#include "llvm/BasicBlock.h"
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
+#include "llvm/GlobalVariable.h"
#include "llvm/Function.h"
#include "llvm/IntrinsicInst.h"
#include "llvm/LLVMContext.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"
-#include "llvm/Analysis/MallocHelper.h"
+#include "llvm/Analysis/ConstantFolding.h"
+#include "llvm/Analysis/Dominators.h"
+#include "llvm/Analysis/MemoryBuiltins.h"
#include "llvm/Analysis/MemoryDependenceAnalysis.h"
+#include "llvm/Analysis/PHITransAddr.h"
#include "llvm/Support/CFG.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
+#include "llvm/Support/IRBuilder.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Transforms/Utils/Local.h"
-#include <cstdio>
+#include "llvm/Transforms/Utils/SSAUpdater.h"
using namespace llvm;
STATISTIC(NumGVNInstr, "Number of instructions deleted");
static cl::opt<bool> EnablePRE("enable-pre",
cl::init(true), cl::Hidden);
static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
+static cl::opt<bool> EnableFullLoadPRE("enable-full-load-pre", cl::init(false));
//===----------------------------------------------------------------------===//
// ValueTable Class
/// two values.
namespace {
struct Expression {
- enum ExpressionOpcode { ADD, FADD, SUB, FSUB, MUL, FMUL,
- 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,
- EMPTY, TOMBSTONE };
+ enum ExpressionOpcode {
+ ADD = Instruction::Add,
+ FADD = Instruction::FAdd,
+ SUB = Instruction::Sub,
+ FSUB = Instruction::FSub,
+ MUL = Instruction::Mul,
+ FMUL = Instruction::FMul,
+ UDIV = Instruction::UDiv,
+ SDIV = Instruction::SDiv,
+ FDIV = Instruction::FDiv,
+ UREM = Instruction::URem,
+ SREM = Instruction::SRem,
+ FREM = Instruction::FRem,
+ SHL = Instruction::Shl,
+ LSHR = Instruction::LShr,
+ ASHR = Instruction::AShr,
+ AND = Instruction::And,
+ OR = Instruction::Or,
+ XOR = Instruction::Xor,
+ TRUNC = Instruction::Trunc,
+ ZEXT = Instruction::ZExt,
+ SEXT = Instruction::SExt,
+ FPTOUI = Instruction::FPToUI,
+ FPTOSI = Instruction::FPToSI,
+ UITOFP = Instruction::UIToFP,
+ SITOFP = Instruction::SIToFP,
+ FPTRUNC = Instruction::FPTrunc,
+ FPEXT = Instruction::FPExt,
+ PTRTOINT = Instruction::PtrToInt,
+ INTTOPTR = Instruction::IntToPtr,
+ BITCAST = Instruction::BitCast,
+ 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, GEP, CALL, CONSTANT,
+ INSERTVALUE, EXTRACTVALUE, EMPTY, TOMBSTONE };
ExpressionOpcode opcode;
const Type* type;
- uint32_t firstVN;
- uint32_t secondVN;
- uint32_t thirdVN;
SmallVector<uint32_t, 4> varargs;
Value *function;
return false;
else if (function != other.function)
return false;
- 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;
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(GetElementPtrInst* G);
Expression create_expression(CallInst* C);
Expression create_expression(Constant* C);
+ Expression create_expression(ExtractValueInst* C);
+ Expression create_expression(InsertValueInst* C);
+
+ uint32_t lookup_or_add_call(CallInst* C);
public:
ValueTable() : nextValueNumber(1) { }
uint32_t lookup_or_add(Value *V);
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;
+ (unsigned)((uintptr_t)e.type >> 9));
for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
E = e.varargs.end(); I != E; ++I)
static bool isEqual(const Expression &LHS, const Expression &RHS) {
return LHS == RHS;
}
- static bool isPod() { return true; }
};
+
+template <>
+struct isPodLike<Expression> { static const bool value = true; };
+
}
//===----------------------------------------------------------------------===//
// ValueTable Internal Functions
//===----------------------------------------------------------------------===//
-Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) {
- switch(BO->getOpcode()) {
- default: // THIS SHOULD NEVER HAPPEN
- llvm_unreachable("Binary operator with unknown opcode?");
- case Instruction::Add: return Expression::ADD;
- case Instruction::FAdd: return Expression::FADD;
- case Instruction::Sub: return Expression::SUB;
- case Instruction::FSub: return Expression::FSUB;
- case Instruction::Mul: return Expression::MUL;
- case Instruction::FMul: return Expression::FMUL;
- 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;
- }
-}
Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) {
if (isa<ICmpInst>(C)) {
}
}
-Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) {
- switch(C->getOpcode()) {
- default: // THIS SHOULD NEVER HAPPEN
- llvm_unreachable("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;
- }
-}
-
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;
Expression ValueTable::create_expression(BinaryOperator* BO) {
Expression e;
-
- e.firstVN = lookup_or_add(BO->getOperand(0));
- e.secondVN = lookup_or_add(BO->getOperand(1));
- e.thirdVN = 0;
+ e.varargs.push_back(lookup_or_add(BO->getOperand(0)));
+ e.varargs.push_back(lookup_or_add(BO->getOperand(1)));
e.function = 0;
e.type = BO->getType();
- e.opcode = getOpcode(BO);
+ e.opcode = static_cast<Expression::ExpressionOpcode>(BO->getOpcode());
return e;
}
Expression ValueTable::create_expression(CmpInst* C) {
Expression e;
- e.firstVN = lookup_or_add(C->getOperand(0));
- e.secondVN = lookup_or_add(C->getOperand(1));
- e.thirdVN = 0;
+ e.varargs.push_back(lookup_or_add(C->getOperand(0)));
+ e.varargs.push_back(lookup_or_add(C->getOperand(1)));
e.function = 0;
e.type = C->getType();
e.opcode = getOpcode(C);
Expression ValueTable::create_expression(CastInst* C) {
Expression e;
- e.firstVN = lookup_or_add(C->getOperand(0));
- e.secondVN = 0;
- e.thirdVN = 0;
+ e.varargs.push_back(lookup_or_add(C->getOperand(0)));
e.function = 0;
e.type = C->getType();
- e.opcode = getOpcode(C);
+ e.opcode = static_cast<Expression::ExpressionOpcode>(C->getOpcode());
return e;
}
Expression ValueTable::create_expression(ShuffleVectorInst* S) {
Expression e;
- 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.varargs.push_back(lookup_or_add(S->getOperand(0)));
+ e.varargs.push_back(lookup_or_add(S->getOperand(1)));
+ e.varargs.push_back(lookup_or_add(S->getOperand(2)));
e.function = 0;
e.type = S->getType();
e.opcode = Expression::SHUFFLE;
Expression ValueTable::create_expression(ExtractElementInst* E) {
Expression e;
- e.firstVN = lookup_or_add(E->getOperand(0));
- e.secondVN = lookup_or_add(E->getOperand(1));
- e.thirdVN = 0;
+ e.varargs.push_back(lookup_or_add(E->getOperand(0)));
+ e.varargs.push_back(lookup_or_add(E->getOperand(1)));
e.function = 0;
e.type = E->getType();
e.opcode = Expression::EXTRACT;
Expression ValueTable::create_expression(InsertElementInst* I) {
Expression e;
- 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.varargs.push_back(lookup_or_add(I->getOperand(0)));
+ e.varargs.push_back(lookup_or_add(I->getOperand(1)));
+ e.varargs.push_back(lookup_or_add(I->getOperand(2)));
e.function = 0;
e.type = I->getType();
e.opcode = Expression::INSERT;
Expression ValueTable::create_expression(SelectInst* I) {
Expression e;
- e.firstVN = lookup_or_add(I->getCondition());
- e.secondVN = lookup_or_add(I->getTrueValue());
- e.thirdVN = lookup_or_add(I->getFalseValue());
+ e.varargs.push_back(lookup_or_add(I->getCondition()));
+ e.varargs.push_back(lookup_or_add(I->getTrueValue()));
+ e.varargs.push_back(lookup_or_add(I->getFalseValue()));
e.function = 0;
e.type = I->getType();
e.opcode = Expression::SELECT;
Expression ValueTable::create_expression(GetElementPtrInst* G) {
Expression e;
- e.firstVN = lookup_or_add(G->getPointerOperand());
- e.secondVN = 0;
- e.thirdVN = 0;
+ e.varargs.push_back(lookup_or_add(G->getPointerOperand()));
e.function = 0;
e.type = G->getType();
e.opcode = Expression::GEP;
return e;
}
+Expression ValueTable::create_expression(ExtractValueInst* E) {
+ Expression e;
+
+ e.varargs.push_back(lookup_or_add(E->getAggregateOperand()));
+ for (ExtractValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
+ II != IE; ++II)
+ e.varargs.push_back(*II);
+ e.function = 0;
+ e.type = E->getType();
+ e.opcode = Expression::EXTRACTVALUE;
+
+ return e;
+}
+
+Expression ValueTable::create_expression(InsertValueInst* E) {
+ Expression e;
+
+ e.varargs.push_back(lookup_or_add(E->getAggregateOperand()));
+ e.varargs.push_back(lookup_or_add(E->getInsertedValueOperand()));
+ for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
+ II != IE; ++II)
+ e.varargs.push_back(*II);
+ e.function = 0;
+ e.type = E->getType();
+ e.opcode = Expression::INSERTVALUE;
+
+ return e;
+}
+
//===----------------------------------------------------------------------===//
// ValueTable External Functions
//===----------------------------------------------------------------------===//
valueNumbering.insert(std::make_pair(V, num));
}
-/// 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)) {
- 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;
- }
-
- // Non-local case.
- 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;
+uint32_t ValueTable::lookup_or_add_call(CallInst* C) {
+ if (AA->doesNotAccessMemory(C)) {
+ Expression exp = create_expression(C);
+ uint32_t& e = expressionNumbering[exp];
+ if (!e) e = nextValueNumber++;
+ valueNumbering[C] = e;
+ return e;
+ } else if (AA->onlyReadsMemory(C)) {
+ Expression exp = create_expression(C);
+ uint32_t& e = expressionNumbering[exp];
+ if (!e) {
+ e = nextValueNumber++;
+ valueNumbering[C] = e;
+ return e;
+ }
+ if (!MD) {
+ e = nextValueNumber++;
+ valueNumbering[C] = e;
+ return e;
+ }
- // We don't handle non-depedencies. If we already have a call, reject
- // instruction dependencies.
- if (I->second.isClobber() || cdep != 0) {
- cdep = 0;
- break;
- }
+ MemDepResult local_dep = MD->getDependency(C);
- 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;
- }
+ if (!local_dep.isDef() && !local_dep.isNonLocal()) {
+ valueNumbering[C] = nextValueNumber;
+ return nextValueNumber++;
+ }
- cdep = 0;
- break;
- }
+ if (local_dep.isDef()) {
+ CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
- if (!cdep) {
- valueNumbering.insert(std::make_pair(V, nextValueNumber));
+ if (local_cdep->getNumOperands() != C->getNumOperands()) {
+ valueNumbering[C] = 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));
+ uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i));
if (c_vn != cd_vn) {
- valueNumbering.insert(std::make_pair(V, nextValueNumber));
+ valueNumbering[C] = nextValueNumber;
return nextValueNumber++;
}
}
- uint32_t v = lookup_or_add(cdep);
- valueNumbering.insert(std::make_pair(V, v));
+ uint32_t v = lookup_or_add(local_cdep);
+ valueNumbering[C] = v;
return v;
-
- } else {
- valueNumbering.insert(std::make_pair(V, nextValueNumber));
- return nextValueNumber++;
- }
- } else if (BinaryOperator* BO = dyn_cast<BinaryOperator>(V)) {
- 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));
+ // Non-local case.
+ 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 NonLocalDepEntry *I = &deps[i];
+ // Ignore non-local dependencies.
+ if (I->getResult().isNonLocal())
+ continue;
- return nextValueNumber++;
- }
- } else if (ShuffleVectorInst* U = dyn_cast<ShuffleVectorInst>(V)) {
- Expression e = create_expression(U);
+ // We don't handle non-depedencies. If we already have a call, reject
+ // instruction dependencies.
+ if (I->getResult().isClobber() || cdep != 0) {
+ cdep = 0;
+ break;
+ }
- 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));
+ CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
+ // FIXME: All duplicated with non-local case.
+ if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
+ cdep = NonLocalDepCall;
+ continue;
+ }
- return nextValueNumber++;
+ cdep = 0;
+ break;
}
- } 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));
+ if (!cdep) {
+ valueNumbering[C] = 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));
+ if (cdep->getNumOperands() != C->getNumOperands()) {
+ valueNumbering[C] = 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++;
+ 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[C] = 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));
+ uint32_t v = lookup_or_add(cdep);
+ valueNumbering[C] = v;
+ return v;
- return nextValueNumber++;
- }
- } else if (GetElementPtrInst* U = dyn_cast<GetElementPtrInst>(V)) {
- Expression e = create_expression(U);
+ } else {
+ valueNumbering[C] = nextValueNumber;
+ return nextValueNumber++;
+ }
+}
- 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));
+/// 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;
- return nextValueNumber++;
- }
- } else {
- valueNumbering.insert(std::make_pair(V, nextValueNumber));
+ if (!isa<Instruction>(V)) {
+ valueNumbering[V] = nextValueNumber;
return nextValueNumber++;
}
+
+ Instruction* I = cast<Instruction>(V);
+ Expression exp;
+ switch (I->getOpcode()) {
+ case Instruction::Call:
+ return lookup_or_add_call(cast<CallInst>(I));
+ case Instruction::Add:
+ case Instruction::FAdd:
+ case Instruction::Sub:
+ case Instruction::FSub:
+ case Instruction::Mul:
+ case Instruction::FMul:
+ case Instruction::UDiv:
+ case Instruction::SDiv:
+ case Instruction::FDiv:
+ case Instruction::URem:
+ case Instruction::SRem:
+ case Instruction::FRem:
+ case Instruction::Shl:
+ case Instruction::LShr:
+ case Instruction::AShr:
+ case Instruction::And:
+ case Instruction::Or :
+ case Instruction::Xor:
+ exp = create_expression(cast<BinaryOperator>(I));
+ break;
+ case Instruction::ICmp:
+ case Instruction::FCmp:
+ exp = create_expression(cast<CmpInst>(I));
+ break;
+ case Instruction::Trunc:
+ case Instruction::ZExt:
+ case Instruction::SExt:
+ case Instruction::FPToUI:
+ case Instruction::FPToSI:
+ case Instruction::UIToFP:
+ case Instruction::SIToFP:
+ case Instruction::FPTrunc:
+ case Instruction::FPExt:
+ case Instruction::PtrToInt:
+ case Instruction::IntToPtr:
+ case Instruction::BitCast:
+ exp = create_expression(cast<CastInst>(I));
+ break;
+ case Instruction::Select:
+ exp = create_expression(cast<SelectInst>(I));
+ break;
+ case Instruction::ExtractElement:
+ exp = create_expression(cast<ExtractElementInst>(I));
+ break;
+ case Instruction::InsertElement:
+ exp = create_expression(cast<InsertElementInst>(I));
+ break;
+ case Instruction::ShuffleVector:
+ exp = create_expression(cast<ShuffleVectorInst>(I));
+ break;
+ case Instruction::ExtractValue:
+ exp = create_expression(cast<ExtractValueInst>(I));
+ break;
+ case Instruction::InsertValue:
+ exp = create_expression(cast<InsertValueInst>(I));
+ break;
+ case Instruction::GetElementPtr:
+ exp = create_expression(cast<GetElementPtrInst>(I));
+ break;
+ default:
+ valueNumbering[V] = nextValueNumber;
+ return nextValueNumber++;
+ }
+
+ uint32_t& e = expressionNumbering[exp];
+ if (!e) e = nextValueNumber++;
+ valueNumbering[V] = e;
+ return e;
}
/// 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);
+ DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
assert(VI != valueNumbering.end() && "Value not numbered?");
return VI->second;
}
/// verifyRemoved - Verify that the value is removed from all internal data
/// structures.
void ValueTable::verifyRemoved(const Value *V) const {
- for (DenseMap<Value*, uint32_t>::iterator
+ for (DenseMap<Value*, uint32_t>::const_iterator
I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
assert(I->first != V && "Inst still occurs in value numbering map!");
}
bool runOnFunction(Function &F);
public:
static char ID; // Pass identification, replacement for typeid
- GVN() : FunctionPass(&ID) { }
+ explicit GVN(bool nopre = false, bool noloads = false)
+ : FunctionPass(&ID), NoPRE(nopre), NoLoads(noloads), MD(0) { }
private:
+ bool NoPRE;
+ bool NoLoads;
MemoryDependenceAnalysis *MD;
DominatorTree *DT;
ValueTable VN;
DenseMap<BasicBlock*, ValueNumberScope*> localAvail;
- typedef DenseMap<Value*, SmallPtrSet<Instruction*, 4> > PhiMapType;
- PhiMapType phiMap;
-
+ // List of critical edges to be split between iterations.
+ SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
// This transformation requires dominator postdominator info
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
AU.addRequired<DominatorTree>();
- AU.addRequired<MemoryDependenceAnalysis>();
+ if (!NoLoads)
+ AU.addRequired<MemoryDependenceAnalysis>();
AU.addRequired<AliasAnalysis>();
AU.addPreserved<DominatorTree>();
bool processNonLocalLoad(LoadInst* L,
SmallVectorImpl<Instruction*> &toErase);
bool processBlock(BasicBlock *BB);
- Value *GetValueForBlock(BasicBlock *BB, Instruction *orig,
- DenseMap<BasicBlock*, Value*> &Phis,
- bool top_level = false);
void dump(DenseMap<uint32_t, Value*>& d);
bool iterateOnFunction(Function &F);
Value *CollapsePhi(PHINode* p);
bool performPRE(Function& F);
Value *lookupNumber(BasicBlock *BB, uint32_t num);
- Value *AttemptRedundancyElimination(Instruction *orig, unsigned valno);
void cleanupGlobalSets();
void verifyRemoved(const Instruction *I) const;
+ bool splitCriticalEdges();
};
char GVN::ID = 0;
}
// createGVNPass - The public interface to this file...
-FunctionPass *llvm::createGVNPass() { return new GVN(); }
+FunctionPass *llvm::createGVNPass(bool NoPRE, bool NoLoads) {
+ return new GVN(NoPRE, NoLoads);
+}
static RegisterPass<GVN> X("gvn",
"Global Value Numbering");
void GVN::dump(DenseMap<uint32_t, Value*>& d) {
- printf("{\n");
+ errs() << "{\n";
for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
E = d.end(); I != E; ++I) {
- printf("%d\n", I->first);
+ errs() << I->first << "\n";
I->second->dump();
}
- printf("}\n");
+ errs() << "}\n";
}
static bool isSafeReplacement(PHINode* p, Instruction *inst) {
return 0;
}
-/// GetValueForBlock - Get the value to use within the specified basic block.
-/// available values are in Phis.
-Value *GVN::GetValueForBlock(BasicBlock *BB, Instruction *Orig,
- DenseMap<BasicBlock*, Value*> &Phis,
- bool TopLevel) {
-
- // If we have already computed this value, return the previously computed val.
- DenseMap<BasicBlock*, Value*>::iterator V = Phis.find(BB);
- if (V != Phis.end() && !TopLevel) 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());
-
- if (BasicBlock *Pred = BB->getSinglePredecessor()) {
- Value *ret = GetValueForBlock(Pred, Orig, Phis);
- 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);
-
- Phis.insert(std::make_pair(BB, PN));
-
- // 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
- if (LoadInst* L = dyn_cast<LoadInst>(Orig))
- phiMap[L->getPointerOperand()].insert(PN);
- else
- phiMap[Orig].insert(PN);
-
- return PN;
- }
-
- PN->replaceAllUsesWith(v);
- if (isa<PointerType>(v->getType()))
- MD->invalidateCachedPointerInfo(v);
-
- for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(),
- E = Phis.end(); I != E; ++I)
- if (I->second == PN)
- I->second = v;
-
- DEBUG(errs() << "GVN removed: " << *PN << '\n');
- MD->removeInstruction(PN);
- PN->eraseFromParent();
- DEBUG(verifyRemoved(PN));
-
- Phis[BB] = v;
- return v;
-}
-
/// 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
SmallVector<BasicBlock*, 32> BBWorklist;
BBWorklist.push_back(BB);
- while (!BBWorklist.empty()) {
+ do {
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.
for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
BBWorklist.push_back(*I);
- }
+ } while (!BBWorklist.empty());
return false;
}
+/// CanCoerceMustAliasedValueToLoad - Return true if
+/// CoerceAvailableValueToLoadType will succeed.
+static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
+ const Type *LoadTy,
+ const TargetData &TD) {
+ // If the loaded or stored value is an first class array or struct, don't try
+ // to transform them. We need to be able to bitcast to integer.
+ if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
+ StoredVal->getType()->isStructTy() ||
+ StoredVal->getType()->isArrayTy())
+ return false;
+
+ // The store has to be at least as big as the load.
+ if (TD.getTypeSizeInBits(StoredVal->getType()) <
+ TD.getTypeSizeInBits(LoadTy))
+ return false;
+
+ return true;
+}
+
+
/// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
/// then a load from a must-aliased pointer of a different type, try to coerce
/// the stored value. LoadedTy is the type of the load we want to replace and
const Type *LoadedTy,
Instruction *InsertPt,
const TargetData &TD) {
+ if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD))
+ return 0;
+
const Type *StoredValTy = StoredVal->getType();
uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy);
// If the store and reload are the same size, we can always reuse it.
if (StoreSize == LoadSize) {
- if (isa<PointerType>(StoredValTy) && isa<PointerType>(LoadedTy)) {
+ if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy()) {
// Pointer to Pointer -> use bitcast.
return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
}
// Convert source pointers to integers, which can be bitcast.
- if (isa<PointerType>(StoredValTy)) {
+ if (StoredValTy->isPointerTy()) {
StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
}
const Type *TypeToCastTo = LoadedTy;
- if (isa<PointerType>(TypeToCastTo))
+ if (TypeToCastTo->isPointerTy())
TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext());
if (StoredValTy != TypeToCastTo)
StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
// Cast to pointer if the load needs a pointer type.
- if (isa<PointerType>(LoadedTy))
+ if (LoadedTy->isPointerTy())
StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
return StoredVal;
// If the loaded value is smaller than the available value, then we can
// extract out a piece from it. If the available value is too small, then we
// can't do anything.
- if (StoreSize < LoadSize)
- return 0;
+ assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
// Convert source pointers to integers, which can be manipulated.
- if (isa<PointerType>(StoredValTy)) {
+ if (StoredValTy->isPointerTy()) {
StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
}
// Convert vectors and fp to integer, which can be manipulated.
- if (!isa<IntegerType>(StoredValTy)) {
+ if (!StoredValTy->isIntegerTy()) {
StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
}
return StoredVal;
// If the result is a pointer, inttoptr.
- if (isa<PointerType>(LoadedTy))
+ if (LoadedTy->isPointerTy())
return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
// Otherwise, bitcast.
/// be expressed as a base pointer plus a constant offset. Return the base and
/// offset to the caller.
static Value *GetBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
- const TargetData *TD) {
+ const TargetData &TD) {
Operator *PtrOp = dyn_cast<Operator>(Ptr);
if (PtrOp == 0) return Ptr;
// Handle a struct and array indices which add their offset to the pointer.
if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
- Offset += TD->getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
+ Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
} else {
- uint64_t Size = TD->getTypeAllocSize(GTI.getIndexedType());
+ uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
Offset += OpC->getSExtValue()*Size;
}
}
// Re-sign extend from the pointer size if needed to get overflow edge cases
// right.
- unsigned PtrSize = TD->getPointerSizeInBits();
+ unsigned PtrSize = TD.getPointerSizeInBits();
if (PtrSize < 64)
Offset = (Offset << (64-PtrSize)) >> (64-PtrSize);
}
-/// AnalyzeLoadFromClobberingStore - This function is called when we have a
-/// memdep query of a load that ends up being a clobbering store. This means
-/// that the store *may* provide bits used by the load but we can't be sure
-/// because the pointers don't mustalias. Check this case to see if there is
-/// anything more we can do before we give up. This returns -1 if we have to
-/// give up, or a byte number in the stored value of the piece that feeds the
-/// load.
-static int AnalyzeLoadFromClobberingStore(LoadInst *L, StoreInst *DepSI,
- const TargetData *TD) {
+/// AnalyzeLoadFromClobberingWrite - This function is called when we have a
+/// memdep query of a load that ends up being a clobbering memory write (store,
+/// memset, memcpy, memmove). This means that the write *may* provide bits used
+/// by the load but we can't be sure because the pointers don't mustalias.
+///
+/// Check this case to see if there is anything more we can do before we give
+/// up. This returns -1 if we have to give up, or a byte number in the stored
+/// value of the piece that feeds the load.
+static int AnalyzeLoadFromClobberingWrite(const Type *LoadTy, Value *LoadPtr,
+ Value *WritePtr,
+ uint64_t WriteSizeInBits,
+ const TargetData &TD) {
+ // If the loaded or stored value is an first class array or struct, don't try
+ // to transform them. We need to be able to bitcast to integer.
+ if (LoadTy->isStructTy() || LoadTy->isArrayTy())
+ return -1;
+
int64_t StoreOffset = 0, LoadOffset = 0;
- Value *StoreBase =
- GetBaseWithConstantOffset(DepSI->getPointerOperand(), StoreOffset, TD);
+ Value *StoreBase = GetBaseWithConstantOffset(WritePtr, StoreOffset, TD);
Value *LoadBase =
- GetBaseWithConstantOffset(L->getPointerOperand(), LoadOffset, TD);
+ GetBaseWithConstantOffset(LoadPtr, LoadOffset, TD);
if (StoreBase != LoadBase)
return -1;
// FIXME: Study to see if/when this happens.
if (LoadOffset == StoreOffset) {
#if 0
- errs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
+ dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
<< "Base = " << *StoreBase << "\n"
- << "Store Ptr = " << *DepSI->getPointerOperand() << "\n"
- << "Store Offs = " << StoreOffset << " - " << *DepSI << "\n"
- << "Load Ptr = " << *L->getPointerOperand() << "\n"
- << "Load Offs = " << LoadOffset << " - " << *L << "\n\n";
- errs() << "'" << L->getParent()->getParent()->getName() << "'"
- << *L->getParent();
+ << "Store Ptr = " << *WritePtr << "\n"
+ << "Store Offs = " << StoreOffset << "\n"
+ << "Load Ptr = " << *LoadPtr << "\n";
+ abort();
#endif
return -1;
}
// must have gotten confused.
// FIXME: Investigate cases where this bails out, e.g. rdar://7238614. Then
// remove this check, as it is duplicated with what we have below.
- uint64_t StoreSize = TD->getTypeSizeInBits(DepSI->getOperand(0)->getType());
- uint64_t LoadSize = TD->getTypeSizeInBits(L->getType());
+ uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy);
- if ((StoreSize & 7) | (LoadSize & 7))
+ if ((WriteSizeInBits & 7) | (LoadSize & 7))
return -1;
- StoreSize >>= 3; // Convert to bytes.
+ uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
LoadSize >>= 3;
}
if (isAAFailure) {
#if 0
- errs() << "STORE LOAD DEP WITH COMMON BASE:\n"
+ dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
<< "Base = " << *StoreBase << "\n"
- << "Store Ptr = " << *DepSI->getPointerOperand() << "\n"
- << "Store Offs = " << StoreOffset << " - " << *DepSI << "\n"
- << "Load Ptr = " << *L->getPointerOperand() << "\n"
- << "Load Offs = " << LoadOffset << " - " << *L << "\n\n";
- errs() << "'" << L->getParent()->getParent()->getName() << "'"
- << *L->getParent();
+ << "Store Ptr = " << *WritePtr << "\n"
+ << "Store Offs = " << StoreOffset << "\n"
+ << "Load Ptr = " << *LoadPtr << "\n";
+ abort();
#endif
return -1;
}
return LoadOffset-StoreOffset;
}
+/// AnalyzeLoadFromClobberingStore - This function is called when we have a
+/// memdep query of a load that ends up being a clobbering store.
+static int AnalyzeLoadFromClobberingStore(const Type *LoadTy, Value *LoadPtr,
+ StoreInst *DepSI,
+ const TargetData &TD) {
+ // Cannot handle reading from store of first-class aggregate yet.
+ if (DepSI->getOperand(0)->getType()->isStructTy() ||
+ DepSI->getOperand(0)->getType()->isArrayTy())
+ return -1;
+
+ Value *StorePtr = DepSI->getPointerOperand();
+ uint64_t StoreSize = TD.getTypeSizeInBits(DepSI->getOperand(0)->getType());
+ return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
+ StorePtr, StoreSize, TD);
+}
+
+static int AnalyzeLoadFromClobberingMemInst(const Type *LoadTy, Value *LoadPtr,
+ MemIntrinsic *MI,
+ const TargetData &TD) {
+ // If the mem operation is a non-constant size, we can't handle it.
+ ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
+ if (SizeCst == 0) return -1;
+ uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
+
+ // If this is memset, we just need to see if the offset is valid in the size
+ // of the memset..
+ if (MI->getIntrinsicID() == Intrinsic::memset)
+ return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
+ MemSizeInBits, TD);
+
+ // If we have a memcpy/memmove, the only case we can handle is if this is a
+ // copy from constant memory. In that case, we can read directly from the
+ // constant memory.
+ MemTransferInst *MTI = cast<MemTransferInst>(MI);
+
+ Constant *Src = dyn_cast<Constant>(MTI->getSource());
+ if (Src == 0) return -1;
+
+ GlobalVariable *GV = dyn_cast<GlobalVariable>(Src->getUnderlyingObject());
+ if (GV == 0 || !GV->isConstant()) return -1;
+
+ // See if the access is within the bounds of the transfer.
+ int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
+ MI->getDest(), MemSizeInBits, TD);
+ if (Offset == -1)
+ return Offset;
+
+ // Otherwise, see if we can constant fold a load from the constant with the
+ // offset applied as appropriate.
+ Src = ConstantExpr::getBitCast(Src,
+ llvm::Type::getInt8PtrTy(Src->getContext()));
+ Constant *OffsetCst =
+ ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
+ Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
+ Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
+ if (ConstantFoldLoadFromConstPtr(Src, &TD))
+ return Offset;
+ return -1;
+}
+
/// GetStoreValueForLoad - This function is called when we have a
/// memdep query of a load that ends up being a clobbering store. This means
/// that the store *may* provide bits used by the load but we can't be sure
/// because the pointers don't mustalias. Check this case to see if there is
/// anything more we can do before we give up.
-static Value *GetStoreValueForLoad(Value *SrcVal, int Offset,const Type *LoadTy,
- Instruction *InsertPt, const TargetData *TD){
+static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
+ const Type *LoadTy,
+ Instruction *InsertPt, const TargetData &TD){
LLVMContext &Ctx = SrcVal->getType()->getContext();
- uint64_t StoreSize = TD->getTypeSizeInBits(SrcVal->getType())/8;
- uint64_t LoadSize = TD->getTypeSizeInBits(LoadTy)/8;
+ uint64_t StoreSize = TD.getTypeSizeInBits(SrcVal->getType())/8;
+ uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
+ IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
// Compute which bits of the stored value are being used by the load. Convert
// to an integer type to start with.
- if (isa<PointerType>(SrcVal->getType()))
- SrcVal = new PtrToIntInst(SrcVal, TD->getIntPtrType(Ctx), "tmp", InsertPt);
- if (!isa<IntegerType>(SrcVal->getType()))
- SrcVal = new BitCastInst(SrcVal, IntegerType::get(Ctx, StoreSize*8),
- "tmp", InsertPt);
+ if (SrcVal->getType()->isPointerTy())
+ SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx), "tmp");
+ if (!SrcVal->getType()->isIntegerTy())
+ SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8),
+ "tmp");
// Shift the bits to the least significant depending on endianness.
unsigned ShiftAmt;
- if (TD->isLittleEndian()) {
+ if (TD.isLittleEndian())
ShiftAmt = Offset*8;
- } else {
- ShiftAmt = StoreSize-LoadSize-Offset;
- }
+ else
+ ShiftAmt = (StoreSize-LoadSize-Offset)*8;
- SrcVal = BinaryOperator::CreateLShr(SrcVal,
- ConstantInt::get(SrcVal->getType(), ShiftAmt), "tmp", InsertPt);
+ if (ShiftAmt)
+ SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt, "tmp");
- SrcVal = new TruncInst(SrcVal, IntegerType::get(Ctx, LoadSize*8),
- "tmp", InsertPt);
+ if (LoadSize != StoreSize)
+ SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8),
+ "tmp");
- return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, *TD);
+ return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
}
-/// GetAvailableBlockValues - Given the ValuesPerBlock list, convert all of the
-/// available values to values of the expected LoadTy in their blocks and insert
-/// the new values into BlockReplValues.
-static void
-GetAvailableBlockValues(DenseMap<BasicBlock*, Value*> &BlockReplValues,
- const SmallVector<std::pair<BasicBlock*,
- Value*>, 16> &ValuesPerBlock,
- const Type *LoadTy,
- const TargetData *TD) {
+/// GetMemInstValueForLoad - This function is called when we have a
+/// memdep query of a load that ends up being a clobbering mem intrinsic.
+static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
+ const Type *LoadTy, Instruction *InsertPt,
+ const TargetData &TD){
+ LLVMContext &Ctx = LoadTy->getContext();
+ uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
- for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
- BasicBlock *BB = ValuesPerBlock[i].first;
- Value *AvailableVal = ValuesPerBlock[i].second;
+ IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
+
+ // We know that this method is only called when the mem transfer fully
+ // provides the bits for the load.
+ if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
+ // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
+ // independently of what the offset is.
+ Value *Val = MSI->getValue();
+ if (LoadSize != 1)
+ Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
- Value *&BlockEntry = BlockReplValues[BB];
- if (BlockEntry) continue;
+ Value *OneElt = Val;
- if (AvailableVal->getType() != LoadTy) {
- assert(TD && "Need target data to handle type mismatch case");
- AvailableVal = CoerceAvailableValueToLoadType(AvailableVal, LoadTy,
- BB->getTerminator(), *TD);
- DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\n"
- << *ValuesPerBlock[i].second << '\n'
- << *AvailableVal << '\n' << "\n\n\n");
+ // Splat the value out to the right number of bits.
+ for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
+ // If we can double the number of bytes set, do it.
+ if (NumBytesSet*2 <= LoadSize) {
+ Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
+ Val = Builder.CreateOr(Val, ShVal);
+ NumBytesSet <<= 1;
+ continue;
+ }
+
+ // Otherwise insert one byte at a time.
+ Value *ShVal = Builder.CreateShl(Val, 1*8);
+ Val = Builder.CreateOr(OneElt, ShVal);
+ ++NumBytesSet;
}
- BlockEntry = AvailableVal;
+
+ return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD);
}
+
+ // Otherwise, this is a memcpy/memmove from a constant global.
+ MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
+ Constant *Src = cast<Constant>(MTI->getSource());
+
+ // Otherwise, see if we can constant fold a load from the constant with the
+ // offset applied as appropriate.
+ Src = ConstantExpr::getBitCast(Src,
+ llvm::Type::getInt8PtrTy(Src->getContext()));
+ Constant *OffsetCst =
+ ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
+ Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
+ Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
+ return ConstantFoldLoadFromConstPtr(Src, &TD);
+}
+
+
+
+struct AvailableValueInBlock {
+ /// BB - The basic block in question.
+ BasicBlock *BB;
+ enum ValType {
+ SimpleVal, // A simple offsetted value that is accessed.
+ MemIntrin // A memory intrinsic which is loaded from.
+ };
+
+ /// V - The value that is live out of the block.
+ PointerIntPair<Value *, 1, ValType> Val;
+
+ /// Offset - The byte offset in Val that is interesting for the load query.
+ unsigned Offset;
+
+ static AvailableValueInBlock get(BasicBlock *BB, Value *V,
+ unsigned Offset = 0) {
+ AvailableValueInBlock Res;
+ Res.BB = BB;
+ Res.Val.setPointer(V);
+ Res.Val.setInt(SimpleVal);
+ Res.Offset = Offset;
+ return Res;
+ }
+
+ static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
+ unsigned Offset = 0) {
+ AvailableValueInBlock Res;
+ Res.BB = BB;
+ Res.Val.setPointer(MI);
+ Res.Val.setInt(MemIntrin);
+ Res.Offset = Offset;
+ return Res;
+ }
+
+ bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
+ Value *getSimpleValue() const {
+ assert(isSimpleValue() && "Wrong accessor");
+ return Val.getPointer();
+ }
+
+ MemIntrinsic *getMemIntrinValue() const {
+ assert(!isSimpleValue() && "Wrong accessor");
+ return cast<MemIntrinsic>(Val.getPointer());
+ }
+
+ /// MaterializeAdjustedValue - Emit code into this block to adjust the value
+ /// defined here to the specified type. This handles various coercion cases.
+ Value *MaterializeAdjustedValue(const Type *LoadTy,
+ const TargetData *TD) const {
+ Value *Res;
+ if (isSimpleValue()) {
+ Res = getSimpleValue();
+ if (Res->getType() != LoadTy) {
+ assert(TD && "Need target data to handle type mismatch case");
+ Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
+ *TD);
+
+ DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
+ << *getSimpleValue() << '\n'
+ << *Res << '\n' << "\n\n\n");
+ }
+ } else {
+ Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
+ LoadTy, BB->getTerminator(), *TD);
+ DEBUG(errs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
+ << " " << *getMemIntrinValue() << '\n'
+ << *Res << '\n' << "\n\n\n");
+ }
+ return Res;
+ }
+};
+
+/// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
+/// construct SSA form, allowing us to eliminate LI. This returns the value
+/// that should be used at LI's definition site.
+static Value *ConstructSSAForLoadSet(LoadInst *LI,
+ SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
+ const TargetData *TD,
+ const DominatorTree &DT,
+ AliasAnalysis *AA) {
+ // Check for the fully redundant, dominating load case. In this case, we can
+ // just use the dominating value directly.
+ if (ValuesPerBlock.size() == 1 &&
+ DT.properlyDominates(ValuesPerBlock[0].BB, LI->getParent()))
+ return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), TD);
+
+ // Otherwise, we have to construct SSA form.
+ SmallVector<PHINode*, 8> NewPHIs;
+ SSAUpdater SSAUpdate(&NewPHIs);
+ SSAUpdate.Initialize(LI);
+
+ const Type *LoadTy = LI->getType();
+
+ for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
+ const AvailableValueInBlock &AV = ValuesPerBlock[i];
+ BasicBlock *BB = AV.BB;
+
+ if (SSAUpdate.HasValueForBlock(BB))
+ continue;
+
+ SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, TD));
+ }
+
+ // Perform PHI construction.
+ Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
+
+ // If new PHI nodes were created, notify alias analysis.
+ if (V->getType()->isPointerTy())
+ for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
+ AA->copyValue(LI, NewPHIs[i]);
+
+ return V;
+}
+
+static bool isLifetimeStart(Instruction *Inst) {
+ if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
+ return II->getIntrinsicID() == Intrinsic::lifetime_start;
+ return false;
}
/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
bool GVN::processNonLocalLoad(LoadInst *LI,
SmallVectorImpl<Instruction*> &toErase) {
// Find the non-local dependencies of the load.
- SmallVector<MemoryDependenceAnalysis::NonLocalDepEntry, 64> Deps;
+ SmallVector<NonLocalDepResult, 64> Deps;
MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(),
Deps);
- //DEBUG(errs() << "INVESTIGATING NONLOCAL LOAD: "
+ //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: "
// << Deps.size() << *LI << '\n');
// If we had to process more than one hundred blocks to find the
// If we had a phi translation failure, we'll have a single entry which is a
// clobber in the current block. Reject this early.
- if (Deps.size() == 1 && Deps[0].second.isClobber()) {
+ if (Deps.size() == 1 && Deps[0].getResult().isClobber()) {
DEBUG(
- errs() << "GVN: non-local load ";
- WriteAsOperand(errs(), LI);
- errs() << " is clobbered by " << *Deps[0].second.getInst() << '\n';
+ dbgs() << "GVN: non-local load ";
+ WriteAsOperand(dbgs(), LI);
+ dbgs() << " is clobbered by " << *Deps[0].getResult().getInst() << '\n';
);
return false;
}
// 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<AvailableValueInBlock, 16> ValuesPerBlock;
SmallVector<BasicBlock*, 16> UnavailableBlocks;
const TargetData *TD = 0;
for (unsigned i = 0, e = Deps.size(); i != e; ++i) {
- BasicBlock *DepBB = Deps[i].first;
- MemDepResult DepInfo = Deps[i].second;
+ BasicBlock *DepBB = Deps[i].getBB();
+ MemDepResult DepInfo = Deps[i].getResult();
if (DepInfo.isClobber()) {
+ // The address being loaded in this non-local block may not be the same as
+ // the pointer operand of the load if PHI translation occurs. Make sure
+ // to consider the right address.
+ Value *Address = Deps[i].getAddress();
+
+ // If the dependence is to a store that writes to a superset of the bits
+ // read by the load, we can extract the bits we need for the load from the
+ // stored value.
+ if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
+ if (TD == 0)
+ TD = getAnalysisIfAvailable<TargetData>();
+ if (TD && Address) {
+ int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
+ DepSI, *TD);
+ if (Offset != -1) {
+ ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
+ DepSI->getOperand(0),
+ Offset));
+ continue;
+ }
+ }
+ }
+
+ // If the clobbering value is a memset/memcpy/memmove, see if we can
+ // forward a value on from it.
+ if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
+ if (TD == 0)
+ TD = getAnalysisIfAvailable<TargetData>();
+ if (TD && Address) {
+ int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
+ DepMI, *TD);
+ if (Offset != -1) {
+ ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
+ Offset));
+ continue;
+ }
+ }
+ }
+
UnavailableBlocks.push_back(DepBB);
continue;
}
Instruction *DepInst = DepInfo.getInst();
// Loading the allocation -> undef.
- if (isa<AllocationInst>(DepInst) || isMalloc(DepInst)) {
- ValuesPerBlock.push_back(std::make_pair(DepBB,
- UndefValue::get(LI->getType())));
+ if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) ||
+ // Loading immediately after lifetime begin -> undef.
+ isLifetimeStart(DepInst)) {
+ ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
+ UndefValue::get(LI->getType())));
continue;
}
-
- if (StoreInst* S = dyn_cast<StoreInst>(DepInst)) {
+
+ if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
// Reject loads and stores that are to the same address but are of
// different types if we have to.
if (S->getOperand(0)->getType() != LI->getType()) {
// If the stored value is larger or equal to the loaded value, we can
// reuse it.
- if (TD == 0 ||
- TD->getTypeSizeInBits(S->getOperand(0)->getType()) <
- TD->getTypeSizeInBits(LI->getType())) {
+ if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getOperand(0),
+ LI->getType(), *TD)) {
UnavailableBlocks.push_back(DepBB);
continue;
}
}
- ValuesPerBlock.push_back(std::make_pair(DepBB, S->getOperand(0)));
-
- } else if (LoadInst* LD = dyn_cast<LoadInst>(DepInst)) {
+ ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
+ S->getOperand(0)));
+ continue;
+ }
+
+ if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
// If the types mismatch and we can't handle it, reject reuse of the load.
if (LD->getType() != LI->getType()) {
if (TD == 0)
// If the stored value is larger or equal to the loaded value, we can
// reuse it.
- if (TD == 0 ||
- TD->getTypeSizeInBits(LD->getType()) <
- TD->getTypeSizeInBits(LI->getType())) {
+ if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){
UnavailableBlocks.push_back(DepBB);
continue;
}
}
- ValuesPerBlock.push_back(std::make_pair(DepBB, LD));
- } else {
- // FIXME: Handle memset/memcpy.
- UnavailableBlocks.push_back(DepBB);
+ ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, LD));
continue;
}
+
+ UnavailableBlocks.push_back(DepBB);
+ continue;
}
// If we have no predecessors that produce a known value for this load, exit
// 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(errs() << "GVN REMOVING NONLOCAL LOAD #1: " << *LI << '\n');
- 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(errs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
-
- // Convert the block information to a map, and insert coersions as needed.
- DenseMap<BasicBlock*, Value*> BlockReplValues;
- GetAvailableBlockValues(BlockReplValues, ValuesPerBlock, LI->getType(), TD);
+ DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
// Perform PHI construction.
- Value *V = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true);
+ Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT,
+ VN.getAliasAnalysis());
LI->replaceAllUsesWith(V);
if (isa<PHINode>(V))
V->takeName(LI);
- if (isa<PointerType>(V->getType()))
+ if (V->getType()->isPointerTy())
MD->invalidateCachedPointerInfo(V);
+ VN.erase(LI);
toErase.push_back(LI);
NumGVNLoad++;
return true;
while (TmpBB->getSinglePredecessor()) {
isSinglePred = true;
TmpBB = TmpBB->getSinglePredecessor();
- if (!TmpBB) // If haven't found any, bail now.
- return false;
if (TmpBB == LoadBB) // Infinite (unreachable) loop.
return false;
if (Blockers.count(TmpBB))
// 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;
+ if (!EnableFullLoadPRE) {
+ for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
+ if (ValuesPerBlock[i].isSimpleValue() &&
+ ValuesPerBlock[i].getSimpleValue() == LI)
+ return false;
+ }
+ // FIXME: It is extremely unclear what this loop is doing, other than
+ // artificially restricting loadpre.
if (isSinglePred) {
bool isHot = false;
- for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
- if (Instruction *I = dyn_cast<Instruction>(ValuesPerBlock[i].second))
+ for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
+ const AvailableValueInBlock &AV = ValuesPerBlock[i];
+ if (AV.isSimpleValue())
// "Hot" Instruction is in some loop (because it dominates its dep.
// instruction).
- if (DT->dominates(LI, I)) {
- isHot = true;
- break;
- }
+ if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue()))
+ if (DT->dominates(LI, I)) {
+ isHot = true;
+ break;
+ }
+ }
// We are interested only in "hot" instructions. We don't want to do any
// mis-optimizations here.
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;
-
+ // Check to see how many predecessors have the loaded value fully
+ // available.
+ DenseMap<BasicBlock*, Value*> PredLoads;
DenseMap<BasicBlock*, char> FullyAvailableBlocks;
for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
- FullyAvailableBlocks[ValuesPerBlock[i].first] = true;
+ FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
FullyAvailableBlocks[UnavailableBlocks[i]] = false;
for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
PI != E; ++PI) {
- if (IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
+ BasicBlock *Pred = *PI;
+ if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) {
continue;
+ }
+ PredLoads[Pred] = 0;
- // If this load is not available in multiple predecessors, reject it.
- if (UnavailablePred && UnavailablePred != *PI)
+ if (Pred->getTerminator()->getNumSuccessors() != 1) {
+ if (isa<IndirectBrInst>(Pred->getTerminator())) {
+ DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
+ << Pred->getName() << "': " << *LI << '\n');
+ return false;
+ }
+ unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB);
+ toSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum));
return false;
- UnavailablePred = *PI;
+ }
}
- assert(UnavailablePred != 0 &&
+ // Decide whether PRE is profitable for this load.
+ unsigned NumUnavailablePreds = PredLoads.size();
+ assert(NumUnavailablePreds != 0 &&
"Fully available value should be eliminated above!");
+ if (!EnableFullLoadPRE) {
+ // If this load is unavailable in multiple predecessors, reject it.
+ // 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.
+ if (NumUnavailablePreds != 1)
+ return false;
+ }
+
+ // Check if the load can safely be moved to all the unavailable predecessors.
+ bool CanDoPRE = true;
+ SmallVector<Instruction*, 8> NewInsts;
+ for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
+ E = PredLoads.end(); I != E; ++I) {
+ BasicBlock *UnavailablePred = I->first;
+
+ // Do PHI translation to get its value in the predecessor if necessary. The
+ // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
+
+ // If all preds have a single successor, then we know it is safe to insert
+ // the load on the pred (?!?), so we can insert code to materialize the
+ // pointer if it is not available.
+ PHITransAddr Address(LI->getOperand(0), TD);
+ Value *LoadPtr = 0;
+ if (allSingleSucc) {
+ LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
+ *DT, NewInsts);
+ } else {
+ Address.PHITranslateValue(LoadBB, UnavailablePred);
+ LoadPtr = Address.getAddr();
+
+ // Make sure the value is live in the predecessor.
+ if (Instruction *Inst = dyn_cast_or_null<Instruction>(LoadPtr))
+ if (!DT->dominates(Inst->getParent(), UnavailablePred))
+ LoadPtr = 0;
+ }
- // 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);
+ // If we couldn't find or insert a computation of this phi translated value,
+ // we fail PRE.
+ if (LoadPtr == 0) {
+ DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
+ << *LI->getOperand(0) << "\n");
+ CanDoPRE = false;
+ break;
+ }
- // 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(errs() << "COULDN'T PRE LOAD BECAUSE PTR IS UNAVAILABLE IN PRED: "
- << *LPInst << '\n' << *LI << "\n");
- return false;
+ // Make sure it is valid to move this load here. We have to watch out for:
+ // @1 = getelementptr (i8* p, ...
+ // test p and branch if == 0
+ // load @1
+ // It is valid to have the getelementptr before the test, even if p can be 0,
+ // as getelementptr only does address arithmetic.
+ // If we are not pushing the value through any multiple-successor blocks
+ // we do not have this case. Otherwise, check that the load is safe to
+ // put anywhere; this can be improved, but should be conservatively safe.
+ if (!allSingleSucc &&
+ // FIXME: REEVALUTE THIS.
+ !isSafeToLoadUnconditionally(LoadPtr,
+ UnavailablePred->getTerminator(),
+ LI->getAlignment(), TD)) {
+ CanDoPRE = false;
+ break;
}
- // We don't currently handle critical edges :(
- if (UnavailablePred->getTerminator()->getNumSuccessors() != 1) {
- DEBUG(errs() << "COULD NOT PRE LOAD BECAUSE OF CRITICAL EDGE '"
- << UnavailablePred->getName() << "': " << *LI << '\n');
- return false;
+ I->second = LoadPtr;
}
- // Make sure it is valid to move this load here. We have to watch out for:
- // @1 = getelementptr (i8* p, ...
- // test p and branch if == 0
- // load @1
- // It is valid to have the getelementptr before the test, even if p can be 0,
- // as getelementptr only does address arithmetic.
- // If we are not pushing the value through any multiple-successor blocks
- // we do not have this case. Otherwise, check that the load is safe to
- // put anywhere; this can be improved, but should be conservatively safe.
- if (!allSingleSucc &&
- !isSafeToLoadUnconditionally(LoadPtr, UnavailablePred->getTerminator()))
+ if (!CanDoPRE) {
+ while (!NewInsts.empty())
+ NewInsts.pop_back_val()->eraseFromParent();
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(errs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
+ DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
+ DEBUG(if (!NewInsts.empty())
+ dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
+ << *NewInsts.back() << '\n');
+
+ // Assign value numbers to the new instructions.
+ for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
+ // FIXME: We really _ought_ to insert these value numbers into their
+ // parent's availability map. However, in doing so, we risk getting into
+ // ordering issues. If a block hasn't been processed yet, we would be
+ // marking a value as AVAIL-IN, which isn't what we intend.
+ VN.lookup_or_add(NewInsts[i]);
+ }
- Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
- LI->getAlignment(),
- UnavailablePred->getTerminator());
+ for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
+ E = PredLoads.end(); I != E; ++I) {
+ BasicBlock *UnavailablePred = I->first;
+ Value *LoadPtr = I->second;
- SmallPtrSet<Instruction*, 4> &p = phiMap[LI->getPointerOperand()];
- for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end();
- I != E; ++I)
- ValuesPerBlock.push_back(std::make_pair((*I)->getParent(), *I));
+ Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
+ LI->getAlignment(),
+ UnavailablePred->getTerminator());
- DenseMap<BasicBlock*, Value*> BlockReplValues;
- GetAvailableBlockValues(BlockReplValues, ValuesPerBlock, LI->getType(), TD);
- BlockReplValues[UnavailablePred] = NewLoad;
+ // Add the newly created load.
+ ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
+ NewLoad));
+ MD->invalidateCachedPointerInfo(LoadPtr);
+ DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
+ }
// Perform PHI construction.
- Value *V = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true);
+ Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT,
+ VN.getAliasAnalysis());
LI->replaceAllUsesWith(V);
if (isa<PHINode>(V))
V->takeName(LI);
- if (isa<PointerType>(V->getType()))
+ if (V->getType()->isPointerTy())
MD->invalidateCachedPointerInfo(V);
+ VN.erase(LI);
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 (!MD)
+ return false;
+
if (L->isVolatile())
return false;
// If the value isn't available, don't do anything!
if (Dep.isClobber()) {
- // FIXME: We should handle memset/memcpy/memmove as dependent instructions
- // to forward the value if available.
- //if (isa<MemIntrinsic>(Dep.getInst()))
- //errs() << "LOAD DEPENDS ON MEM: " << *L << "\n" << *Dep.getInst()<<"\n\n";
-
// Check to see if we have something like this:
// store i32 123, i32* %P
// %A = bitcast i32* %P to i8*
// a common base + constant offset, and if the previous store (or memset)
// completely covers this load. This sort of thing can happen in bitfield
// access code.
+ Value *AvailVal = 0;
if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst()))
if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) {
- int Offset = AnalyzeLoadFromClobberingStore(L, DepSI, TD);
- if (Offset != -1) {
- Value *AvailVal = GetStoreValueForLoad(DepSI->getOperand(0), Offset,
- L->getType(), L, TD);
- DEBUG(errs() << "GVN COERCED STORE BITS:\n" << *DepSI << '\n'
- << *AvailVal << '\n' << *L << "\n\n\n");
-
- // Replace the load!
- L->replaceAllUsesWith(AvailVal);
- if (isa<PointerType>(AvailVal->getType()))
- MD->invalidateCachedPointerInfo(AvailVal);
- toErase.push_back(L);
- NumGVNLoad++;
- return true;
- }
+ int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
+ L->getPointerOperand(),
+ DepSI, *TD);
+ if (Offset != -1)
+ AvailVal = GetStoreValueForLoad(DepSI->getOperand(0), Offset,
+ L->getType(), L, *TD);
}
+ // If the clobbering value is a memset/memcpy/memmove, see if we can forward
+ // a value on from it.
+ if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
+ if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) {
+ int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
+ L->getPointerOperand(),
+ DepMI, *TD);
+ if (Offset != -1)
+ AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L,*TD);
+ }
+ }
+
+ if (AvailVal) {
+ DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
+ << *AvailVal << '\n' << *L << "\n\n\n");
+
+ // Replace the load!
+ L->replaceAllUsesWith(AvailVal);
+ if (AvailVal->getType()->isPointerTy())
+ MD->invalidateCachedPointerInfo(AvailVal);
+ VN.erase(L);
+ toErase.push_back(L);
+ NumGVNLoad++;
+ return true;
+ }
+
DEBUG(
// fast print dep, using operator<< on instruction would be too slow
- errs() << "GVN: load ";
- WriteAsOperand(errs(), L);
+ dbgs() << "GVN: load ";
+ WriteAsOperand(dbgs(), L);
Instruction *I = Dep.getInst();
- errs() << " is clobbered by " << *I << '\n';
+ dbgs() << " is clobbered by " << *I << '\n';
);
return false;
}
// actually have the same type. See if we know how to reuse the stored
// value (depending on its type).
const TargetData *TD = 0;
- if (StoredVal->getType() != L->getType() &&
- (TD = getAnalysisIfAvailable<TargetData>())) {
- StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), L, *TD);
- if (StoredVal == 0)
+ if (StoredVal->getType() != L->getType()) {
+ if ((TD = getAnalysisIfAvailable<TargetData>())) {
+ StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
+ L, *TD);
+ if (StoredVal == 0)
+ return false;
+
+ DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
+ << '\n' << *L << "\n\n\n");
+ }
+ else
return false;
-
- DEBUG(errs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
- << '\n' << *L << "\n\n\n");
}
// Remove it!
L->replaceAllUsesWith(StoredVal);
- if (isa<PointerType>(StoredVal->getType()))
+ if (StoredVal->getType()->isPointerTy())
MD->invalidateCachedPointerInfo(StoredVal);
+ VN.erase(L);
toErase.push_back(L);
NumGVNLoad++;
return true;
// the same type. See if we know how to reuse the previously loaded value
// (depending on its type).
const TargetData *TD = 0;
- if (DepLI->getType() != L->getType() &&
- (TD = getAnalysisIfAvailable<TargetData>())) {
- AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), L, *TD);
- if (AvailableVal == 0)
- return false;
+ if (DepLI->getType() != L->getType()) {
+ if ((TD = getAnalysisIfAvailable<TargetData>())) {
+ AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), L,*TD);
+ if (AvailableVal == 0)
+ return false;
- DEBUG(errs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
- << "\n" << *L << "\n\n\n");
+ DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
+ << "\n" << *L << "\n\n\n");
+ }
+ else
+ return false;
}
// Remove it!
L->replaceAllUsesWith(AvailableVal);
- if (isa<PointerType>(DepLI->getType()))
+ if (DepLI->getType()->isPointerTy())
MD->invalidateCachedPointerInfo(DepLI);
+ VN.erase(L);
toErase.push_back(L);
NumGVNLoad++;
return true;
// 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) || isMalloc(DepInst)) {
+ if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) {
L->replaceAllUsesWith(UndefValue::get(L->getType()));
+ VN.erase(L);
toErase.push_back(L);
NumGVNLoad++;
return true;
}
+
+ // If this load occurs either right after a lifetime begin,
+ // then the loaded value is undefined.
+ if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(DepInst)) {
+ if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
+ L->replaceAllUsesWith(UndefValue::get(L->getType()));
+ VN.erase(L);
+ toErase.push_back(L);
+ NumGVNLoad++;
+ return true;
+ }
+ }
return false;
}
return 0;
}
-/// AttemptRedundancyElimination - If the "fast path" of redundancy elimination
-/// by inheritance from the dominator fails, see if we can perform phi
-/// construction to eliminate the redundancy.
-Value *GVN::AttemptRedundancyElimination(Instruction *orig, unsigned valno) {
- BasicBlock *BaseBlock = orig->getParent();
-
- SmallPtrSet<BasicBlock*, 4> Visited;
- SmallVector<BasicBlock*, 8> Stack;
- Stack.push_back(BaseBlock);
-
- DenseMap<BasicBlock*, Value*> Results;
-
- // Walk backwards through our predecessors, looking for instances of the
- // value number we're looking for. Instances are recorded in the Results
- // map, which is then used to perform phi construction.
- while (!Stack.empty()) {
- BasicBlock *Current = Stack.back();
- Stack.pop_back();
-
- // If we've walked all the way to a proper dominator, then give up. Cases
- // where the instance is in the dominator will have been caught by the fast
- // path, and any cases that require phi construction further than this are
- // probably not worth it anyways. Note that this is a SIGNIFICANT compile
- // time improvement.
- if (DT->properlyDominates(Current, orig->getParent())) return 0;
-
- DenseMap<BasicBlock*, ValueNumberScope*>::iterator LA =
- localAvail.find(Current);
- if (LA == localAvail.end()) return 0;
- DenseMap<uint32_t, Value*>::iterator V = LA->second->table.find(valno);
-
- if (V != LA->second->table.end()) {
- // Found an instance, record it.
- Results.insert(std::make_pair(Current, V->second));
- continue;
- }
-
- // If we reach the beginning of the function, then give up.
- if (pred_begin(Current) == pred_end(Current))
- return 0;
-
- for (pred_iterator PI = pred_begin(Current), PE = pred_end(Current);
- PI != PE; ++PI)
- if (Visited.insert(*PI))
- Stack.push_back(*PI);
- }
-
- // If we didn't find instances, give up. Otherwise, perform phi construction.
- if (Results.size() == 0)
- return 0;
- else
- return GetValueForBlock(BaseBlock, orig, Results, true);
-}
/// processInstruction - When calculating availability, handle an instruction
/// by inserting it into the appropriate sets
bool GVN::processInstruction(Instruction *I,
SmallVectorImpl<Instruction*> &toErase) {
+ // Ignore dbg info intrinsics.
+ if (isa<DbgInfoIntrinsic>(I))
+ return false;
+
if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
bool Changed = processLoad(LI, toErase);
// Allocations are always uniquely numbered, so we can save time and memory
// by fast failing them.
- } else if (isa<AllocationInst>(I) || isMalloc(I) || isa<TerminatorInst>(I)) {
+ } else if (isa<AllocaInst>(I) || isa<TerminatorInst>(I)) {
localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
return false;
}
Value *constVal = CollapsePhi(p);
if (constVal) {
- for (PhiMapType::iterator PI = phiMap.begin(), PE = phiMap.end();
- PI != PE; ++PI)
- PI->second.erase(p);
-
p->replaceAllUsesWith(constVal);
- if (isa<PointerType>(constVal->getType()))
+ if (MD && constVal->getType()->isPointerTy())
MD->invalidateCachedPointerInfo(constVal);
VN.erase(p);
// Remove it!
VN.erase(I);
I->replaceAllUsesWith(repl);
- if (isa<PointerType>(repl->getType()))
+ if (MD && repl->getType()->isPointerTy())
MD->invalidateCachedPointerInfo(repl);
toErase.push_back(I);
return true;
-#if 0
- // Perform slow-pathvalue-number based elimination with phi construction.
- } else if (Value *repl = AttemptRedundancyElimination(I, Num)) {
- // Remove it!
- VN.erase(I);
- I->replaceAllUsesWith(repl);
- if (isa<PointerType>(repl->getType()))
- MD->invalidateCachedPointerInfo(repl);
- toErase.push_back(I);
- return true;
-#endif
} else {
localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
}
/// runOnFunction - This is the main transformation entry point for a function.
bool GVN::runOnFunction(Function& F) {
- MD = &getAnalysis<MemoryDependenceAnalysis>();
+ if (!NoLoads)
+ MD = &getAnalysis<MemoryDependenceAnalysis>();
DT = &getAnalysis<DominatorTree>();
VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
VN.setMemDep(MD);
unsigned Iteration = 0;
while (ShouldContinue) {
- DEBUG(errs() << "GVN iteration: " << Iteration << "\n");
+ DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
ShouldContinue = iterateOnFunction(F);
+ if (splitCriticalEdges())
+ ShouldContinue = true;
Changed |= ShouldContinue;
++Iteration;
}
for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
E = toErase.end(); I != E; ++I) {
- DEBUG(errs() << "GVN removed: " << **I << '\n');
- MD->removeInstruction(*I);
+ DEBUG(dbgs() << "GVN removed: " << **I << '\n');
+ if (MD) MD->removeInstruction(*I);
(*I)->eraseFromParent();
DEBUG(verifyRemoved(*I));
}
/// 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) {
+bool GVN::performPRE(Function &F) {
bool Changed = false;
- SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
DenseMap<BasicBlock*, Value*> predMap;
for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
BE = CurrentBlock->end(); BI != BE; ) {
Instruction *CurInst = BI++;
- if (isa<AllocationInst>(CurInst) || isMalloc(CurInst) ||
+ if (isa<AllocaInst>(CurInst) ||
isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
- (CurInst->getType() == Type::getVoidTy(F.getContext())) ||
+ CurInst->getType()->isVoidTy() ||
CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
isa<DbgInfoIntrinsic>(CurInst))
continue;
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
- // own predecessor, on in blocks with predecessors
+ // own predecessor, or in blocks with predecessors
// that are not reachable.
if (*PI == CurrentBlock) {
NumWithout = 2;
// we would need to insert instructions in more than one pred.
if (NumWithout != 1 || NumWith == 0)
continue;
+
+ // Don't do PRE across indirect branch.
+ if (isa<IndirectBrInst>(PREPred->getTerminator()))
+ 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;
- }
-
+ unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
continue;
}
- // Instantiate the expression the in predecessor that lacked it.
+ // Instantiate the expression in the 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
+ // that weren't originally present will have been instantiated earlier
// in this loop.
- Instruction *PREInstr = CurInst->clone(CurInst->getContext());
+ Instruction *PREInstr = CurInst->clone();
bool success = true;
for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
Value *Op = PREInstr->getOperand(i);
localAvail[CurrentBlock]->table[ValNo] = Phi;
CurInst->replaceAllUsesWith(Phi);
- if (isa<PointerType>(Phi->getType()))
+ if (MD && Phi->getType()->isPointerTy())
MD->invalidateCachedPointerInfo(Phi);
VN.erase(CurInst);
- DEBUG(errs() << "GVN PRE removed: " << *CurInst << '\n');
- MD->removeInstruction(CurInst);
+ DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
+ if (MD) MD->removeInstruction(CurInst);
CurInst->eraseFromParent();
DEBUG(verifyRemoved(CurInst));
Changed = true;
}
}
- for (SmallVector<std::pair<TerminatorInst*, unsigned>, 4>::iterator
- I = toSplit.begin(), E = toSplit.end(); I != E; ++I)
- SplitCriticalEdge(I->first, I->second, this);
+ if (splitCriticalEdges())
+ Changed = true;
- return Changed || toSplit.size();
+ return Changed;
+}
+
+/// splitCriticalEdges - Split critical edges found during the previous
+/// iteration that may enable further optimization.
+bool GVN::splitCriticalEdges() {
+ if (toSplit.empty())
+ return false;
+ do {
+ std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
+ SplitCriticalEdge(Edge.first, Edge.second, this);
+ } while (!toSplit.empty());
+ MD->invalidateCachedPredecessors();
+ return true;
}
/// iterateOnFunction - Executes one iteration of GVN
void GVN::cleanupGlobalSets() {
VN.clear();
- phiMap.clear();
for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator
I = localAvail.begin(), E = localAvail.end(); I != E; ++I)
void GVN::verifyRemoved(const Instruction *Inst) const {
VN.verifyRemoved(Inst);
- // Walk through the PHI map to make sure the instruction isn't hiding in there
- // somewhere.
- for (PhiMapType::iterator
- I = phiMap.begin(), E = phiMap.end(); I != E; ++I) {
- assert(I->first != Inst && "Inst is still a key in PHI map!");
-
- for (SmallPtrSet<Instruction*, 4>::iterator
- II = I->second.begin(), IE = I->second.end(); II != IE; ++II) {
- assert(*II != Inst && "Inst is still a value in PHI map!");
- }
- }
-
// Walk through the value number scope to make sure the instruction isn't
// ferreted away in it.
- for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator
+ for (DenseMap<BasicBlock*, ValueNumberScope*>::const_iterator
I = localAvail.begin(), E = localAvail.end(); I != E; ++I) {
const ValueNumberScope *VNS = I->second;
while (VNS) {
- for (DenseMap<uint32_t, Value*>::iterator
+ for (DenseMap<uint32_t, Value*>::const_iterator
II = VNS->table.begin(), IE = VNS->table.end(); II != IE; ++II) {
assert(II->second != Inst && "Inst still in value numbering scope!");
}
projects / oota-llvm.git / blobdiff