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/Target/TargetData.h"
17 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
18 #include "llvm/Transforms/Utils/Local.h"
19 #include "llvm/ADT/Statistic.h"
22 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
24 Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) {
25 // Convert: alloca Ty, C - where C is a constant != 1 into: alloca [C x Ty], 1
26 if (AI.isArrayAllocation()) { // Check C != 1
27 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
29 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
30 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
31 AllocaInst *New = Builder->CreateAlloca(NewTy, 0, AI.getName());
32 New->setAlignment(AI.getAlignment());
34 // Scan to the end of the allocation instructions, to skip over a block of
35 // allocas if possible...also skip interleaved debug info
37 BasicBlock::iterator It = New;
38 while (isa<AllocaInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
40 // Now that I is pointing to the first non-allocation-inst in the block,
41 // insert our getelementptr instruction...
43 Value *NullIdx =Constant::getNullValue(Type::getInt32Ty(AI.getContext()));
47 Value *V = GetElementPtrInst::CreateInBounds(New, Idx, Idx + 2,
48 New->getName()+".sub", It);
50 // Now make everything use the getelementptr instead of the original
52 return ReplaceInstUsesWith(AI, V);
53 } else if (isa<UndefValue>(AI.getArraySize())) {
54 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
58 if (TD && isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized()) {
59 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
60 // Note that we only do this for alloca's, because malloc should allocate
61 // and return a unique pointer, even for a zero byte allocation.
62 if (TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
63 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
65 // If the alignment is 0 (unspecified), assign it the preferred alignment.
66 if (AI.getAlignment() == 0)
67 AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
74 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
75 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
76 const TargetData *TD) {
77 User *CI = cast<User>(LI.getOperand(0));
78 Value *CastOp = CI->getOperand(0);
80 const PointerType *DestTy = cast<PointerType>(CI->getType());
81 const Type *DestPTy = DestTy->getElementType();
82 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
84 // If the address spaces don't match, don't eliminate the cast.
85 if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
88 const Type *SrcPTy = SrcTy->getElementType();
90 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
91 isa<VectorType>(DestPTy)) {
92 // If the source is an array, the code below will not succeed. Check to
93 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
95 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
96 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
97 if (ASrcTy->getNumElements() != 0) {
99 Idxs[0] = Constant::getNullValue(Type::getInt32Ty(LI.getContext()));
101 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
102 SrcTy = cast<PointerType>(CastOp->getType());
103 SrcPTy = SrcTy->getElementType();
106 if (IC.getTargetData() &&
107 (SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
108 isa<VectorType>(SrcPTy)) &&
109 // Do not allow turning this into a load of an integer, which is then
110 // casted to a pointer, this pessimizes pointer analysis a lot.
111 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
112 IC.getTargetData()->getTypeSizeInBits(SrcPTy) ==
113 IC.getTargetData()->getTypeSizeInBits(DestPTy)) {
115 // Okay, we are casting from one integer or pointer type to another of
116 // the same size. Instead of casting the pointer before the load, cast
117 // the result of the loaded value.
119 IC.Builder->CreateLoad(CastOp, LI.isVolatile(), CI->getName());
120 // Now cast the result of the load.
121 return new BitCastInst(NewLoad, LI.getType());
128 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
129 Value *Op = LI.getOperand(0);
131 // Attempt to improve the alignment.
133 unsigned KnownAlign =
134 GetOrEnforceKnownAlignment(Op, TD->getPrefTypeAlignment(LI.getType()));
136 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
138 LI.setAlignment(KnownAlign);
141 // load (cast X) --> cast (load X) iff safe.
142 if (isa<CastInst>(Op))
143 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
146 // None of the following transforms are legal for volatile loads.
147 if (LI.isVolatile()) return 0;
149 // Do really simple store-to-load forwarding and load CSE, to catch cases
150 // where there are several consequtive memory accesses to the same location,
151 // separated by a few arithmetic operations.
152 BasicBlock::iterator BBI = &LI;
153 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
154 return ReplaceInstUsesWith(LI, AvailableVal);
156 // load(gep null, ...) -> unreachable
157 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
158 const Value *GEPI0 = GEPI->getOperand(0);
159 // TODO: Consider a target hook for valid address spaces for this xform.
160 if (isa<ConstantPointerNull>(GEPI0) && GEPI->getPointerAddressSpace() == 0){
161 // Insert a new store to null instruction before the load to indicate
162 // that this code is not reachable. We do this instead of inserting
163 // an unreachable instruction directly because we cannot modify the
165 new StoreInst(UndefValue::get(LI.getType()),
166 Constant::getNullValue(Op->getType()), &LI);
167 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
171 // load null/undef -> unreachable
172 // TODO: Consider a target hook for valid address spaces for this xform.
173 if (isa<UndefValue>(Op) ||
174 (isa<ConstantPointerNull>(Op) && LI.getPointerAddressSpace() == 0)) {
175 // Insert a new store to null instruction before the load to indicate that
176 // this code is not reachable. We do this instead of inserting an
177 // unreachable instruction directly because we cannot modify the CFG.
178 new StoreInst(UndefValue::get(LI.getType()),
179 Constant::getNullValue(Op->getType()), &LI);
180 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
183 // Instcombine load (constantexpr_cast global) -> cast (load global)
184 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
186 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
189 if (Op->hasOneUse()) {
190 // Change select and PHI nodes to select values instead of addresses: this
191 // helps alias analysis out a lot, allows many others simplifications, and
192 // exposes redundancy in the code.
194 // Note that we cannot do the transformation unless we know that the
195 // introduced loads cannot trap! Something like this is valid as long as
196 // the condition is always false: load (select bool %C, int* null, int* %G),
197 // but it would not be valid if we transformed it to load from null
200 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
201 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
202 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
203 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
204 Value *V1 = Builder->CreateLoad(SI->getOperand(1),
205 SI->getOperand(1)->getName()+".val");
206 Value *V2 = Builder->CreateLoad(SI->getOperand(2),
207 SI->getOperand(2)->getName()+".val");
208 return SelectInst::Create(SI->getCondition(), V1, V2);
211 // load (select (cond, null, P)) -> load P
212 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
213 if (C->isNullValue()) {
214 LI.setOperand(0, SI->getOperand(2));
218 // load (select (cond, P, null)) -> load P
219 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
220 if (C->isNullValue()) {
221 LI.setOperand(0, SI->getOperand(1));
229 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
230 /// when possible. This makes it generally easy to do alias analysis and/or
231 /// SROA/mem2reg of the memory object.
232 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
233 User *CI = cast<User>(SI.getOperand(1));
234 Value *CastOp = CI->getOperand(0);
236 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
237 const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
238 if (SrcTy == 0) return 0;
240 const Type *SrcPTy = SrcTy->getElementType();
242 if (!DestPTy->isInteger() && !isa<PointerType>(DestPTy))
245 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
246 /// to its first element. This allows us to handle things like:
247 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
249 SmallVector<Value*, 4> NewGEPIndices;
251 // If the source is an array, the code below will not succeed. Check to
252 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
254 if (isa<ArrayType>(SrcPTy) || isa<StructType>(SrcPTy)) {
255 // Index through pointer.
256 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(SI.getContext()));
257 NewGEPIndices.push_back(Zero);
260 if (const StructType *STy = dyn_cast<StructType>(SrcPTy)) {
261 if (!STy->getNumElements()) /* Struct can be empty {} */
263 NewGEPIndices.push_back(Zero);
264 SrcPTy = STy->getElementType(0);
265 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
266 NewGEPIndices.push_back(Zero);
267 SrcPTy = ATy->getElementType();
273 SrcTy = PointerType::get(SrcPTy, SrcTy->getAddressSpace());
276 if (!SrcPTy->isInteger() && !isa<PointerType>(SrcPTy))
279 // If the pointers point into different address spaces or if they point to
280 // values with different sizes, we can't do the transformation.
281 if (!IC.getTargetData() ||
282 SrcTy->getAddressSpace() !=
283 cast<PointerType>(CI->getType())->getAddressSpace() ||
284 IC.getTargetData()->getTypeSizeInBits(SrcPTy) !=
285 IC.getTargetData()->getTypeSizeInBits(DestPTy))
288 // Okay, we are casting from one integer or pointer type to another of
289 // the same size. Instead of casting the pointer before
290 // the store, cast the value to be stored.
292 Value *SIOp0 = SI.getOperand(0);
293 Instruction::CastOps opcode = Instruction::BitCast;
294 const Type* CastSrcTy = SIOp0->getType();
295 const Type* CastDstTy = SrcPTy;
296 if (isa<PointerType>(CastDstTy)) {
297 if (CastSrcTy->isInteger())
298 opcode = Instruction::IntToPtr;
299 } else if (isa<IntegerType>(CastDstTy)) {
300 if (isa<PointerType>(SIOp0->getType()))
301 opcode = Instruction::PtrToInt;
304 // SIOp0 is a pointer to aggregate and this is a store to the first field,
305 // emit a GEP to index into its first field.
306 if (!NewGEPIndices.empty())
307 CastOp = IC.Builder->CreateInBoundsGEP(CastOp, NewGEPIndices.begin(),
308 NewGEPIndices.end());
310 NewCast = IC.Builder->CreateCast(opcode, SIOp0, CastDstTy,
311 SIOp0->getName()+".c");
312 return new StoreInst(NewCast, CastOp);
315 /// equivalentAddressValues - Test if A and B will obviously have the same
316 /// value. This includes recognizing that %t0 and %t1 will have the same
317 /// value in code like this:
318 /// %t0 = getelementptr \@a, 0, 3
319 /// store i32 0, i32* %t0
320 /// %t1 = getelementptr \@a, 0, 3
321 /// %t2 = load i32* %t1
323 static bool equivalentAddressValues(Value *A, Value *B) {
324 // Test if the values are trivially equivalent.
325 if (A == B) return true;
327 // Test if the values come form identical arithmetic instructions.
328 // This uses isIdenticalToWhenDefined instead of isIdenticalTo because
329 // its only used to compare two uses within the same basic block, which
330 // means that they'll always either have the same value or one of them
331 // will have an undefined value.
332 if (isa<BinaryOperator>(A) ||
335 isa<GetElementPtrInst>(A))
336 if (Instruction *BI = dyn_cast<Instruction>(B))
337 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
340 // Otherwise they may not be equivalent.
344 // If this instruction has two uses, one of which is a llvm.dbg.declare,
345 // return the llvm.dbg.declare.
346 DbgDeclareInst *InstCombiner::hasOneUsePlusDeclare(Value *V) {
349 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
351 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI))
353 if (isa<BitCastInst>(UI) && UI->hasOneUse()) {
354 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI->use_begin()))
361 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
362 Value *Val = SI.getOperand(0);
363 Value *Ptr = SI.getOperand(1);
365 // If the RHS is an alloca with a single use, zapify the store, making the
367 // If the RHS is an alloca with a two uses, the other one being a
368 // llvm.dbg.declare, zapify the store and the declare, making the
369 // alloca dead. We must do this to prevent declares from affecting
371 if (!SI.isVolatile()) {
372 if (Ptr->hasOneUse()) {
373 if (isa<AllocaInst>(Ptr))
374 return EraseInstFromFunction(SI);
375 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
376 if (isa<AllocaInst>(GEP->getOperand(0))) {
377 if (GEP->getOperand(0)->hasOneUse())
378 return EraseInstFromFunction(SI);
379 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(GEP->getOperand(0))) {
380 EraseInstFromFunction(*DI);
381 return EraseInstFromFunction(SI);
386 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(Ptr)) {
387 EraseInstFromFunction(*DI);
388 return EraseInstFromFunction(SI);
392 // Attempt to improve the alignment.
394 unsigned KnownAlign =
395 GetOrEnforceKnownAlignment(Ptr, TD->getPrefTypeAlignment(Val->getType()));
397 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
399 SI.setAlignment(KnownAlign);
402 // Do really simple DSE, to catch cases where there are several consecutive
403 // stores to the same location, separated by a few arithmetic operations. This
404 // situation often occurs with bitfield accesses.
405 BasicBlock::iterator BBI = &SI;
406 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
409 // Don't count debug info directives, lest they affect codegen,
410 // and we skip pointer-to-pointer bitcasts, which are NOPs.
411 // It is necessary for correctness to skip those that feed into a
412 // llvm.dbg.declare, as these are not present when debugging is off.
413 if (isa<DbgInfoIntrinsic>(BBI) ||
414 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
419 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
420 // Prev store isn't volatile, and stores to the same location?
421 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
425 EraseInstFromFunction(*PrevSI);
431 // If this is a load, we have to stop. However, if the loaded value is from
432 // the pointer we're loading and is producing the pointer we're storing,
433 // then *this* store is dead (X = load P; store X -> P).
434 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
435 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
437 return EraseInstFromFunction(SI);
439 // Otherwise, this is a load from some other location. Stores before it
444 // Don't skip over loads or things that can modify memory.
445 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
450 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
452 // store X, null -> turns into 'unreachable' in SimplifyCFG
453 if (isa<ConstantPointerNull>(Ptr) && SI.getPointerAddressSpace() == 0) {
454 if (!isa<UndefValue>(Val)) {
455 SI.setOperand(0, UndefValue::get(Val->getType()));
456 if (Instruction *U = dyn_cast<Instruction>(Val))
457 Worklist.Add(U); // Dropped a use.
459 return 0; // Do not modify these!
462 // store undef, Ptr -> noop
463 if (isa<UndefValue>(Val))
464 return EraseInstFromFunction(SI);
466 // If the pointer destination is a cast, see if we can fold the cast into the
468 if (isa<CastInst>(Ptr))
469 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
471 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
473 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
477 // If this store is the last instruction in the basic block (possibly
478 // excepting debug info instructions and the pointer bitcasts that feed
479 // into them), and if the block ends with an unconditional branch, try
480 // to move it to the successor block.
484 } while (isa<DbgInfoIntrinsic>(BBI) ||
485 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType())));
486 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
487 if (BI->isUnconditional())
488 if (SimplifyStoreAtEndOfBlock(SI))
489 return 0; // xform done!
494 /// SimplifyStoreAtEndOfBlock - Turn things like:
495 /// if () { *P = v1; } else { *P = v2 }
496 /// into a phi node with a store in the successor.
498 /// Simplify things like:
499 /// *P = v1; if () { *P = v2; }
500 /// into a phi node with a store in the successor.
502 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
503 BasicBlock *StoreBB = SI.getParent();
505 // Check to see if the successor block has exactly two incoming edges. If
506 // so, see if the other predecessor contains a store to the same location.
507 // if so, insert a PHI node (if needed) and move the stores down.
508 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
510 // Determine whether Dest has exactly two predecessors and, if so, compute
511 // the other predecessor.
512 pred_iterator PI = pred_begin(DestBB);
513 BasicBlock *OtherBB = 0;
517 if (PI == pred_end(DestBB))
520 if (*PI != StoreBB) {
525 if (++PI != pred_end(DestBB))
528 // Bail out if all the relevant blocks aren't distinct (this can happen,
529 // for example, if SI is in an infinite loop)
530 if (StoreBB == DestBB || OtherBB == DestBB)
533 // Verify that the other block ends in a branch and is not otherwise empty.
534 BasicBlock::iterator BBI = OtherBB->getTerminator();
535 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
536 if (!OtherBr || BBI == OtherBB->begin())
539 // If the other block ends in an unconditional branch, check for the 'if then
540 // else' case. there is an instruction before the branch.
541 StoreInst *OtherStore = 0;
542 if (OtherBr->isUnconditional()) {
544 // Skip over debugging info.
545 while (isa<DbgInfoIntrinsic>(BBI) ||
546 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
547 if (BBI==OtherBB->begin())
551 // If this isn't a store, isn't a store to the same location, or if the
552 // alignments differ, bail out.
553 OtherStore = dyn_cast<StoreInst>(BBI);
554 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1) ||
555 OtherStore->getAlignment() != SI.getAlignment())
558 // Otherwise, the other block ended with a conditional branch. If one of the
559 // destinations is StoreBB, then we have the if/then case.
560 if (OtherBr->getSuccessor(0) != StoreBB &&
561 OtherBr->getSuccessor(1) != StoreBB)
564 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
565 // if/then triangle. See if there is a store to the same ptr as SI that
568 // Check to see if we find the matching store.
569 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
570 if (OtherStore->getOperand(1) != SI.getOperand(1) ||
571 OtherStore->getAlignment() != SI.getAlignment())
575 // If we find something that may be using or overwriting the stored
576 // value, or if we run out of instructions, we can't do the xform.
577 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
578 BBI == OtherBB->begin())
582 // In order to eliminate the store in OtherBr, we have to
583 // make sure nothing reads or overwrites the stored value in
585 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
586 // FIXME: This should really be AA driven.
587 if (I->mayReadFromMemory() || I->mayWriteToMemory())
592 // Insert a PHI node now if we need it.
593 Value *MergedVal = OtherStore->getOperand(0);
594 if (MergedVal != SI.getOperand(0)) {
595 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
596 PN->reserveOperandSpace(2);
597 PN->addIncoming(SI.getOperand(0), SI.getParent());
598 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
599 MergedVal = InsertNewInstBefore(PN, DestBB->front());
602 // Advance to a place where it is safe to insert the new store and
604 BBI = DestBB->getFirstNonPHI();
605 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
606 OtherStore->isVolatile(),
607 SI.getAlignment()), *BBI);
609 // Nuke the old stores.
610 EraseInstFromFunction(SI);
611 EraseInstFromFunction(*OtherStore);