The LLVM Compiler

Framework and Infrastructure

15-411: Compiler Design
Slides by David Koes

Substantial portions courtesy Chris Lattner and Vikram Adve

LLVM Compiler System

The LLVM Compiler Infrastructure

Provides reusable components for building compilers
Reduce the time/cost to build a new compiler
Build static compilers, JITs, trace-based optimizers, ...

The LLVM Compiler Framework

End-to-end compilers using the LLVM infrastructure
C and C++ gcc frontend
Backends for C, X86, Sparc, PowerPC, Alpha, Arm,

Thumb, IA-64
2

Three primary LLVM components

The LLVM Virtual Instruction Set

The common language- and target-independent IR
Internal (IR) and external (persistent) representation

A collection of well-integrated libraries

Analyses, optimizations, code generators, JIT

compiler, garbage collection support, profiling,

A collection of tools built from the libraries

Assemblers, automatic debugger, linker, code

generator, compiler driver, modular optimizer,

3

Tutorial Overview

Introduction to the running example

LLVM C/C++ Compiler Overview
High-level view of an example LLVM compiler

The LLVM Virtual Instruction Set

IR overview and type-system

LLVM C++ IR and important APIs

Basics, PassManager, dataflow, ArgPromotion

Important LLVM Tools

Running example: arg promotion

Consider use of by-reference parameters:
int callee(const int &X) {
return X+1;
}
int caller() {
return callee(4);
}

We want:
int callee(int X) {
return X+1;
}
int caller() {
return callee(4);
}

compiles to

int callee(const int *X) {

return *X+1; // memory load
}
int caller() {
int tmp;
// stack object
tmp = 4;
// memory store
return callee(&tmp);
}

Eliminated load in callee

Eliminated store in caller
Eliminated stack slot for tmp
5

Why is this hard?

Requires interprocedural analysis:

Must change the prototype of the callee
Must update all call sites we must know all callers
What about callers outside the translation unit?

Requires alias analysis:

Reference could alias other pointers in callee
Must know that loaded value doesnt change from

function entry to the load

Must know the pointer is not being stored through

Reference might not be to a stack object!

6

Tutorial Overview

Introduction to the running example

LLVM C/C++ Compiler Overview
High-level view of an example LLVM compiler

The LLVM Virtual Instruction Set

IR overview and type-system

LLVM C++ IR and important APIs

Basics, PassManager, dataflow, ArgPromotion

Important LLVM Tools

The LLVM C/C++ Compiler

From the high level, it is a standard compiler:

Compatible with standard makefiles
Uses GCC 4.2 C and C++ parser
Generates native executables/object files/assembly

Distinguishing features:
Uses LLVM optimizers, not GCC optimizers
Pass -emit-llvm to output LLVM IR

-S: human readable assembly

-c: efficient bitcode binary
8

Looking into events at compile-time

C/C++ file

IR

GENERIC

llvm-gcc/llvm-g++ -O -S

GIMPLE
(tree-ssa)

LLVM IR

assembly

Machine
Code IR

-emit-llvm

LLVM
asm

>50 LLVM Analysis & Optimization Passes:

Dead Global Elimination, IP Constant Propagation, Dead
Argument Elimination, Inlining, Reassociation, LICM, Loop
Opts, Memory Promotion, Dead Store Elimination, ADCE,
9

Looking into events at link-time

LLVM bitcode .o file
LLVM bitcode .o file

LLVM
Linker

llvm-ld

Link-time
Optimizer

executable

.bc file for LLVM JIT

-native

Native executable

>30 LLVM Analysis &

Optimization Passes
Optionally internalizes:
marks most functions as
internal, to improve IPO
Perfect place for argument
promotion optimization!
10

Goals of the compiler design

Analyze and optimize as early as possible:

Compile-time opts reduce modify-rebuild-execute cycle
Compile-time optimizations reduce work at link-time (by

shrinking the program)

All IPA/IPO make an open-world assumption

Thus, they all work on libraries and at compile-time
Internalize pass enables whole program optzn

One IR (without lowering) for analysis & optzn

Compile-time optzns can be run at link-time too!
The same IR is used as input to the JIT

IR design is the key to these goals!

11

Tutorial Overview

Introduction to the running example

LLVM C/C++ Compiler Overview
High-level view of an example LLVM compiler

The LLVM Virtual Instruction Set

IR overview and type-system

LLVM C++ IR and important APIs

Basics, PassManager, dataflow, ArgPromotion

Important LLVM Tools

12

Goals of LLVM IR

Easy to produce, understand, and define!

Language- and Target-Independent
AST-level IR (e.g. ANDF, UNCOL) is not very feasible

Every analysis/xform must know about all languages

One IR for analysis and optimization

IR must be able to support aggressive IPO, loop opts,

scalar opts, high- and low-level optimization!

Optimize as much as early as possible

Cant postpone everything until link or runtime
No lowering in the IR!
13

LLVM Instruction Set Overview #1

Low-level and target-independent semantics

RISC-like three address code
Infinite virtual register set in SSA form
Simple, low-level control flow constructs
Load/store instructions with typed-pointers

IR has text, binary, and in-memory forms

bb:
; preds = %bb, %entry
%i.1 = phi i32 [ 0, %entry ], [ %i.2, %bb ]
%AiAddr = getelementptr float* %A, i32 %i.1
for (i = 0; i < N;
call void @Sum( float* %AiAddr, %pair* %P )
%i.2 = add i32 %i.1, 1
++i)
%exitcond = icmp eq i32 %i.2, %N
Sum(&A[i], &P);
br i1 %exitcond, label %return, label %bb
14

LLVM Instruction Set Overview #2

High-level information exposed in the code

Explicit dataflow through SSA form
Explicit control-flow graph (even for exceptions)
Explicit language-independent type-information
Explicit typed pointer arithmetic

Preserve array subscript and structure indexing

bb:
; preds = %bb, %entry
%i.1 = phi i32 [ 0, %entry ], [ %i.2, %bb ]
%AiAddr = getelementptr float* %A, i32 %i.1
for (i = 0; i < N;
call void @Sum( float* %AiAddr, %pair* %P )
%i.2 = add i32 %i.1, 1
++i)
%exitcond = icmp eq i32 %i.2, %N
Sum(&A[i], &P);
br i1 %exitcond, label %return, label %bb
15

LLVM Type System Details

The entire type system consists of:

Primitives: integer, floating point, label, void

no signed integer types

arbitrary bitwidth integers (i32, i64, i1)

Derived: pointer, array, structure, function, vector,

No high-level types: type-system is language neutral!

Type system allows arbitrary casts:

Allows expressing weakly-typed languages, like C
Front-ends can implement safe languages
Also easy to define a type-safe subset of LLVM
See also: docs/LangRef.html
16

Lowering source-level types to LLVM

Source language types are lowered:

Rich type systems expanded to simple type system
Implicit & abstract types are made explicit & concrete

Examples of lowering:
References turn into pointers: T& T*
Complex numbers: complex float { float, float }
Bitfields: struct X { int Y:4; int Z:2; } { i32 }
Inheritance: class T : S { int X; } { S, i32 }
Methods: class T { void foo(); } void foo(T*)

Same idea as lowering to machine code

17

LLVM Program Structure

Module contains Functions/GlobalVariables

Module is unit of compilation/analysis/optimization

Function contains BasicBlocks/Arguments

Functions roughly correspond to functions in C

BasicBlock contains list of instructions

Each block ends in a control flow instruction

Instruction is opcode + vector of operands

All operands have types
Instruction result is typed

18

Our example, compiled to LLVM

int callee(const int *X) {
return *X+1; // load
}
int caller() {
int T;
// on stack
T = 4;
// store
return callee(&T);
}

All loads/stores are

Stack allocation is
explicit in the LLVM
explicit in LLVM
representation

define internal i32 @callee(i32* %X) {

entry:
%tmp2 = load i32* %X
%tmp3 = add i32 %tmp2, 1
ret i32 %tmp3
}
define internal i32 @caller() {
entry:
%T = alloca i32
store i32 4, i32* %T
%tmp1 = call i32 @callee( i32* %T )
ret i32 %tmp1
}

19

Our example, desired transformation

define i32 @callee(i32* %X) {
%tmp2 = load i32* %X
%tmp3 = add i32 %tmp2, 1
ret i32 %tmp3
}

define internal i32 @callee1(i32 %X.val)

{
%tmp3 = add i32 %X.val, 1
ret i32 %tmp3
}

define i32 @caller() {

%T = alloca i32
store i32 4, i32* %T
%tmp1 = call i32 @callee( i32* %T )
ret i32 %tmp1
}

define internal i32 @caller() {

%T = alloca i32
store i32 4, i32* %T
%Tval = load i32* %T
%tmp1 = call i32 @callee1( i32 %Tval )
ret i32 %tmp1
}

Other transformation
Update
all
call
sites of
Change
theinstructions
prototype
Insert
load
(-mem2reg) cleans up
callee
for
intothe
allfunction
callers
the rest

define internal i32 @caller() {

%tmp1 = call i32 @callee1( i32 4 )
ret i32 %tmp1
}

20

Tutorial Overview

Introduction to the running example

LLVM C/C++ Compiler Overview
High-level view of an example LLVM compiler

The LLVM Virtual Instruction Set

IR overview and type-system

LLVM C++ IR and important APIs

Basics, PassManager, dataflow, ArgPromotion

Important LLVM Tools

21

LLVM Coding Basics

Written in modern C++, uses the STL:

Particularly the vector, set, and map classes

LLVM IR is almost all doubly-linked lists:

Module contains lists of Functions & GlobalVariables
Function contains lists of BasicBlocks & Arguments
BasicBlock contains list of Instructions

Linked lists are traversed with iterators:

Function *M =
for (Function::iterator I = M->begin(); I != M->end(); ++I) {
BasicBlock &BB = *I;
...

See also: docs/ProgrammersManual.html

22

LLVM Coding Basics cont.

BasicBlock doesnt provide a reverse iterator

Highly obnoxious when doing the assignment
for(BasicBlock::iterator I = bb->end(); I != bb->begin(); ) {
--I;
Instruction *insn = I;

Traversing successors of a BasicBlock:

for (succ_iterator SI = succ_begin(bb), E = succ_end(bb);
SI != E; ++SI) {
BasicBlock *Succ = *SI;

C++ is not Java

valgrind toinitialized
the rescue!
primitive class variable not automatically
http://valgrind.org
you must manage memory
virtual vs. non-virtual functions
and much much more
23

LLVM Pass Manager

Compiler is organized as a series of passes:

Each pass is one analysis or transformation

Types of Pass:
ModulePass: general interprocedural pass
CallGraphSCCPass: bottom-up on the call graph
FunctionPass: process a function at a time
LoopPass: process a natural loop at a time
BasicBlockPass: process a basic block at a time

Constraints imposed (e.g. FunctionPass):

FunctionPass can only look at current function
Cannot maintain state across functions
See also: docs/WritingAnLLVMPass.html
24

Services provided by PassManager

Optimization of pass execution:

Process a function at a time instead of a pass at a time
Example: If F, G, H are three functions in input pgm:

FFFFGGGGHHHH not FGHFGHFGHFGH

Process functions in parallel on an SMP (future work)

Declarative dependency management:

Automatically fulfill and manage analysis pass lifetimes
Share analyses between passes when safe:

e.g. DominatorSet live unless pass modifies CFG

Avoid boilerplate for traversal of program

See also: docs/WritingAnLLVMPass.html
25

Pass Manager + Arg Promotion #1/2

Arg Promotion is a CallGraphSCCPass:

Naturally operates bottom-up on the CallGraph

Bubble pointers from callees out to callers

24: #include "llvm/CallGraphSCCPass.h"

47: struct SimpleArgPromotion : public CallGraphSCCPass {

Arg Promotion requires AliasAnalysis info

To prove safety of transformation

Works with any alias analysis algorithm though

48: virtual void getAnalysisUsage(AnalysisUsage &AU)

AU.addRequired<AliasAnalysis>();
// Get
AU.addRequired<TargetData>();
// Get
CallGraphSCCPass::getAnalysisUsage(AU); // Get
}

const {
aliases
data layout
CallGraph
26

Pass Manager + Arg Promotion #2/2

Finally, implement runOnSCC (line 65):

bool SimpleArgPromotion::
runOnSCC(const std::vector<CallGraphNode*> &SCC) {
bool Changed = false, LocalChange;
do {
// Iterate until we stop promoting from this SCC.
LocalChange = false;
// Attempt to promote arguments from all functions in this SCC.
for (unsigned i = 0, e = SCC.size(); i != e; ++i)
LocalChange |= PromoteArguments(SCC[i]);
Changed |= LocalChange; // Remember that we changed something.
} while (LocalChange);
return Changed;
// Passes return true if something changed.
}

static int foo(int ***P) {

return ***P;
}

static int foo(int P_val_val_val) {

return P_val_val_val;
}
27

LLVM Dataflow Analysis

LLVM IR is in SSA form:

use-def and def-use chains are always available
All objects have user/use info, even functions

Control Flow Graph is always available:

Exposed as BasicBlock predecessor/successor lists
Many generic graph algorithms usable with the CFG

Higher-level info implemented as passes:

Dominators, CallGraph, induction vars, aliasing, GVN,
See also: docs/ProgrammersManual.html
28

Arg Promotion: safety check #1/4

#1: Function must be internal (aka static)
88: if (!F || !F->hasInternalLinkage()) return false;

#2: Make sure address of F is not taken

In LLVM, check that there are only direct calls using F
99: for (Value::use_iterator UI = F->use_begin();
UI != F->use_end(); ++UI) {
CallSite CS = CallSite::get(*UI);
if (!CS.getInstruction()) // "Taking the address" of F.
return false;

#3: Check to see if any args are promotable:

114: for (unsigned i = 0; i != PointerArgs.size(); ++i)
if (!isSafeToPromoteArgument(PointerArgs[i]))
PointerArgs.erase(PointerArgs.begin()+i);
if (PointerArgs.empty()) return false; // no args promotable
29

Arg Promotion: safety check #2/4

#4: Argument pointer can only be loaded from:
No stores through argument pointer allowed!
// Loop over all uses of the argument (use-def chains).
138: for (Value::use_iterator UI = Arg->use_begin();
UI != Arg->use_end(); ++UI) {
// If the user is a load:
if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
// Don't modify volatile loads.
if (LI->isVolatile()) return false;
Loads.push_back(LI);
} else {
return false; // Not a load.
}
}
30

Arg Promotion: safety check #3/4

#5: Value of *P must not change in the BB
We move load out to the caller, value cannot change!

load P

Modifie
s *P?

// Get AliasAnalysis implementation from the pass manager.

156: AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
// Ensure *P is not modified from start of block to load
169: if (AA.canInstructionRangeModify(BB->front(), *Load,
Arg, LoadSize))
return false; // Pointer is invalidated!

See also: docs/AliasAnalysis.html

31

Arg Promotion: safety check #4/4

#6: *P cannot change from Fn entry to BB
Entry

Modifie
s *P?

Modifie
s *P?

Entry
load P

load P
175: for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);
PI != E; ++PI)
// Loop over predecessors of BB.
// Check each block from BB to entry (DF search on inverse graph).
for (idf_iterator<BasicBlock*> I = idf_begin(*PI);
I != idf_end(*PI); ++I)
// Might *P be modified in this basic block?
if (AA.canBasicBlockModify(**I, Arg, LoadSize))
return false;
32

Arg Promotion: xform outline #1/4

#1: Make prototype with new arg types: #197
Basically just replaces int* with int in prototype

#2: Create function with new prototype:

214: Function *NF = new Function(NFTy, F->getLinkage(),
F->getName());
F->getParent()->getFunctionList().insert(F, NF);

#3: Change all callers of F to call NF:

// If there are uses of F, then calls to it remain.
221: while (!F->use_empty()) {
// Get a caller of F.
CallSite CS = CallSite::get(F->use_back());

33

Arg Promotion: xform outline #2/4

#4: For each caller, add loads, determine args
Loop over the args, inserting the loads in the caller
220: std::vector<Value*> Args;
226: CallSite::arg_iterator AI = CS.arg_begin();
for (Function::aiterator I = F->abegin(); I != F->aend();
++I, ++AI)
if (!ArgsToPromote.count(I))
// Unmodified argument.
Args.push_back(*AI);
else {
// Insert the load before the call.
LoadInst *LI = new LoadInst(*AI, (*AI)->getName()+".val",
Call); // Insertion point
Args.push_back(LI);
}

34

Arg Promotion: xform outline #3/4

#5: Replace the call site of F with call of NF
// Create the call to NF with the adjusted arguments.
242: Instruction *New = new CallInst(NF, Args, "", Call);
// If the return value of the old call was used, use the retval of the new call.
if (!Call->use_empty())
Call->replaceAllUsesWith(New);
// Finally, remove the old call from the program, reducing the use-count of F.
Call->getParent()->getInstList().erase(Call);

#6: Move code from old function to new Fn

259: NF->getBasicBlockList().splice(NF->begin(),
F->getBasicBlockList());

35

Arg Promotion: xform outline #4/4

#7: Change users of Fs arguments to use NFs
264: for (Function::aiterator I = F->abegin(), I2 = NF->abegin();
I != F->aend(); ++I, ++I2)
if (!ArgsToPromote.count(I)) { // Not promoting this arg?
I->replaceAllUsesWith(I2);
// Use new arg, not old arg.
} else {
while (!I->use_empty()) {
// Only users can be loads.
LoadInst *LI = cast<LoadInst>(I->use_back());
LI->replaceAllUsesWith(I2);
LI->getParent()->getInstList().erase(LI);
}
}

#8: Delete old function:

286: F->getParent()->getFunctionList().erase(F);
36

Tutorial Overview

Introduction to the running example

LLVM C/C++ Compiler Overview
High-level view of an example LLVM compiler

The LLVM Virtual Instruction Set

IR overview and type-system

LLVM C++ IR and important APIs

Basics, PassManager, dataflow, ArgPromotion

Important LLVM Tools

37

LLVM tools: two flavors

Primitive tools: do a single job

llvm-as: Convert from .ll (text) to .bc (binary)
llvm-dis: Convert from .bc (binary) to .ll (text)
llvm-link: Link multiple .bc files together
llvm-prof: Print profile output to human readers
llvmc: Configurable compiler driver

Aggregate tools: pull in multiple features

bugpoint: automatic compiler debugger
llvm-gcc/llvm-g++: C/C++ compilers

See also: docs/CommandGuide/

38

opt tool: LLVM modular optimizer

Invoke arbitrary sequence of passes:

Completely control PassManager from command line
Supports loading passes as plugins from .so files

opt -load foo.so -pass1 -pass2 -pass3 x.bc -o y.bc

Passes register themselves:

61: RegisterOpt<SimpleArgPromotion> X("simpleargpromotion",

"Promote 'by reference' arguments to 'by value'");

From this, they are exposed through opt:

> opt -load libsimpleargpromote.so help
...
-sccp
- Sparse Conditional Constant Propagation
-simpleargpromotion - Promote 'by reference' arguments to 'by
-simplifycfg
- Simplify the CFG
...
39

Running Arg Promotion with opt

Basic execution with opt:

opt -simpleargpromotion in.bc -o out.bc

Load .bc file, run pass, write out results

Use -load filename.so if compiled into a library
PassManager resolves all dependencies

Optionally choose an alias analysis to use:

opt basicaa simpleargpromotion

(default)

Alternatively, steens-aa, anders-aa, ds-aa,

Other useful options available:

-stats: Print statistics collected from the passes
-time-passes: Time each pass being run, print output
40

Example -stats output (176.gcc)

===-------------------------------------------------------------------------===
... Statistics Collected ...
===-------------------------------------------------------------------------===
23426 adce
- Number of instructions removed
1663 adce
- Number of basic blocks removed
5052592 bytecodewriter
- Number of bytecode bytes written
57489 cfgsimplify
- Number of blocks simplified
4186 constmerge
- Number of global constants merged
211 dse
- Number of stores deleted
15943 gcse
- Number of loads removed
54245 gcse
- Number of instructions removed
253 inline
- Number of functions deleted because all callers found
3952 inline
- Number of functions inlined
9425 instcombine
- Number of constant folds
160469 instcombine
- Number of insts combined
208 licm
- Number of load insts hoisted or sunk
4982 licm
- Number of instructions hoisted out of loop
350 loop-unroll
- Number of loops completely unrolled
30156 mem2reg
- Number of alloca's promoted
2934 reassociate
- Number of insts with operands swapped
650 reassociate
- Number of insts reassociated
67 scalarrepl
- Number of allocas broken up
279 tailcallelim
- Number of tail calls removed
25395 tailduplicate
- Number of unconditional branches eliminated
..........................

41

Example -time-passes (176.gcc)

===-------------------------------------------------------------------------===
... Pass execution timing report ...
===-------------------------------------------------------------------------===
---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Name --16.2400 ( 23.0%) 0.0000 ( 0.0%) 16.2400 ( 22.9%) 16.2192 ( 22.9%) Global Common Subexpression Elimination
11.1200 ( 15.8%) 0.0499 ( 13.8%) 11.1700 ( 15.8%) 11.1028 ( 15.7%) Reassociate expressions
6.5499 ( 9.3%) 0.0300 ( 8.3%) 6.5799 ( 9.3%) 6.5824 ( 9.3%) Bytecode Writer
3.2499 ( 4.6%) 0.0100 ( 2.7%) 3.2599 ( 4.6%) 3.2140 ( 4.5%) Scalar Replacement of Aggregates
3.0300 ( 4.3%) 0.0499 ( 13.8%) 3.0800 ( 4.3%) 3.0382 ( 4.2%) Combine redundant instructions
2.6599 ( 3.7%) 0.0100 ( 2.7%) 2.6699 ( 3.7%) 2.7339 ( 3.8%) Dead Store Elimination
2.1600 ( 3.0%) 0.0300 ( 8.3%) 2.1900 ( 3.0%) 2.1924 ( 3.1%) Function Integration/Inlining
2.1600 ( 3.0%) 0.0100 ( 2.7%) 2.1700 ( 3.0%) 2.1125 ( 2.9%) Sparse Conditional Constant Propagation
1.6600 ( 2.3%) 0.0000 ( 0.0%) 1.6600 ( 2.3%) 1.6389 ( 2.3%) Aggressive Dead Code Elimination
1.4999 ( 2.1%) 0.0100 ( 2.7%) 1.5099 ( 2.1%) 1.4462 ( 2.0%) Tail Duplication
1.5000 ( 2.1%) 0.0000 ( 0.0%) 1.5000 ( 2.1%) 1.4410 ( 2.0%) Post-Dominator Set Construction
1.3200 ( 1.8%) 0.0000 ( 0.0%) 1.3200 ( 1.8%) 1.3722 ( 1.9%) Canonicalize natural loops
1.2700 ( 1.8%) 0.0000 ( 0.0%) 1.2700 ( 1.7%) 1.2717 ( 1.7%) Merge Duplicate Global Constants
1.0300 ( 1.4%) 0.0000 ( 0.0%) 1.0300 ( 1.4%) 1.1418 ( 1.6%) Combine redundant instructions
0.9499 ( 1.3%) 0.0400 ( 11.1%) 0.9899 ( 1.4%) 0.9979 ( 1.4%) Raise Pointer References
0.9399 ( 1.3%) 0.0100 ( 2.7%) 0.9499 ( 1.3%) 0.9688 ( 1.3%) Simplify the CFG
0.9199 ( 1.3%) 0.0300 ( 8.3%) 0.9499 ( 1.3%) 0.8993 ( 1.2%) Promote Memory to Register
0.9600 ( 1.3%) 0.0000 ( 0.0%) 0.9600 ( 1.3%) 0.8742 ( 1.2%) Loop Invariant Code Motion
0.5600 ( 0.7%) 0.0000 ( 0.0%) 0.5600 ( 0.7%) 0.6022 ( 0.8%) Module Verifier

42

LLC Tool: Static code generator

Compiles LLVM native assembly language

llc file.bc -o file.s -march=x86
as file.s o file.o

Compiles LLVM portable C code

llc file.bc -o file.c -march=c
gcc c file.c o file.o

Targets are modular & dynamically loadable:

llc load libarm.so file.bc -march=arm

43

LLI Tool: LLVM Execution Engine

LLI allows direct execution of .bc files

E.g.: lli grep.bc -i foo *.c

LLI uses a Just-In-Time compiler if available:

Uses same code generator as LLC

Optionally uses faster components than LLC

Emits machine code to memory instead of .s file

JIT is a library that can be embedded in other tools

Otherwise, it uses the LLVM interpreter:

Interpreter is extremely simple and very slow
Interpreter is portable though!
44