#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/Dominators.h"
#include "llvm/Analysis/AliasAnalysis.h"
-#include "llvm/Analysis/MallocHelper.h"
+#include "llvm/Analysis/MallocFreeHelper.h"
#include "llvm/Analysis/MemoryDependenceAnalysis.h"
#include "llvm/Support/CFG.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Transforms/Utils/Local.h"
+#include "llvm/Transforms/Utils/SSAUpdater.h"
#include <cstdio>
using namespace llvm;
SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI,
FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT,
PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, CONSTANT,
- EMPTY, TOMBSTONE };
+ 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;
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)
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);
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);
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;
+ }
- // 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 MemoryDependenceAnalysis::NonLocalDepEntry *I = &deps[i];
+ // Ignore non-local dependencies.
+ if (I->second.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->second.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->second.getInst());
+ // FIXME: All duplicated with non-local case.
+ if (NonLocalDepCall && DT->properlyDominates(I->first, 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
ValueTable VN;
DenseMap<BasicBlock*, ValueNumberScope*> localAvail;
- typedef DenseMap<Value*, SmallPtrSet<Instruction*, 4> > PhiMapType;
- PhiMapType phiMap;
-
-
// This transformation requires dominator postdominator info
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
AU.addRequired<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;
};
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
}
+/// 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 (isa<StructType>(LoadTy) || isa<ArrayType>(LoadTy) ||
+ isa<StructType>(StoredVal->getType()) ||
+ isa<ArrayType>(StoredVal->getType()))
+ 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 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)) {
return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
}
-static void
-GetAvailableBlockValues(DenseMap<BasicBlock*, Value*> &BlockReplValues,
- SmallVector<std::pair<BasicBlock*,
- Value*>, 16> &ValuesPerBlock,
- const Type *LoadTy,
- const TargetData *TD) {
+/// GetBaseWithConstantOffset - Analyze the specified pointer to see if it can
+/// 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) {
+ Operator *PtrOp = dyn_cast<Operator>(Ptr);
+ if (PtrOp == 0) return Ptr;
+
+ // Just look through bitcasts.
+ if (PtrOp->getOpcode() == Instruction::BitCast)
+ return GetBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
+
+ // If this is a GEP with constant indices, we can look through it.
+ GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
+ if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
+
+ gep_type_iterator GTI = gep_type_begin(GEP);
+ for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
+ ++I, ++GTI) {
+ ConstantInt *OpC = cast<ConstantInt>(*I);
+ if (OpC->isZero()) continue;
+
+ // 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());
+ } else {
+ 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();
+ if (PtrSize < 64)
+ Offset = (Offset << (64-PtrSize)) >> (64-PtrSize);
+
+ return GetBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
+}
+
+/// 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) {
+ // 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 (isa<StructType>(L->getType()) || isa<ArrayType>(L->getType()) ||
+ isa<StructType>(DepSI->getOperand(0)->getType()) ||
+ isa<ArrayType>(DepSI->getOperand(0)->getType()))
+ return -1;
+
+ int64_t StoreOffset = 0, LoadOffset = 0;
+ Value *StoreBase =
+ GetBaseWithConstantOffset(DepSI->getPointerOperand(), StoreOffset, TD);
+ Value *LoadBase =
+ GetBaseWithConstantOffset(L->getPointerOperand(), LoadOffset, TD);
+ if (StoreBase != LoadBase)
+ return -1;
+
+ // If the load and store are to the exact same address, they should have been
+ // a must alias. AA must have gotten confused.
+ // FIXME: Study to see if/when this happens.
+ if (LoadOffset == StoreOffset) {
+#if 0
+ errs() << "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();
+#endif
+ return -1;
+ }
+
+ // If the load and store don't overlap at all, the store doesn't provide
+ // anything to the load. In this case, they really don't alias at all, AA
+ // 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());
+
+ if ((StoreSize & 7) | (LoadSize & 7))
+ return -1;
+ StoreSize >>= 3; // Convert to bytes.
+ LoadSize >>= 3;
+
+
+ bool isAAFailure = false;
+ if (StoreOffset < LoadOffset) {
+ isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
+ } else {
+ isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
+ }
+ if (isAAFailure) {
+#if 0
+ errs() << "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();
+#endif
+ return -1;
+ }
+
+ // If the Load isn't completely contained within the stored bits, we don't
+ // have all the bits to feed it. We could do something crazy in the future
+ // (issue a smaller load then merge the bits in) but this seems unlikely to be
+ // valuable.
+ if (StoreOffset > LoadOffset ||
+ StoreOffset+StoreSize < LoadOffset+LoadSize)
+ return -1;
+
+ // Okay, we can do this transformation. Return the number of bytes into the
+ // store that the load is.
+ return LoadOffset-StoreOffset;
+}
+
+
+/// 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, 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;
+
+
+ // 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);
+
+ // Shift the bits to the least significant depending on endianness.
+ unsigned ShiftAmt;
+ if (TD.isLittleEndian()) {
+ ShiftAmt = Offset*8;
+ } else {
+ ShiftAmt = (StoreSize-LoadSize-Offset)*8;
+ }
+
+ if (ShiftAmt)
+ SrcVal = BinaryOperator::CreateLShr(SrcVal,
+ ConstantInt::get(SrcVal->getType(), ShiftAmt), "tmp", InsertPt);
+
+ if (LoadSize != StoreSize)
+ SrcVal = new TruncInst(SrcVal, IntegerType::get(Ctx, LoadSize*8),
+ "tmp", InsertPt);
+
+ return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
+}
+
+struct AvailableValueInBlock {
+ /// BB - The basic block in question.
+ BasicBlock *BB;
+ /// V - The value that is live out of the block.
+ Value *V;
+ /// Offset - The byte offset in V 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.V = V;
+ Res.Offset = Offset;
+ 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,
+ AliasAnalysis *AA) {
+ 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) {
- BasicBlock *BB = ValuesPerBlock[i].first;
- Value *AvailableVal = ValuesPerBlock[i].second;
+ BasicBlock *BB = ValuesPerBlock[i].BB;
+ Value *AvailableVal = ValuesPerBlock[i].V;
+ unsigned Offset = ValuesPerBlock[i].Offset;
- Value *&BlockEntry = BlockReplValues[BB];
- if (BlockEntry) continue;
+ if (SSAUpdate.HasValueForBlock(BB))
+ continue;
if (AvailableVal->getType() != LoadTy) {
assert(TD && "Need target data to handle type mismatch case");
- AvailableVal = CoerceAvailableValueToLoadType(AvailableVal, LoadTy,
- BB->getTerminator(), *TD);
+ AvailableVal = GetStoreValueForLoad(AvailableVal, Offset, LoadTy,
+ BB->getTerminator(), *TD);
+
+ if (Offset) {
+ DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\n"
+ << *ValuesPerBlock[i].V << '\n'
+ << *AvailableVal << '\n' << "\n\n\n");
+ }
+
+
DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\n"
- << *ValuesPerBlock[i].second << '\n'
- << *AvailableVal << '\n' << "\n\n\n");
+ << *ValuesPerBlock[i].V << '\n'
+ << *AvailableVal << '\n' << "\n\n\n");
}
- BlockEntry = AvailableVal;
+
+ SSAUpdate.AddAvailableValue(BB, AvailableVal);
}
+
+ // Perform PHI construction.
+ Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
+
+ // If new PHI nodes were created, notify alias analysis.
+ if (isa<PointerType>(V->getType()))
+ for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
+ AA->copyValue(LI, NewPHIs[i]);
+
+ return V;
}
/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
// 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;
MemDepResult DepInfo = Deps[i].second;
if (DepInfo.isClobber()) {
+ // 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) {
+ int Offset = AnalyzeLoadFromClobberingStore(LI, DepSI, *TD);
+ if (Offset != -1) {
+ ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
+ DepSI->getOperand(0),
+ Offset));
+ continue;
+ }
+ }
+ }
+
+ // FIXME: Handle memset/memcpy.
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)) {
+ 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);
// Perform PHI construction.
- Value *V = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true);
+ Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD,
+ VN.getAliasAnalysis());
LI->replaceAllUsesWith(V);
if (isa<PHINode>(V))
// 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)
+ if (ValuesPerBlock[i].V == LI)
return false;
if (isSinglePred) {
bool isHot = false;
for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
- if (Instruction *I = dyn_cast<Instruction>(ValuesPerBlock[i].second))
+ if (Instruction *I = dyn_cast<Instruction>(ValuesPerBlock[i].V))
// "Hot" Instruction is in some loop (because it dominates its dep.
// instruction).
if (DT->dominates(LI, I)) {
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;
LI->getAlignment(),
UnavailablePred->getTerminator());
- 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));
-
- 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));
// Perform PHI construction.
- Value *V = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true);
+ Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD,
+ VN.getAliasAnalysis());
LI->replaceAllUsesWith(V);
if (isa<PHINode>(V))
V->takeName(LI);
return true;
}
-/// GetBaseWithConstantOffset - Analyze the specified pointer to see if it can
-/// 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) {
- Operator *PtrOp = dyn_cast<Operator>(Ptr);
- if (PtrOp == 0) return Ptr;
-
- // Just look through bitcasts.
- if (PtrOp->getOpcode() == Instruction::BitCast)
- return GetBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
-
- // If this is a GEP with constant indices, we can look through it.
- GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
- if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
-
- gep_type_iterator GTI = gep_type_begin(GEP);
- for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
- ++I, ++GTI) {
- ConstantInt *OpC = cast<ConstantInt>(*I);
- if (OpC->isZero()) continue;
-
- // 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());
- } else {
- 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();
- if (PtrSize < 64)
- Offset = (Offset << (64-PtrSize)) >> (64-PtrSize);
-
- return GetBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
-}
-
-
-/// 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) {
- int64_t StoreOffset = 0, LoadOffset = 0;
- Value *StoreBase =
- GetBaseWithConstantOffset(DepSI->getPointerOperand(), StoreOffset, TD);
- Value *LoadBase =
- GetBaseWithConstantOffset(L->getPointerOperand(), LoadOffset, TD);
- if (StoreBase != LoadBase)
- return -1;
-
- // If the load and store are to the exact same address, they should have been
- // a must alias. AA must have gotten confused.
- // FIXME: Study to see if/when this happens.
- if (LoadOffset == StoreOffset) {
-#if 0
- errs() << "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();
-#endif
- return -1;
- }
-
- // If the load and store don't overlap at all, the store doesn't provide
- // anything to the load. In this case, they really don't alias at all, AA
- // 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());
-
- if ((StoreSize & 7) | (LoadSize & 7))
- return -1;
- StoreSize >>= 3; // Convert to bytes.
- LoadSize >>= 3;
-
-
- bool isAAFailure = false;
- if (StoreOffset < LoadOffset) {
- isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
- } else {
- isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
- }
- if (isAAFailure) {
-#if 0
- errs() << "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();
-#endif
- return -1;
- }
-
- // If the Load isn't completely contained within the stored bits, we don't
- // have all the bits to feed it. We could do something crazy in the future
- // (issue a smaller load then merge the bits in) but this seems unlikely to be
- // valuable.
- if (StoreOffset > LoadOffset ||
- StoreOffset+StoreSize < LoadOffset+LoadSize)
- return -1;
-
- // Okay, we can do this transformation. Return the number of bytes into the
- // store that the load is.
- return LoadOffset-StoreOffset;
-}
-
-
-/// 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){
- LLVMContext &Ctx = SrcVal->getType()->getContext();
-
- uint64_t StoreSize = TD->getTypeSizeInBits(SrcVal->getType())/8;
- uint64_t LoadSize = TD->getTypeSizeInBits(LoadTy)/8;
-
-
- // 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);
-
- // Shift the bits to the least significant depending on endianness.
- unsigned ShiftAmt;
- if (TD->isLittleEndian()) {
- ShiftAmt = Offset*8;
- } else {
- ShiftAmt = StoreSize-LoadSize-Offset;
- }
-
- SrcVal = BinaryOperator::CreateLShr(SrcVal,
- ConstantInt::get(SrcVal->getType(), ShiftAmt), "tmp", InsertPt);
-
- SrcVal = new TruncInst(SrcVal, IntegerType::get(Ctx, LoadSize*8),
- "tmp", InsertPt);
-
- return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, *TD);
-}
-
-
-
/// 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) {
// access code.
if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst()))
if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) {
- int Offset = AnalyzeLoadFromClobberingStore(L, DepSI, TD);
+ int Offset = AnalyzeLoadFromClobberingStore(L, DepSI, *TD);
if (Offset != -1) {
Value *AvailVal = GetStoreValueForLoad(DepSI->getOperand(0), Offset,
- L->getType(), L, TD);
+ L->getType(), L, *TD);
DEBUG(errs() << "GVN COERCED STORE BITS:\n" << *DepSI << '\n'
<< *AvailVal << '\n' << *L << "\n\n\n");
// 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(errs() << "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!
// 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(errs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
+ << "\n" << *L << "\n\n\n");
+ }
+ else
+ return false;
}
// Remove it!
// 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()));
toErase.push_back(L);
NumGVNLoad++;
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
// 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()))
MD->invalidateCachedPointerInfo(constVal);
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));
}
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;
// 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.
- 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);
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
projects / oota-llvm.git / blobdiff