//===----------------------------------------------------------------------===//
#include "InstCombine.h"
-#include "llvm/IntrinsicInst.h"
+#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/Loads.h"
-#include "llvm/Target/TargetData.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/IntrinsicInst.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Transforms/Utils/Local.h"
-#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;
+}
-// Try to kill dead allocas by walking through its uses until we see some use
-// that could escape. This is a conservative analysis which tries to handle
-// GEPs, bitcasts, stores, and no-op intrinsics. These tend to be the things
-// left after inlining and SROA finish chewing on an alloca.
-static Instruction *removeDeadAlloca(InstCombiner &IC, AllocaInst &AI) {
- SmallVector<Instruction *, 4> Worklist, DeadStores;
- Worklist.push_back(&AI);
- do {
- Instruction *PI = Worklist.pop_back_val();
- for (Value::use_iterator UI = PI->use_begin(), UE = PI->use_end();
- UI != UE; ++UI) {
- Instruction *I = cast<Instruction>(*UI);
- switch (I->getOpcode()) {
- default:
- // Give up the moment we see something we can't handle.
- return 0;
-
- case Instruction::GetElementPtr:
- case Instruction::BitCast:
- Worklist.push_back(I);
+/// 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;
- case Instruction::Call:
- // We can handle a limited subset of calls to no-op intrinsics.
- if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
- switch (II->getIntrinsicID()) {
- case Intrinsic::dbg_declare:
- case Intrinsic::dbg_value:
- case Intrinsic::invariant_start:
- case Intrinsic::invariant_end:
- case Intrinsic::lifetime_start:
- case Intrinsic::lifetime_end:
- continue;
- default:
- return 0;
- }
- }
- // Reject everything else.
- return 0;
-
- case Instruction::Store: {
- // Stores into the alloca are only live if the alloca is live.
- StoreInst *SI = cast<StoreInst>(I);
- // We can eliminate atomic stores, but not volatile.
- if (SI->isVolatile())
- return 0;
- // The store is only trivially safe if the poniter is the destination
- // as opposed to the value. We're conservative here and don't check for
- // the case where we store the address of a dead alloca into a dead
- // alloca.
- if (SI->getPointerOperand() != PI)
- return 0;
- DeadStores.push_back(I);
+ // 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;
}
- }
}
- } while (!Worklist.empty());
- // The alloca is dead. Kill off all the stores to it, and then replace it
- // with undef.
- while (!DeadStores.empty())
- IC.EraseInstFromFunction(*DeadStores.pop_back_val());
- return IC.ReplaceInstUsesWith(AI, UndefValue::get(AI.getType()));
+ // 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;
}
Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) {
// 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())) {
- 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());
}
}
- 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);
+ }
+ }
+ }
+
+ if (AI.getAlignment()) {
+ // 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)) {
+ unsigned SourceAlign = getOrEnforceKnownAlignment(Copy->getSource(),
+ AI.getAlignment(), TD);
+ if (AI.getAlignment() <= SourceAlign) {
+ 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;
+ }
+ }
}
- // Try to aggressively remove allocas which are only used for GEPs, lifetime
- // markers, and stores. This happens when SROA iteratively promotes stores
- // out of the alloca, and we need to cleanup after it.
- return removeDeadAlloca(*this, AI);
+ // At last, use the generic allocation site handler to aggressively remove
+ // unused allocas.
+ return visitAllocSite(AI);
}
/// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
- const TargetData *TD) {
+ const DataLayout *TD) {
User *CI = cast<User>(LI.getOperand(0));
Value *CastOp = CI->getOperand(0);
Type *SrcPTy = SrcTy->getElementType();
- if (DestPTy->isIntegerTy() || DestPTy->isPointerTy() ||
+ 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
SrcPTy = SrcTy->getElementType();
}
- if (IC.getTargetData() &&
- (SrcPTy->isIntegerTy() || SrcPTy->isPointerTy() ||
+ if (IC.getDataLayout() &&
+ (SrcPTy->isIntegerTy() || SrcPTy->isPointerTy() ||
SrcPTy->isVectorTy()) &&
// Do not allow turning this into a load of an integer, which is then
// casted to a pointer, this pessimizes pointer analysis a lot.
(SrcPTy->isPointerTy() == LI.getType()->isPointerTy()) &&
- IC.getTargetData()->getTypeSizeInBits(SrcPTy) ==
- IC.getTargetData()->getTypeSizeInBits(DestPTy)) {
+ IC.getDataLayout()->getTypeSizeInBits(SrcPTy) ==
+ IC.getDataLayout()->getTypeSizeInBits(DestPTy)) {
// Okay, we are casting from one integer or pointer type to another of
// the same size. Instead of casting the pointer before the load, cast
// the result of the loaded value.
- LoadInst *NewLoad =
+ LoadInst *NewLoad =
IC.Builder->CreateLoad(CastOp, LI.isVolatile(), CI->getName());
NewLoad->setAlignment(LI.getAlignment());
NewLoad->setAtomic(LI.getOrdering(), LI.getSynchScope());
// 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,
// separated by a few arithmetic operations.
Constant::getNullValue(Op->getType()), &LI);
return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
}
- }
+ }
// load null/undef -> unreachable
// TODO: Consider a target hook for valid address spaces for this xform.
if (CE->isCast())
if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
return Res;
-
+
if (Op->hasOneUse()) {
// Change select and PHI nodes to select values instead of addresses: this
// helps alias analysis out a lot, allows many others simplifications, and
Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
if (SrcTy == 0) return 0;
-
+
Type *SrcPTy = SrcTy->getElementType();
if (!DestPTy->isIntegerTy() && !DestPTy->isPointerTy())
return 0;
-
+
/// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
/// to its first element. This allows us to handle things like:
/// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
/// on 32-bit hosts.
SmallVector<Value*, 4> NewGEPIndices;
-
+
// 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.
// Index through pointer.
Constant *Zero = Constant::getNullValue(Type::getInt32Ty(SI.getContext()));
NewGEPIndices.push_back(Zero);
-
+
while (1) {
if (StructType *STy = dyn_cast<StructType>(SrcPTy)) {
if (!STy->getNumElements()) /* Struct can be empty {} */
break;
}
}
-
+
SrcTy = PointerType::get(SrcPTy, SrcTy->getAddressSpace());
}
if (!SrcPTy->isIntegerTy() && !SrcPTy->isPointerTy())
return 0;
-
+
// If the pointers point into different address spaces or if they point to
// values with different sizes, we can't do the transformation.
- if (!IC.getTargetData() ||
- SrcTy->getAddressSpace() !=
+ if (!IC.getDataLayout() ||
+ SrcTy->getAddressSpace() !=
cast<PointerType>(CI->getType())->getAddressSpace() ||
- IC.getTargetData()->getTypeSizeInBits(SrcPTy) !=
- IC.getTargetData()->getTypeSizeInBits(DestPTy))
+ IC.getDataLayout()->getTypeSizeInBits(SrcPTy) !=
+ IC.getDataLayout()->getTypeSizeInBits(DestPTy))
return 0;
// Okay, we are casting from one integer or pointer type to another of
- // the same size. Instead of casting the pointer before
+ // the same size. Instead of casting the pointer before
// the store, cast the value to be stored.
Value *NewCast;
Value *SIOp0 = SI.getOperand(0);
if (SIOp0->getType()->isPointerTy())
opcode = Instruction::PtrToInt;
}
-
+
// 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);
-
+
NewCast = IC.Builder->CreateCast(opcode, SIOp0, CastDstTy,
SIOp0->getName()+".c");
SI.setOperand(0, NewCast);
static bool equivalentAddressValues(Value *A, Value *B) {
// Test if the values are trivially equivalent.
if (A == B) return true;
-
+
// Test if the values come form identical arithmetic instructions.
// This uses isIdenticalToWhenDefined instead of isIdenticalTo because
// its only used to compare two uses within the same basic block, which
if (Instruction *BI = dyn_cast<Instruction>(B))
if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
return true;
-
+
// Otherwise they may not be equivalent.
return false;
}
// If the RHS is an alloca with a single use, zapify the store, making the
// alloca dead.
if (Ptr->hasOneUse()) {
- if (isa<AllocaInst>(Ptr))
+ if (isa<AllocaInst>(Ptr))
return EraseInstFromFunction(SI);
if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
if (isa<AllocaInst>(GEP->getOperand(0))) {
(isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
ScanInsts++;
continue;
- }
-
+ }
+
if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
// Prev store isn't volatile, and stores to the same location?
if (PrevSI->isSimple() && equivalentAddressValues(PrevSI->getOperand(1),
}
break;
}
-
+
// If this is a load, we have to stop. However, if the loaded value is from
// the pointer we're loading and is producing the pointer we're storing,
// then *this* store is dead (X = load P; store X -> P).
if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
LI->isSimple())
return EraseInstFromFunction(SI);
-
+
// Otherwise, this is a load from some other location. Stores before it
// may not be dead.
break;
}
-
+
// Don't skip over loads or things that can modify memory.
if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
break;
if (Instruction *Res = InstCombineStoreToCast(*this, SI))
return Res;
-
+
// If this store is the last instruction in the basic block (possibly
// excepting debug info instructions), and if the block ends with an
// unconditional branch, try to move it to the successor block.
- BBI = &SI;
+ BBI = &SI;
do {
++BBI;
} while (isa<DbgInfoIntrinsic>(BBI) ||
if (BI->isUnconditional())
if (SimplifyStoreAtEndOfBlock(SI))
return 0; // xform done!
-
+
return 0;
}
///
bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
BasicBlock *StoreBB = SI.getParent();
-
+
// Check to see if the successor block has exactly two incoming edges. If
// so, see if the other predecessor contains a store to the same location.
// if so, insert a PHI node (if needed) and move the stores down.
BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
-
+
// Determine whether Dest has exactly two predecessors and, if so, compute
// the other predecessor.
pred_iterator PI = pred_begin(DestBB);
if (++PI == pred_end(DestBB))
return false;
-
+
P = *PI;
if (P != StoreBB) {
if (OtherBB)
BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
if (!OtherBr || BBI == OtherBB->begin())
return false;
-
+
// If the other block ends in an unconditional branch, check for the 'if then
// else' case. there is an instruction before the branch.
StoreInst *OtherStore = 0;
} else {
// Otherwise, the other block ended with a conditional branch. If one of the
// destinations is StoreBB, then we have the if/then case.
- if (OtherBr->getSuccessor(0) != StoreBB &&
+ if (OtherBr->getSuccessor(0) != StoreBB &&
OtherBr->getSuccessor(1) != StoreBB)
return false;
-
+
// Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
// if/then triangle. See if there is a store to the same ptr as SI that
// lives in OtherBB.
BBI == OtherBB->begin())
return false;
}
-
+
// In order to eliminate the store in OtherBr, we have to
// make sure nothing reads or overwrites the stored value in
// StoreBB.
return false;
}
}
-
+
// Insert a PHI node now if we need it.
Value *MergedVal = OtherStore->getOperand(0);
if (MergedVal != SI.getOperand(0)) {
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->getFirstInsertionPt();
SI.getOrdering(),
SI.getSynchScope());
InsertNewInstBefore(NewSI, *BBI);
- NewSI->setDebugLoc(OtherStore->getDebugLoc());
+ NewSI->setDebugLoc(OtherStore->getDebugLoc());
+
+ // If the two stores had the same TBAA tag, preserve it.
+ if (MDNode *TBAATag = SI.getMetadata(LLVMContext::MD_tbaa))
+ if ((TBAATag = MDNode::getMostGenericTBAA(TBAATag,
+ OtherStore->getMetadata(LLVMContext::MD_tbaa))))
+ NewSI->setMetadata(LLVMContext::MD_tbaa, TBAATag);
+
// Nuke the old stores.
EraseInstFromFunction(SI);
projects / oota-llvm.git / blobdiff