#include "llvm/ADT/Statistic.h"
using namespace llvm;
-STATISTIC(NumDeadStore, "Number of dead stores eliminated");
+STATISTIC(NumDeadStore, "Number of dead stores eliminated");
+STATISTIC(NumGlobalCopies, "Number of allocas copied from constant global");
+
+/// pointsToConstantGlobal - Return true if V (possibly indirectly) points to
+/// some part of a constant global variable. This intentionally only accepts
+/// constant expressions because we can't rewrite arbitrary instructions.
+static bool pointsToConstantGlobal(Value *V) {
+ if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
+ return GV->isConstant();
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
+ if (CE->getOpcode() == Instruction::BitCast ||
+ CE->getOpcode() == Instruction::GetElementPtr)
+ return pointsToConstantGlobal(CE->getOperand(0));
+ return false;
+}
+
+/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
+/// pointer to an alloca. Ignore any reads of the pointer, return false if we
+/// see any stores or other unknown uses. If we see pointer arithmetic, keep
+/// track of whether it moves the pointer (with IsOffset) but otherwise traverse
+/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
+/// the alloca, and if the source pointer is a pointer to a constant global, we
+/// can optimize this.
+static bool
+isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
+ SmallVectorImpl<Instruction *> &ToDelete,
+ bool IsOffset = false) {
+ // We track lifetime intrinsics as we encounter them. If we decide to go
+ // ahead and replace the value with the global, this lets the caller quickly
+ // eliminate the markers.
+
+ for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
+ User *U = cast<Instruction>(*UI);
+
+ if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
+ // Ignore non-volatile loads, they are always ok.
+ if (!LI->isSimple()) return false;
+ continue;
+ }
+
+ if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
+ // If uses of the bitcast are ok, we are ok.
+ if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, ToDelete, IsOffset))
+ return false;
+ continue;
+ }
+ if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
+ // If the GEP has all zero indices, it doesn't offset the pointer. If it
+ // doesn't, it does.
+ if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, ToDelete,
+ IsOffset || !GEP->hasAllZeroIndices()))
+ return false;
+ continue;
+ }
+
+ if (CallSite CS = U) {
+ // If this is the function being called then we treat it like a load and
+ // ignore it.
+ if (CS.isCallee(UI))
+ continue;
+
+ // If this is a readonly/readnone call site, then we know it is just a
+ // load (but one that potentially returns the value itself), so we can
+ // ignore it if we know that the value isn't captured.
+ unsigned ArgNo = CS.getArgumentNo(UI);
+ if (CS.onlyReadsMemory() &&
+ (CS.getInstruction()->use_empty() || CS.doesNotCapture(ArgNo)))
+ continue;
+
+ // If this is being passed as a byval argument, the caller is making a
+ // copy, so it is only a read of the alloca.
+ if (CS.isByValArgument(ArgNo))
+ continue;
+ }
+
+ // Lifetime intrinsics can be handled by the caller.
+ if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
+ if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
+ II->getIntrinsicID() == Intrinsic::lifetime_end) {
+ assert(II->use_empty() && "Lifetime markers have no result to use!");
+ ToDelete.push_back(II);
+ continue;
+ }
+ }
+
+ // If this is isn't our memcpy/memmove, reject it as something we can't
+ // handle.
+ MemTransferInst *MI = dyn_cast<MemTransferInst>(U);
+ if (MI == 0)
+ return false;
+
+ // If the transfer is using the alloca as a source of the transfer, then
+ // ignore it since it is a load (unless the transfer is volatile).
+ if (UI.getOperandNo() == 1) {
+ if (MI->isVolatile()) return false;
+ continue;
+ }
+
+ // If we already have seen a copy, reject the second one.
+ if (TheCopy) return false;
+
+ // If the pointer has been offset from the start of the alloca, we can't
+ // safely handle this.
+ if (IsOffset) return false;
+
+ // If the memintrinsic isn't using the alloca as the dest, reject it.
+ if (UI.getOperandNo() != 0) return false;
+
+ // If the source of the memcpy/move is not a constant global, reject it.
+ if (!pointsToConstantGlobal(MI->getSource()))
+ return false;
+
+ // Otherwise, the transform is safe. Remember the copy instruction.
+ TheCopy = MI;
+ }
+ return true;
+}
+
+/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
+/// modified by a copy from a constant global. If we can prove this, we can
+/// replace any uses of the alloca with uses of the global directly.
+static MemTransferInst *
+isOnlyCopiedFromConstantGlobal(AllocaInst *AI,
+ SmallVectorImpl<Instruction *> &ToDelete) {
+ MemTransferInst *TheCopy = 0;
+ if (isOnlyCopiedFromConstantGlobal(AI, TheCopy, ToDelete))
+ return TheCopy;
+ return 0;
+}
+
+/// getPointeeAlignment - Compute the minimum alignment of the value pointed
+/// to by the given pointer.
+static unsigned getPointeeAlignment(Value *V, const TargetData &TD) {
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
+ if (CE->getOpcode() == Instruction::BitCast ||
+ (CE->getOpcode() == Instruction::GetElementPtr &&
+ cast<GEPOperator>(CE)->hasAllZeroIndices()))
+ return getPointeeAlignment(CE->getOperand(0), TD);
+
+ if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
+ if (!GV->isDeclaration())
+ return TD.getPreferredAlignment(GV);
+
+ if (PointerType *PT = dyn_cast<PointerType>(V->getType()))
+ return TD.getABITypeAlignment(PT->getElementType());
+
+ return 0;
+}
Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) {
// Ensure that the alloca array size argument has type intptr_t, so that
// any casting is exposed early.
if (TD) {
- const Type *IntPtrTy = TD->getIntPtrType(AI.getContext());
+ Type *IntPtrTy = TD->getIntPtrType(AI.getContext());
if (AI.getArraySize()->getType() != IntPtrTy) {
Value *V = Builder->CreateIntCast(AI.getArraySize(),
IntPtrTy, false);
// Convert: alloca Ty, C - where C is a constant != 1 into: alloca [C x Ty], 1
if (AI.isArrayAllocation()) { // Check C != 1
if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
- const Type *NewTy =
+ Type *NewTy =
ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
- assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
AllocaInst *New = Builder->CreateAlloca(NewTy, 0, AI.getName());
New->setAlignment(AI.getAlignment());
Value *Idx[2];
Idx[0] = NullIdx;
Idx[1] = NullIdx;
- Value *V = GetElementPtrInst::CreateInBounds(New, Idx, Idx + 2,
- New->getName()+".sub", It);
+ Instruction *GEP =
+ GetElementPtrInst::CreateInBounds(New, Idx, New->getName()+".sub");
+ InsertNewInstBefore(GEP, *It);
// Now make everything use the getelementptr instead of the original
// allocation.
- return ReplaceInstUsesWith(AI, V);
+ return ReplaceInstUsesWith(AI, GEP);
} else if (isa<UndefValue>(AI.getArraySize())) {
return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
}
}
- if (TD && isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized()) {
- // If alloca'ing a zero byte object, replace the alloca with a null pointer.
- // Note that we only do this for alloca's, because malloc should allocate
- // and return a unique pointer, even for a zero byte allocation.
- if (TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
- return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
-
+ if (TD && AI.getAllocatedType()->isSized()) {
// If the alignment is 0 (unspecified), assign it the preferred alignment.
if (AI.getAlignment() == 0)
AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
+
+ // Move all alloca's of zero byte objects to the entry block and merge them
+ // together. Note that we only do this for alloca's, because malloc should
+ // allocate and return a unique pointer, even for a zero byte allocation.
+ if (TD->getTypeAllocSize(AI.getAllocatedType()) == 0) {
+ // For a zero sized alloca there is no point in doing an array allocation.
+ // This is helpful if the array size is a complicated expression not used
+ // elsewhere.
+ if (AI.isArrayAllocation()) {
+ AI.setOperand(0, ConstantInt::get(AI.getArraySize()->getType(), 1));
+ return &AI;
+ }
+
+ // Get the first instruction in the entry block.
+ BasicBlock &EntryBlock = AI.getParent()->getParent()->getEntryBlock();
+ Instruction *FirstInst = EntryBlock.getFirstNonPHIOrDbg();
+ if (FirstInst != &AI) {
+ // If the entry block doesn't start with a zero-size alloca then move
+ // this one to the start of the entry block. There is no problem with
+ // dominance as the array size was forced to a constant earlier already.
+ AllocaInst *EntryAI = dyn_cast<AllocaInst>(FirstInst);
+ if (!EntryAI || !EntryAI->getAllocatedType()->isSized() ||
+ TD->getTypeAllocSize(EntryAI->getAllocatedType()) != 0) {
+ AI.moveBefore(FirstInst);
+ return &AI;
+ }
+
+ // If the alignment of the entry block alloca is 0 (unspecified),
+ // assign it the preferred alignment.
+ if (EntryAI->getAlignment() == 0)
+ EntryAI->setAlignment(
+ TD->getPrefTypeAlignment(EntryAI->getAllocatedType()));
+ // Replace this zero-sized alloca with the one at the start of the entry
+ // block after ensuring that the address will be aligned enough for both
+ // types.
+ unsigned MaxAlign = std::max(EntryAI->getAlignment(),
+ AI.getAlignment());
+ EntryAI->setAlignment(MaxAlign);
+ if (AI.getType() != EntryAI->getType())
+ return new BitCastInst(EntryAI, AI.getType());
+ return ReplaceInstUsesWith(AI, EntryAI);
+ }
+ }
}
- return 0;
+ if (TD) {
+ // Check to see if this allocation is only modified by a memcpy/memmove from
+ // a constant global whose alignment is equal to or exceeds that of the
+ // allocation. If this is the case, we can change all users to use
+ // the constant global instead. This is commonly produced by the CFE by
+ // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
+ // is only subsequently read.
+ SmallVector<Instruction *, 4> ToDelete;
+ if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(&AI, ToDelete)) {
+ if (AI.getAlignment() <= getPointeeAlignment(Copy->getSource(), *TD)) {
+ DEBUG(dbgs() << "Found alloca equal to global: " << AI << '\n');
+ DEBUG(dbgs() << " memcpy = " << *Copy << '\n');
+ for (unsigned i = 0, e = ToDelete.size(); i != e; ++i)
+ EraseInstFromFunction(*ToDelete[i]);
+ Constant *TheSrc = cast<Constant>(Copy->getSource());
+ Instruction *NewI
+ = ReplaceInstUsesWith(AI, ConstantExpr::getBitCast(TheSrc,
+ AI.getType()));
+ EraseInstFromFunction(*Copy);
+ ++NumGlobalCopies;
+ return NewI;
+ }
+ }
+ }
+
+ // At last, use the generic allocation site handler to aggressively remove
+ // unused allocas.
+ return visitAllocSite(AI);
}
User *CI = cast<User>(LI.getOperand(0));
Value *CastOp = CI->getOperand(0);
-
const PointerType *DestTy = cast<PointerType>(CI->getType());
- const Type *DestPTy = DestTy->getElementType();
- if (
const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
+ PointerType *DestTy = cast<PointerType>(CI->getType());
+ Type *DestPTy = DestTy->getElementType();
+ if (PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
// If the address spaces don't match, don't eliminate the cast.
if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
return 0;
- const Type *SrcPTy = SrcTy->getElementType();
+ Type *SrcPTy = SrcTy->getElementType();
if (DestPTy->isIntegerTy() || DestPTy->isPointerTy() ||
DestPTy->isVectorTy()) {
// If the source is an array, the code below will not succeed. Check to
// see if a trivial 'gep P, 0, 0' will help matters. Only do this for
// constants.
- if (
const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
+ if (ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
if (Constant *CSrc = dyn_cast<Constant>(CastOp))
if (ASrcTy->getNumElements() != 0) {
Value *Idxs[2];
Idxs[0] = Constant::getNullValue(Type::getInt32Ty(LI.getContext()));
Idxs[1] = Idxs[0];
- CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
+ CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
SrcTy = cast<PointerType>(CastOp->getType());
SrcPTy = SrcTy->getElementType();
}
LoadInst *NewLoad =
IC.Builder->CreateLoad(CastOp, LI.isVolatile(), CI->getName());
NewLoad->setAlignment(LI.getAlignment());
+ NewLoad->setAtomic(LI.getOrdering(), LI.getSynchScope());
// Now cast the result of the load.
return new BitCastInst(NewLoad, LI.getType());
}
if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
return Res;
- // None of the following transforms are legal for volatile loads.
- if (LI.isVolatile()) return 0;
+ // None of the following transforms are legal for volatile/atomic loads.
+ // FIXME: Some of it is okay for atomic loads; needs refactoring.
+ if (!LI.isSimple()) return 0;
// Do really simple store-to-load forwarding and load CSE, to catch cases
// where there are several consecutive memory accesses to the same location,
User *CI = cast<User>(SI.getOperand(1));
Value *CastOp = CI->getOperand(0);
-
const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
-
const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
+ Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
+ PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
if (SrcTy == 0) return 0;
- const Type *SrcPTy = SrcTy->getElementType();
+ Type *SrcPTy = SrcTy->getElementType();
if (!DestPTy->isIntegerTy() && !DestPTy->isPointerTy())
return 0;
NewGEPIndices.push_back(Zero);
while (1) {
- if (const StructType *STy = dyn_cast<StructType>(SrcPTy)) {
+ if (StructType *STy = dyn_cast<StructType>(SrcPTy)) {
if (!STy->getNumElements()) /* Struct can be empty {} */
break;
NewGEPIndices.push_back(Zero);
SrcPTy = STy->getElementType(0);
- } else if (
const ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
+ } else if (ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
NewGEPIndices.push_back(Zero);
SrcPTy = ATy->getElementType();
} else {
Value *NewCast;
Value *SIOp0 = SI.getOperand(0);
Instruction::CastOps opcode = Instruction::BitCast;
- const Type* CastSrcTy = SIOp0->getType();
- const Type* CastDstTy = SrcPTy;
+ Type* CastSrcTy = SIOp0->getType();
+ Type* CastDstTy = SrcPTy;
if (CastDstTy->isPointerTy()) {
if (CastSrcTy->isIntegerTy())
opcode = Instruction::IntToPtr;
// SIOp0 is a pointer to aggregate and this is a store to the first field,
// emit a GEP to index into its first field.
if (!NewGEPIndices.empty())
- CastOp = IC.Builder->CreateInBoundsGEP(CastOp, NewGEPIndices.begin(),
- NewGEPIndices.end());
+ CastOp = IC.Builder->CreateInBoundsGEP(CastOp, NewGEPIndices);
NewCast = IC.Builder->CreateCast(opcode, SIOp0, CastDstTy,
SIOp0->getName()+".c");
return false;
}
-// If this instruction has two uses, one of which is a llvm.dbg.declare,
-// return the llvm.dbg.declare.
-DbgDeclareInst *InstCombiner::hasOneUsePlusDeclare(Value *V) {
- if (!V->hasNUses(2))
- return 0;
- for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
- UI != E; ++UI) {
- User *U = *UI;
- if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(U))
- return DI;
- if (isa<BitCastInst>(U) && U->hasOneUse()) {
- if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(*U->use_begin()))
- return DI;
- }
- }
- return 0;
-}
-
Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
Value *Val = SI.getOperand(0);
Value *Ptr = SI.getOperand(1);
- // If the RHS is an alloca with a single use, zapify the store, making the
- // alloca dead.
- // If the RHS is an alloca with a two uses, the other one being a
- // llvm.dbg.declare, zapify the store and the declare, making the
- // alloca dead. We must do this to prevent declares from affecting
- // codegen.
- if (!SI.isVolatile()) {
- if (Ptr->hasOneUse()) {
- if (isa<AllocaInst>(Ptr))
- return EraseInstFromFunction(SI);
- if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
- if (isa<AllocaInst>(GEP->getOperand(0))) {
- if (GEP->getOperand(0)->hasOneUse())
- return EraseInstFromFunction(SI);
- if (DbgDeclareInst *DI = hasOneUsePlusDeclare(GEP->getOperand(0))) {
- EraseInstFromFunction(*DI);
- return EraseInstFromFunction(SI);
- }
- }
- }
- }
- if (DbgDeclareInst *DI = hasOneUsePlusDeclare(Ptr)) {
- EraseInstFromFunction(*DI);
- return EraseInstFromFunction(SI);
- }
- }
-
// Attempt to improve the alignment.
if (TD) {
unsigned KnownAlign =
SI.setAlignment(EffectiveStoreAlign);
}
+ // Don't hack volatile/atomic stores.
+ // FIXME: Some bits are legal for atomic stores; needs refactoring.
+ if (!SI.isSimple()) return 0;
+
+ // If the RHS is an alloca with a single use, zapify the store, making the
+ // alloca dead.
+ if (Ptr->hasOneUse()) {
+ if (isa<AllocaInst>(Ptr))
+ return EraseInstFromFunction(SI);
+ if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
+ if (isa<AllocaInst>(GEP->getOperand(0))) {
+ if (GEP->getOperand(0)->hasOneUse())
+ return EraseInstFromFunction(SI);
+ }
+ }
+ }
+
// Do really simple DSE, to catch cases where there are several consecutive
// stores to the same location, separated by a few arithmetic operations. This
// situation often occurs with bitfield accesses.
if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
// Prev store isn't volatile, and stores to the same location?
- if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
- SI.getOperand(1))) {
+ if (PrevSI->isSimple() && equivalentAddressValues(PrevSI->getOperand(1),
+ SI.getOperand(1))) {
++NumDeadStore;
++BBI;
EraseInstFromFunction(*PrevSI);
// then *this* store is dead (X = load P; store X -> P).
if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
- !SI.isVolatile())
+ LI->isSimple())
return EraseInstFromFunction(SI);
// Otherwise, this is a load from some other location. Stores before it
if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
break;
}
-
-
- if (SI.isVolatile()) return 0; // Don't hack volatile stores.
// store X, null -> turns into 'unreachable' in SimplifyCFG
if (isa<ConstantPointerNull>(Ptr) && SI.getPointerAddressSpace() == 0) {
return false;
--BBI;
}
- // If this isn't a store, isn't a store to the same location, or if the
- // alignments differ, bail out.
+ // If this isn't a store, isn't a store to the same location, or is not the
+ // right kind of store, bail out.
OtherStore = dyn_cast<StoreInst>(BBI);
if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1) ||
- OtherStore->getAlignment() != SI.getAlignment())
+ !SI.isSameOperationAs(OtherStore))
return false;
} else {
// Otherwise, the other block ended with a conditional branch. If one of the
// Check to see if we find the matching store.
if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
if (OtherStore->getOperand(1) != SI.getOperand(1) ||
- OtherStore->getAlignment() != SI.getAlignment())
+ !SI.isSameOperationAs(OtherStore))
return false;
break;
}
// Insert a PHI node now if we need it.
Value *MergedVal = OtherStore->getOperand(0);
if (MergedVal != SI.getOperand(0)) {
- PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
- PN->reserveOperandSpace(2);
+ PHINode *PN = PHINode::Create(MergedVal->getType(), 2, "storemerge");
PN->addIncoming(SI.getOperand(0), SI.getParent());
PN->addIncoming(OtherStore->getOperand(0), OtherBB);
MergedVal = InsertNewInstBefore(PN, DestBB->front());
// Advance to a place where it is safe to insert the new store and
// insert it.
- BBI = DestBB->getFirstNonPHI();
- InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
- OtherStore->isVolatile(),
- SI.getAlignment()), *BBI);
-
+ BBI = DestBB->getFirstInsertionPt();
+ StoreInst *NewSI = new StoreInst(MergedVal, SI.getOperand(1),
+ SI.isVolatile(),
+ SI.getAlignment(),
+ SI.getOrdering(),
+ SI.getSynchScope());
+ InsertNewInstBefore(NewSI, *BBI);
+ NewSI->setDebugLoc(OtherStore->getDebugLoc());
+
// Nuke the old stores.
EraseInstFromFunction(SI);
EraseInstFromFunction(*OtherStore);
projects / oota-llvm.git / blobdiff