1 //===- InstCombineLoadStoreAlloca.cpp -------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the visit functions for load, store and alloca.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombine.h"
15 #include "llvm/IntrinsicInst.h"
16 #include "llvm/Analysis/Loads.h"
17 #include "llvm/DataLayout.h"
18 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
19 #include "llvm/Transforms/Utils/Local.h"
20 #include "llvm/ADT/Statistic.h"
23 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
24 STATISTIC(NumGlobalCopies, "Number of allocas copied from constant global");
26 /// pointsToConstantGlobal - Return true if V (possibly indirectly) points to
27 /// some part of a constant global variable. This intentionally only accepts
28 /// constant expressions because we can't rewrite arbitrary instructions.
29 static bool pointsToConstantGlobal(Value *V) {
30 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
31 return GV->isConstant();
32 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
33 if (CE->getOpcode() == Instruction::BitCast ||
34 CE->getOpcode() == Instruction::GetElementPtr)
35 return pointsToConstantGlobal(CE->getOperand(0));
39 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
40 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
41 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
42 /// track of whether it moves the pointer (with IsOffset) but otherwise traverse
43 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
44 /// the alloca, and if the source pointer is a pointer to a constant global, we
45 /// can optimize this.
47 isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
48 SmallVectorImpl<Instruction *> &ToDelete,
49 bool IsOffset = false) {
50 // We track lifetime intrinsics as we encounter them. If we decide to go
51 // ahead and replace the value with the global, this lets the caller quickly
52 // eliminate the markers.
54 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
55 User *U = cast<Instruction>(*UI);
57 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
58 // Ignore non-volatile loads, they are always ok.
59 if (!LI->isSimple()) return false;
63 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
64 // If uses of the bitcast are ok, we are ok.
65 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, ToDelete, IsOffset))
69 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
70 // If the GEP has all zero indices, it doesn't offset the pointer. If it
72 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, ToDelete,
73 IsOffset || !GEP->hasAllZeroIndices()))
78 if (CallSite CS = U) {
79 // If this is the function being called then we treat it like a load and
84 // If this is a readonly/readnone call site, then we know it is just a
85 // load (but one that potentially returns the value itself), so we can
86 // ignore it if we know that the value isn't captured.
87 unsigned ArgNo = CS.getArgumentNo(UI);
88 if (CS.onlyReadsMemory() &&
89 (CS.getInstruction()->use_empty() || CS.doesNotCapture(ArgNo)))
92 // If this is being passed as a byval argument, the caller is making a
93 // copy, so it is only a read of the alloca.
94 if (CS.isByValArgument(ArgNo))
98 // Lifetime intrinsics can be handled by the caller.
99 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
100 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
101 II->getIntrinsicID() == Intrinsic::lifetime_end) {
102 assert(II->use_empty() && "Lifetime markers have no result to use!");
103 ToDelete.push_back(II);
108 // If this is isn't our memcpy/memmove, reject it as something we can't
110 MemTransferInst *MI = dyn_cast<MemTransferInst>(U);
114 // If the transfer is using the alloca as a source of the transfer, then
115 // ignore it since it is a load (unless the transfer is volatile).
116 if (UI.getOperandNo() == 1) {
117 if (MI->isVolatile()) return false;
121 // If we already have seen a copy, reject the second one.
122 if (TheCopy) return false;
124 // If the pointer has been offset from the start of the alloca, we can't
125 // safely handle this.
126 if (IsOffset) return false;
128 // If the memintrinsic isn't using the alloca as the dest, reject it.
129 if (UI.getOperandNo() != 0) return false;
131 // If the source of the memcpy/move is not a constant global, reject it.
132 if (!pointsToConstantGlobal(MI->getSource()))
135 // Otherwise, the transform is safe. Remember the copy instruction.
141 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
142 /// modified by a copy from a constant global. If we can prove this, we can
143 /// replace any uses of the alloca with uses of the global directly.
144 static MemTransferInst *
145 isOnlyCopiedFromConstantGlobal(AllocaInst *AI,
146 SmallVectorImpl<Instruction *> &ToDelete) {
147 MemTransferInst *TheCopy = 0;
148 if (isOnlyCopiedFromConstantGlobal(AI, TheCopy, ToDelete))
153 /// getPointeeAlignment - Compute the minimum alignment of the value pointed
154 /// to by the given pointer.
155 static unsigned getPointeeAlignment(Value *V, const DataLayout &TD) {
156 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
157 if (CE->getOpcode() == Instruction::BitCast ||
158 (CE->getOpcode() == Instruction::GetElementPtr &&
159 cast<GEPOperator>(CE)->hasAllZeroIndices()))
160 return getPointeeAlignment(CE->getOperand(0), TD);
162 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
163 if (!GV->isDeclaration())
164 return TD.getPreferredAlignment(GV);
166 if (PointerType *PT = dyn_cast<PointerType>(V->getType()))
167 if (PT->getElementType()->isSized())
168 return TD.getABITypeAlignment(PT->getElementType());
173 Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) {
174 // Ensure that the alloca array size argument has type intptr_t, so that
175 // any casting is exposed early.
177 Type *IntPtrTy = TD->getIntPtrType(AI.getContext());
178 if (AI.getArraySize()->getType() != IntPtrTy) {
179 Value *V = Builder->CreateIntCast(AI.getArraySize(),
186 // Convert: alloca Ty, C - where C is a constant != 1 into: alloca [C x Ty], 1
187 if (AI.isArrayAllocation()) { // Check C != 1
188 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
190 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
191 AllocaInst *New = Builder->CreateAlloca(NewTy, 0, AI.getName());
192 New->setAlignment(AI.getAlignment());
194 // Scan to the end of the allocation instructions, to skip over a block of
195 // allocas if possible...also skip interleaved debug info
197 BasicBlock::iterator It = New;
198 while (isa<AllocaInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
200 // Now that I is pointing to the first non-allocation-inst in the block,
201 // insert our getelementptr instruction...
203 Value *NullIdx =Constant::getNullValue(Type::getInt32Ty(AI.getContext()));
208 GetElementPtrInst::CreateInBounds(New, Idx, New->getName()+".sub");
209 InsertNewInstBefore(GEP, *It);
211 // Now make everything use the getelementptr instead of the original
213 return ReplaceInstUsesWith(AI, GEP);
214 } else if (isa<UndefValue>(AI.getArraySize())) {
215 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
219 if (TD && AI.getAllocatedType()->isSized()) {
220 // If the alignment is 0 (unspecified), assign it the preferred alignment.
221 if (AI.getAlignment() == 0)
222 AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
224 // Move all alloca's of zero byte objects to the entry block and merge them
225 // together. Note that we only do this for alloca's, because malloc should
226 // allocate and return a unique pointer, even for a zero byte allocation.
227 if (TD->getTypeAllocSize(AI.getAllocatedType()) == 0) {
228 // For a zero sized alloca there is no point in doing an array allocation.
229 // This is helpful if the array size is a complicated expression not used
231 if (AI.isArrayAllocation()) {
232 AI.setOperand(0, ConstantInt::get(AI.getArraySize()->getType(), 1));
236 // Get the first instruction in the entry block.
237 BasicBlock &EntryBlock = AI.getParent()->getParent()->getEntryBlock();
238 Instruction *FirstInst = EntryBlock.getFirstNonPHIOrDbg();
239 if (FirstInst != &AI) {
240 // If the entry block doesn't start with a zero-size alloca then move
241 // this one to the start of the entry block. There is no problem with
242 // dominance as the array size was forced to a constant earlier already.
243 AllocaInst *EntryAI = dyn_cast<AllocaInst>(FirstInst);
244 if (!EntryAI || !EntryAI->getAllocatedType()->isSized() ||
245 TD->getTypeAllocSize(EntryAI->getAllocatedType()) != 0) {
246 AI.moveBefore(FirstInst);
250 // If the alignment of the entry block alloca is 0 (unspecified),
251 // assign it the preferred alignment.
252 if (EntryAI->getAlignment() == 0)
253 EntryAI->setAlignment(
254 TD->getPrefTypeAlignment(EntryAI->getAllocatedType()));
255 // Replace this zero-sized alloca with the one at the start of the entry
256 // block after ensuring that the address will be aligned enough for both
258 unsigned MaxAlign = std::max(EntryAI->getAlignment(),
260 EntryAI->setAlignment(MaxAlign);
261 if (AI.getType() != EntryAI->getType())
262 return new BitCastInst(EntryAI, AI.getType());
263 return ReplaceInstUsesWith(AI, EntryAI);
269 // Check to see if this allocation is only modified by a memcpy/memmove from
270 // a constant global whose alignment is equal to or exceeds that of the
271 // allocation. If this is the case, we can change all users to use
272 // the constant global instead. This is commonly produced by the CFE by
273 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
274 // is only subsequently read.
275 SmallVector<Instruction *, 4> ToDelete;
276 if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(&AI, ToDelete)) {
277 if (AI.getAlignment() <= getPointeeAlignment(Copy->getSource(), *TD)) {
278 DEBUG(dbgs() << "Found alloca equal to global: " << AI << '\n');
279 DEBUG(dbgs() << " memcpy = " << *Copy << '\n');
280 for (unsigned i = 0, e = ToDelete.size(); i != e; ++i)
281 EraseInstFromFunction(*ToDelete[i]);
282 Constant *TheSrc = cast<Constant>(Copy->getSource());
284 = ReplaceInstUsesWith(AI, ConstantExpr::getBitCast(TheSrc,
286 EraseInstFromFunction(*Copy);
293 // At last, use the generic allocation site handler to aggressively remove
295 return visitAllocSite(AI);
299 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
300 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
301 const DataLayout *TD) {
302 User *CI = cast<User>(LI.getOperand(0));
303 Value *CastOp = CI->getOperand(0);
305 PointerType *DestTy = cast<PointerType>(CI->getType());
306 Type *DestPTy = DestTy->getElementType();
307 if (PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
309 // If the address spaces don't match, don't eliminate the cast.
310 if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
313 Type *SrcPTy = SrcTy->getElementType();
315 if (DestPTy->isIntegerTy() || DestPTy->isPointerTy() ||
316 DestPTy->isVectorTy()) {
317 // If the source is an array, the code below will not succeed. Check to
318 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
320 if (ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
321 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
322 if (ASrcTy->getNumElements() != 0) {
324 Idxs[0] = Constant::getNullValue(Type::getInt32Ty(LI.getContext()));
326 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
327 SrcTy = cast<PointerType>(CastOp->getType());
328 SrcPTy = SrcTy->getElementType();
331 if (IC.getDataLayout() &&
332 (SrcPTy->isIntegerTy() || SrcPTy->isPointerTy() ||
333 SrcPTy->isVectorTy()) &&
334 // Do not allow turning this into a load of an integer, which is then
335 // casted to a pointer, this pessimizes pointer analysis a lot.
336 (SrcPTy->isPointerTy() == LI.getType()->isPointerTy()) &&
337 IC.getDataLayout()->getTypeSizeInBits(SrcPTy) ==
338 IC.getDataLayout()->getTypeSizeInBits(DestPTy)) {
340 // Okay, we are casting from one integer or pointer type to another of
341 // the same size. Instead of casting the pointer before the load, cast
342 // the result of the loaded value.
344 IC.Builder->CreateLoad(CastOp, LI.isVolatile(), CI->getName());
345 NewLoad->setAlignment(LI.getAlignment());
346 NewLoad->setAtomic(LI.getOrdering(), LI.getSynchScope());
347 // Now cast the result of the load.
348 return new BitCastInst(NewLoad, LI.getType());
355 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
356 Value *Op = LI.getOperand(0);
358 // Attempt to improve the alignment.
360 unsigned KnownAlign =
361 getOrEnforceKnownAlignment(Op, TD->getPrefTypeAlignment(LI.getType()),TD);
362 unsigned LoadAlign = LI.getAlignment();
363 unsigned EffectiveLoadAlign = LoadAlign != 0 ? LoadAlign :
364 TD->getABITypeAlignment(LI.getType());
366 if (KnownAlign > EffectiveLoadAlign)
367 LI.setAlignment(KnownAlign);
368 else if (LoadAlign == 0)
369 LI.setAlignment(EffectiveLoadAlign);
372 // load (cast X) --> cast (load X) iff safe.
373 if (isa<CastInst>(Op))
374 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
377 // None of the following transforms are legal for volatile/atomic loads.
378 // FIXME: Some of it is okay for atomic loads; needs refactoring.
379 if (!LI.isSimple()) return 0;
381 // Do really simple store-to-load forwarding and load CSE, to catch cases
382 // where there are several consecutive memory accesses to the same location,
383 // separated by a few arithmetic operations.
384 BasicBlock::iterator BBI = &LI;
385 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
386 return ReplaceInstUsesWith(LI, AvailableVal);
388 // load(gep null, ...) -> unreachable
389 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
390 const Value *GEPI0 = GEPI->getOperand(0);
391 // TODO: Consider a target hook for valid address spaces for this xform.
392 if (isa<ConstantPointerNull>(GEPI0) && GEPI->getPointerAddressSpace() == 0){
393 // Insert a new store to null instruction before the load to indicate
394 // that this code is not reachable. We do this instead of inserting
395 // an unreachable instruction directly because we cannot modify the
397 new StoreInst(UndefValue::get(LI.getType()),
398 Constant::getNullValue(Op->getType()), &LI);
399 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
403 // load null/undef -> unreachable
404 // TODO: Consider a target hook for valid address spaces for this xform.
405 if (isa<UndefValue>(Op) ||
406 (isa<ConstantPointerNull>(Op) && LI.getPointerAddressSpace() == 0)) {
407 // Insert a new store to null instruction before the load to indicate that
408 // this code is not reachable. We do this instead of inserting an
409 // unreachable instruction directly because we cannot modify the CFG.
410 new StoreInst(UndefValue::get(LI.getType()),
411 Constant::getNullValue(Op->getType()), &LI);
412 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
415 // Instcombine load (constantexpr_cast global) -> cast (load global)
416 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
418 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
421 if (Op->hasOneUse()) {
422 // Change select and PHI nodes to select values instead of addresses: this
423 // helps alias analysis out a lot, allows many others simplifications, and
424 // exposes redundancy in the code.
426 // Note that we cannot do the transformation unless we know that the
427 // introduced loads cannot trap! Something like this is valid as long as
428 // the condition is always false: load (select bool %C, int* null, int* %G),
429 // but it would not be valid if we transformed it to load from null
432 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
433 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
434 unsigned Align = LI.getAlignment();
435 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI, Align, TD) &&
436 isSafeToLoadUnconditionally(SI->getOperand(2), SI, Align, TD)) {
437 LoadInst *V1 = Builder->CreateLoad(SI->getOperand(1),
438 SI->getOperand(1)->getName()+".val");
439 LoadInst *V2 = Builder->CreateLoad(SI->getOperand(2),
440 SI->getOperand(2)->getName()+".val");
441 V1->setAlignment(Align);
442 V2->setAlignment(Align);
443 return SelectInst::Create(SI->getCondition(), V1, V2);
446 // load (select (cond, null, P)) -> load P
447 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
448 if (C->isNullValue()) {
449 LI.setOperand(0, SI->getOperand(2));
453 // load (select (cond, P, null)) -> load P
454 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
455 if (C->isNullValue()) {
456 LI.setOperand(0, SI->getOperand(1));
464 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
465 /// when possible. This makes it generally easy to do alias analysis and/or
466 /// SROA/mem2reg of the memory object.
467 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
468 User *CI = cast<User>(SI.getOperand(1));
469 Value *CastOp = CI->getOperand(0);
471 Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
472 PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
473 if (SrcTy == 0) return 0;
475 Type *SrcPTy = SrcTy->getElementType();
477 if (!DestPTy->isIntegerTy() && !DestPTy->isPointerTy())
480 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
481 /// to its first element. This allows us to handle things like:
482 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
484 SmallVector<Value*, 4> NewGEPIndices;
486 // If the source is an array, the code below will not succeed. Check to
487 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
489 if (SrcPTy->isArrayTy() || SrcPTy->isStructTy()) {
490 // Index through pointer.
491 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(SI.getContext()));
492 NewGEPIndices.push_back(Zero);
495 if (StructType *STy = dyn_cast<StructType>(SrcPTy)) {
496 if (!STy->getNumElements()) /* Struct can be empty {} */
498 NewGEPIndices.push_back(Zero);
499 SrcPTy = STy->getElementType(0);
500 } else if (ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
501 NewGEPIndices.push_back(Zero);
502 SrcPTy = ATy->getElementType();
508 SrcTy = PointerType::get(SrcPTy, SrcTy->getAddressSpace());
511 if (!SrcPTy->isIntegerTy() && !SrcPTy->isPointerTy())
514 // If the pointers point into different address spaces or if they point to
515 // values with different sizes, we can't do the transformation.
516 if (!IC.getDataLayout() ||
517 SrcTy->getAddressSpace() !=
518 cast<PointerType>(CI->getType())->getAddressSpace() ||
519 IC.getDataLayout()->getTypeSizeInBits(SrcPTy) !=
520 IC.getDataLayout()->getTypeSizeInBits(DestPTy))
523 // Okay, we are casting from one integer or pointer type to another of
524 // the same size. Instead of casting the pointer before
525 // the store, cast the value to be stored.
527 Value *SIOp0 = SI.getOperand(0);
528 Instruction::CastOps opcode = Instruction::BitCast;
529 Type* CastSrcTy = SIOp0->getType();
530 Type* CastDstTy = SrcPTy;
531 if (CastDstTy->isPointerTy()) {
532 if (CastSrcTy->isIntegerTy())
533 opcode = Instruction::IntToPtr;
534 } else if (CastDstTy->isIntegerTy()) {
535 if (SIOp0->getType()->isPointerTy())
536 opcode = Instruction::PtrToInt;
539 // SIOp0 is a pointer to aggregate and this is a store to the first field,
540 // emit a GEP to index into its first field.
541 if (!NewGEPIndices.empty())
542 CastOp = IC.Builder->CreateInBoundsGEP(CastOp, NewGEPIndices);
544 NewCast = IC.Builder->CreateCast(opcode, SIOp0, CastDstTy,
545 SIOp0->getName()+".c");
546 SI.setOperand(0, NewCast);
547 SI.setOperand(1, CastOp);
551 /// equivalentAddressValues - Test if A and B will obviously have the same
552 /// value. This includes recognizing that %t0 and %t1 will have the same
553 /// value in code like this:
554 /// %t0 = getelementptr \@a, 0, 3
555 /// store i32 0, i32* %t0
556 /// %t1 = getelementptr \@a, 0, 3
557 /// %t2 = load i32* %t1
559 static bool equivalentAddressValues(Value *A, Value *B) {
560 // Test if the values are trivially equivalent.
561 if (A == B) return true;
563 // Test if the values come form identical arithmetic instructions.
564 // This uses isIdenticalToWhenDefined instead of isIdenticalTo because
565 // its only used to compare two uses within the same basic block, which
566 // means that they'll always either have the same value or one of them
567 // will have an undefined value.
568 if (isa<BinaryOperator>(A) ||
571 isa<GetElementPtrInst>(A))
572 if (Instruction *BI = dyn_cast<Instruction>(B))
573 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
576 // Otherwise they may not be equivalent.
580 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
581 Value *Val = SI.getOperand(0);
582 Value *Ptr = SI.getOperand(1);
584 // Attempt to improve the alignment.
586 unsigned KnownAlign =
587 getOrEnforceKnownAlignment(Ptr, TD->getPrefTypeAlignment(Val->getType()),
589 unsigned StoreAlign = SI.getAlignment();
590 unsigned EffectiveStoreAlign = StoreAlign != 0 ? StoreAlign :
591 TD->getABITypeAlignment(Val->getType());
593 if (KnownAlign > EffectiveStoreAlign)
594 SI.setAlignment(KnownAlign);
595 else if (StoreAlign == 0)
596 SI.setAlignment(EffectiveStoreAlign);
599 // Don't hack volatile/atomic stores.
600 // FIXME: Some bits are legal for atomic stores; needs refactoring.
601 if (!SI.isSimple()) return 0;
603 // If the RHS is an alloca with a single use, zapify the store, making the
605 if (Ptr->hasOneUse()) {
606 if (isa<AllocaInst>(Ptr))
607 return EraseInstFromFunction(SI);
608 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
609 if (isa<AllocaInst>(GEP->getOperand(0))) {
610 if (GEP->getOperand(0)->hasOneUse())
611 return EraseInstFromFunction(SI);
616 // Do really simple DSE, to catch cases where there are several consecutive
617 // stores to the same location, separated by a few arithmetic operations. This
618 // situation often occurs with bitfield accesses.
619 BasicBlock::iterator BBI = &SI;
620 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
623 // Don't count debug info directives, lest they affect codegen,
624 // and we skip pointer-to-pointer bitcasts, which are NOPs.
625 if (isa<DbgInfoIntrinsic>(BBI) ||
626 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
631 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
632 // Prev store isn't volatile, and stores to the same location?
633 if (PrevSI->isSimple() && equivalentAddressValues(PrevSI->getOperand(1),
637 EraseInstFromFunction(*PrevSI);
643 // If this is a load, we have to stop. However, if the loaded value is from
644 // the pointer we're loading and is producing the pointer we're storing,
645 // then *this* store is dead (X = load P; store X -> P).
646 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
647 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
649 return EraseInstFromFunction(SI);
651 // Otherwise, this is a load from some other location. Stores before it
656 // Don't skip over loads or things that can modify memory.
657 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
661 // store X, null -> turns into 'unreachable' in SimplifyCFG
662 if (isa<ConstantPointerNull>(Ptr) && SI.getPointerAddressSpace() == 0) {
663 if (!isa<UndefValue>(Val)) {
664 SI.setOperand(0, UndefValue::get(Val->getType()));
665 if (Instruction *U = dyn_cast<Instruction>(Val))
666 Worklist.Add(U); // Dropped a use.
668 return 0; // Do not modify these!
671 // store undef, Ptr -> noop
672 if (isa<UndefValue>(Val))
673 return EraseInstFromFunction(SI);
675 // If the pointer destination is a cast, see if we can fold the cast into the
677 if (isa<CastInst>(Ptr))
678 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
680 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
682 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
686 // If this store is the last instruction in the basic block (possibly
687 // excepting debug info instructions), and if the block ends with an
688 // unconditional branch, try to move it to the successor block.
692 } while (isa<DbgInfoIntrinsic>(BBI) ||
693 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy()));
694 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
695 if (BI->isUnconditional())
696 if (SimplifyStoreAtEndOfBlock(SI))
697 return 0; // xform done!
702 /// SimplifyStoreAtEndOfBlock - Turn things like:
703 /// if () { *P = v1; } else { *P = v2 }
704 /// into a phi node with a store in the successor.
706 /// Simplify things like:
707 /// *P = v1; if () { *P = v2; }
708 /// into a phi node with a store in the successor.
710 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
711 BasicBlock *StoreBB = SI.getParent();
713 // Check to see if the successor block has exactly two incoming edges. If
714 // so, see if the other predecessor contains a store to the same location.
715 // if so, insert a PHI node (if needed) and move the stores down.
716 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
718 // Determine whether Dest has exactly two predecessors and, if so, compute
719 // the other predecessor.
720 pred_iterator PI = pred_begin(DestBB);
722 BasicBlock *OtherBB = 0;
727 if (++PI == pred_end(DestBB))
736 if (++PI != pred_end(DestBB))
739 // Bail out if all the relevant blocks aren't distinct (this can happen,
740 // for example, if SI is in an infinite loop)
741 if (StoreBB == DestBB || OtherBB == DestBB)
744 // Verify that the other block ends in a branch and is not otherwise empty.
745 BasicBlock::iterator BBI = OtherBB->getTerminator();
746 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
747 if (!OtherBr || BBI == OtherBB->begin())
750 // If the other block ends in an unconditional branch, check for the 'if then
751 // else' case. there is an instruction before the branch.
752 StoreInst *OtherStore = 0;
753 if (OtherBr->isUnconditional()) {
755 // Skip over debugging info.
756 while (isa<DbgInfoIntrinsic>(BBI) ||
757 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
758 if (BBI==OtherBB->begin())
762 // If this isn't a store, isn't a store to the same location, or is not the
763 // right kind of store, bail out.
764 OtherStore = dyn_cast<StoreInst>(BBI);
765 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1) ||
766 !SI.isSameOperationAs(OtherStore))
769 // Otherwise, the other block ended with a conditional branch. If one of the
770 // destinations is StoreBB, then we have the if/then case.
771 if (OtherBr->getSuccessor(0) != StoreBB &&
772 OtherBr->getSuccessor(1) != StoreBB)
775 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
776 // if/then triangle. See if there is a store to the same ptr as SI that
779 // Check to see if we find the matching store.
780 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
781 if (OtherStore->getOperand(1) != SI.getOperand(1) ||
782 !SI.isSameOperationAs(OtherStore))
786 // If we find something that may be using or overwriting the stored
787 // value, or if we run out of instructions, we can't do the xform.
788 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
789 BBI == OtherBB->begin())
793 // In order to eliminate the store in OtherBr, we have to
794 // make sure nothing reads or overwrites the stored value in
796 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
797 // FIXME: This should really be AA driven.
798 if (I->mayReadFromMemory() || I->mayWriteToMemory())
803 // Insert a PHI node now if we need it.
804 Value *MergedVal = OtherStore->getOperand(0);
805 if (MergedVal != SI.getOperand(0)) {
806 PHINode *PN = PHINode::Create(MergedVal->getType(), 2, "storemerge");
807 PN->addIncoming(SI.getOperand(0), SI.getParent());
808 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
809 MergedVal = InsertNewInstBefore(PN, DestBB->front());
812 // Advance to a place where it is safe to insert the new store and
814 BBI = DestBB->getFirstInsertionPt();
815 StoreInst *NewSI = new StoreInst(MergedVal, SI.getOperand(1),
820 InsertNewInstBefore(NewSI, *BBI);
821 NewSI->setDebugLoc(OtherStore->getDebugLoc());
823 // Nuke the old stores.
824 EraseInstFromFunction(SI);
825 EraseInstFromFunction(*OtherStore);