache reads, the two-way ICACHE uses the state of the LRU bit to determine which
element should be returned to the CPU. In the following cycle, the ICACHE determines if the correct element
was returned. If not, the ICACHE takes an extra cycle to return the correct element to the CPU and inverts the
LRU bit.

Table 28: ICACHE Configurations

Configuration ICACHE_INST RAM ICACHE_TAG RAM

no instruction cache no RAM required no RAM required

1K bytes, 2-way 2 x 64 x 64 bits 32 x 24 and 32 x 26 bits

2K bytes, 2-way 2 x 128 x 64 bits 64 x 23 and 64 x 25 bits

4K bytes, 2-way 2 x 256 x 64 bits 128 x 22 and 128 x 24 bits

8K bytes, 2-way 2 x 512 x 64 bits 256 x 21 and 256 x 23 bits

16K bytes, 2-way 2 x 1,024 x 64 bits 512 x 20 and 512 x 22 bits

32K bytes, 2-way 2 x 2,048 x 64 bits 1,024 x 19 and 1,024 x 21 bits

64K bytes, 2-way 2 x 4,096 x 64 bits 2,048 x 18 and 2,048 x 20 bits

1K bytes, direct mapped 128 x 64 bits 64 x 23 bits

2K bytes, direct mapped 256 x 64 bits 128 x22 bits

4K bytes, direct mapped 512 x 64 bits 256 x 21 bits

8K bytes, direct mapped 1,024 x 64 bits 512 x 20 bits

Lexra Proprietary & Confidential - 73 - Release 1.9

 April 30, 2001 LX5280

Configuration ICACHE_INST RAM ICACHE_TAG RAM

16K bytes, direct mapped 2,048 x 64 bits 1,024 x 19 bits

32K bytes, direct mapped 4,096 x 64 bits 2,048 x 18 bits

64K bytes, direct mapped 8,192 x 64 bits 4,096 x 17 bits

Table 29 lists the ICACHE signals that are connected to application specific modules. The IC_ prefix
indicates signals that are driven by the ICACHE LMI module and received by the RAMs. The ICR_ prefix
indicates signals that are driven by the ICACHE RAMs and received by the ICACHE LMI. Lexra supplies
the Verilog module that makes all required connections to these wires. The width of the index and data lines
depends upon the RAM connected to the LMI, and can be inferred from the Table 28.

Table 29: ICACHE RAM Interfaces

Signal Description

IC_TAGINDEX Tag and state RAM address (line).

ICR_TAGRD0 Tag and state RAM element 0 read path.

IC_TAGWR0 Tag and state RAM element 0 write path.

ICR_TAGRD1 Tag and state RAM element 1 read path.

IC_TAGWR1 Tag and state RAM element 1 write path.

IC_TAG0WE<N> Tag 0 RAM write enable.

IC_TAG0RE<N> Tag 0 RAM read enable.

IC_TAG0CS<N> Tag 0 RAM chip select.

IC_TAG1WE<N> Tag 1 RAM write enable.

IC_TAG1RE<N> Tag 1 RAM read enable.

IC_TAG1CS<N> Tag 1 RAM chip select.

IC_INSTINDEX Instruction RAM address (word).

ICR_INST0RD Instruction RAM element 0 read path.

ICR_INST1RD Instruction RAM element 1 read path.

IC_INSTWR Instruction RAM write path (to both elements).

IC_INST0WE<N>[1:0] Instruction RAM 0 write enable.

IC_INST0RE<N> Instruction RAM 0 read enable.

IC_INST0CS<N> Instruction RAM 0 chip select.

IC_INST1WE<N>[1:0] Instruction RAM 1 write enable.

IC_INST1RE<N> Instruction RAM 1 read enable.

IC_INST1CS<N> Instruction RAM 1 chip select.

Note: <N> designates an available active-low version of a signal.

Lexra Proprietary & Confidential - 74 - Release 1.9

 April 30, 2001 LX5280

7.4. Instruction Memory (IMEM) LMI

The IMEM LMI supplies the interface for an optional local instruction store. The IMEM serves a fixed range
of the physical address space, determined by configuration settings in lconfig. The IMEM contents are filled
and invalidated under the control of the CP0 CCTL register, described in Section 7.2, Cache Control
Register: CCTL. The IMEM module services instruction fetches that falls within its configured range. The
IMEM is a convenient, low-cost alternative to a cache that makes instruction memory available to the core for
high-speed access.

The configurations supported by IMEM, and the synchronous RAMs required for each, are summarized in
Table 30.

Table 30: IMEM Configurations

Configuration IMEM_INST RAM

no local instruction RAM no RAM required

1K bytes 128 x 64 bits

2K bytes 256 x 64 bits

4K bytes 512 x 64 bits

8K bytes 1,024 x 64 bits

16K bytes 2,048 x 64 bits

32K bytes 4,096 x 64 bits

64K bytes 8,192 x 64 bits

128K bytes 16,384 x 64 bits

256K bytes 32,768 x 64 bits

Table 31 lists the IMEM signals that are connected to application specific modules. The IW_ prefix indicates
signals that are driven by the IMEM LMI module and received by RAMs. The IWR_ prefix indicates signals
that are driven by RAMs and received by the IMEM LMI. The CFG_ prefix identifies configuration ports on
the IMEM LMI that are typically wired to constant values. The width of the index and data lines depends
upon the RAM connected to the LMI, and can be inferred from Table 30.

The CFG_ wires define where the IMEM is mapped into the physical address space. This configuration
information defines the local bus address region of the IMEM. It also determines the address of the external
resources which are accessed when an IMEM miss occurs. The lconfig utility supplied by Lexra will verify
that the configured address range does not interfere with other regions defined for LX5280. The size of the
memory region must be a power of two, and must be naturally aligned.

Table 31: IMEM RAM Interfaces

Signal Description

IW_INSTINDEX IMEM index.

IWR_INSTRD Instruction read data.

IW_INSTWR Instruction write data.

IW_INSTWE<N>[1:0] Instruction RAM write enable.

Lexra Proprietary & Confidential - 75 - Release 1.9

 April 30, 2001 LX5280

Signal Description

IW_INSTRE<N> Instruction RAM read enable.

IW_INSTCS<N> Instruction RAM chip select.

CFG_IWBASE[31:10] Configured base address (modulo 1K bytes).

CFG_IWTOP[17:10] Configured top address (bits that may differ from base).

Note: <N> designates an available active-low version of a signal.

7.5. Instruction ROM (IROM) LMI

The IROM LMI supplies the interface for an optional read-only local instruction store. The IROM serves a
fixed range of the physical address space, determined by configuration settings in lconfig. IROM may be
disabled via a hardware configuration pin, CFG_IROFF. IROM may also be enabled and disabled under
software control as described in Section 7.2, Cache Control Register: CCTL. The IROM is a convenient,
low-cost alternative to a cache that makes read-only instruction memory available to the core for high-speed
access.

The configurations supported by IROM, and the synchronous ROMs required for each, are summarized in
Table 32.

Table 32: IROM Configurations

Configuration IROM_DATA

no local instruction RAM no ROM required

1K bytes, direct mapped 128 x 64 bits

2K bytes, direct mapped 256 x 64 bits

4K bytes, direct mapped 512 x 64 bits

8K bytes, direct mapped 1,024 x 64 bits

16K bytes, direct mapped 2,048 x 64 bits

32K bytes, direct mapped 4,096 x 64 bits

64K bytes, direct mapped 8,192 x 64 bits

128K bytes, direct mapped 16,384 x 64 bits

256K bytes, direct mapped 32,768 x 64 bits

Table 33 lists the IROM signals that are connected to application specific modules. The IR_ prefix indicates
signals that are driven by the IROM LMI module and received by the ROM. The IRR_ prefix indicates
signals that are driven by ROM and received by the IROM LMI. The CFG_ prefix identifies configuration
ports on the IROM LMI that are typically wired to constant values. Lexra supplies the Verilog module that
makes all required connections to these wires. The width of the index and data lines depends upon the ROM
connected to the LMI, and can be inferred from Table 31.

The CFG_ wires define where IROM is mapped into the physical address space. This configuration
information defines the local bus address region of the IROM. It also determines the address of the external
resources which are accessed when an IROM miss occurs. The lconfig utility supplied by Lexra will verify
that the configured address range does not interfere with other regions defined by the LX5280. Note that the

Lexra Proprietary & Confidential - 76 - Release 1.9

 April 30, 2001 LX5280

size of the memory region must be a power of two, and must be naturally aligned.

Table 33: IROM ROM Interfaces

Signal Description

IR_INSTINDEX IROM index.

IRR_INSTRD Instruction read data.

IR_INSTRE<N> Instruction ROM read enable.

IR_INSTCS<N> Instruction ROM chip select.

CFG_IRBASE[31:10] Configured base address (modulo 1K bytes).

CFG_IRTOP[17:10] Configured top address (bits that may differ from base).

Note: <N> designates an available active-low version of a signal.

7.6. Direct Mapped Write Through Data Cache (DCACHE) LMI

The DCACHE LMI supplies the interface for a direct mapped, write through data cache attached to the
LX5280 local bus. The DCACHE LMI participates in cacheable data reads and writes, but only if the address
is not claimed by the DMEM LMI. The configurations supported by DCACHE, and the synchronous RAMs
required for each, are summarized in Table 34.

The direct mapped DCACHE module services word or twin-word read requests from the core in one cycle
when the request hits the cache. Byte or half-word reads that hit the data cache require an extra cycle for
alignment. The data cache can stream word and twin-word reads or writes that hit the cache at the rate of one
per cycle. If the LX5280 is configured to work with RAMs that have word write granularity, byte or half-
word writes that follow any write by one cycle and hit the cache require an extra cycle to merge the data with
the current cache contents. Alternatively, the LX5280 can be configured to work with RAMs support byte
write granularity, which eliminates the extra cycle. See Appendix C, LX5280 Pipeline Stalls, for detailed
descriptions of these and other pipeline stall conditions.

Writes that are serviced by the data cache may require extra time to be serviced by the LBC if its write buffer
is full. Also, when a cache write operation is immediately followed by a cache read, the cache must delay the
read for one cycle while the write completes.

When a miss occurs, the cache obtains a cache line (4, 8, 16, or 32 words) of data from the Lexra Bus
Controller (LBC). Write operations that hit the data cache are simultaneously written into the cache and
forwarded to the write buffer of the LBC. Thus, if the core subsequently reads the data, it will likely be
available from the cache. For main memory systems that support byte writes, all data writes that miss the
cache are forwarded to the write buffer of the LBC, without disturbing any data currently in the cache. For
main memory systems that can only write with word granularity, a byte or half-word write that misses the
cache causes the cache to perform a line fill from main memory. The cache then merges the partial write data
with the full word data obtained from memory, and writes the word to the system bus.

Lexra Proprietary & Confidential - 77 - Release 1.9

 April 30, 2001 LX5280

Table 34: DCACHE Configurations

Configuration DCACHE_DATA RAM DCACHE_TAG RAM

no data cache no RAM required no RAM required

1K bytes, direct mapped 128 x 64 bits 64 x 23 bits

2K bytes, direct mapped 256 x 64 bits 128 x 22 bits

4K bytes, direct mapped 512 x 64 bits 256 x 21 bits

8K bytes, direct mapped 1,024 x 64 bits 512 x 20 bits

16K bytes, direct mapped 2,048 x 64 bits 1,024 x 19 bits

32K bytes, direct mapped 4,096 x 64 bits 2,048 x 18 bits

64K bytes, direct mapped 8,192 x 64 bits 4,096 x 17 bits

Table 35 lists the DCACHE signals that are connected to application specific modules. The DC_ prefix
indicates signals that are driven by the DCACHE LMI module and received by the RAMs. The DCR_ prefix
indicates signals that are driven by the DCACHE RAMs and received by the DCACHE LMI. Lexra supplies
the Verilog module that makes all required connections to these wires. The width of the index and data lines
depends upon the RAM connected to the LMI, and can be inferred from Table 34.

Table 35: DCACHE RAM Interfaces

Signal Description

DC_TAGINDEX Tag and state RAM address.

DCR_TAGRD Tag and state RAM read path.

DC_TAGWR Tag and state RAM write path.

DC_TAGWE<N> Tag and state RAM write enable.

DC_TAGRE<N> Tag and state RAM read enable.

DC_TAGCS<N> Tag and state RAM chip select.

DC_DATAINDEX Data RAM address (word).

DCR_DATARD Data RAM read path.

DC_DATAWR Data RAM write path.

DC_DATAWE<N>[1:0] Data RAM write enable.

DC_DATARE<N> Data RAM read enable.

DC_DATACS<N> Data RAM chip select.

Note: <N> designates an available active-low version of a signal.

7.7. Scratch Pad Data Memory (DMEM) LMI

The DMEM LMI supplies the interface for a scratch pad data RAM attached to the LX5280 local bus. The
DMEM module services in any cacheable or uncacheable data read or write operation that falls within its
configured range.

Lexra Proprietary & Confidential - 78 - Release 1.9

 April 30, 2001 LX5280

Byte or half-word reads that hit the DMEM require an extra cycle for alignment. DMEM can stream word
and twin-word reads or writes that hit DMEM at the rate of one per cycle. If the LX5280 is configured to
work with RAMs that have word write granularity, byte or half-word writes that follow any write by one
cycle and hit DMEM require an extra cycle to merge the data with the current DMEM contents. Alternatively,
the LX5280 can be configured to work with RAMs support byte write granularity, which eliminates the extra
cycle. See Appendix C, LX5280 Pipeline Stalls, for detailed descriptions of these and other pipeline stall
conditions. Also, because a write operation to the DMEM is never sent to the LBC, writes to DMEM will not
cause the LBC to stall the processor due to a full write buffer condition.

The DMEM configurations and the synchronous RAMs required for each are summarized in the Table 36.

Table 36: DMEM Configurations

Configuration DMEM_DATA RAM (64-bit) DMEM_DATA RAM (128-bit)

no local data RAM no RAM required no RAM required

1K bytes 128 x 64 bits 64 x 128 bits

2K bytes 256 x 64 bits 128 x 128 bits

4K bytes 512 x 64 bits 256 x 128 bits

8K bytes 1,024 x 64 bits 512 x 128 bits

16K bytes 2,048 x 64 bits 1,024 x 128 bits

32K bytes 4,096 x 64 bits 2,048 x 128 bits

64K bytes 8,192 x 64 bits 4,096 x 128 bits

128K bytes 16,384 x 64 bits 8,192 x 128 bits

256K bytes 32,768 x 64 bits 16,384 x 128 bits

Table 37 lists the DMEM signals that are connected to application specific modules. The DW_ prefix
indicates signals that are driven by the DMEM LMI module and received by RAMs. The DWR_ prefix
indicates signals that are driven by RAMs and received by the DMEM LMI. The CFG_ prefix identifies
configuration ports on the DMEM LMI that are typically wired to constant values. The width of the index and
data lines depends upon the RAM connected to the LMI, and can be inferred from Table 36.

The CFG_ wires define where DMEM is mapped into the physical address space. It is not possible for any
DMEM reference to result in an operation on the system bus. The lconfig utility supplied by Lexra will verify
that the configured address range does not interfere with other regions defined for LX5280. The size of the
memory region must be a power of two, and must be naturally aligned.

The DMEM LMI can also be used as a ROM controller simply by tying off the write enable and data input
lines in the RAM wrapper, and instancing a ROM in the RAM wrapper.

Table 37: DMEM RAM Interfaces

Signal Description

DW_DATAINDEX Decoded data RAM index.

DWR_DATARD Data RAM read data.

DW_DATAWR Data RAM write data.

Lexra Proprietary & Confidential - 79 - Release 1.9

 April 30, 2001 LX5280

Signal Description

DW_DATAWE<N> Data RAM write enable.

DW_DATARE<N> Data RAM read enable

DW_DATACS<N> Data RAM chip select

CFG_DWBASE[31:10] Configured base address (modulo 1K bytes).

CFG_DWTOP[17:10] Configured top address (bits that may differ from base).

Note: <N> designates an available active-low version of a signal.

Lexra Proprietary & Confidential - 80 - Release 1.9

 April 30, 2001 LX5280

8. LX5280 System Bus

8.1. Connecting the LX5280 to internal devices

The Lexra System Bus (LBus) is the connection between the LX5280 and other internal devices, such as
system memory, USB, IEEE-1394 (Firewire), and an external bus interface. The LBC uses a protocol similar
to that of the Peripheral Component Interface (PCI) bus. This is a well-known and proven architecture.
Adding new devices to the Lexra Bus is straightforward and the performance approaches the highest that can
be achieved without adding a great deal of complexity to the protocol.

Lexra Lexra System Bus

Bus Bus
Controller Bridge
Bus to CPU (LBC) External Bus
and Local (e.g. PCI)
Memory FireWire USB
Interfaces

Figure 6: Lexra System Bus Diagram

The Lexra bus supports multiple masters. This allows for mastering I/O controllers with DMA engines to be
connected to the bus. The bus has a pended architecture, in which a master holds the bus until all the data is
transferred. This simplifies the design of user-supplied bus agents and reduces latency for cache miss
servicing.

The Lexra bus is a synchronous bus. Signals are registered and sampled at the positive edge of the bus clock.
Certain logical operations may be made to the sampled signals and then new signals can be driven
immediately, such as for address decoding. This allows for same-cycle turn-around. The LBC provides an
optional asynchronous interface between the CPU and the Lexra bus, allowing the Lexra bus speed can be set
to be any speed equal to or less than the CPU clock frequency.

The Lexra bus data path for the LX5280 is 32 bits wide. Therefore, the bus can transfer one word, halfword,
or byte in one bus clock. The bus supports line and burst transfers in which several words of data are
transferred. The Lexra bus accomplishes this by transferring words of data from incremental addresses on
successive clock cycles.

The LBC contains a write buffer. When the CPU issues a write request to a Lexra Bus device, the address and
data are saved in the buffer and sent to the device sometime later. The CPU can continue processing, having
safely assumed that the write will eventually happen. This is described more thoroughly in Section 8.7.2.

The LBC drives enabling signals to control muxes or tristate buffers. This allows the Lexra bus to have either
a bi-directional or point-to-point topology.

8.2. Terminology

The Lexra bus borrows terminology from the PCI bus specification, on which the Lexra bus is partially based.

Bus transactions take place between two bus agents. One bus agent requests the bus and initiates a transfer.
The second responds to the transfer.

Lexra Proprietary & Confidential - 81 - Release 1.9

 April 30, 2001 LX5280

The agent initiating a transfer is called the bus initiator. It is also referred to as the bus master. Both terms are
used interchangeably in this document.

The responding agent is known as the bus target. It samples the address when it is valid, and determines if the
address is within the domain of the device. If so, indicates as such to the initiator and becomes the target.

A read transfer is a bus operation whereby the master requests data from the target.

A write transfer is a bus operation whereby the master requests to send data to the target.

A single-cycle bus operation is used to transfer one word, halfword, or byte of data. This amount of data can
be transferred in one bus cycle, not including the address cycle and device latencies.

A line transfer is a read or write operation where an entire cache line of data is transferred in successive
cycles as fast as the initiator and target can send/receive the data.

A burst transfer is a read or write operation where a large amount of data needs to be sent. The initiator
presents a starting address and data is transferred starting at that address in successive cycles; for each word
transferred, the address is incremented by the devices internally.

Some signals on the Lexra bus are active low. That is, they are considered logically true when they are
electrically low and logically false when electrically high. A device asserts a signal when it drives it to its
logical true electrical state.

8.3. Bus Operations

The purpose of the Lexra bus is to connect together the various components of the system, including the
LX5280 CPU, main system memory, I/O devices, and external bus bridges. Different devices have different
transfer requirements. For example, the LX5280 CPU will request the bus to fetch a cache line of data from
memory. I/O devices will request large blocks of data to be sent to and from memory. The Lexra bus supports
the various types of transfers needed by both I/O and the processor.

The six types of bus operations are single-cycle read, line read, burst read, single-cycle write, line write
(though this won’t be used by the LX5280 core) and burst write.

8.3.1. Single-Cycle Read

The single-cycle read operation reads a single word, halfword, or byte from the target device. This operation
is usually used by the CPU to read data from uncachable address space. (If the read address was in cacheable
address space, either a hit would occur resulting in no bus activity, or a miss would occur resulting in a read
line transaction.)

8.3.2. Read Line

The read line operation reads a sequence of data from memory corresponding to the size of a cache line. The
cache line size affects how many cycles are required to transfer the full line. The LX5280 and the Lexra bus
support a configurable line size, specified through lconfig. The default line size of four words (16 bytes) is
assumed here.

There are two ways that the target could transfer the data back to the initiator. The conventional way is to
transfer four words of data in sequence, starting at the nearest 16-byte-aligned address smaller or equal to the
address that the initiator drives. In other words, the target starts the transfer at the beginning of the line
containing the requested address.

Some memory devices may implement a performance optimization called desired-word-first. If the address is

Lexra Proprietary & Confidential - 82 - Release 1.9

 April 30, 2001 LX5280

not aligned to a 16-byte boundary, then the first data returned by the target is the word corresponding to the
address instead of the first word of the line. The second word is the next sequential word of data and so on. At
the end of the line, the target wraps around and returns the first word of line.

The LX5280 supports two ways of incrementing the address of a line refill. One is by linear wrap, where the
address is simply incremented by one. The other is by interleaved wrap, where the next address is determined
by the logical xor of the cycle count and the first word address. The interleave sequence is shown in the table
below. The low-order address bits 3:2 for the first data beat are the obtained from the address of the line read
request. The low order address bits for the subsequent data indicate the corresponding interleave order.

Table 38: Line Read Interleave Order

Interleaved Address[3:2]

1st data beat 00 01 10 11

2nd data beat 01 00 11 10

3rd data beat 10 11 00 01

4th data beat 11 10 01 00

8.3.3. Burst Read

The burst read operation transfers an arbitrary amount of data from the target to the initiator. The initiator first
presents a starting address to the target. The target responds by providing multiple cycles of data words in
sequence, starting at the initial address. The initiator indicates to the target when to stop providing data.

Burst read operations are used by I/O devices for block DMA transfers. The LX5280 will never issue a burst
read operation.

Note that there is a difference between a 4-cycles burst and a line read. A line read may use a desired-word-
first increment and wrap. A burst will always increment and will never wrap.

8.3.4. Single-Cycle Write

The single-cycle write operation writes a single word, a halfword, or a byte to the target.

The LX5280 uses a cache with a write-through policy. All CPU instructions that write to memory generate a
single-cycle write operation. (Unless the address is in the local scratchpad memory, in which case the write
operation will not make it out to the Lexra bus).

8.3.5. Line Write

The line write operation is not used by the LX5280. This operation could be used by a processor that has a
data cache that implements a write-back policy.

8.3.6. Burst Write

A burst write is an operation where the initiator sends an address and then an indefinite sequence of data to
the target. The initiator will inform the target when it has finished sending data. This operation is used by I/O
devices for DMA transfers. It is not used by the processor.

Lexra Proprietary & Confidential - 83 - Release 1.9

 April 30, 2001 LX5280

8.4. Signal Descriptions

Table 39: LBus Signal Description

Source
Signal Name Description
(Initiator/Target/Ctrl)

BCLOCK Ctrl Bus Clock

BCMD[6:0] Initiator Encoded command. Active during first cycle that

BFRAME is asserted.

BADDR[31:0] Initiator Address; Target indicates valid address by asserting

BFRAME.

BFRAME Initiator Asserted by initiator a beginning of operation with

address and command signals; de-asserted when
initiator is ready to accept or send last piece of data.
Other bus masters sample this and BIRDY to indicate
that the bus will be available on the next cycle.

BIRDY Initiator For writes, indicates that initiator is driving valid data;
on reads, indicates that initiator is ready to accept
data.

BDATA[31:0] Initiator on write/Target on Data; if driven by initiator, BIRDY indicates valid data
read on bus; if driven by target, BTRDY indicates valid data
on bus.

BTRDY Target For writes, indicates that target is ready to accept

data; on reads, indicates that target is driving valid
data.

BSEL Target Asserted by selected target after initiator asserts

BFRAME; indicates that target has decoded address
and will respond to the transaction (i.e. has been
selected).

8.5. LBus Commands

The initiator drives BCMD during the cycle that BFRAME is asserted.

BCMD[6] 0=read, 1=write

BCMD[5:4] 54
00 burst, fixed length1
01 burst, unlimited number of words
10 line, interleaved wrap2
11 line, linear wrap

1. The number of words comes from BCMD[2:0]

2. Length is determined by the Line size, not BCMD[3:0]

Lexra Proprietary & Confidential - 84 - Release 1.9

 April 30, 2001 LX5280

BCMD[3:0] 3210
1000 1 byte
1001 2 bytes
1010 3 bytes
1011 1 word
1100 2 words
1101 reserved
111x reserved
0000 4 words
0001 8 words
0010 16 words
0011 32 words
01xx reserved

8.6. Byte Alignment

The Lexra Bus is a big endian bus. Transactions must have their data driven to the appropriate bus rails. The
bus mapping is as shown in Table 40.

Table 40: LBus Byte Lane Assignment

Lexra Bus data byte lanes used

BCMD[1:0] ADDR[1:0] 31:24 23:16 15:8 7:0

00 00 X

00 01 X

00 10 X

00 11 X

01 00 X X

01 10 X X

10 00 X X X

10 01 X X X

11 00 X X X X

The Lexra Bus does not define unaligned data transfers, such as a halfword transfer that starts at
ADDR[1:0]=01, or transfers that would need to wrap to the next word.

8.7. Lexra Bus Controller

The Lexra Bus Controller (LBC) is the element of the LX5280 that connects to the Lexra Bus. It forwards all
transaction requests from the LX5280 CPU to the Lexra Bus. It is an initiator and will never respond to
requests from other Lexra Bus initiators.

8.7.1. LBC Commands

The LBC issues the only the LBus commands listed in the table below.

Lexra Proprietary & Confidential - 85 - Release 1.9

 April 30, 2001 LX5280

Table 41: LBus Commands Issued by the LBC

Command BCMD[5:4] BCMD[3:0] Circumstances

Read Line 10 or 11, 0000 A cache miss during a read by the

depending on CPU
configuration

Read Single 00 10xx A read by the CPU from an

(word/halfword/byte) address in uncachable address
space

Write Single 00 10xx A write by the CPU into cacheable

(word/halfword/byte) or uncachable address space

8.7.2. LBC Write Buffer and Out-of-Order Processing

The LBC contains a write buffer with a depth that is configurable with lconfig. All write requests from the
CPU are posted in the write buffer. The CPU will not wait for the write to complete. Write operations
complete in the order they are entered into the queue. If the queue fills, then the CPU must wait until an entry
becomes available.

When the CPU issues a read operation, the LBC will attempt to forward that request to the Lexra Bus ahead
of any pending write operations. This significantly improves performance since the CPU needs to wait for the
read operation to complete and would waste time if it had to also wait for unnecessary or irrelevant writes to
complete.

There are a few cases when the LBC will not allow the read operation to pass pending writes:

1. The address of a pending write is within the same cache line as the read. The LBC will hold the
read operation until the matching write operation, and all write operations ahead of it, com-
plete. If the read is for an instruction fetch, it can still pass a pending write that is inside the
same cache line.

2. The read is to uncacheable address space. All writes will complete before the read is issued.
This avoids any problems with I/O devices and their associated control/status registers.

3. A pending write is to uncachable address space. The LBC will hold the read operation until all
writes up to and including the write to uncacheable address space complete. This further avoids
I/O device problems.

The write buffer bypass feature can be disabled so that reads will never pass writes.

8.7.3. LBC Read Buffer

The LBC contains a read buffer with a depth that is configurable with lconfig. All incoming read data from
the system bus passes through the read buffer. This allows the LBC to accept incoming data as a result of a
cache line fill operation without having to hold the bus.

When the LBC is configured with an asynchronous interface, a larger read buffer improves system and
processor performance in the event of cache miss. When the LBC is configured with a synchronous interface,
the cache can accept the data as fast as the LBC can read it. Therefore, there is no need for a large read buffer.
Customers may reduce the size of the read buffer to a minimum size of two 32-bit entries.

Lexra Proprietary & Confidential - 86 - Release 1.9

 April 30, 2001 LX5280

In some cases, there is a need to minimize the number of gates. The read buffer size may be reduced to two or
four entries for the asynchronous case. This causes a penalty in terms of Lbus utilization since now the LBC
may have to de-assert IRDY if it cannot hold part of the line of data. When the read buffer is the size of a
cache line, this will be relatively rare since simultaneous instruction cache and data cache misses are
relatively rare. For a smaller read buffer, IRDY deassertion is almost a certainty.

8.7.4. Transfer Descriptions

This section describes the various types of read and write transfers in detail. These operations follow certain
patterns and rules. The rules for driving and sampling the bus are as follows:

1. Agents that drive the bus do so as early as possible after the rising edge of the bus clock. There
is some time to perform some combinational logic after the bus clock goes high, but the
amount of time is determined by the speed of the bus clock and the number of devices on the
bus.

2. Agents sample signals on the bus at the rising edge of the bus clock.

3. All bus signals must be driven at all times. If the bus is not owned, and external device must
drive the bus to a legal level.

4. A change in signal ownership requires one dead cycle. If an initiator gives up the bus, another
initiator needs to wait for one dead cycle before it can drive the bus. If the same initiator issues
a read operation and then needs to issue a write operation, it also must wait one extra cycle for
the data bus to turn around.

5. Agents that own signals must drive the signals to a logical true or logical false; all other agents
must disable (tristate) their output buffers.

The Lexra Bus protocol is based on the PCI Bus protocol1. The Lexra Bus signals BFRAME, BTRY, BIRDY,
and BSEL have a similar function to the PCI signals FRAME#, TRDY#, IRDY#, and DEVSEL#,
respectively. In general, the protocol for the Lexra bus is as follows:

1. The initiator gains control of the bus through arbitration (described later in this chapter).

2. During the first bus cycle of its ownership (before the first rising clock edge), the initiator
drives the address for the bus transaction onto BADDR. At the same time, it asserts BFRAME
to indicate that the bus is in use. It will de-assert BFRAME before it send or accepts the last
word of data. In most cases, the initiator will asserts BIRDY to indicate that it is ready to
receive data (or read operations) or is driving valid data (for write operations). If the operation
is a write, the initiator will drive valid data onto BDATA.

3. At the rising edge of the first clock, all agents sample BADDR and decode it to determine
which agent will be the target.

4. The agent that determines that the address is within its address space asserts BSEL sometime
after the first rising edge of the bus clock. BSEL stays asserted until the transaction is com-
plete.

5. The initiator and the target transfer data either in one cycle or in successive cycles. The agent
driving data (the initiator for a write, the target for a read) indicates valid data by asserting its
ready signal (IRDY or TRDY for writes and reads, respectively). The agent receiving data (tar-
get for a write, initiator for a read) indicates its ability to receive the data by asserting its ready

1. The Lexra Bus is not PCI compatible; it merely borrows concepts from the PCI Bus specification.

Lexra Proprietary & Confidential - 87 - Release 1.9

 April 30, 2001 LX5280

signal. Either agent may de-assert its ready signal to indicate that it cannot source or accept
data on this particular clock edge.

6. When the initiator is ready to send or receive the last word of data, that is, when it asserts
BIRDY for the last time, it also de-asserts BFRAME. It will deassert BIRDY when the last
word of data is transferred.

7. The arbiter grants the bus to the next initiator, and may do so during a bus transfer by a differ-
ent initiator. The new initiator must sample BFRAME and BIRDY. When both BIRDY and
BFRAME is sampled de-asserted and the new initiator has been given grant, it can assert
BFRAME the next cycle to start a new transaction.

NOTE: in the examples below, the signals BADDR and BDATA are often shown to be in a high-impedance
state. In reality, internal bus signals should always be driven, even if they are not being sampled. The Hi-Z
states are shown for conceptual purposes only.

8.7.5. Single Cycle Read with No Waits

This operation is used to read a word, halfword or byte from memory, usually in uncachable address space.

CLOCK
BFRAME
BADDR
BDATA
BIRDY
BTRDY
BSEL
D0000

This is a simple read operation where the target responds immediately with data. This is unlikely, since most
bus memory will require one or more cycles to fetch data. This example illustrates the most basic read
operation without waits.

1. Initiator asserts BFRAME and drives BADDR.

2. Target asserts BSEL to indicate to initiator that a target is responding. In this example, there is
an immediate fetch of data, so Target drives data and asserts BTRDY to indicate to target that it
is driving data. The Initiator de-asserts BFRAME and asserts BIRDY to indicate that the next
piece of data received will be the last.

3. Initiator de-asserts IBIRDY and the target de-asserts BSEL and BTRDY to indicate the end of
the transaction. The Initiator that has been given grant owns the bus this cycle.

Lexra Proprietary & Confidential - 88 - Release 1.9

 April 30, 2001 LX5280

8.7.6. Single Cycle Read with Target Wait

This is the same as the single-cycle read, except that the target needs time to fetch the data from memory.

CLOCK
BFRAME
BADDR
BDATA
BIRDY
BTRDY
BSEL
D0001

This is a common single-cycle read operation.

1. Initiator asserts BFRAME and drives BADDR.

2. Target asserts BSEL to indicate that it has decoded the address and is acknowledging that it is
the target device. However, it is not ready to send data, so it does not assert BTRDY. Initiator
de-asserts BFRAME and asserts BIRDY to indicate that the next piece of data will be the last it
wants.

3. Target has not asserted BTRDY so no data is transferred.

4. After a second wait cycle, target drives data and asserts BTRDY to indicate that data is on the
bus.

5. Target de-asserts BSEL and BTRDY. Initiator de-asserts BIRDY. Another initiator may drive
the bus this cycle.

8.7.7. Line Read with No Waits

This operation is used to service a cache miss. Four words of data are transferred in sequence. In this
example, the target is supplying four words of data without any waits.

CLOCK
BFRAME
BADDR
BDATA
BIRDY
BTRDY
BSEL
D0002

1. Initiator drives BADDR and asserts BFRAME to indicate beginning of transaction.

Lexra Proprietary & Confidential - 89 - Release 1.9

 April 30, 2001 LX5280

2. Target asserts BSEL to indicate that it had decoded the address and will send data when it is
ready. Initiator asserts BIRDY to indicate that it is ready to receive data.

3. Target drives data and asserts BTRDY.

4. Target drives second word of data and continues to assert BTRDY.

5. Target drives third word of data and continues to assert BTRDY.

6. Target drives last word of data. Initiator de-asserts BFRAME to indicate that the next word of
data it receives will be the last it needs.

7. Target de-asserts BTRDY and BSEL; initiator de-asserts BIRDY. Another master may gain
ownership of the bus this cycle.

8.7.8. Line Read with Target Waits

This illustrates what happens when a target needs extra time to fetch data it needs to service a cache miss.

CLOCK
BFRAME
BADDR
BDATA
BIRDY
BTRDY
BSEL
D0003

1. Initiator asserts BFRAME and drives BADDR.

2. Target asserts BSEL to indicate that it is acknowledging the operation. Initiator asserts BIRDY
to indicate that it is ready to receive data.

3. Target waits until it has the data.

4. Target drives first word of data and asserts BTDRY.

5. Target drives second word of data and asserts BTRDY.

6. Target cannot get third word of data, so it de-asserts BTRDY.

7. Target drives third word of data and asserts BTRDY.

8. Target cannot get fourth word of data, so it de-asserts BTRDY.

9. Target drives fourth word of data and asserts BTRDY.

8.7.9. Line Read with Initiator Waits

This occurs when a line of data is requested from the target and the initiator cannot accept all of the data in

Lexra Proprietary & Confidential - 90 - Release 1.9

 April 30, 2001 LX5280

successive cycles.

CLOCK
BFRAME
BADDR
BDATA
BIRDY
BTRDY
BSEL
D0004

1. Initiator drives address and asserts BFRAME.

2. Target asserts BSEL. It doesn’t have data, so it does not assert BTRDY. Initiator asserts BIRDY
to indicate that it can accept data

3. Target now has data, so it drives the data and asserts BTRDY.

4. Target drives second word of data; initiator cannot accept it, so it de-asserts BIRDY.

5. Target holds second word of data; initiator can accept it and asserts BIRDY.

6. Target drives third word of data; initiator accepts it.

7. Target drives fourth word of data; initiator cannot accept it and de-asserts BIRDY. initiator hold
BFRAME until it can assert BIRDY.

8. Initiator asserts BIRDY to accept fourth word of data. It de-asserts BFRAME to indicate this is
the last word of data.

8.7.10. Burst Read

This is identical to the read line.

8.7.11. Single-Cycle Write with No Waits

A single-cycle write operation occurs almost every time the LX5280 processor executes a store instruction.
This is because the cache used in the processor uses a write-through policy. Of course, writes to uncacheable
address space and to an I/O device will also generate a single-word write. Single-word write operations are
used to write words, halfwords and bytes.

Lexra Proprietary & Confidential - 91 - Release 1.9

 April 30, 2001 LX5280

A single-word write without waits requires two cycles.

CLOCK
BFRAME
BADDR
BDATA
BIRDY
BTRDY
BSEL
D0005

1. Initiator asserts BFRAME and drives address.

2. Target samples address and asserts BSEL. Initiator drives data and asserts BIRDY. In this case,
target is also able to accept data, so it asserts BTRDY. Initiator also de-asserts BFRAME to
indicate that it is ready to send the last (and only) word of data.

3. Target accepts data, de-asserts BTRDY and BSEL. Initiator de-asserts BIRDY.

8.7.12. Single-Cycle Write with Waits

This is an example of a single-cycle write operation where the target cannot immediately accept data and
must insert wait states.

CLOCK
BFRAME
BADDR
BDATA
BIRDY
BTRDY
BSEL
D0006

This is the same description as the above example, except that the target inserts two wait states until it asserts
BIRDY to indicate acceptance of data.

8.7.13. Burst Write with No Waits

A burst write operation is generally used to transfer large amounts of data from an I/O device to memory via

Lexra Proprietary & Confidential - 92 - Release 1.9

 April 30, 2001 LX5280

a DMA transfer. The following illustrates a best-case scenario with no wait states.

CLOCK
BFRAME
BADDR
BDATA
BIRDY
BTRDY
BSEL
D0007

1. Initiator drives address and asserts BFRAME.

2. Target asserts BSEL and BTRDY to indicate it will accept data. Initiator drive data and asserts
BIRDY.

3. Initiator drives next word of data; target continues to accept data and indicates as such by con-
tinuing to assert BTRDY.

4. Initiator drives third word of data; target continues to accept.

5. Initiator drives fourth word of data and de-asserts BFRAME to indicate that this will be its last
word sent; target accepts data.

6. Target de-asserts BTRDY and BSEL; initiator gives up control of the bus by de-asserting
BIRDY.

8.7.14. Burst Write with Target Waits

This example is similar to the above example, except that during the third and fourth data word transfer, the
target cannot accept the data quickly enough, so it de-asserts BTRDY which indicates to the initiator that it
should hold the data for an additional cycle.

CLOCK
BFRAME
BADDR
BDATA
BIRDY
BTRDY
BSEL
D0008

8.7.15. Burst Write with Initiator Waits

The example illustrates what happens when the initiator cannot supply data fast enough and has to insert

Lexra Proprietary & Confidential - 93 - Release 1.9

 April 30, 2001 LX5280

waits.

CLOCK
BFRAME
BADDR
BDATA
BIRDY
BTRDY
BSEL
D0009

8.8. LBC Signals

The table below summarizes the LX5280 LBC ports. The "LBC Port" column indicates the name of the port
supplied by the LBC. The "Bus Signal" column indicates the corresponding Lexra bus signal. The LBC ports
are strictly uni-directional, while the bus signals (at least conceptually) include multiple sources and sinks.
The manner in which LBC ports are connected to bus signals is technology dependent, and may employ tri-
state drivers or logic gating in conjunction with the LBC’s LCoe, LDoe and LToe outputs.

Table 42: LBC Interface Signals

I/O LBC Port Bus Signal Description

output LAddrO[31:0] BADDR[31:0] LBC address

output LDataO[31:0] BDATA[31:0] LBC data

input LDataI[31:0] BDATA[31:0] System data

output LIrdy BIRDY LBC initiator ready

input LIrdyI BIRDY System initiator ready

output LFrame BRAME LBC transaction frame

input LFrameI BFRAME System transaction frame

input LSel BSEL System slave select

input LTrdy BTRDY System target ready

output LCmd[6:0] BCMD[6:0] LBC command

output LReq - LBC bus request

input LGnt - System bus grant

output LCoe[9:0] - LBC command output enable terms

output LDoe[7:0] - LBC data output enable terms

output LToe - LBC transaction output enable terms

Lexra Proprietary & Confidential - 94 - Release 1.9

 April 30, 2001 LX5280

8.9. Arbitration

8.9.1. Rules

The following are the rules for arbitration (GNT=grant, REQ=request):

1. Master asserts REQ at the beginning of a cycle and may start sampling for asserted GNT in the
same cycle (in case GNT is already asserting in the case of a “park”).

2. If bus is idle or it is the last data phase of the previous transaction when master samples
asserted GNT, master may assert FRAME on next cycle.

3. If the bus is busy when the master samples GNT, is must also snoop FRAME, IRDY and Trdy.
One cycle after FRAME is not asserted and both IRDY and TRDY are asserted (indicating the
last data phase), if GNT is still asserted, master may now drive FRAME (i.e. GNT &
~Frame_R & (Irdy_R & Trdy_R)).

8.9.2. LBC behavior

The LBC, when it need access to the bus, asserts REQ and in the same cycle samples GNT, ~FRAME, and
either ~IRDY or (IRDY & TRDY). If these are true, then the LBC will on the next cycle take ownership of
the bus. REQ is deasserted on the cycle after LBC asserts FRAME. If the bus is busy, LBC continues to
snoop these four signals for this condition. All other Lbus arbitration rules can be based on this behavior of
the LBC.

8.10. Connecting Devices to the Bus

There are three sets of output enables: TOE(valid for the length of the transaction), COE (valid for only the
first cycle of a transaction), and DOE (valid for data transfers, asserted by the master for writes and by the
slave for reads).

TOE is intended to qualify:

FRAME
IRDY

COE is intended to qualify:

CMD
ADDR

DOE is intended to qualify:

DATA

There is no output enable to qualify TRDY and SEL. These are defined by customer logic for slave devices.

Instead of using TOE it may be desirable to instead OR all of the FRAME signals, either centrally or one OR
gate for each target and master. The same holds true for IRDY, TRDY, and SEL. This simplifies the
connections when a relatively few number of devices are used and there are no off-chip devices connected
directly to the Lexra Bus.

Therefore, it is defined that masters and slaves not taking part in a transaction always keep FRAME, IRDY,
TRDY, and SEL driven and de-asserted.

Lexra Proprietary & Confidential - 95 - Release 1.9

 April 30, 2001 LX5280

Lexra Proprietary & Confidential - 96 - Release 1.9

 April 30, 2001 LX5280

9. LX5280 Coprocessor Interface

The LX5280 processor provides customer access points for the Coprocessor Interfaces. This section provides
a description of these access points. Attachment of memory devices to the LMIs, the System Bus, and the
EJTAG interface are described in separate chapters.

9.1. Attaching a Coprocessor Using the Coprocessor Interface (CI)

A coprocessor may contain up to 32 general registers and up to 32 control registers. Each of these registers is
up to 32 bits wide. Typically, programs use the general registers for loading and storing data on which the
coprocessor operates. Data is moved to the coprocessor’s general registers from the core’s general registers
with the MTCz instruction. Data is moved from the coprocessor’s general registers to the core’s general
registers with the MFCz instruction. Main memory data is loaded into or stored from the coprocessor’s
general registers with the LWCz and SWCz instructions.

Programs may load and store the coprocessor’s control registers from the core’s general registers with the
CTCz and CFCz instructions respectively. Programs may not load or store the control registers directly from
main memory.

The coprocessor may also provide a condition flag to the core. The condition flag can be a bit of a control
register or a logical function of several control register values. The condition flag is tested with the BCzT and
BCzF instructions. These instructions indicate that the program should branch if the condition is true (BCzT)
or false (BCzF).

9.2. Coprocessor Interface (CI) Signals

The CI provides the mechanism to attach the custom coprocessor to the core. The CI snoops the instruction
bus for coprocessor instructions and then gives the coprocessor the signals necessary for reading or writing
the general and control registers.

Table 43: Coprocessor Interface Signals

Signal I/O Description

C<z>condin input Cop branch flag.

C<z>rd_addr[4:0] output Cop read address.

C<z>rhold output Cop hold condition, one stalls coprocessor.

C<z>rd_gen output Cop general register read command.

C<z>rd_con output Cop control register read command.

C<z>rd_data[31:0] input Cop read data.

C<z>wr_addr[4:0] output Cop write address.

C<z>wr_gen output Cop general register write command.

C<z>wr_con output Cop control write address command.

C<z>wr_data[31:0] output Cop write data.

C<z>invld_M output Cop invalid instruction flag, one indicates

invalid instruction in M stage.

Lexra Proprietary & Confidential - 97 - Release 1.9

 April 30, 2001 LX5280

Signal I/O Description

C<z>xcpn_M output Cop exception flag, one indicates exception

in M stage.

The addresses, output data, and control signals are supplied to the user’s Coprocessor on the rising edge of the
system clock. In the case of a read cycle, the coprocessor must supply the data from either the control or
general register on C<z>rd_data by the end of the same cycle. Similarly, the write of data from C<z>wr_data
to the addressed control or general register must be complete by the end of the cycle.

The CI incorporates a forwarding path so that data which is written in instruction (N) can be read in
instruction (N + 2). The Coprocessor registers should be implemented as positive-edge flip-flops using the
LX5280 system clock.

9.3. Coprocessor Write Operations

During a coprocessor write, the CI sends C<z>wr_addr and C<z>wr_data, and asserts either C<z>wr_gen or
C<z>wr_con. The coprocessor must ensure that the coprocessor completes the write to the appropriate
register on the subsequent rising edge of the clock. The target register is a decoding of C<z>wr_addr,
C<z>wr_gen and C<z>wr_con. Use these instructions to cause a coprocessor write: LWCz, MTCz, and
CTCz.

9.4. Coprocessor Read Operations

During a coprocessor read, the CI sends C<z>rd_addr and asserts either C<z>rd_gen or C<z>rd_con. The
coprocessor must return valid data through C<z>rd_data in the following clock cycle. If the core asserts
C<z>rhold, indicating that it is not ready to accept the coprocessor data, the coprocessor must hold the
previous value of C<z>rd_data. The target register for the read is a decoding of C<z>rd_addr, C<z>rd_gen,
and C<z>rd_con. The instructions causing a coprocessor read are SWCz, MFCz, and CFCz.

The CPU stalls the pipeline so that the program can access data read by a coprocessor instruction in the
immediately following instruction. For example, if an MFCz instruction reads data from the coprocessor and
stores it in the core’s general register $4, the program can get access to that data in the following instruction:

mfc2 $4, $3 # Move from COP2 to CPU register $4

subu $5, $4, $2 # Subtract $R2 from $R4 and store in $5

When the core initiates a coprocessor read, the coprocessor must return valid data in the following clock
cycle. The coprocessor cannot stall the CPU. Applications must ensure that the source code does not access
invalid coprocessor data if the coprocessor operations take several clock cycles to complete. This is done in
one of three ways:

• Ensure that code does not access data from the coprocessor until N instructions after the
coprocessor operation has stared. This is the least desirable method as it depends on the
relative execution of the core and coprocessor. It can also complicate software debug.

• Have the coprocessor send an interrupt to the core, and the service routine for that
interrupt accesses the appropriate coprocessor registers.

• Have the coprocessor set the C<z>condin flag when its operation is complete. The source

Lexra Proprietary & Confidential - 98 - Release 1.9

 April 30, 2001 LX5280

code can poll the flag as shown in the example below:

mtc2 $2, $3 # store data to COP2 general register $3

ctc2 $3, $5 # set COP2 control register $5 to start
nop
loop: bc2f loop # branch back to loop if C<z>condin bit off
nop # branch delay slot
mfc2 $4, $7 # get results from COP2 general register $7

9.5. Coprocessor Interface and Pipeline Stages

Coprocessor writes occur in the W stage of the instruction pipeline. For coprocessor reads, the core generates
address, rd_gen, and rd_con signals during the S stage, and the coprocessor returns data during the E stage
which is passed by the CI to the core in the M stage. The core introduces a pipeline bubble after coprocessor
instructions to ensure that the result of a MTCz instruction can be used by the immediately following
instruction.

In particular, if there are back-to-back MTCz and MFCz instructions that access the same coprocessor
register, the pipeline bubble still does not allow a cycle between the W stage write and E stage read as
required. In this case a special forwarding path within the CI is used. That is, the “true” data from the
coprocessor is ignored. Instead the exact data from the MTCz is used.

mtc2 I D S E M W
bubble I D . . . .
mfc2 I D S E M W # data forwarded by CI from mtc2
wr_gen (W) X
rd_gen (S) X
rd_data(E) X

The forwarding path can cause side effects if the coprocessor does not implement all of the bits of a register,
contains read-only bits, or updates the register value upon reading the register. In such cases, the mfc2
instruction returns different data from what it would if the core did not activate the forwarding path. To avoid
the forwarding path, another instruction must be inserted between the mtc2 and mfc2:

mtc2 I D S E M W
bubble I D . . . .
foo I D S E M W
mfc2 I D S E M W # read data from coprocessor
wr_gen (W) X
rd_data(E) X

9.5.1. Pipeline Holds

The coprocessor must register the read address and the control signals rd_gen and rd_con. It must hold the (E
stage) registered values of these signals when C<z>_rhold is active high, and should make the read data
output a function of the (E stage) registered read address and control signals.

The wr_addr, wr_data, wr_gen and wr_con signals need not be registered. The coprocessor may decode these
(W stage) signals directly to the appropriate register.

9.5.2. Pipeline Invalidation

Under certain circumstances the instruction pipeline can contain an instruction that must be discarded. This
can be due to mispredicted branches, cache misses, exceptions, inserted pipeline bubbles etc. In such cases,

Lexra Proprietary & Confidential - 99 - Release 1.9

 April 30, 2001 LX5280

the CI may decode an instruction that must actually be discarded.

For the coprocessor write-type instructions, the CI will only issue the W stage control signals wr_gen and
wr_con for valid instructions. The coprocessor does not need to qualify these controls.

For the coprocessor read-type instructions, the CI may issue the S stage control signals rd_gen and rd_con for
instructions that must be discarded. If the coprocessor can tolerate speculative reads then it need not qualify
those signals. However, if the coprocessor performs “destructive” reads, such as updating a FIFO pointer
upon read, then it must use the qualifying signals C<z>_xcpn_m and C<z>_invld_m as follows:

The signal C<z>_xcpn_m signal is used to discard any S stage (from CI) rd_gen or rd_con signal and any E
stage (registered in the coprocessor) rd_gen or rd_con signal. It indicates that a preceding instruction in the
pipe has taken an exception and that subsequent instructions in the pipe must be discarded.

The signal C<z>_invld_m signal is used to invalidate the operation of the current instruction in the M stage.
This can be for various reasons not limited to an exception on a preceding instruction. If the coprocessor
cannot tolerate speculative reads, it must register an M stage version of rd_gen and rd_con. The coprocessor
must use the C<z>_rhold signal to hold this M stage version (as well as the E stage version). If
C<z>_invld_m is asserted, then any such M stage signals must be discarded. To summarize, a rd_gen or
rd_con instruction can “retire” only if it reaches the M stage and neither C<z>_rhold nor C<z>_invld_m is
asserted.

Lexra Proprietary & Confidential - 100 - Release 1.9

 April 30, 2001 LX5280

10. LX5280 EJTAG

10.1. Introduction

Given the increasing complexity of SoC designs, the nature of embedded processor-design debug, hardware
and software, and the time-to-market requirements of embedded systems, a debug solution is needed which
allows on-chip processor visibility in a cost-effect, I/O constrained manner.

Lexra’s EJTAG solution meets all such requirements. It uses existing IEEE JTAG pins as well as fast bring-up
on new designs. It provides a way of debugging all devices accessible to the processor in the same way the
processor would access those devices itself. Using EJTAG, a debug probe can access all the processor
internal registers and caches. It can also access devices connected to the Lexra Bus, bypassing internal caches
and memories.

Software debug is enhanced by EJTAG features that allow single-stepping through code and halting on
breakpoints (hardware and software, address and data with masking). For debugging problems that are
artifacts of real-time interactions, EJTAG gives real-time Program Counter trace capabilities from which an
accurate program execution history is derived. For the code-system perspective, PC profiling provides
statistical analysis of code usage to aim code optimization.

10.2. Overview

A debug host computer communicates to the EJTAG probe through either a serial or parallel port or Ethernet
connection. The probe, in turn, communicates to the LX5280 EJTAG hardware via the included IEEE 1149.1
JTAG interface. Through the use of the JTAG TAP controller, probe data is shifted into to the EJTAG data and
control registers in the LX5280 to respond to processor requests, DMA into system memory, configure the
EJTAG control logic, enable single-step mode, or configure the EJTAG breakpoint registers. Through the use
of the EJTAG control registers, the user can set hardware breakpoints on the instruction cache address, data
cache address or data cache data values.

Physical address range 0xFF20_0000 to 0xFF3F_FFFF is reserved for EJTAG use only and should not be
mapped to any other device.

Currently, Embedded Performance Inc. (EPI) and Green Hills Inc. provide EJTAG debuggers and probes for
the LX5280. Information on these products is available at the following web sites.

EPI Inc.: http://www.epitools.com

Green Hills Inc.: http://www.ghs.com

LX5280 EJTAG implements all required features of version 2.0.0 of the EJTAG specification, and includes
support for the following features:

• Processor access of host via addressing of probe memory space.

• Host probe can DMA directly into system memory or I/O devices.

• Hardware breakpoints on internal instruction and data busses.

• Single-step execution mode.

• Real-time Program Counter Trace.

• Debug exception and two new debug instructions: one for raising a debug exception via
software, and one for returning from a debug exception.

Lexra Proprietary & Confidential - 101 - Release 1.9

 April 30, 2001 LX5280

10.2.1. IEEE JTAG-specific Pinout

IEEE JTAG pins used by EJTAG are shown below. These are required for all EJTAG implementations.
JTAG_TRST_N is an optional pin.

Table 44: EJTAG Pinout

Signal Name I/O Description

JTAG_TDO_NR output Serial output of EJTAG TAP scan chain.

JTAG_TDI Input Serial input to EJTAG TAP scan chain.

JTAG_TMS Input Test Mode Select. Connected to each EJTAG TAP controller.

JTAG_CLOCK Input JTAG clock. Connected to each EJTAG TAP controller

JTAG_TRST_N Input TAP controller reset. Connected to each EJTAG TAP controller.a

a. This pin is optional in multiprocessor configurations

Table 45: EJTAG AC Characteristics1

Signal Parameter Condition Min Max Unit

JTAG_CLOCK Frequency <1 40 MHz

Duty Cycle 40/60 60/40 %

JTAG_TMS Setup to TCK rising edge 1.8V 5 ns

Hold after TCK rising edge 1.8V 5 ns

JTAG_TDI Setup to TCK rising edge 1.8V 5 ns

Hold after TCK rising edge 1.8V 5 ns

JTAG_TDO_NR Output Delay TCK falling edge to TDO 1.8V 0 7 ns

Table 46: EJTAG Synthesis Constraints2

Signal Name Probe Budget Core Budget Slack remaining for other logic

JTAG_TDO_NR 0 to -7ns 11.5ns 13.5 to 20.5ns

JTAG_TDI 5ns 13.5ns 6.5ns

JTAG_TMS 5ns 13.5ns 6.5ns

10.3. Single Processor PC Trace

The LX5280 EJTAG includes support for real-time Program Counter Trace (PC Trace). When in PC Trace

1. Based on EPI Interface Specifications for MAJICTM and MAJICPLUS TM

2. Based on 25ns JTAG clock period.

Lexra Proprietary & Confidential - 102 - Release 1.9

 April 30, 2001 LX5280

mode, the LX5280 will serially output a new value of the program counter whenever a change in program
control occurs (i.e. branch or jump instruction, or an exception).

When the PC Trace option is set to EXPORT in lconfig, the following signals will be output from the
LX5280: DCLK, PCST, and TPC. These are described in more detail in the following subsections.

The DCLK output is used to synchronize the probe with the LX5280’s SYSCLK.

The PCST (PC Trace Status) signals are used to indicate the status of program execution. Example status
indications are sequential instruction, pipeline stall, branch, or exception.

The TPC pins output the value of the PC every time there is a change of program control.

10.3.1. PC Trace DCLK - Debug Clock

The maximum speed allowed for the Debug Clock (DCLK) output is 100MHz (as an EPI probe
requirement). As cores typically run in excess of this speed DCLK can be set to a divided down value of
SYSCLK. This is set by the DCLK N parameter in lconfig, which indicates the ratio of SYSCLK frequency
to DCLK: 1, 2, 3 or 4.

10.3.2. PC Trace PCST - Program Counter Status Trace

The Program Counter Status (PCST) output comprises N sets of 3-bit PCST values, where N is configurable
as 1, 2, 3 or 4 via lconfig. A PCST value is generated every SYSCLK cycle. When DCLK is slower than the
LX5280’s SYSCLK, up to N PCST values are output simultaneously.

10.3.3. PC Trace TPC - Target Program Counter

The bus width of the Target Program Counter (TPC) output is user configured in lconfig via the “M”
parameter to be one of 1, 2, 4 or 8 bits. When change in program flow occurs the current PC value is sent out
of TPC. As the PC is 32-bits wide, the number of TPC pins affects how quickly the PC is sent. For example,
if the TPC is 4 bits wide the PC will take 8 DCLK cycles to be sent. If another change in flow occurs while
the PC of the previous change is being transmitted, the new PC will be sent and the remainder of the previous
PC will be lost.

The TPC bus also outputs the exception type when an exception occurs. The exception type field-width is
either 3- or 4-bits depending on whether or not vectored interrupts are present. This is covered in more detail
below.

To reduce pinout, the TDO output is used for the least significant bit of TPC (or the only bit if “M” is set to 1).

10.3.4. Dual Pipe PC Trace

The EJTAG PC Trace facility specifies that a PCST (PC Trace Status) code is issued if the instruction pipeline
has stalled, sequentially completed an instruction, or taken an branch or jump. In order to accommodate the
two pipelines in the LX5280, the capability of emitting more than one PCST code per cycle is employed.
Specifically, to the external EJTAG probe, the LX5280 appears to be a single pipe machine running at twice
the speed that it actually does.

Since there must be an even number of PCST codes made available at every DCLK rising edge (in the
EJTAG nomenclature), the DCLK parameter “N” must be set to 2 or 4. Setting the DCLK N parameter to 2
results in DCLK running at the same frequency of SYSCLK; setting the parameter to 4 results in DCLK
running at one-half the frequency of SYSCLK.

The maximum value of the N parameter is 4, and the maximum DCLK frequency is 100MHz. Therefore,

Lexra Proprietary & Confidential - 103 - Release 1.9

 April 30, 2001 LX5280

until the EJTAG specification is extended beyond N=4 or a maximum DCLK of 100MHz, the maximum
SYSCLK frequency for which dual-pipe PC Trace can be used is 200 MHz.

10.3.5. Single-Processor PC Trace Pinout

Table 47: Single-Processor PC Trace Pinout.

Signal Name I/O Description

JPT_TPC_DR O/P The PC value is output on these pins when a PC-discontinuity occursa
M bits

JPT_PCST_DR O/P PC Trace Status: Outputs current instruction type every DCLK
N*3 bits

JPT_DCLK O/P PCST and TPC clock. Frequency determined as a fraction of SYSCLK
via the N parameter. Maximum frequency of DCLK is 100MHz.

a. TPC[0] is multiplexed with TDO in the single-processor PC Trace solution.

Table 48: Single-Processor PC Trace AC Characteristics1

Signal Parameter Min Max Unit

JTAG_DCLK Frequency DC 100 MHz

DCLK High Time 4 ns

Low Time 4 ns

TPC Setup to DCLK falling edge at probe 0 ns

Hold after DCLK falling edge 4 ns

PCST Setup to DCLK falling edge at probe 0 ns

Hold after DCLK falling edge 4 ns

10.3.6. Vectored Interrupts and PC Trace

The EJTAG PC Trace facility specifies a 3-bit code be output on the TPC output when an exception occurs
(the PCST pins give the EXP code). In order to distinguish the eight vectored interrupts in the LX5280 from
all other exceptions, a 4-bit code is used instead.

For all exceptions other than vectored interrupts, the most significant bit of the 4-bit code is zero and the
remaining 3-bits are the standard 3-bit code. Note that this includes the standard software and hardware
interrupts numbered 0 through 7.

For vectored interrupts, the most significant bit is always 1. The 4-bit code is simply the number of the
vectored interrupt (from 8 through 15) being taken.

Since the target of the vectored interrupt is determined by the contents of the INTVEC register, the debug
software which monitors the EJTAG PC Trace codes must be aware of the contents of this register in order to
trace the code after the vectored interrupt is taken.

1. Based on EPI Interface Specifications for MAJICTM and MAJICPLUS TM

Lexra Proprietary & Confidential - 104 - Release 1.9

 April 30, 2001 LX5280

For probes that do not support a 4-bit exception code, the LX5280 can be configured via the
EJTAG_XV_BITS lconfig option to use only the 3-bit standard codes. In that case, if a vectored interrupt is
taken, the 3-bit code for RESET will be presented.

10.3.7. Demultiplexing of TDO and TDI During PC Trace

In normal EJTAG PC Trace, TDI and TDO are multiplexed with the debug interrupt (DINT) and the lsb of
the TPC (TPC[0]) when in PC Trace mode. This reduces the number of pins required by PC Trace, but has
the unfortunate side-affect of preventing any access to EJTAG registers during PC Trace.

In order to allow access to EJTAG registers during PC Trace, and to facilitate PC Trace in multiprocessor
environments, the lconfig option JTAG_TRST_IS_TPC=YES causes TDI and TDO to be demultiplexed
such that TRST is used as TPC[0] and DINT is generated via EJTAG registers. Note: setting this option may
require changes in EJTAG probe hardware. Check with probe manufacturer for details.

Lexra Proprietary & Confidential - 105 - Release 1.9

 April 30, 2001 LX5280

Lexra Proprietary & Confidential - 106 - Release 1.9

 April 30, 2001 LX5280

Appendix A. LX5280 Lconfig Forms

A.1. Configuration Options for the LX5280 Processor

This section provides a summary of the configuration options available with lconfig. Refer to lconfig forms
for a detailed description of these form options.

PRODUCT -- Lexra Processor name

PRODUCT_TYPE -- indicates product type
TECHNOLOGY -- identifies target technology
TESTBED_ENV -- identifies simulation testbed environment type
RESET_TYPE -- flip-flop reset method
RESET_DIST -- reset distribution method
SLEEP -- include clock SLEEP support
RESET_BUFFERS -- reset buffers at top-level module
CLOCK_BUFFERS -- clock buffers at top-level module
RAM_CLOCK_BUFFERS -- LMI RAM clock distribution method
COP1 -- coprocessor interface 1
COP2 -- coprocessor interface 2
COP3 -- coprocessor interface 3
CE0 -- custom engine 0
CE1 -- custom engine 1
M16_SUPPORT -- 16-bit opcode support
MEM_LINE_ORDER -- cache line fill beat ordering
MEM_FIRST_WORD -- cache line fill first word
MEM_GRANULARITY -- main memory system partial word write support
SYSTEM_INTERFACE -- system bus interface type
LBC_WBUF -- Lexra Bus Controller write buffer depth
LBC_RBUF -- Lexra Bus Controller read buffer depth
LBC_RDBYPASS -- Lexra Bus Controller read bypass enable
LBC_SYNC_MODE -- LBC synchronous/asynchronous selection
LINE_SIZE -- cache line size, in words
ICACHE -- instruction cache size
DCACHE -- data cache size
IMEM -- local instruction RAM with line valid bits
IROM -- local instruction ROM
DMEM -- local scratch pad data RAM
LMI_DATA_GRANULARITY -- DCACHE and DMEM write granularity
LMI_RANGE_SOURCE -- source of LMI address ranges
LMI_RAM_ARB -- allow external agents to arbitrate for LMI RAMs
JTAG -- Internal JTAG Tap controller with EJTAG support
EJTAG -- EJTAG Debug Support
EJTAG_INST_BREAK -- Number of instruction breaks to be compiled
EJTAG_DATA_BREAK -- Number of data breaks to be compiled
JTAG_TRST_IS_TPC -- TRST pin is TPC out, instead of TDO/TPC mux
PC_TRACE -- EJTAG PC trace pins
EJTAG_DCLK_N -- EJTAG PCTrace DCLK N parameter
EJTAG_TPC_M -- EJTAG PCTrace TPC M parameter
EJTAG_XV_BITS -- EJTAG PCTrace number of Exception Vector bits
EJTAG_PC_ISABIT -- EJTAG PCTrace include ISA as PC Bit0
SCAN_INSERT -- Controls scan insertion and synthesis
SCAN_MIX_CLOCKS -- scan chains can cross clock boundaries with
lock-up latches
SCAN_NUM_CHAINS -- number of scan chains

Lexra Proprietary & Confidential - 107 - Release 1.9

 April 30, 2001 LX5280

SCAN_SCL -- scan collar insertion on RAM interfaces

SEN_DIST -- scan enable distribution method
SEN_BUFFERS -- scan enable buffering
RAM_BIST_MUX -- include test RAM mux and ports

Lexra Proprietary & Confidential - 108 - Release 1.9

 April 30, 2001 LX5280

Appendix B. LX5280 Port Descriptions

All ports must be connected to valid logic-level sources.

The timing information indicates the point within a cycle when the signal is stable, in terms of percent. The
timing information also includes parenthetical references to these notes:

1. Clocked in the JTAG_CLOCK domain.

2. Clocked in the BUSCLK domain if crossbar or LBC are asynchronous. Otherwise, clocked in
the SYSCLK domain.

3. Does not require a constraint (e.g., a clock).

4. A constant that is treated as false path for timing analysis. These inputs must not change after
the processor is taken out of reset.

5. Timing is specified with a symbol in techvars.scr script (e.g. RAM timing).

6. A test-related input or output that is treated as false path for timing analysis. Such inputs must
not change during normal at-speed operation.

7. An asynchronous input.

If no clock domain is specified, the signal is clocked in the SYSCLK domain.

The table below shows the possible port connections for the top level module of the LX5280 processor,
known as lx2. The actual ports that are present depends upon lconfig settings. The timing information and
notes have the same meaning as for the previous table.

Names that include _N indicate active low signals. All other signals are active high unless otherwise
indicated.

For single bit signals, the signal name and signal description indicate the action or function when the signal is
in the active state.

Table 49: LX5280 Processor Port Summary

Port Name I/O Timing Description

Clocking, Reset, Interrupts and Control

SYSCLK input (3) Processor clock.

SYSCLKF input (3) Free running processor clock, if proces-

sor is configured with sleep support.

SL_SLEEPSYS_R output 30% Clock gating term for SYSCLK, if pro-

cessor is configured with sleep support.

BUSCLK input (3) Bus clock, if processor is configured

with async LBC.

BUSCLKF input (3) Free running bus clock, if processor is

configured with async LBC sleep sup-
port.

Lexra Proprietary & Confidential - 109 - Release 1.9

 April 30, 2001 LX5280

Port Name I/O Timing Description

SL_SLEEPBUS_BR output 30% Clock gating term for BUSCLK, if pro-

cessor is configured with async LBC
and sleep support.

ResetN input 10% Warm reset (or reset "button"), active

low.

CResetN input 10% Cold reset (or power on), active low.

RESET_D1_R_N input 30% SYSCLK domain reset combination of

ResetN, CResetN, EJTAG.

RESET_D1_BR_N input 30% BUSCLK domain reset combination of

ResetN, CResetN, EJTAG.

RESET_PWRON_C1_N input 30% Power on reset copy for JTAG.

RESET_PWRON_D1_LR_N input 30% SYSCLK domain power on reset for

EJTAG.

RESET_D1_R_N_O output 30% SYSCLK domain reset combination of

ResetN, CResetN, EJTAG.

RESET_D1_BR_N_O output 30%, (2) BUSCLK domain reset combination of

ResetN, CResetN, EJTAG.

RESET_PWRON_C1_N_O output 30% Power on reset copy for JTAG.

RESET_PWRON_D1_LR_N_O output 30% SYSCLK domain power on reset for

EJTAG.

INTREQ_N[15:2] input (7) Interrupt requests.

EXT_HALT_P input 50% External stall line.

EXT_SLEEPREQ_R input 30% External sleep request.

Configuration

CFG_TLB_DISABLE input (4) Disable TLB mappings even if tlb is

present.

CFG_SLEEPENABLE input (4) Sleep enable configuration.

CFG_RAD_LEXOP[5:0] input (4) LEXOP encoding. Must be 011111 for

LX5280.

CFG_RAD_DISABLE input (4) LEXOP disable configuration. Must be

one for LX5280.

CFG_SINGLEISSUE input (4) Forces single instruction issue.

CFG_HLENABLE input (4) Strap to one to enable internal HI/LO

registers.

CFG_MACENABLE input (4) Strap to one to enable internal MAC (if

present).

CFG_MEMSEQUENTIAL input (4) Strap to one if line reads return words

in sequential order, zero if interleave
order.

CFG_MEMZEROFIRST input (4) Strap to one if line reads return word

zero first, zero if desired word first.

Lexra Proprietary & Confidential - 110 - Release 1.9

 April 30, 2001 LX5280

Port Name I/O Timing Description

CFG_MEMFULLWORD input (4) Strap to one if main memory must be

written with 32-bit words, zero if byte
and halfword writes are allowed.

CFG_LBCWBDISABLE input (4) Strap to one to disable read bypass of

LBC write buffer, zero to allow read
bypass.

CFG_EJTNMINUS1[1:0] input (4) Strap with EJTAG DCLK N minus 1

configuration (0-3=1-4).

CFG_EJTMLOG2[1:0] input (4) Strap with EJTAG M log2 (0-3=1,2,4,8)

configuration.

CFG_EJT3BITXVTPC input (4) Strap with ETJAG 3-bit TPC configura-

tion.

CFG_EJTBIT0M16 input (4) Strap with EJTAG PC bit0 in TPC con-

figuration.

CFG_DWBASE[31:10] input 30% Strapped with DMEM base address

configuration value.

CFG_DWTOP[23:10] input 30% Strapped with DMEM top address con-

figuration value.

CFG_IWBASE[31:10] input 30% Strapped with IMEM base address con-

figuration value.

CFG_IWTOP[`23:10] input 30% Strapped with IMEM top address con-

figuration value.

CFG_IWROM input (4) Strap to one to treat IMEM like a ROM.

(Note, new applications should use
IROM instead of ROM-like IMEM.)

CFG_IROFF input (4) Strap to one to disable IROM.

CFG_DWDISW input (4) Strap to one to disable processor

DMEM writes. Must be zero for
LX5280.

CFG_EJDIS input (4) Must be strapped to zero.

Test and Debug

JTAG_RESET_O output 20%, (1) JTAG is in TEST-LOGIC-RESET state,

active low.

JTAG_RESET input (6) JTAG is in TEST-LOGIC-RESET state,

active low.

TAP_RESET_N_O output 20%, (1) TAP controller reset.

TAP_RESET_N input (6) TAP controller reset.

JTAG_TDO_NR output 50%, (1) Test data out, active low.

JTAG_TDI input 60%, (1) Test data in.

JTAG_TMS input 60%, (1) Test mode select.

JTAG_CLOCK input (3) Test clock.

Lexra Proprietary & Confidential - 111 - Release 1.9

 April 30, 2001 LX5280

Port Name I/O Timing Description

JTAG_TRST_N input (6) Test reset.

JTAG_CAPTURE output 20%, (1) JTAG is in DATA REGISTER CAP-

TURE state

JTAG_SCANIN output 50%, (1) Scan input to chain

JTAG_SCANOUT input 50%, (1) Scan output from chain

JTAG_IR[4:0] output 20%, (1) Contents of INSTRUCTION REGIS-

TER

JTAG_SHIFT_IR output 20%, (1) JTAG is in SHIFT INSTRUCTION REG-

ISTER state

JTAG_SHIFT_DR output 20%, (1) JTAG is in SHIFT DATA REGISTER

state

JTAG_RUNTEST output 20%, (1) JTAG is in RUN-TEST state

JTAG_UPDATE output 20%, (1) JTAG is in DATA REGISTER UPDATE

state

EJC_ECRPROBEEN_R output 30% One indicates EJTAG probe is active.

JPT_PCST_DR[M-1:0] output 30% EJTAG PC trace status; M= 1, 2, 4 or 8.

JPT_TPC_DR(N*3-1:0] output 30% EJTAG PC trace value, N= 1, 2, 3 or 4.

JPT_DCLK output (3) EJTAG PC trace clock.

SEN input (6) Scan enable, active high.

TMODE input (6) Test mode, active high.

SIN[<k>:0] input (6) Scan Input. <k> can range from 7 to 0.

SOUT[<k>:0] output (6) Scan Output. <k> can range from 7 to

0.

RBC_SEL[7:0] input (6) RAM BIST RAM select code:

10000000 - instruction MEM
01000000 - data MEM
00100000 - dcache data store
00010000 - dcache tag store
00001000 - icache tag store, set 1
00000100 - icache inst store, set 1
00000010 - icache tag store, set 0
00000001 - icache inst store, set 0

RBC_WE[<k>:0] input (6) RAM BIST write enable, where <k> is 1

for word write granularity, 7 for byte
write granularity.

RBC_RE input (6) RAM BIST read enable.

RBC_CS input (6) RAM BIST select.

RBC_ADDR[15:0] input (6) RAM BIST address.

RBC_DATAWR[63:0] input (6) RAM BIST write data.

RBM_DATARD[63:0] output (6) RAM BIST read data.

Lexra Proprietary & Confidential - 112 - Release 1.9

 April 30, 2001 LX5280

Port Name I/O Timing Description

LBC Interface (to LBus)

LAddrO[31:0] output (2), 20% Address.

LCmdO[6:0] output (2), 20% LBC command.

LDataO[31:0] output (2), 20% LBC data.

LDataI[31:0] input (2), 50% System data.

LIrdyO output (2), 20% LBC initiator ready.

LIrdyI input (2), 30% System initiator ready.

LFrameO output (2), 20% LBC transaction frame.

LFrameI input (2), 30% System transaction frame.

LSel input (2), 30% System slave select.

LTrdyI input (2), 30% System target ready.

XBRdVld input (2), 30% Crossbar read data valid.

XBRdSize input (2), 30% Split read data size.

SpltRdFull output (2), 30% Read data queue full.

LId output (2), 20% Instruction/data.

LUc output (2), 20% Bus request.

LCoe[9:0] output (2), 20% Command output enable.

LToe output (2), 20% Transaction output enable.

LDoe[7:0] output (2), 20% Data output enable.

LReq output (2), 50% Bus request.

LGnt input (2), 30% Bus grant.

Shared RAM Request/Grant Interface

EXT_IWREQRAM_R input 30% External hardware drives to one to

request access to IMEM.

IW_GNTRAM_R output 30% Cpu drives to one to grant external

IMEM access request.

EXT_DWREQRAM_R input 30% External hardware drives to one to

request access to DMEM.

DW_GNTRAM_R output 30% Cpu drives to one to grant external

DMEM access request.

EXT_ICREQRAM_R input 30% External hardware drives to one to

request access to ICACHE.

IC_GNTRAM_R output 30% Cpu drives to one to grant external

ICACHE access request.

EXT_DCREQRAM_R input 30% External hardware drive to one to

request access to DCACHE.

Lexra Proprietary & Confidential - 113 - Release 1.9

 April 30, 2001 LX5280

Port Name I/O Timing Description

DC_GNTRAM_R output 30% Cpu drives to one to grant external

DCACHE access request.

Coprocessor Interface

C<z>condin input 80% Cop branch flag.

C<z>rd_addr[4:0] output 50% Cop read address.

C<z>rhold output 45% Cop hold condition, one stalls copro-

cessor.

C<z>rd_gen output 50% Cop general register read command.

C<z>rd_con output 50% Cop control register read command.

C<z>rd_data[31:0] input 80% Cop read data.

C<z>wr_addr[4:0] output 20% Cop write address.

C<z>wr_gen output 20% Cop general register write command.

C<z>wr_con output 20% Cop control write address command.

C<z>wr_data[31:0] output 30% Cop write data.

C<z>invld_M output 60% Cop invalid instruction flag, one indi-

cates invalid instruction in M stage.

C<z>xcpn_M output 60% Cop exception flag, one indicates

exception in M stage.

C3cnt_iparet output 20% Count instructions retired Pipe A

C3cnt_ipbret output 20% Count instructions retired Pipe B

C3cnt_ifetch output 20% Count instruction fetches

C3cnt_imiss output 20% Count icache misses

C3cnt_istall output 20% Count icache stalls

C3cnt_dmiss output 20% Count dcache misses

C3cnt_dstall output 20% Count dcache stalls

C3cnt_dload output 20% Count data load operations

C3cnt_dstore output 20% Count data store operations

Custom Engine Interface

CEI_CE1HOLD output 45% CPU is halting Custom Engine.

CEI_CE1INVLD_M output 40% Instruction is not valid, M stage.

CEI_CE1INVLDP_S_R output 30% Instruction is not valid, S stage.

CEI_XCPN_M_C1 output 40% CPU reports exception.

CEI_CE1OP_S_R[11:0] output 30% Custom Engine op code.

CEI_INSTM32_S_R_C1_N output 30% One indicates 32-bit instruction mode;

zero indicates 16-bit instruction mode.

Lexra Proprietary & Confidential - 114 - Release 1.9

 April 30, 2001 LX5280

Port Name I/O Timing Description

CEI_CE1AOP_E_R[31:0] output 35% A operand.

CEI_CE1BOP_E_R[31:0] output 35% B operand.

CE1_RES_E[31:0] input 45% Result from Custom Engine.

CE1_SEL_E_R input 30% One indicates Custom Engine opcode

is present in E stage.

CE1_HALT_E_R[2:0] input 20% Custom Engine stalls processor by

driving to ones, allows processor to run
by driving to zeros. (Copies must be
supplied from multiple registers to meet
timing requirements.)

Lexra Proprietary & Confidential - 115 - Release 1.9

 April 30, 2001 LX5280

Lexra Proprietary & Confidential - 116 - Release 1.9

 April 30, 2001 LX5280

Appendix C. LX5280 Pipeline Stalls

This section documents stall conditions that may arise in the LX5280.

C.1. Stall Definitions

Issue stall: an invalid instruction enters the pipe, while any other valid instructions in the pipe advance.

Pipeline stall: All instructions in either pipe stay in the same stage, and do not advance.

Dual-issue interlock: Only one of the potential pair of instructions enters a pipe, the other instruction of the
pair waits for the next cycle to enter.

Stall: if not otherwise qualified, means pipeline stall.

C.2. Instruction Groupings

These instruction groupings are used to describe stall conditions that are based on the type of instructions in
the pipeline.

Table 50: Instruction Groupings For Stall Definition

Group Name Instructions in Group

M-I-LoadStore: LB, LH, LW, LBU, LHU, LWC1, LWC2, LWC3

SB, SH, SW, SWC1, SWC2, SWC3

M-I-Mac MULT(U),DIV(U),MFHI,MFLO,MTHI,MTLO

M-I-Control J, JAL(X), JR, JALR

BLTZAL, BGEZAL, (linked branches)
SYSCALL, BREAK
All COPz (MFCz, CFCz, MTCz, CTCz, BCFz, BCTz, RFE)
LWCz, SWCz (also in LoadStore group)

M-I-UnlinkedBranch BEQ, BNE, BLEZ, BGTZ, BLTZ, BGEZ

M-I-General All remaining M-I instructions.

MIV-CMove MOVZ, MOVN

M16-LoadStore LB, LH, LWSP, LW, LBU, LHU, LWPC, SB, SH, SWSP, SW,
SWRASP

M16-Mac MULT(U), DIV(U), MFHI, MFLO

M16-Control JAL(X), JR, JALR, BREAK

M16-UnlinkedBranch B, BEQZ, BNEZ, BTEQZ, BTNEZ

M16-General All remaining M16 instructions

RAD-Mac MTA2, MFA, MFA2, MULTA(U), MULTA2, MULNA2, CMULTA,

MADDA(U), MSUBA(U), ADDMA, SUBMA, DIVA(U), MADDA2,
MSUBA2, RNDA2

RAD-LoadStore LT, LTP, LWP, LHP, LBP, LHPU, LBPU, ST, STP, SWP, SHP, SBP

RAD-Control MTRU, MFRU, MTRK, MFRK, MTLXC0, MFLXC0

Lexra Proprietary & Confidential - 117 - Release 1.9

 April 30, 2001 LX5280

Group Name Instructions in Group

RAD-General All remaining RAD instructions

EJTAG-Control DERET, SDBBP, M16SDBBP

C.3. Dual Pipe Issue Rules

These instruction groups must issue to Pipe A:

M-I-LoadStore, M-I-Control, M-I-UnlinkedBranch,

M16-LoadStore, M16-Control, M16-UnlinkedBranch,
RAD-LoadStore, RAD-Control,
EJTAG-Control

These instruction groups must issue to Pipe B:

M-I-Mac, M16-Mac, RAD-Mac

These instruction groups must single issue:

M-I-Control, RAD-Control, EJTAG-Control,

ALL M16 instructions

Instruction doubleword issue rule:

In order for a pair of instructions to dual-issue, they must be found in the same aligned doubleword.

UnlinkedBranch-delay slot rules:

An UnlinkedBranch can dual issue with the preceding instruction, if no other rules are violated. The
delay slot instruction of an M-I-UnlinkedBranch single issues in the cycle following the
UnlinkedBranch.

Producer-consumer Read-After-Write (RAW) hazard:

A pair of instructions will NOT dual issue if the second instruction uses a register updated by the
first instruction. This does not apply to register 0, which never causes an interlock.

Producer-Producer Write-After-Write (WAW) hazard:

A pair of instructions will NOT dual issue if the second instruction updates a register updated by the
first instruction. Unless the common target register is also a source register of the second instruction
(in which case the RAW interlock applies), no useful program is expected to include such a pair of
instructions, since the results of the first update are lost. This does not apply to register 0, which
never causes an interlock.

Examples:

# Both RAW and WAW apply, causing single issue

00: add s0,s1,s2 ; 04: add s0,s0,s3 ; 2xsingle issue (s0 RAW)

# First instruction does no useful work (visible only in case of exception)

00: add s0,s1,s2 ; 04: add s0,s4,s3 ; 2xsingle issue (s0 WAW)

Lexra Proprietary & Confidential - 118 - Release 1.9

 April 30, 2001 LX5280

C.4. M16 32-bit Instructions

M16-JAL(X) issues in two consecutive cycles.

M16 Extended instructions issue in two consecutive cycles.

C.5. Non-Sequential Program Flow Issue Stalls

M-I JR, JALR and M16 JR, JALR, AL(X):

Two issue stalls after the delay slot instruction.

(The delay slot instruction always single issues.)

M-I J, JAL(X), and M-I taken branches:

NO stall cycles after the delay slot instruction.

(The delay slot instruction always single issues.)

M16 taken branches:

One issue stall after the branch.

M-I not-taken branches:

Two issue stalls after the delay slot instruction.

(The delay slot instruction always single issues.)

M16 not-taken branches:

Three issue stalls after the branch.

The branch rules are a consequence of the fact that all branches are assumed to be taken.

C.6. Load/Store Rules

Store twinword dual-issue interlock:

The Store Twinword instructions (ST,STP) always single issue. (Because they use 3 of the 4 register
file read ports, leaving only one for the other instruction, which usually needs two read ports.)

M16 Load slot issue stall:

There is one unconditional issue stall after any M16 Load instruction. (there is no M16 target
register analysis).

Load-use single cycle issue stall:

After a Load instruction to a target register, an instruction which follows the load by one CYCLE
and uses the target register of the load will stall issue for one cycle.

Note: The architectural load-delay slot has been eliminated. This issue stall applies even to the

Lexra Proprietary & Confidential - 119 - Release 1.9

 April 30, 2001 LX5280

instruction immediately following the load.

This does NOT apply to M16 Loads, since they are always followed by a single cycle issue stall.

Examples:

# this executes in 3 cycles:

00: lw s0,0(a0) ; 04: addi a0,4 ; dual issue
08: add s1,s0 ; 0c: add t1,t2 ; stall(s0 Load-Use), dual issue

# this executes in 3 cycles:

00: lw s0,0(a0) ; 04: addi a0,4 ; dual issue
08: add t1,t2 ; 0c: add s1,s0 ; 2xsingle issue (s0 Load-Use)

# this executes in 3 cycles:

00: add t1,t2 ; 04: lw s0,0(a0) ; dual issue
08: addi a0,4 ; 0c: add s1,s0 ; 2xsingle issue (s0 Load-Use)

# this executes in 3 cycles:

00: lw s0,0(a0) ; 04: lw s2,4(a0) ; 2xPipeA single issue
08: add s1,s0 ; 0c: addi a0,8 ; dual issue

# this executes in 3 cycles NOTE ELIMINATION OF ARCHITECTURAL DELAY SLOT!

00: add t1,t2 ; 04: lw s0,0(a0) ; dual issue
08: add s1,s0 ; 0c: addi a0,4 ; stall(s0 Load-Use), dual issue

# this executes in 2 cycles:

00: add t1,t2 ; 04: lw s0,0(a0) ; dual issue
08: add s2,s1 ; 0c: addi a0,4 ; dual issue (s0 not used)

For Twinword Loads (LT, LTP) this rule applies to both of the target registers in the register-pair
operand.

For Radiax Pointer-Update Load instructions, (LBP,LHP,LTP,LWP,LBPU) this rule does NOT
apply to the updated pointer register, which is covered by the RAW and WAW hazard dual-issue
interlocks.

Load sub-word stall:

Load instructions which have Byte or Halfword operands always cause a one-cycle stall.

Store-Load stall:

A Load instruction which follows a Store instruction by one CYCLE always causes a one-cycle
stall.

Note: This stall only applies if the Store instruction hits in the Dcache or has a Byte or Halfword
operand.

Examples:

# this executes in 3 cycles:

00: sw s0,4(a0) ; 04: addi a0,8 ; dual issue
08: add s0,s1 ; 0c: lw s2,0(a0) ; dual issue (and sw-lw stall)

Lexra Proprietary & Confidential - 120 - Release 1.9

 April 30, 2001 LX5280

# this executes in 3 cycles:

00: sw s0,4(a0) ; 04: addi a0,8 ; dual issue
08: lw s2,0(a0) ; 0c: add s0,s1 ; dual issue (and sw-lw stall)

# this executes in 4 cycles:

00: sw s0,4(a0) ; 04: lw s2,8(a0) ; 2xPipeA sing issue (and sw-lw stall)
08: addi a0,8 ; 0c: add s0,s1 ; dual issue

# this executes in 2 cycles:

00: lw s2,0(a0) ; 04: add s0,s1 ; dual issue
08: sw s0,4(a0) ; 0c: addi a0,8 ; dual issue (lw-sw okay)

StoreTwin - StoreAny stall:

Any store instruction which follows a Store Twinword instruction (ST,STP) by one CYCLE, always
causes a single cycle stall.

Examples:

# this executes in 4 cycles:

00: nop ; 04: st s0,8(a0) ; 2xsingle issue (st single issue)
08: add s0,s1 ; 0c: sw s2,0(a0) ; dual issue (and st-sw stall)

# this executes in 3 cycles:

00: st s2,0(a0) ; 04: st s0,8(a0) ; 2xsingle issue (and st-st stall)

StoreAny - StoreSubword stall:

A Store instruction which has a Byte or Halfword operand, and which follows any Store instruction
by one CYCLE, always causes a one-cycle stall. This cycle includes any potential StoreTwin-
StoreAny stall.

Examples:

# this executes in 3 cycles:

00: sw s0,4(a0) ; 04: addi a0,8 ; dual issue
08: add s0,s1 ; 0c: sb s2,0(a0) ; dual issue (and sw-sb stall)

# this executes in 4 cycles:

00: sh s0,4(a0) ; 04: addi a0,8 ; dual issue (and subword stall)
08: sb s2,0(a0) ; 0c: add s0,s1 ; dual issue (and sh-sb stall)

# this executes in 3 cycles:

00: st s2,0(a0) ; 04: sb s0,3(a0) ; 2xsingle issue (and st-sb stall)

# this executes in 4 cycles:

00: nop ; 04: st s0,8(a0) ; 2xsingle issue (st single issue)
08: add s0,s1 ; 0c: sb s2,0(a0) ; dual issue (and st-sb stall)

C.7. Load/Store Ops Stall Matrix

The following table summarizes the stall rules related to Load and Store instructions described above. This

Lexra Proprietary & Confidential - 121 - Release 1.9

 April 30, 2001 LX5280

table does NOT include the RAW and WAW dual-issue interlocks. In this table, the "2nd OP" refers to an
instruction which issues in the CYCLE after the "1st OP".

Table 51: Load/Store Ops Stall Matrix

1st op

M16 M-I/RAD M-I/RAD

2nd op M16 LB(U) LB(U) LBP(U) SB, SBP SW ST
LW LH(U) LH(U) LHP(U) SH, SHP SWP STP

non 1U
1U 2U 1S - - -
load-store 2S

LT, LTP
LW, LWP
LB(U)
1U 2U 1S 2S 1U 1U 1U
LH(U)
LBP(U)
LHP(U)

SB, SBP, SH, 1U

1U 2U 1S 1U 1U 1U
SHP 2S

SW, SWP
1U 2U 1S 2S - - 1U
ST, STP

Notes:

- means no stalls
xU indicates unconditional stall for the indicated number of cycles
xS indicates stall only if 2ndOp Source = 1stOp Load-target
xW indicates stall if data RAMs have word-write granularity

C.8. Mac Ops Interlock Matrix

The Mac in the 5280 eliminates all programming hazards between Mac instructions by stalling the pipeline
as necessary. This is done both to avoid resource conflicts as well as to wait for results of a first instruction
that is needed by a second instruction.

The following table indicates the number of cycles that must be inserted between the first indicated
instruction and the second. A zero (or dash) indicates that the instructions can issue back-to-back to the Mac
pipe with no stalls. A non-zero number indicates the number of stall cycles that will occur if the instructions
are issued in consecutive cycles. These stall cycles are available for any other non-Mac instructions, but
should NOT be filled with NOPs since that would only increase the code footprint without improving
performance.

Lexra Proprietary & Confidential - 122 - Release 1.9

 April 30, 2001 LX5280

Table 52: Cycles Required Between Dual MAC Instructions

MADDA2[.S]
MSUBA2[.S]
ADDMA[.S]
SUBMA[.S]
MULTA2
1st Op MULNA2
MADDA(U) RNDA2 MFA
MSUBA(U) MTA2 MFA2
MULTA(U) MAD(U) DIVA(U) MTHI MFHI
2nd Op MULT(U) MSUB(U) CMULTA DIV(U) MTLO MFLO

MULTA(U)
MADDA(U)
MSUBA(U) 1U 1U 1U (19T) - -
MAD(U)
MSUB(U)

DIVA(U)
(3T) (4T) (1T) 19U - -
DIV(U)

CMULTA
MADDA2[.S]
MSUBA2[.S]
MULTA2
3U 4U 1U (19T) - -
MULNA2
MTA2
MTHI
MTLO

ADDMA[.S] LO HI
1S (19S)
SUBMA[.S] 2S 3S 4U - -
1T (19T)
RNDA2 2T 3T

MFA
MFA2 LO HI LO HI
MFHI 4S 5S 5S 6S 3S 19S 2S -
MFLO

Nomenclature:

- = The two ops can be issued back-to-back.

xU = Unconditional delay of the indicated number of cycles.
xS = Delay only if (any) 2nd Op source is the same as (any) 1st Op tar-
get (producer-consumer dependency).
xT = Delay only if (any) 2nd Op target is the same as (any) 1s Op tar-
get (preserve write after write order).
( ) = Items in parenthesis are unlikely to occur in any useful program,
which would probably have an intervening MFA.
LO/HI = For the 72-bit result of a 32x32 MULT or MADDA, the LO 32-bits
(m0l, m1l, etc.) are available one cycle earlier.

Delay of “x” cycles means that if the 1st Op issues in cycle N, then the 2nd Op may
issue in cycle N+x+1.

Lexra Proprietary & Confidential - 123 - Release 1.9

 April 30, 2001 LX5280

Examples:

# this executes in 7 cycles:

00: mult s0,s1 ; 04: addi a0,8 ; dual issue
08: lw s0,0(a0) ; 0c: lw s1,0(a0) ; 2xPipeA sing issue
10: addi a1,4 ; 14: mflo v0 ; dual issue (and stall2)
18: sw v0,0(a1) ; 1c: nop ; dual issue

C.9. MVCz Stall

The coprocessor move instructions (M-I: LWCz, MTCz, MFCz, and Radiax: MTLXC0, MFLXC0, MTRU,
MFRU, MTRK, MFRK) always single issue and are always followed by a single cycle issue stall.

C.10. ZovLoop Rules

There are no special ZovLoop rules but the execution of a ZovLoop must follow all of the other rules as it
wraps from the loop end back to the loop start. Unless one of these other rules require it, there are NO stall
cycles between loop end and loop start.

Examples:

lps = 00, lpe = 0c unless otherwise noted

# executes in 2 cycles per loop BUT GETS THE WRONG ANSWER!

00: add s1,s0 ; 04: addi a0,8 ; dual issue
08: add t1,t2 ; 0c: lw s0,0(a0) ; dual issue
00: add s1,s0 ; 04: addi a0,8 ; dual issue BAD s1 RESULT!
08: add t1,t2 ; 0c: lw s0,0(a0) ; dual issue

# executes in 3 cycles per loop

00: addi a0,8 ; 04: add s1,s0 ; dual issue
08: add t1,t2 ; 0c: lw s0,0(a0) ; dual issue
00: addi a0,8 ; 04: add s1,s0 ; 2x single issue (s0 Load-Use)
08: add t1,t2 ; 0c: lw s0,0(a0) ; dual issue

# executes in 2 cycles per loop

00: addi a0,8 ; 04: add t1,t2 ; dual issue
08: add s1,s0 ; 0c: lw s0,0(a0) ; dual issue
00: addi a0,8 ; 04: add t1,t2 ; dual issue
08: add s1,s0 ; 0c: lw s0,0(a0) ; dual issue

# executes in 4 cycles per loop (block copy, one word per loop)
00: addi a0,4 ; 04: lw s0,0(a1) ; dual issue
08: addi a1,4 ; 0c: sw s0,0(a0) ; 2x single issue (s0 Load-use)
addi a0,4 ; 04: lw s0,0(a1) ; 2x single issue (sw-lw stall)
08: addi a1,4 ; 0c: sw s0,0(a0) ; 2x single issue (s0 Load-use)

# executes in 5 cycles per loop (block copy, two words per loop)
(lps = 00, lpe = 14)
00: lw s0,0(a1) ; 04: lw s1,4(a1) ; 2x PipeA single issue
08: sw s0,0(a0) ; 0c: sw s1,4(a0) ; 2x PipeA single issue
10: addi a0,8 ; 14: addi a1,8 ; dual issue
00: lw s0,0(a1) ; 04: lw s1,4(a1) ; 2x PipeA single issue
08: sw s0,0(a0) ; 0c: sw s1,4(a0) ; 2x PipeA single issue
10: addi a0,8 ; 14: addi a1,8 ; dual issue

Lexra Proprietary & Confidential - 124 - Release 1.9

 April 30, 2001 LX5280

Use pointer-update instructions

# executes in 3 cycles per loop (poor)

00: lwp s0,(a0)8; 04: add s1,s2 ; dual issue
08: lw s2,4(a0); 0c: add s1,s0 ; 2x single issue (s0 Load-Use)
00: lwp s0,(a0)8; 04: add s1,s2 ; dual issue (s2 already delayed)
08: lw s2,4(a0); 0c: add s1,s0 ; 2x single issue (s0 Load-Use)

# executes in 2 cycles per loop (optimum)

00: add s1,s0 ; 04: lwp s0,(a0)8; dual issue
08: add s1,s2 ; 0c: lw s2,4(a0); dual issue
00: add s1,s0 ; 04: lwp s0,(a0)8; dual issue
08: add s1,s2 ; 0c: lw s2,4(a0); dual issue

lps = 04, lpe = 10 (poor alignment)

# executes in 3 cycles per loop

00: ; 04: add s1,s0 ; single issue
08: add t1,t2 ; 0c: lw s0,0(a0) ; dual issue
10: addi a0,8 ; 14: single issue

00: ; 04: add s1,s0 ; single issue

08: add t1,t2 ; 0c: lw s0,0(a0) ; dual issue
10: addi a0,8 ; 14: single issue

C.11. IMMU Stalls

IMMU stall:

When the program jumps, branches, or increments between the two most recently used pages, a
single cycle stall is incurred.

When the program jumps, branches or increments to a third page a two-cycle stall is incurred.

IMMU Issue Stall

When an IMMU stall occurs due to incrementing across a page boundary, AND there is any of the
following instructions found anywhere in the last doubleword of the page, then there is one issue
stall in addition to the IMMU stalls:

M-I or M16 branch of any kind

M-I J, JAL(X)
EJTAG DERET
M16 EXTEND
M16 JALX first half

C.12. Cache Miss Stalls

Instruction cache miss stall:

When an instruction cache miss occurs, the processor is stalled for the duration of the cache line fill
operation.

The number of cycles required to complete the line fill is system dependent.

Instruction cache 2-way soft miss stall:

Lexra Proprietary & Confidential - 125 - Release 1.9

 April 30, 2001 LX5280

When a 2-way set associative instruction cache is in use, a soft-miss is defined as a hit in the
unpredicted set, with set prediction defined as follows:

If not running in Lock mode, or if the current cache index has no Locked line, set prediction is based
onthe LRU bit (predict the non-least recently used set at the current cache index.)

If running in Lock mode, and the current cache index has a Locked line, set prediction is based on
the previous Icache access (predict the Locked set if the previous Icache access hit a Locked line
and vice versa).

A soft miss always causes a two-cycle stall.

Data cache miss stall:

When a data cache miss occurs as the result of a load instruction, the processor stalls while it waits
for the data. The data cache releases the stall condition after the required word is supplied to the
processor, even if additional words must still be filled into the data cache. However, if the processor
issues another load or store operation to the data cache while the remainder of the line fill is in
progress, the cache will again stall the processor until the line fill operation is completed.

When a data cache miss occurs as a result of a load byte or load halfword, the processor stalls for
the duration of the cache line fill operation.

The number of cycles required to complete the line fill is system dependent.

C.13. Non-Sequential Program Flow Issue Stall Pipeline Diagrams

M-I JR,JALR and M16 JR,JALR,JAL(X):

JR I D S E M W
delayslot I D S E M W
notvld I . . .
notvld I . .
target I D S E
I D S E

M-I J, JAL(X), and M-I taken branches:

J I D S E M W
delayslot I D S E M W
target I D S E M
I D S E M

M16 taken branches:

B I D S E M W
notvld I . . . . .
target I D S E M

M-I not-taken branches:

B-ntkn I D S E M W
delayslot I D S E M W
notvld I . . .
notvld I . . .
delay+4 I D S
I D S

Lexra Proprietary & Confidential - 126 - Release 1.9

 April 30, 2001 LX5280

M16 not-taken branches:

B-ntkn I D S E M W
notvld I . . . . .
notvld I . . . .
notvld I . . .
delay+4 I D S

C.14. Load/Store Stall Pipeline Diagrams

M16 Load slot issue stall:

Load I D S E M W
notvld I . . . . .
Load+2 I D S E M

Load-Use single cycle issue stall:

00: lw s0,0(a0) I D S E M W
04: addi a0,4 I D S E M W
08: add s1,s0 I d D S E M W
0c: add t1,t2 I d D S E M W

00: lw s0,0(a0) I D S E M W
04: addi a0,4 I D S E M W
08: add t1,t2 I D S E M W
0c: add s1,s0 I d D S E M W

00: add t1,t2 I D S E M W

04: lw s0,0(a0) I D S E M W
08: addi a0,4 I D S E M W
0c: add s1,s0 I d D S E M W

00: lw s0,0(a0) I D S E M W
04: lw s2,4(a0) I d D S E M W
08: add s1,s0 I D S E M W
0c: addi a0,8 I D S E M W

Load Subword stall:

00: lb I D S E M M W
04: foo1 I D S E M M W
08: foo2 I D S E E M W
0c: foo3 I D S E E M W
10: foo4 I D S S E M W
14: foo5 I D S S E M W

RHOLD X

Store-Load stall:

00: sw s0,4(a0) I D S E M W
04: addi a0,8 I D S E M W
08: add s0,s1 I D S E M M W
0c: lw s2,0(a0) I D S E M M W
10: foo2 I D S E E M W
14: foo3 I D S E E M W

RHOLD X

Lexra Proprietary & Confidential - 127 - Release 1.9

 April 30, 2001 LX5280

00: sw s0,4(a0) I D S E M W
04: lw s2,8(a0) I d D S E M M W
08: addi a0,8 I D S E E M W
0c: add s0,s2 I D S E E M W
10: foo2 I D S S E M W
14: foo3 I D S S E M W

RHOLD X

StoreTwin - StoreAny stall:

00: nop I D S E M W
04: st s0,8(a0) I d D S E M W1 W2
08: add s0,s1 I D S E M M W
0c: sw s2,0(a0) I D S E M M W
10: foo2 I D S E E M W
14: foo3 I D S E E M W

RHOLD X

00: st s2,0(a0) I D S E M W1 W2
04: st s0,8(a0) I d D S E M M W1 W2
08: foo2 I D S E E M W
0c: foo3 I D S E E M W
10: foo4 I D S S E M W
14: foo5 I D S S E M W

RHOLD X

StoreAny - StoreSubword stall:

00: sw s0,4(a0) I D S E M W
04: addi a0,8 I D S E M W
08: add s0,s1 I D S E M M W
0c: sb s2,0(a0) I D S E M M W
10: foo2 I D S E E M W
14: foo3 I D S E E M W

RHOLD X

00: sh s0,4(a0) I D S E M M W
04: addi a0,8 I D S E M M W
08: sb s2,0(a0) I D S E E M M W
0c: add s0,s1 I D S E E M M W
10: foo2 I D S S E E M W
14: foo3 I D S S E E M W

RHOLD X X

00: st s2,0(a0) I D S E M W1 W2
04: sb s0,3(a0) I d D S E M M W
08: foo2 I D S E E M W
0c: foo3 I D S E E M W
10: foo4 I D S S E M W
14: foo5 I D S S E M W

RHOLD X

Lexra Proprietary & Confidential - 128 - Release 1.9

 April 30, 2001 LX5280

00: nop I D S E M W
04: st s0,8(a0) I d D S E M W1 W2
08: add s0,s1 I D S E M M W
0c: sb s2,0(a0) I D S E M M W
10: foo2 I D S E E M W
14: foo3 I D S E E M W

RHOLD X

C.15. Mac Ops Interlock Pipeline Diagram

00: mult s0,s1 I D S E M -

04: addi a0,8 I D S E M W
08: lw s0,0(a0) I D S E M W
0c: lw s1,0(a0) I D S E M M M W
10: addi a1,4 I D S E E E M W
14: mflo v0 I D S E E E M W
18: sw v0,0(a1) I D S S S E M W
1c: nop I D S S S E M W

multcount(4S) 0 1 2 3 4
RHOLD X X

C.16. MVCz Stall Pipeline Diagrams

00: mtc0 I D S E M W
notvld I . . . . . .
04: foo I d d D S E M W
08: foo1 I D S E M W
0c: foo2 I D S E M W

00: nop I D S E M W
04: mtc0 I d D S E M W
notvld I . . . . .
08: foo1 I d D S E M W
0c: foo2 I d D S E M W
10: foo3 I D S E M W
14: foo4 I D S E M W

C.17. ZovLoop Pipeline Diagrams

00: addi a0,4 I D S E M W

04: lw s0,0(a1) I D S E M W
08: addi a1,4 I D S E M W
0c: sw s0,0(a0) I d D S E M W
00: addi a0,4 I D S E M W
04: lw s0,0(a1) I D S E M M W
08: addi a1,4 I D S E E M W
0c: sw s0,0(a0) I d D S S E M W

RHOLD X

Lexra Proprietary & Confidential - 129 - Release 1.9

 April 30, 2001 LX5280

00: lw s0,0(a1) I D S E M W
04: lw s1,4(a1) I d D S E M W
08: sw s0,0(a0) I D S E M W
0c: sw s1,4(a0) I d D S E M W
10: addi a0,8 I D S E M W
14: addi a1,8 I D S E M W

00: lwp s0,(a0)8 I D S E M W

04: add s1,s2 I D S E M W
08: lw s2,4(a0) I D S E M W
0c: add s1,s0 I d D S E M W
00: lwp s0,(a0)8 I D S E M W
04: add s1,s2 I D S E M W
08: lw s2,4(a0) I D S E M W
0c: add s1,s0 I d D S E M W

C.18. Cache Miss Pipeline Diagrams

Icache miss pipeline diagram:

00: foo0 I D S E M M M M M M W
04: foo1 I D S E M M M M M M W
08: foo2 I D S E E E E E E M W
0c: foo3 I D S E E E E E E M W
10: foo4 I ~d . . . I D S E M W
14: foo5 I ~d . . . I D S E M W

RHOLD X X X X X

Icache 2-way soft miss pipeline diagram:

00: foo0 I D S E M M M W
04: foo1 I D S E M M M W
08: foo2 I D S E E E M W
0c: foo3 I D S E E E M W
10: foo4 I ~d I D S E M W
14: foo5 I ~d I D S E M W
18: foo6 I D S E M W
1c: foo7 I D S E M W

RHOLD X X

Dcache miss pipeline diagram:

00: foo I D S E M W
04: lw I D S E M . . . . W
08: foo1 I D S E M M M M M W
0c: foo2 I D S E M M M M M W
08: foo3 I D S E E E E E M W
0c: foo4 I D S E E E E E M W

RHOLD X X X X

Lexra Proprietary & Confidential - 130 - Release 1.9