Newer
Older
//===- GVN.cpp - Eliminate redundant values and loads ---------------------===//
Owen Anderson
committed
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
Owen Anderson
committed
//
//===----------------------------------------------------------------------===//
//
// This pass performs global value numbering to eliminate fully redundant
// instructions. It also performs simple dead load elimination.
//
Matthijs Kooijman
committed
// Note that this pass does the value numbering itself, it does not use the
// ValueNumbering analysis passes.
//
Owen Anderson
committed
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "gvn"
#include "llvm/Transforms/Scalar.h"
#include "llvm/BasicBlock.h"
#include "llvm/Constants.h"
Owen Anderson
committed
#include "llvm/DerivedTypes.h"
#include "llvm/Function.h"
#include "llvm/Instructions.h"
#include "llvm/Value.h"
Owen Anderson
committed
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/DepthFirstIterator.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/Dominators.h"
#include "llvm/Analysis/AliasAnalysis.h"
Owen Anderson
committed
#include "llvm/Analysis/MemoryDependenceAnalysis.h"
#include "llvm/Support/CFG.h"
#include "llvm/Support/CommandLine.h"
Owen Anderson
committed
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
Owen Anderson
committed
using namespace llvm;
STATISTIC(NumGVNInstr, "Number of instructions deleted");
STATISTIC(NumGVNLoad, "Number of loads deleted");
Owen Anderson
committed
STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
Owen Anderson
committed
STATISTIC(NumGVNBlocks, "Number of blocks merged");
STATISTIC(NumPRELoad, "Number of loads PRE'd");
static cl::opt<bool> EnablePRE("enable-pre",
cl::init(true), cl::Hidden);
cl::opt<bool> EnableLoadPRE("enable-load-pre"/*, cl::init(true)*/);
Owen Anderson
committed
//===----------------------------------------------------------------------===//
// ValueTable Class
//===----------------------------------------------------------------------===//
/// This class holds the mapping between values and value numbers. It is used
/// as an efficient mechanism to determine the expression-wise equivalence of
/// two values.
namespace {
struct VISIBILITY_HIDDEN Expression {
enum ExpressionOpcode { ADD, SUB, MUL, UDIV, SDIV, FDIV, UREM, SREM,
FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ,
ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE,
ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ,
FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE,
FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE,
FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT,
SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI,
FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT,
PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, CONSTANT,
Owen Anderson
committed
EMPTY, TOMBSTONE };
Owen Anderson
committed
ExpressionOpcode opcode;
const Type* type;
uint32_t firstVN;
uint32_t secondVN;
uint32_t thirdVN;
SmallVector<uint32_t, 4> varargs;
Value* function;
Owen Anderson
committed
Expression() { }
Expression(ExpressionOpcode o) : opcode(o) { }
bool operator==(const Expression &other) const {
if (opcode != other.opcode)
return false;
else if (opcode == EMPTY || opcode == TOMBSTONE)
return true;
else if (type != other.type)
return false;
else if (function != other.function)
return false;
Owen Anderson
committed
else if (firstVN != other.firstVN)
return false;
else if (secondVN != other.secondVN)
return false;
else if (thirdVN != other.thirdVN)
return false;
else {
if (varargs.size() != other.varargs.size())
return false;
for (size_t i = 0; i < varargs.size(); ++i)
if (varargs[i] != other.varargs[i])
return false;
return true;
}
}
bool operator!=(const Expression &other) const {
if (opcode != other.opcode)
return true;
else if (opcode == EMPTY || opcode == TOMBSTONE)
return false;
else if (type != other.type)
return true;
else if (function != other.function)
return true;
Owen Anderson
committed
else if (firstVN != other.firstVN)
return true;
else if (secondVN != other.secondVN)
return true;
else if (thirdVN != other.thirdVN)
return true;
else {
if (varargs.size() != other.varargs.size())
return true;
for (size_t i = 0; i < varargs.size(); ++i)
if (varargs[i] != other.varargs[i])
return true;
return false;
}
}
};
class VISIBILITY_HIDDEN ValueTable {
private:
DenseMap<Value*, uint32_t> valueNumbering;
DenseMap<Expression, uint32_t> expressionNumbering;
AliasAnalysis* AA;
MemoryDependenceAnalysis* MD;
DominatorTree* DT;
Owen Anderson
committed
uint32_t nextValueNumber;
Expression::ExpressionOpcode getOpcode(BinaryOperator* BO);
Expression::ExpressionOpcode getOpcode(CmpInst* C);
Expression::ExpressionOpcode getOpcode(CastInst* C);
Expression create_expression(BinaryOperator* BO);
Expression create_expression(CmpInst* C);
Expression create_expression(ShuffleVectorInst* V);
Expression create_expression(ExtractElementInst* C);
Expression create_expression(InsertElementInst* V);
Expression create_expression(SelectInst* V);
Expression create_expression(CastInst* C);
Expression create_expression(GetElementPtrInst* G);
Expression create_expression(CallInst* C);
Expression create_expression(Constant* C);
Owen Anderson
committed
public:
ValueTable() : nextValueNumber(1) { }
Owen Anderson
committed
uint32_t lookup_or_add(Value* V);
uint32_t lookup(Value* V) const;
void add(Value* V, uint32_t num);
void clear();
void erase(Value* v);
unsigned size();
void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
AliasAnalysis *getAliasAnalysis() const { return AA; }
void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
void setDomTree(DominatorTree* D) { DT = D; }
Owen Anderson
committed
uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
Owen Anderson
committed
};
}
namespace llvm {
template <> struct DenseMapInfo<Expression> {
static inline Expression getEmptyKey() {
return Expression(Expression::EMPTY);
}
static inline Expression getTombstoneKey() {
return Expression(Expression::TOMBSTONE);
}
Owen Anderson
committed
static unsigned getHashValue(const Expression e) {
unsigned hash = e.opcode;
hash = e.firstVN + hash * 37;
hash = e.secondVN + hash * 37;
hash = e.thirdVN + hash * 37;
hash = ((unsigned)((uintptr_t)e.type >> 4) ^
(unsigned)((uintptr_t)e.type >> 9)) +
hash * 37;
Owen Anderson
committed
for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
E = e.varargs.end(); I != E; ++I)
Owen Anderson
committed
hash = *I + hash * 37;
hash = ((unsigned)((uintptr_t)e.function >> 4) ^
(unsigned)((uintptr_t)e.function >> 9)) +
hash * 37;
Owen Anderson
committed
return hash;
}
static bool isEqual(const Expression &LHS, const Expression &RHS) {
return LHS == RHS;
}
Owen Anderson
committed
static bool isPod() { return true; }
};
}
//===----------------------------------------------------------------------===//
// ValueTable Internal Functions
//===----------------------------------------------------------------------===//
Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) {
Owen Anderson
committed
switch(BO->getOpcode()) {
default: // THIS SHOULD NEVER HAPPEN
assert(0 && "Binary operator with unknown opcode?");
case Instruction::Add: return Expression::ADD;
case Instruction::Sub: return Expression::SUB;
case Instruction::Mul: return Expression::MUL;
case Instruction::UDiv: return Expression::UDIV;
case Instruction::SDiv: return Expression::SDIV;
case Instruction::FDiv: return Expression::FDIV;
case Instruction::URem: return Expression::UREM;
case Instruction::SRem: return Expression::SREM;
case Instruction::FRem: return Expression::FREM;
case Instruction::Shl: return Expression::SHL;
case Instruction::LShr: return Expression::LSHR;
case Instruction::AShr: return Expression::ASHR;
case Instruction::And: return Expression::AND;
case Instruction::Or: return Expression::OR;
case Instruction::Xor: return Expression::XOR;
Owen Anderson
committed
}
}
Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) {
if (isa<ICmpInst>(C) || isa<VICmpInst>(C)) {
Owen Anderson
committed
switch (C->getPredicate()) {
default: // THIS SHOULD NEVER HAPPEN
assert(0 && "Comparison with unknown predicate?");
case ICmpInst::ICMP_EQ: return Expression::ICMPEQ;
case ICmpInst::ICMP_NE: return Expression::ICMPNE;
case ICmpInst::ICMP_UGT: return Expression::ICMPUGT;
case ICmpInst::ICMP_UGE: return Expression::ICMPUGE;
case ICmpInst::ICMP_ULT: return Expression::ICMPULT;
case ICmpInst::ICMP_ULE: return Expression::ICMPULE;
case ICmpInst::ICMP_SGT: return Expression::ICMPSGT;
case ICmpInst::ICMP_SGE: return Expression::ICMPSGE;
case ICmpInst::ICMP_SLT: return Expression::ICMPSLT;
case ICmpInst::ICMP_SLE: return Expression::ICMPSLE;
Owen Anderson
committed
}
}
assert((isa<FCmpInst>(C) || isa<VFCmpInst>(C)) && "Unknown compare");
switch (C->getPredicate()) {
default: // THIS SHOULD NEVER HAPPEN
assert(0 && "Comparison with unknown predicate?");
case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ;
case FCmpInst::FCMP_OGT: return Expression::FCMPOGT;
case FCmpInst::FCMP_OGE: return Expression::FCMPOGE;
case FCmpInst::FCMP_OLT: return Expression::FCMPOLT;
case FCmpInst::FCMP_OLE: return Expression::FCMPOLE;
case FCmpInst::FCMP_ONE: return Expression::FCMPONE;
case FCmpInst::FCMP_ORD: return Expression::FCMPORD;
case FCmpInst::FCMP_UNO: return Expression::FCMPUNO;
case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ;
case FCmpInst::FCMP_UGT: return Expression::FCMPUGT;
case FCmpInst::FCMP_UGE: return Expression::FCMPUGE;
case FCmpInst::FCMP_ULT: return Expression::FCMPULT;
case FCmpInst::FCMP_ULE: return Expression::FCMPULE;
case FCmpInst::FCMP_UNE: return Expression::FCMPUNE;
}
Owen Anderson
committed
}
Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) {
Owen Anderson
committed
switch(C->getOpcode()) {
default: // THIS SHOULD NEVER HAPPEN
assert(0 && "Cast operator with unknown opcode?");
case Instruction::Trunc: return Expression::TRUNC;
case Instruction::ZExt: return Expression::ZEXT;
case Instruction::SExt: return Expression::SEXT;
case Instruction::FPToUI: return Expression::FPTOUI;
case Instruction::FPToSI: return Expression::FPTOSI;
case Instruction::UIToFP: return Expression::UITOFP;
case Instruction::SIToFP: return Expression::SITOFP;
case Instruction::FPTrunc: return Expression::FPTRUNC;
case Instruction::FPExt: return Expression::FPEXT;
case Instruction::PtrToInt: return Expression::PTRTOINT;
case Instruction::IntToPtr: return Expression::INTTOPTR;
case Instruction::BitCast: return Expression::BITCAST;
Owen Anderson
committed
}
}
Expression ValueTable::create_expression(CallInst* C) {
Expression e;
e.type = C->getType();
e.firstVN = 0;
e.secondVN = 0;
e.thirdVN = 0;
e.function = C->getCalledFunction();
e.opcode = Expression::CALL;
for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end();
I != E; ++I)
Owen Anderson
committed
e.varargs.push_back(lookup_or_add(*I));
return e;
}
Owen Anderson
committed
Expression ValueTable::create_expression(BinaryOperator* BO) {
Expression e;
Owen Anderson
committed
e.firstVN = lookup_or_add(BO->getOperand(0));
e.secondVN = lookup_or_add(BO->getOperand(1));
Owen Anderson
committed
e.thirdVN = 0;
e.function = 0;
Owen Anderson
committed
e.type = BO->getType();
e.opcode = getOpcode(BO);
return e;
}
Expression ValueTable::create_expression(CmpInst* C) {
Expression e;
Owen Anderson
committed
e.firstVN = lookup_or_add(C->getOperand(0));
e.secondVN = lookup_or_add(C->getOperand(1));
Owen Anderson
committed
e.thirdVN = 0;
e.function = 0;
Owen Anderson
committed
e.type = C->getType();
e.opcode = getOpcode(C);
return e;
}
Expression ValueTable::create_expression(CastInst* C) {
Expression e;
Owen Anderson
committed
e.firstVN = lookup_or_add(C->getOperand(0));
Owen Anderson
committed
e.secondVN = 0;
e.thirdVN = 0;
e.function = 0;
Owen Anderson
committed
e.type = C->getType();
e.opcode = getOpcode(C);
return e;
}
Expression ValueTable::create_expression(ShuffleVectorInst* S) {
Expression e;
Owen Anderson
committed
e.firstVN = lookup_or_add(S->getOperand(0));
e.secondVN = lookup_or_add(S->getOperand(1));
e.thirdVN = lookup_or_add(S->getOperand(2));
e.function = 0;
Owen Anderson
committed
e.type = S->getType();
e.opcode = Expression::SHUFFLE;
return e;
}
Expression ValueTable::create_expression(ExtractElementInst* E) {
Expression e;
Owen Anderson
committed
e.firstVN = lookup_or_add(E->getOperand(0));
e.secondVN = lookup_or_add(E->getOperand(1));
Owen Anderson
committed
e.thirdVN = 0;
e.function = 0;
Owen Anderson
committed
e.type = E->getType();
e.opcode = Expression::EXTRACT;
return e;
}
Expression ValueTable::create_expression(InsertElementInst* I) {
Expression e;
Owen Anderson
committed
e.firstVN = lookup_or_add(I->getOperand(0));
e.secondVN = lookup_or_add(I->getOperand(1));
e.thirdVN = lookup_or_add(I->getOperand(2));
e.function = 0;
Owen Anderson
committed
e.type = I->getType();
e.opcode = Expression::INSERT;
return e;
}
Expression ValueTable::create_expression(SelectInst* I) {
Expression e;
Owen Anderson
committed
e.firstVN = lookup_or_add(I->getCondition());
e.secondVN = lookup_or_add(I->getTrueValue());
e.thirdVN = lookup_or_add(I->getFalseValue());
e.function = 0;
Owen Anderson
committed
e.type = I->getType();
e.opcode = Expression::SELECT;
return e;
}
Expression ValueTable::create_expression(GetElementPtrInst* G) {
Expression e;
Owen Anderson
committed
e.firstVN = lookup_or_add(G->getPointerOperand());
Owen Anderson
committed
e.secondVN = 0;
e.thirdVN = 0;
e.function = 0;
Owen Anderson
committed
e.type = G->getType();
e.opcode = Expression::GEP;
for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end();
I != E; ++I)
Owen Anderson
committed
e.varargs.push_back(lookup_or_add(*I));
Owen Anderson
committed
return e;
}
//===----------------------------------------------------------------------===//
// ValueTable External Functions
//===----------------------------------------------------------------------===//
Owen Anderson
committed
/// add - Insert a value into the table with a specified value number.
void ValueTable::add(Value* V, uint32_t num) {
valueNumbering.insert(std::make_pair(V, num));
}
Owen Anderson
committed
/// lookup_or_add - Returns the value number for the specified value, assigning
/// it a new number if it did not have one before.
uint32_t ValueTable::lookup_or_add(Value* V) {
DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
if (VI != valueNumbering.end())
return VI->second;
if (CallInst* C = dyn_cast<CallInst>(V)) {
Owen Anderson
committed
if (AA->doesNotAccessMemory(C)) {
Expression e = create_expression(C);
DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
if (EI != expressionNumbering.end()) {
valueNumbering.insert(std::make_pair(V, EI->second));
return EI->second;
} else {
expressionNumbering.insert(std::make_pair(e, nextValueNumber));
valueNumbering.insert(std::make_pair(V, nextValueNumber));
return nextValueNumber++;
}
} else if (AA->onlyReadsMemory(C)) {
Expression e = create_expression(C);
if (expressionNumbering.find(e) == expressionNumbering.end()) {
expressionNumbering.insert(std::make_pair(e, nextValueNumber));
valueNumbering.insert(std::make_pair(V, nextValueNumber));
return nextValueNumber++;
}
MemDepResult local_dep = MD->getDependency(C);
if (!local_dep.isDef() && !local_dep.isNonLocal()) {
valueNumbering.insert(std::make_pair(V, nextValueNumber));
return nextValueNumber++;
if (local_dep.isDef()) {
CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
if (local_cdep->getNumOperands() != C->getNumOperands()) {
valueNumbering.insert(std::make_pair(V, nextValueNumber));
return nextValueNumber++;
for (unsigned i = 1; i < C->getNumOperands(); ++i) {
uint32_t c_vn = lookup_or_add(C->getOperand(i));
uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i));
if (c_vn != cd_vn) {
valueNumbering.insert(std::make_pair(V, nextValueNumber));
return nextValueNumber++;
}
}
uint32_t v = lookup_or_add(local_cdep);
valueNumbering.insert(std::make_pair(V, v));
return v;
}
const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
MD->getNonLocalCallDependency(CallSite(C));
// FIXME: call/call dependencies for readonly calls should return def, not
// clobber! Move the checking logic to MemDep!
CallInst* cdep = 0;
// Check to see if we have a single dominating call instruction that is
// identical to C.
for (unsigned i = 0, e = deps.size(); i != e; ++i) {
const MemoryDependenceAnalysis::NonLocalDepEntry *I = &deps[i];
// Ignore non-local dependencies.
if (I->second.isNonLocal())
continue;
// We don't handle non-depedencies. If we already have a call, reject
// instruction dependencies.
if (I->second.isClobber() || cdep != 0) {
cdep = 0;
break;
}
CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->second.getInst());
// FIXME: All duplicated with non-local case.
if (NonLocalDepCall && DT->properlyDominates(I->first, C->getParent())){
cdep = NonLocalDepCall;
continue;
cdep = 0;
break;
if (!cdep) {
valueNumbering.insert(std::make_pair(V, nextValueNumber));
return nextValueNumber++;
}
if (cdep->getNumOperands() != C->getNumOperands()) {
valueNumbering.insert(std::make_pair(V, nextValueNumber));
return nextValueNumber++;
}
for (unsigned i = 1; i < C->getNumOperands(); ++i) {
uint32_t c_vn = lookup_or_add(C->getOperand(i));
uint32_t cd_vn = lookup_or_add(cdep->getOperand(i));
if (c_vn != cd_vn) {
valueNumbering.insert(std::make_pair(V, nextValueNumber));
return nextValueNumber++;
}
uint32_t v = lookup_or_add(cdep);
valueNumbering.insert(std::make_pair(V, v));
return v;
} else {
valueNumbering.insert(std::make_pair(V, nextValueNumber));
return nextValueNumber++;
}
} else if (BinaryOperator* BO = dyn_cast<BinaryOperator>(V)) {
Owen Anderson
committed
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
Expression e = create_expression(BO);
DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
if (EI != expressionNumbering.end()) {
valueNumbering.insert(std::make_pair(V, EI->second));
return EI->second;
} else {
expressionNumbering.insert(std::make_pair(e, nextValueNumber));
valueNumbering.insert(std::make_pair(V, nextValueNumber));
return nextValueNumber++;
}
} else if (CmpInst* C = dyn_cast<CmpInst>(V)) {
Expression e = create_expression(C);
DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
if (EI != expressionNumbering.end()) {
valueNumbering.insert(std::make_pair(V, EI->second));
return EI->second;
} else {
expressionNumbering.insert(std::make_pair(e, nextValueNumber));
valueNumbering.insert(std::make_pair(V, nextValueNumber));
return nextValueNumber++;
}
} else if (ShuffleVectorInst* U = dyn_cast<ShuffleVectorInst>(V)) {
Expression e = create_expression(U);
DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
if (EI != expressionNumbering.end()) {
valueNumbering.insert(std::make_pair(V, EI->second));
return EI->second;
} else {
expressionNumbering.insert(std::make_pair(e, nextValueNumber));
valueNumbering.insert(std::make_pair(V, nextValueNumber));
return nextValueNumber++;
}
} else if (ExtractElementInst* U = dyn_cast<ExtractElementInst>(V)) {
Expression e = create_expression(U);
DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
if (EI != expressionNumbering.end()) {
valueNumbering.insert(std::make_pair(V, EI->second));
return EI->second;
} else {
expressionNumbering.insert(std::make_pair(e, nextValueNumber));
valueNumbering.insert(std::make_pair(V, nextValueNumber));
return nextValueNumber++;
}
} else if (InsertElementInst* U = dyn_cast<InsertElementInst>(V)) {
Expression e = create_expression(U);
DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
if (EI != expressionNumbering.end()) {
valueNumbering.insert(std::make_pair(V, EI->second));
return EI->second;
} else {
expressionNumbering.insert(std::make_pair(e, nextValueNumber));
valueNumbering.insert(std::make_pair(V, nextValueNumber));
return nextValueNumber++;
}
} else if (SelectInst* U = dyn_cast<SelectInst>(V)) {
Expression e = create_expression(U);
DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
if (EI != expressionNumbering.end()) {
valueNumbering.insert(std::make_pair(V, EI->second));
return EI->second;
} else {
expressionNumbering.insert(std::make_pair(e, nextValueNumber));
valueNumbering.insert(std::make_pair(V, nextValueNumber));
return nextValueNumber++;
}
} else if (CastInst* U = dyn_cast<CastInst>(V)) {
Expression e = create_expression(U);
DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
if (EI != expressionNumbering.end()) {
valueNumbering.insert(std::make_pair(V, EI->second));
return EI->second;
} else {
expressionNumbering.insert(std::make_pair(e, nextValueNumber));
valueNumbering.insert(std::make_pair(V, nextValueNumber));
return nextValueNumber++;
}
} else if (GetElementPtrInst* U = dyn_cast<GetElementPtrInst>(V)) {
Expression e = create_expression(U);
DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
if (EI != expressionNumbering.end()) {
valueNumbering.insert(std::make_pair(V, EI->second));
return EI->second;
} else {
expressionNumbering.insert(std::make_pair(e, nextValueNumber));
valueNumbering.insert(std::make_pair(V, nextValueNumber));
return nextValueNumber++;
}
} else {
valueNumbering.insert(std::make_pair(V, nextValueNumber));
return nextValueNumber++;
}
}
/// lookup - Returns the value number of the specified value. Fails if
/// the value has not yet been numbered.
uint32_t ValueTable::lookup(Value* V) const {
DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
assert(VI != valueNumbering.end() && "Value not numbered?");
return VI->second;
Owen Anderson
committed
}
/// clear - Remove all entries from the ValueTable
void ValueTable::clear() {
valueNumbering.clear();
expressionNumbering.clear();
nextValueNumber = 1;
}
/// erase - Remove a value from the value numbering
void ValueTable::erase(Value* V) {
valueNumbering.erase(V);
}
Owen Anderson
committed
//===----------------------------------------------------------------------===//
Owen Anderson
committed
//===----------------------------------------------------------------------===//
Owen Anderson
committed
namespace {
struct VISIBILITY_HIDDEN ValueNumberScope {
ValueNumberScope* parent;
DenseMap<uint32_t, Value*> table;
ValueNumberScope(ValueNumberScope* p) : parent(p) { }
};
}
Owen Anderson
committed
namespace {
class VISIBILITY_HIDDEN GVN : public FunctionPass {
bool runOnFunction(Function &F);
public:
static char ID; // Pass identification, replacement for typeid
GVN() : FunctionPass(&ID) { }
Owen Anderson
committed
private:
MemoryDependenceAnalysis *MD;
DominatorTree *DT;
Owen Anderson
committed
ValueTable VN;
Owen Anderson
committed
DenseMap<BasicBlock*, ValueNumberScope*> localAvail;
Owen Anderson
committed
typedef DenseMap<Value*, SmallPtrSet<Instruction*, 4> > PhiMapType;
PhiMapType phiMap;
Owen Anderson
committed
// This transformation requires dominator postdominator info
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
AU.addRequired<DominatorTree>();
AU.addRequired<MemoryDependenceAnalysis>();
AU.addRequired<AliasAnalysis>();
Owen Anderson
committed
AU.addPreserved<DominatorTree>();
AU.addPreserved<AliasAnalysis>();
Owen Anderson
committed
}
// Helper fuctions
// FIXME: eliminate or document these better
bool processLoad(LoadInst* L,
Chris Lattner
committed
SmallVectorImpl<Instruction*> &toErase);
Owen Anderson
committed
bool processInstruction(Instruction* I,
Chris Lattner
committed
SmallVectorImpl<Instruction*> &toErase);
Chris Lattner
committed
SmallVectorImpl<Instruction*> &toErase);
Value *GetValueForBlock(BasicBlock *BB, LoadInst* orig,
DenseMap<BasicBlock*, Value*> &Phis,
bool top_level = false);
Owen Anderson
committed
void dump(DenseMap<uint32_t, Value*>& d);
Value* CollapsePhi(PHINode* p);
bool isSafeReplacement(PHINode* p, Instruction* inst);
Owen Anderson
committed
bool performPRE(Function& F);
Owen Anderson
committed
Value* lookupNumber(BasicBlock* BB, uint32_t num);
Owen Anderson
committed
bool mergeBlockIntoPredecessor(BasicBlock* BB);
void cleanupGlobalSets();
Owen Anderson
committed
};
char GVN::ID = 0;
}
// createGVNPass - The public interface to this file...
FunctionPass *llvm::createGVNPass() { return new GVN(); }
static RegisterPass<GVN> X("gvn",
"Global Value Numbering");
Owen Anderson
committed
void GVN::dump(DenseMap<uint32_t, Value*>& d) {
Owen Anderson
committed
for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
E = d.end(); I != E; ++I) {
Owen Anderson
committed
printf("%d\n", I->first);
I->second->dump();
}
printf("}\n");
}
Value* GVN::CollapsePhi(PHINode* p) {
Value* constVal = p->hasConstantValue();
Instruction* inst = dyn_cast<Instruction>(constVal);
if (!inst)
return constVal;
if (DT->dominates(inst, p))
if (isSafeReplacement(p, inst))
return inst;
bool GVN::isSafeReplacement(PHINode* p, Instruction* inst) {
if (!isa<PHINode>(inst))
return true;
for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end();
UI != E; ++UI)
if (PHINode* use_phi = dyn_cast<PHINode>(UI))
if (use_phi->getParent() == inst->getParent())
return false;
return true;
}
/// GetValueForBlock - Get the value to use within the specified basic block.
/// available values are in Phis.
Value *GVN::GetValueForBlock(BasicBlock *BB, LoadInst* orig,
DenseMap<BasicBlock*, Value*> &Phis,
bool top_level) {
// If we have already computed this value, return the previously computed val.
Owen Anderson
committed
DenseMap<BasicBlock*, Value*>::iterator V = Phis.find(BB);
if (V != Phis.end() && !top_level) return V->second;
// If the block is unreachable, just return undef, since this path
// can't actually occur at runtime.
if (!DT->isReachableFromEntry(BB))
return Phis[BB] = UndefValue::get(orig->getType());
Owen Anderson
committed
if (BasicBlock *Pred = BB->getSinglePredecessor()) {
Value *ret = GetValueForBlock(Pred, orig, Phis);
Owen Anderson
committed
Phis[BB] = ret;
return ret;
// Get the number of predecessors of this block so we can reserve space later.
// If there is already a PHI in it, use the #preds from it, otherwise count.
// Getting it from the PHI is constant time.
unsigned NumPreds;
if (PHINode *ExistingPN = dyn_cast<PHINode>(BB->begin()))
NumPreds = ExistingPN->getNumIncomingValues();
else
NumPreds = std::distance(pred_begin(BB), pred_end(BB));
// Otherwise, the idom is the loop, so we need to insert a PHI node. Do so
// now, then get values to fill in the incoming values for the PHI.
PHINode *PN = PHINode::Create(orig->getType(), orig->getName()+".rle",
BB->begin());
PN->reserveOperandSpace(NumPreds);
Owen Anderson
committed
Phis.insert(std::make_pair(BB, PN));
Owen Anderson
committed
// Fill in the incoming values for the block.
for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
Value* val = GetValueForBlock(*PI, orig, Phis);
PN->addIncoming(val, *PI);
}
VN.getAliasAnalysis()->copyValue(orig, PN);
// Attempt to collapse PHI nodes that are trivially redundant
Value* v = CollapsePhi(PN);
if (!v) {
// Cache our phi construction results
phiMap[orig->getPointerOperand()].insert(PN);
return PN;
}
if (isa<PointerType>(v->getType()))
MD->invalidateCachedPointerInfo(v);
Owen Anderson
committed
for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(),
E = Phis.end(); I != E; ++I)
if (I->second == PN)
I->second = v;
Owen Anderson
committed
DEBUG(cerr << "GVN removed: " << *PN);
MD->removeInstruction(PN);
/// IsValueFullyAvailableInBlock - Return true if we can prove that the value
/// we're analyzing is fully available in the specified block. As we go, keep
/// track of which blocks we know are fully alive in FullyAvailableBlocks. This
/// map is actually a tri-state map with the following values:
/// 0) we know the block *is not* fully available.
/// 1) we know the block *is* fully available.
/// 2) we do not know whether the block is fully available or not, but we are
/// currently speculating that it will be.
/// 3) we are speculating for this block and have used that to speculate for
/// other blocks.
static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
DenseMap<BasicBlock*, char> &FullyAvailableBlocks) {
// Optimistically assume that the block is fully available and check to see
// if we already know about this block in one lookup.
std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
FullyAvailableBlocks.insert(std::make_pair(BB, 2));
// If the entry already existed for this block, return the precomputed value.
if (!IV.second) {
// If this is a speculative "available" value, mark it as being used for
// speculation of other blocks.
if (IV.first->second == 2)
IV.first->second = 3;
return IV.first->second != 0;
}
// Otherwise, see if it is fully available in all predecessors.
pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
// If this block has no predecessors, it isn't live-in here.
if (PI == PE)
for (; PI != PE; ++PI)
// If the value isn't fully available in one of our predecessors, then it
// isn't fully available in this block either. Undo our previous
// optimistic assumption and bail out.
if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
return true;
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
// SpeculationFailure - If we get here, we found out that this is not, after
// all, a fully-available block. We have a problem if we speculated on this and
// used the speculation to mark other blocks as available.
SpeculationFailure:
char &BBVal = FullyAvailableBlocks[BB];
// If we didn't speculate on this, just return with it set to false.
if (BBVal == 2) {
BBVal = 0;
return false;
}
// If we did speculate on this value, we could have blocks set to 1 that are
// incorrect. Walk the (transitive) successors of this block and mark them as
// 0 if set to one.
SmallVector<BasicBlock*, 32> BBWorklist;
BBWorklist.push_back(BB);
while (!BBWorklist.empty()) {
BasicBlock *Entry = BBWorklist.pop_back_val();
// Note that this sets blocks to 0 (unavailable) if they happen to not
// already be in FullyAvailableBlocks. This is safe.
char &EntryVal = FullyAvailableBlocks[Entry];
if (EntryVal == 0) continue; // Already unavailable.
// Mark as unavailable.
EntryVal = 0;
for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
BBWorklist.push_back(*I);
}
return false;
/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
/// non-local by performing PHI construction.
bool GVN::processNonLocalLoad(LoadInst *LI,
Chris Lattner
committed
SmallVectorImpl<Instruction*> &toErase) {
// Find the non-local dependencies of the load.
SmallVector<MemoryDependenceAnalysis::NonLocalDepEntry, 64> Deps;
MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(),
Deps);
//DEBUG(cerr << "INVESTIGATING NONLOCAL LOAD: " << Deps.size() << *LI);
// If we had to process more than one hundred blocks to find the
// dependencies, this load isn't worth worrying about. Optimizing
// it will be too expensive.
return false;
// Filter out useless results (non-locals, etc). Keep track of the blocks
// where we have a value available in repl, also keep track of whether we see
// dependencies that produce an unknown value for the load (such as a call
// that could potentially clobber the load).
SmallVector<std::pair<BasicBlock*, Value*>, 16> ValuesPerBlock;
SmallVector<BasicBlock*, 16> UnavailableBlocks;
for (unsigned i = 0, e = Deps.size(); i != e; ++i) {
BasicBlock *DepBB = Deps[i].first;
MemDepResult DepInfo = Deps[i].second;
if (DepInfo.isClobber()) {
UnavailableBlocks.push_back(DepBB);
continue;
}
Instruction *DepInst = DepInfo.getInst();
// Loading the allocation -> undef.
if (isa<AllocationInst>(DepInst)) {
ValuesPerBlock.push_back(std::make_pair(DepBB,
UndefValue::get(LI->getType())));
continue;
}
if (StoreInst* S = dyn_cast<StoreInst>(DepInst)) {
// Reject loads and stores that are to the same address but are of
// different types.
// NOTE: 403.gcc does have this case (e.g. in readonly_fields_p) because
// of bitfield access, it would be interesting to optimize for it at some
// point.
if (S->getOperand(0)->getType() != LI->getType()) {
UnavailableBlocks.push_back(DepBB);
continue;
}
ValuesPerBlock.push_back(std::make_pair(DepBB, S->getOperand(0)));
} else if (LoadInst* LD = dyn_cast<LoadInst>(DepInst)) {
if (LD->getType() != LI->getType()) {
UnavailableBlocks.push_back(DepBB);
continue;
}
ValuesPerBlock.push_back(std::make_pair(DepBB, LD));
UnavailableBlocks.push_back(DepBB);
continue;
// If we have no predecessors that produce a known value for this load, exit
// early.
if (ValuesPerBlock.empty()) return false;
// If all of the instructions we depend on produce a known value for this
// load, then it is fully redundant and we can use PHI insertion to compute
// its value. Insert PHIs and remove the fully redundant value now.
if (UnavailableBlocks.empty()) {
// Use cached PHI construction information from previous runs
SmallPtrSet<Instruction*, 4> &p = phiMap[LI->getPointerOperand()];
// FIXME: What does phiMap do? Are we positive it isn't getting invalidated?
for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end();
I != E; ++I) {
if ((*I)->getParent() == LI->getParent()) {
DEBUG(cerr << "GVN REMOVING NONLOCAL LOAD #1: " << *LI);
LI->replaceAllUsesWith(*I);
if (isa<PointerType>((*I)->getType()))
MD->invalidateCachedPointerInfo(*I);
toErase.push_back(LI);
NumGVNLoad++;
return true;
}
ValuesPerBlock.push_back(std::make_pair((*I)->getParent(), *I));
DEBUG(cerr << "GVN REMOVING NONLOCAL LOAD: " << *LI);
DenseMap<BasicBlock*, Value*> BlockReplValues;
BlockReplValues.insert(ValuesPerBlock.begin(), ValuesPerBlock.end());
// Perform PHI construction.
Value* v = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true);
LI->replaceAllUsesWith(v);
if (isa<PointerType>(v->getType()))
MD->invalidateCachedPointerInfo(v);
toErase.push_back(LI);
NumGVNLoad++;
return true;
if (!EnablePRE || !EnableLoadPRE)
return false;
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
// Okay, we have *some* definitions of the value. This means that the value
// is available in some of our (transitive) predecessors. Lets think about
// doing PRE of this load. This will involve inserting a new load into the
// predecessor when it's not available. We could do this in general, but
// prefer to not increase code size. As such, we only do this when we know
// that we only have to insert *one* load (which means we're basically moving
// the load, not inserting a new one).
// Everything we do here is based on local predecessors of LI's block. If it
// only has one predecessor, bail now.
BasicBlock *LoadBB = LI->getParent();
if (LoadBB->getSinglePredecessor())
return false;
// If we have a repl set with LI itself in it, this means we have a loop where
// at least one of the values is LI. Since this means that we won't be able
// to eliminate LI even if we insert uses in the other predecessors, we will
// end up increasing code size. Reject this by scanning for LI.
for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
if (ValuesPerBlock[i].second == LI)
return false;
// Okay, we have some hope :). Check to see if the loaded value is fully
// available in all but one predecessor.
// FIXME: If we could restructure the CFG, we could make a common pred with
// all the preds that don't have an available LI and insert a new load into
// that one block.
BasicBlock *UnavailablePred = 0;
DenseMap<BasicBlock*, char> FullyAvailableBlocks;
for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
FullyAvailableBlocks[ValuesPerBlock[i].first] = true;
for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
PI != E; ++PI) {
if (IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
continue;
// If this load is not available in multiple predecessors, reject it.
if (UnavailablePred && UnavailablePred != *PI)
return false;
UnavailablePred = *PI;
}
assert(UnavailablePred != 0 &&
"Fully available value should be eliminated above!");
// If the loaded pointer is PHI node defined in this block, do PHI translation
// to get its value in the predecessor.
Value *LoadPtr = LI->getOperand(0)->DoPHITranslation(LoadBB, UnavailablePred);
// Make sure the value is live in the predecessor. If it was defined by a
// non-PHI instruction in this block, we don't know how to recompute it above.
if (Instruction *LPInst = dyn_cast<Instruction>(LoadPtr))
if (!DT->dominates(LPInst->getParent(), UnavailablePred)) {
DEBUG(cerr << "COULDN'T PRE LOAD BECAUSE PTR IS UNAVAILABLE IN PRED: "
<< *LPInst << *LI << "\n");
return false;
}
// We don't currently handle critical edges :(
if (UnavailablePred->getTerminator()->getNumSuccessors() != 1) {
DEBUG(cerr << "COULD NOT PRE LOAD BECAUSE OF CRITICAL EDGE '"
<< UnavailablePred->getName() << "': " << *LI);
return false;
}
// Okay, we can eliminate this load by inserting a reload in the predecessor
// and using PHI construction to get the value in the other predecessors, do
// it.
DEBUG(cerr << "GVN REMOVING PRE LOAD: " << *LI);
Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
LI->getAlignment(),
UnavailablePred->getTerminator());
DenseMap<BasicBlock*, Value*> BlockReplValues;
BlockReplValues.insert(ValuesPerBlock.begin(), ValuesPerBlock.end());
BlockReplValues[UnavailablePred] = NewLoad;
// Perform PHI construction.
Value* v = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true);
LI->replaceAllUsesWith(v);
if (isa<PointerType>(v->getType()))
MD->invalidateCachedPointerInfo(v);
toErase.push_back(LI);
NumPRELoad++;
return true;
}
/// processLoad - Attempt to eliminate a load, first by eliminating it
/// locally, and then attempting non-local elimination if that fails.
bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) {
if (L->isVolatile())
Owen Anderson
committed
return false;
Value* pointer = L->getPointerOperand();
Owen Anderson
committed
// ... to a pointer that has been loaded from before...
MemDepResult dep = MD->getDependency(L);
// If the value isn't available, don't do anything!
if (dep.isClobber())
return false;
// If it is defined in another block, try harder.
return processNonLocalLoad(L, toErase);
Instruction *DepInst = dep.getInst();
if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
// Only forward substitute stores to loads of the same type.
// FIXME: Could do better!
if (DepSI->getPointerOperand()->getType() != pointer->getType())
return false;
Owen Anderson
committed
// Remove it!
L->replaceAllUsesWith(DepSI->getOperand(0));
if (isa<PointerType>(DepSI->getOperand(0)->getType()))
MD->invalidateCachedPointerInfo(DepSI->getOperand(0));
toErase.push_back(L);
NumGVNLoad++;
return true;
}
if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
// Only forward substitute stores to loads of the same type.
// FIXME: Could do better! load i32 -> load i8 -> truncate on little endian.
if (DepLI->getType() != L->getType())
return false;
Owen Anderson
committed
// Remove it!
L->replaceAllUsesWith(DepLI);
if (isa<PointerType>(DepLI->getType()))
MD->invalidateCachedPointerInfo(DepLI);
toErase.push_back(L);
NumGVNLoad++;
return true;
Owen Anderson
committed
}
// If this load really doesn't depend on anything, then we must be loading an
// undef value. This can happen when loading for a fresh allocation with no
// intervening stores, for example.
if (isa<AllocationInst>(DepInst)) {
L->replaceAllUsesWith(UndefValue::get(L->getType()));
toErase.push_back(L);
NumGVNLoad++;
}
Owen Anderson
committed
}
Owen Anderson
committed
Value* GVN::lookupNumber(BasicBlock* BB, uint32_t num) {
Owen Anderson
committed
DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB);
if (I == localAvail.end())
return 0;
ValueNumberScope* locals = I->second;
Owen Anderson
committed
while (locals) {
DenseMap<uint32_t, Value*>::iterator I = locals->table.find(num);
if (I != locals->table.end())
return I->second;
else
locals = locals->parent;
}
return 0;
}
Owen Anderson
committed
/// processInstruction - When calculating availability, handle an instruction
Owen Anderson
committed
/// by inserting it into the appropriate sets
bool GVN::processInstruction(Instruction *I,
Chris Lattner
committed
SmallVectorImpl<Instruction*> &toErase) {
Owen Anderson
committed
if (LoadInst* L = dyn_cast<LoadInst>(I)) {
bool changed = processLoad(L, toErase);
Owen Anderson
committed
if (!changed) {
unsigned num = VN.lookup_or_add(L);
Owen Anderson
committed
localAvail[I->getParent()]->table.insert(std::make_pair(num, L));
Owen Anderson
committed
}
return changed;
}
Owen Anderson
committed
uint32_t nextNum = VN.getNextUnusedValueNumber();
Owen Anderson
committed
unsigned num = VN.lookup_or_add(I);
Chris Lattner
committed
Owen Anderson
committed
// Allocations are always uniquely numbered, so we can save time and memory
// by fast failing them.
Owen Anderson
committed
if (isa<AllocationInst>(I) || isa<TerminatorInst>(I)) {
Owen Anderson
committed
localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
Owen Anderson
committed
return false;
Owen Anderson
committed
}
Owen Anderson
committed
Owen Anderson
committed
if (PHINode* p = dyn_cast<PHINode>(I)) {
Value* constVal = CollapsePhi(p);
Owen Anderson
committed
if (constVal) {
for (PhiMapType::iterator PI = phiMap.begin(), PE = phiMap.end();
PI != PE; ++PI)
Owen Anderson
committed
p->replaceAllUsesWith(constVal);
if (isa<PointerType>(constVal->getType()))
MD->invalidateCachedPointerInfo(constVal);
Owen Anderson
committed
} else {
Owen Anderson
committed
localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
Owen Anderson
committed
}
Owen Anderson
committed
// If the number we were assigned was a brand new VN, then we don't
// need to do a lookup to see if the number already exists
// somewhere in the domtree: it can't!
} else if (num == nextNum) {
localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
// Perform value-number based elimination
Owen Anderson
committed
} else if (Value* repl = lookupNumber(I->getParent(), num)) {
Owen Anderson
committed
// Remove it!
Owen Anderson
committed
I->replaceAllUsesWith(repl);
if (isa<PointerType>(repl->getType()))
MD->invalidateCachedPointerInfo(repl);
Owen Anderson
committed
toErase.push_back(I);
return true;
Owen Anderson
committed
} else {
Owen Anderson
committed
localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
Owen Anderson
committed
}
return false;
}
// GVN::runOnFunction - This is the main transformation entry point for a
// function.
//
MD = &getAnalysis<MemoryDependenceAnalysis>();
DT = &getAnalysis<DominatorTree>();
VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
VN.setMemDep(MD);
VN.setDomTree(DT);
bool changed = false;
bool shouldContinue = true;
Owen Anderson
committed
// Merge unconditional branches, allowing PRE to catch more
// optimization opportunities.
for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
BasicBlock* BB = FI;
++FI;
bool removedBlock = MergeBlockIntoPredecessor(BB, this);
if (removedBlock) NumGVNBlocks++;
changed |= removedBlock;
Owen Anderson
committed
}
DEBUG(cerr << "GVN iteration: " << Iteration << "\n");
shouldContinue = iterateOnFunction(F);
changed |= shouldContinue;
Owen Anderson
committed
if (EnablePRE) {
bool PREChanged = true;
while (PREChanged) {
PREChanged = performPRE(F);
Owen Anderson
committed
changed |= PREChanged;
Owen Anderson
committed
}
// FIXME: Should perform GVN again after PRE does something. PRE can move
// computations into blocks where they become fully redundant. Note that
// we can't do this until PRE's critical edge splitting updates memdep.
// Actually, when this happens, we should just fully integrate PRE into GVN.
cleanupGlobalSets();
bool GVN::processBlock(DomTreeNode* DTN) {
BasicBlock* BB = DTN->getBlock();
// FIXME: Kill off toErase by doing erasing eagerly in a helper function (and
// incrementing BI before processing an instruction).
SmallVector<Instruction*, 8> toErase;
bool changed_function = false;
Owen Anderson
committed
if (DTN->getIDom())
Owen Anderson
committed
localAvail[BB] =
new ValueNumberScope(localAvail[DTN->getIDom()->getBlock()]);
else
localAvail[BB] = new ValueNumberScope(0);
Owen Anderson
committed
for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
BI != BE;) {
changed_function |= processInstruction(BI, toErase);
if (toErase.empty()) {
++BI;
continue;
}
// If we need some instructions deleted, do it now.
NumGVNInstr += toErase.size();
// Avoid iterator invalidation.
bool AtStart = BI == BB->begin();
if (!AtStart)
--BI;
for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
E = toErase.end(); I != E; ++I) {
DEBUG(cerr << "GVN removed: " << **I);
MD->removeInstruction(*I);
if (AtStart)
BI = BB->begin();
else
++BI;
}
return changed_function;
}
Owen Anderson
committed
/// performPRE - Perform a purely local form of PRE that looks for diamond
/// control flow patterns and attempts to perform simple PRE at the join point.
bool GVN::performPRE(Function& F) {
Chris Lattner
committed
bool Changed = false;
SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
DenseMap<BasicBlock*, Value*> predMap;
Owen Anderson
committed
for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
BasicBlock* CurrentBlock = *DI;
// Nothing to PRE in the entry block.
if (CurrentBlock == &F.getEntryBlock()) continue;
for (BasicBlock::iterator BI = CurrentBlock->begin(),
BE = CurrentBlock->end(); BI != BE; ) {
Chris Lattner
committed
Instruction *CurInst = BI++;
if (isa<AllocationInst>(CurInst) || isa<TerminatorInst>(CurInst) ||
isa<PHINode>(CurInst) || CurInst->mayReadFromMemory() ||
CurInst->mayWriteToMemory())
Owen Anderson
committed
continue;
Chris Lattner
committed
uint32_t valno = VN.lookup(CurInst);
Owen Anderson
committed
// Look for the predecessors for PRE opportunities. We're
// only trying to solve the basic diamond case, where
// a value is computed in the successor and one predecessor,
// but not the other. We also explicitly disallow cases
// where the successor is its own predecessor, because they're
// more complicated to get right.
unsigned numWith = 0;
unsigned numWithout = 0;
BasicBlock* PREPred = 0;
predMap.clear();
Owen Anderson
committed
for (pred_iterator PI = pred_begin(CurrentBlock),
PE = pred_end(CurrentBlock); PI != PE; ++PI) {
// We're not interested in PRE where the block is its
Owen Anderson
committed
// own predecessor, on in blocks with predecessors
// that are not reachable.
if (*PI == CurrentBlock) {
Owen Anderson
committed
numWithout = 2;
Owen Anderson
committed
break;
} else if (!localAvail.count(*PI)) {
numWithout = 2;
break;
}
DenseMap<uint32_t, Value*>::iterator predV =
localAvail[*PI]->table.find(valno);
if (predV == localAvail[*PI]->table.end()) {
Owen Anderson
committed
PREPred = *PI;
numWithout++;
Chris Lattner
committed
} else if (predV->second == CurInst) {
Owen Anderson
committed
numWithout = 2;
} else {
Owen Anderson
committed
predMap[*PI] = predV->second;
Owen Anderson
committed
numWith++;
}
}
// Don't do PRE when it might increase code size, i.e. when
// we would need to insert instructions in more than one pred.
Chris Lattner
committed
if (numWithout != 1 || numWith == 0)
Owen Anderson
committed
continue;
// We can't do PRE safely on a critical edge, so instead we schedule
// the edge to be split and perform the PRE the next time we iterate
// on the function.
unsigned succNum = 0;
for (unsigned i = 0, e = PREPred->getTerminator()->getNumSuccessors();
i != e; ++i)
if (PREPred->getTerminator()->getSuccessor(i) == CurrentBlock) {
succNum = i;
break;
}
if (isCriticalEdge(PREPred->getTerminator(), succNum)) {
toSplit.push_back(std::make_pair(PREPred->getTerminator(), succNum));
continue;
}
Owen Anderson
committed
// Instantiate the expression the in predecessor that lacked it.
// Because we are going top-down through the block, all value numbers
// will be available in the predecessor by the time we need them. Any
// that weren't original present will have been instantiated earlier
// in this loop.
Chris Lattner
committed
Instruction* PREInstr = CurInst->clone();
Owen Anderson
committed
bool success = true;
Chris Lattner
committed
for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
Value *Op = PREInstr->getOperand(i);
if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
continue;
if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) {
PREInstr->setOperand(i, V);
} else {
success = false;
break;
Owen Anderson
committed
}
// Fail out if we encounter an operand that is not available in
// the PRE predecessor. This is typically because of loads which
// are not value numbered precisely.
if (!success) {
delete PREInstr;
continue;
}
PREInstr->insertBefore(PREPred->getTerminator());
Chris Lattner
committed
PREInstr->setName(CurInst->getName() + ".pre");
Owen Anderson
committed
predMap[PREPred] = PREInstr;
Owen Anderson
committed
VN.add(PREInstr, valno);
NumGVNPRE++;
// Update the availability map to include the new instruction.
Owen Anderson
committed
localAvail[PREPred]->table.insert(std::make_pair(valno, PREInstr));
Owen Anderson
committed
// Create a PHI to make the value available in this block.
Chris Lattner
committed
PHINode* Phi = PHINode::Create(CurInst->getType(),
CurInst->getName() + ".pre-phi",
Owen Anderson
committed
CurrentBlock->begin());
for (pred_iterator PI = pred_begin(CurrentBlock),
PE = pred_end(CurrentBlock); PI != PE; ++PI)
Owen Anderson
committed
Phi->addIncoming(predMap[*PI], *PI);
Owen Anderson
committed
VN.add(Phi, valno);
Owen Anderson
committed
localAvail[CurrentBlock]->table[valno] = Phi;
Owen Anderson
committed
Chris Lattner
committed
CurInst->replaceAllUsesWith(Phi);
if (isa<PointerType>(Phi->getType()))
MD->invalidateCachedPointerInfo(Phi);
Chris Lattner
committed
VN.erase(CurInst);
Owen Anderson
committed
Chris Lattner
committed
DEBUG(cerr << "GVN PRE removed: " << *CurInst);
MD->removeInstruction(CurInst);
CurInst->eraseFromParent();
Changed = true;
Owen Anderson
committed
}
}
for (SmallVector<std::pair<TerminatorInst*, unsigned>, 4>::iterator
Anton Korobeynikov
committed
I = toSplit.begin(), E = toSplit.end(); I != E; ++I)
SplitCriticalEdge(I->first, I->second, this);
Anton Korobeynikov
committed
return Changed || toSplit.size();
Owen Anderson
committed
}
Owen Anderson
committed
// iterateOnFunction - Executes one iteration of GVN
Owen Anderson
committed
// Top-down walk of the dominator tree
Owen Anderson
committed
bool changed = false;
for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
DE = df_end(DT->getRootNode()); DI != DE; ++DI)
Owen Anderson
committed
changed |= processBlock(*DI);
Owen Anderson
committed
return changed;
Owen Anderson
committed
}
void GVN::cleanupGlobalSets() {
VN.clear();
phiMap.clear();
for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator
I = localAvail.begin(), E = localAvail.end(); I != E; ++I)
delete I->second;
localAvail.clear();
}