1 //===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===//
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 transformation implements the well known scalar replacement of
11 // aggregates transformation. This xform breaks up alloca instructions of
12 // aggregate type (structure or array) into individual alloca instructions for
13 // each member (if possible). Then, if possible, it transforms the individual
14 // alloca instructions into nice clean scalar SSA form.
16 // This combines a simple SRoA algorithm with the Mem2Reg algorithm because
17 // often interact, especially for C++ programs. As such, iterating between
18 // SRoA, then Mem2Reg until we run out of things to promote works well.
20 //===----------------------------------------------------------------------===//
22 #define DEBUG_TYPE "scalarrepl"
23 #include "llvm/Transforms/Scalar.h"
24 #include "llvm/Constants.h"
25 #include "llvm/DerivedTypes.h"
26 #include "llvm/Function.h"
27 #include "llvm/GlobalVariable.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/IntrinsicInst.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/Pass.h"
32 #include "llvm/Analysis/Dominators.h"
33 #include "llvm/Target/TargetData.h"
34 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
35 #include "llvm/Transforms/Utils/Local.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/ErrorHandling.h"
38 #include "llvm/Support/GetElementPtrTypeIterator.h"
39 #include "llvm/Support/IRBuilder.h"
40 #include "llvm/Support/MathExtras.h"
41 #include "llvm/Support/raw_ostream.h"
42 #include "llvm/ADT/SmallVector.h"
43 #include "llvm/ADT/Statistic.h"
46 STATISTIC(NumReplaced, "Number of allocas broken up");
47 STATISTIC(NumPromoted, "Number of allocas promoted");
48 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
49 STATISTIC(NumGlobals, "Number of allocas copied from constant global");
52 struct SROA : public FunctionPass {
53 static char ID; // Pass identification, replacement for typeid
54 explicit SROA(signed T = -1) : FunctionPass(&ID) {
61 bool runOnFunction(Function &F);
63 bool performScalarRepl(Function &F);
64 bool performPromotion(Function &F);
66 // getAnalysisUsage - This pass does not require any passes, but we know it
67 // will not alter the CFG, so say so.
68 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
69 AU.addRequired<DominatorTree>();
70 AU.addRequired<DominanceFrontier>();
77 /// DeadInsts - Keep track of instructions we have made dead, so that
78 /// we can remove them after we are done working.
79 SmallVector<Value*, 32> DeadInsts;
81 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
82 /// information about the uses. All these fields are initialized to false
83 /// and set to true when something is learned.
85 /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
88 /// needsCleanup - This is set to true if there is some use of the alloca
89 /// that requires cleanup.
90 bool needsCleanup : 1;
92 /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
95 /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
99 : isUnsafe(false), needsCleanup(false),
100 isMemCpySrc(false), isMemCpyDst(false) {}
103 unsigned SRThreshold;
105 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; }
107 int isSafeAllocaToScalarRepl(AllocaInst *AI);
109 void isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
111 void isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t &Offset,
113 void isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize,
114 const Type *MemOpType, bool isStore, AllocaInfo &Info);
115 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size);
116 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset,
119 void DoScalarReplacement(AllocaInst *AI,
120 std::vector<AllocaInst*> &WorkList);
121 void DeleteDeadInstructions();
122 void CleanupAllocaUsers(Value *V);
123 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocaInst *Base);
125 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
126 SmallVector<AllocaInst*, 32> &NewElts);
127 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
128 SmallVector<AllocaInst*, 32> &NewElts);
129 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
130 SmallVector<AllocaInst*, 32> &NewElts);
131 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
133 SmallVector<AllocaInst*, 32> &NewElts);
134 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
135 SmallVector<AllocaInst*, 32> &NewElts);
136 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
137 SmallVector<AllocaInst*, 32> &NewElts);
139 bool CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy,
140 bool &SawVec, uint64_t Offset, unsigned AllocaSize);
141 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset);
142 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType,
143 uint64_t Offset, IRBuilder<> &Builder);
144 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
145 uint64_t Offset, IRBuilder<> &Builder);
146 static Instruction *isOnlyCopiedFromConstantGlobal(AllocaInst *AI);
151 static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates");
153 // Public interface to the ScalarReplAggregates pass
154 FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) {
155 return new SROA(Threshold);
159 bool SROA::runOnFunction(Function &F) {
160 TD = getAnalysisIfAvailable<TargetData>();
162 bool Changed = performPromotion(F);
164 // FIXME: ScalarRepl currently depends on TargetData more than it
165 // theoretically needs to. It should be refactored in order to support
166 // target-independent IR. Until this is done, just skip the actual
167 // scalar-replacement portion of this pass.
168 if (!TD) return Changed;
171 bool LocalChange = performScalarRepl(F);
172 if (!LocalChange) break; // No need to repromote if no scalarrepl
174 LocalChange = performPromotion(F);
175 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
182 bool SROA::performPromotion(Function &F) {
183 std::vector<AllocaInst*> Allocas;
184 DominatorTree &DT = getAnalysis<DominatorTree>();
185 DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
187 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
189 bool Changed = false;
194 // Find allocas that are safe to promote, by looking at all instructions in
196 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
197 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
198 if (isAllocaPromotable(AI))
199 Allocas.push_back(AI);
201 if (Allocas.empty()) break;
203 PromoteMemToReg(Allocas, DT, DF);
204 NumPromoted += Allocas.size();
211 /// getNumSAElements - Return the number of elements in the specific struct or
213 static uint64_t getNumSAElements(const Type *T) {
214 if (const StructType *ST = dyn_cast<StructType>(T))
215 return ST->getNumElements();
216 return cast<ArrayType>(T)->getNumElements();
219 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
220 // which runs on all of the malloc/alloca instructions in the function, removing
221 // them if they are only used by getelementptr instructions.
223 bool SROA::performScalarRepl(Function &F) {
224 std::vector<AllocaInst*> WorkList;
226 // Scan the entry basic block, adding any alloca's and mallocs to the worklist
227 BasicBlock &BB = F.getEntryBlock();
228 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
229 if (AllocaInst *A = dyn_cast<AllocaInst>(I))
230 WorkList.push_back(A);
232 // Process the worklist
233 bool Changed = false;
234 while (!WorkList.empty()) {
235 AllocaInst *AI = WorkList.back();
238 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
239 // with unused elements.
240 if (AI->use_empty()) {
241 AI->eraseFromParent();
245 // If this alloca is impossible for us to promote, reject it early.
246 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
249 // Check to see if this allocation is only modified by a memcpy/memmove from
250 // a constant global. If this is the case, we can change all users to use
251 // the constant global instead. This is commonly produced by the CFE by
252 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
253 // is only subsequently read.
254 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) {
255 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n');
256 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n');
257 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2));
258 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
259 TheCopy->eraseFromParent(); // Don't mutate the global.
260 AI->eraseFromParent();
266 // Check to see if we can perform the core SROA transformation. We cannot
267 // transform the allocation instruction if it is an array allocation
268 // (allocations OF arrays are ok though), and an allocation of a scalar
269 // value cannot be decomposed at all.
270 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
272 // Do not promote [0 x %struct].
273 if (AllocaSize == 0) continue;
275 // Do not promote any struct whose size is too big.
276 if (AllocaSize > SRThreshold) continue;
278 if ((isa<StructType>(AI->getAllocatedType()) ||
279 isa<ArrayType>(AI->getAllocatedType())) &&
280 // Do not promote any struct into more than "32" separate vars.
281 getNumSAElements(AI->getAllocatedType()) <= SRThreshold/4) {
282 // Check that all of the users of the allocation are capable of being
284 switch (isSafeAllocaToScalarRepl(AI)) {
285 default: llvm_unreachable("Unexpected value!");
286 case 0: // Not safe to scalar replace.
288 case 1: // Safe, but requires cleanup/canonicalizations first
289 CleanupAllocaUsers(AI);
291 case 3: // Safe to scalar replace.
292 DoScalarReplacement(AI, WorkList);
298 // If we can turn this aggregate value (potentially with casts) into a
299 // simple scalar value that can be mem2reg'd into a register value.
300 // IsNotTrivial tracks whether this is something that mem2reg could have
301 // promoted itself. If so, we don't want to transform it needlessly. Note
302 // that we can't just check based on the type: the alloca may be of an i32
303 // but that has pointer arithmetic to set byte 3 of it or something.
304 bool IsNotTrivial = false;
305 const Type *VectorTy = 0;
306 bool HadAVector = false;
307 if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector,
308 0, unsigned(AllocaSize)) && IsNotTrivial) {
310 // If we were able to find a vector type that can handle this with
311 // insert/extract elements, and if there was at least one use that had
312 // a vector type, promote this to a vector. We don't want to promote
313 // random stuff that doesn't use vectors (e.g. <9 x double>) because then
314 // we just get a lot of insert/extracts. If at least one vector is
315 // involved, then we probably really do have a union of vector/array.
316 if (VectorTy && isa<VectorType>(VectorTy) && HadAVector) {
317 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = "
318 << *VectorTy << '\n');
320 // Create and insert the vector alloca.
321 NewAI = new AllocaInst(VectorTy, 0, "", AI->getParent()->begin());
322 ConvertUsesToScalar(AI, NewAI, 0);
324 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n");
326 // Create and insert the integer alloca.
327 const Type *NewTy = IntegerType::get(AI->getContext(), AllocaSize*8);
328 NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
329 ConvertUsesToScalar(AI, NewAI, 0);
332 AI->eraseFromParent();
338 // Otherwise, couldn't process this alloca.
344 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
345 /// predicate, do SROA now.
346 void SROA::DoScalarReplacement(AllocaInst *AI,
347 std::vector<AllocaInst*> &WorkList) {
348 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n');
349 SmallVector<AllocaInst*, 32> ElementAllocas;
350 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
351 ElementAllocas.reserve(ST->getNumContainedTypes());
352 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
353 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
355 AI->getName() + "." + Twine(i), AI);
356 ElementAllocas.push_back(NA);
357 WorkList.push_back(NA); // Add to worklist for recursive processing
360 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
361 ElementAllocas.reserve(AT->getNumElements());
362 const Type *ElTy = AT->getElementType();
363 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
364 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
365 AI->getName() + "." + Twine(i), AI);
366 ElementAllocas.push_back(NA);
367 WorkList.push_back(NA); // Add to worklist for recursive processing
371 // Now that we have created the new alloca instructions, rewrite all the
372 // uses of the old alloca.
373 RewriteForScalarRepl(AI, AI, 0, ElementAllocas);
375 // Now erase any instructions that were made dead while rewriting the alloca.
376 DeleteDeadInstructions();
377 AI->eraseFromParent();
382 /// DeleteDeadInstructions - Erase instructions on the DeadInstrs list,
383 /// recursively including all their operands that become trivially dead.
384 void SROA::DeleteDeadInstructions() {
385 while (!DeadInsts.empty()) {
386 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val());
388 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
389 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
390 // Zero out the operand and see if it becomes trivially dead.
391 // (But, don't add allocas to the dead instruction list -- they are
392 // already on the worklist and will be deleted separately.)
394 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U))
395 DeadInsts.push_back(U);
398 I->eraseFromParent();
402 /// isSafeForScalarRepl - Check if instruction I is a safe use with regard to
403 /// performing scalar replacement of alloca AI. The results are flagged in
404 /// the Info parameter. Offset indicates the position within AI that is
405 /// referenced by this instruction.
406 void SROA::isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
408 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
409 Instruction *User = cast<Instruction>(*UI);
411 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
412 isSafeForScalarRepl(BC, AI, Offset, Info);
413 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
414 uint64_t GEPOffset = Offset;
415 isSafeGEP(GEPI, AI, GEPOffset, Info);
417 isSafeForScalarRepl(GEPI, AI, GEPOffset, Info);
418 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) {
419 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
421 isSafeMemAccess(AI, Offset, Length->getZExtValue(), 0,
422 UI.getOperandNo() == 1, Info);
425 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
426 if (!LI->isVolatile()) {
427 const Type *LIType = LI->getType();
428 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(LIType),
429 LIType, false, Info);
432 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
433 // Store is ok if storing INTO the pointer, not storing the pointer
434 if (!SI->isVolatile() && SI->getOperand(0) != I) {
435 const Type *SIType = SI->getOperand(0)->getType();
436 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(SIType),
440 } else if (isa<DbgInfoIntrinsic>(UI)) {
441 // If one user is DbgInfoIntrinsic then check if all users are
442 // DbgInfoIntrinsics.
443 if (OnlyUsedByDbgInfoIntrinsics(I)) {
444 Info.needsCleanup = true;
449 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n');
452 if (Info.isUnsafe) return;
456 /// isSafeGEP - Check if a GEP instruction can be handled for scalar
457 /// replacement. It is safe when all the indices are constant, in-bounds
458 /// references, and when the resulting offset corresponds to an element within
459 /// the alloca type. The results are flagged in the Info parameter. Upon
460 /// return, Offset is adjusted as specified by the GEP indices.
461 void SROA::isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI,
462 uint64_t &Offset, AllocaInfo &Info) {
463 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI);
467 // Walk through the GEP type indices, checking the types that this indexes
469 for (; GEPIt != E; ++GEPIt) {
470 // Ignore struct elements, no extra checking needed for these.
471 if (isa<StructType>(*GEPIt))
474 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand());
476 return MarkUnsafe(Info);
479 // Compute the offset due to this GEP and check if the alloca has a
480 // component element at that offset.
481 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
482 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(),
483 &Indices[0], Indices.size());
484 if (!TypeHasComponent(AI->getAllocatedType(), Offset, 0))
488 /// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI
489 /// alloca or has an offset and size that corresponds to a component element
490 /// within it. The offset checked here may have been formed from a GEP with a
491 /// pointer bitcasted to a different type.
492 void SROA::isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize,
493 const Type *MemOpType, bool isStore,
495 // Check if this is a load/store of the entire alloca.
496 if (Offset == 0 && MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) {
497 bool UsesAggregateType = (MemOpType == AI->getAllocatedType());
498 // This is safe for MemIntrinsics (where MemOpType is 0), integer types
499 // (which are essentially the same as the MemIntrinsics, especially with
500 // regard to copying padding between elements), or references using the
501 // aggregate type of the alloca.
502 if (!MemOpType || isa<IntegerType>(MemOpType) || UsesAggregateType) {
503 if (!UsesAggregateType) {
505 Info.isMemCpyDst = true;
507 Info.isMemCpySrc = true;
512 // Check if the offset/size correspond to a component within the alloca type.
513 const Type *T = AI->getAllocatedType();
514 if (TypeHasComponent(T, Offset, MemSize))
517 return MarkUnsafe(Info);
520 /// TypeHasComponent - Return true if T has a component type with the
521 /// specified offset and size. If Size is zero, do not check the size.
522 bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) {
525 if (const StructType *ST = dyn_cast<StructType>(T)) {
526 const StructLayout *Layout = TD->getStructLayout(ST);
527 unsigned EltIdx = Layout->getElementContainingOffset(Offset);
528 EltTy = ST->getContainedType(EltIdx);
529 EltSize = TD->getTypeAllocSize(EltTy);
530 Offset -= Layout->getElementOffset(EltIdx);
531 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) {
532 EltTy = AT->getElementType();
533 EltSize = TD->getTypeAllocSize(EltTy);
534 if (Offset >= AT->getNumElements() * EltSize)
540 if (Offset == 0 && (Size == 0 || EltSize == Size))
542 // Check if the component spans multiple elements.
543 if (Offset + Size > EltSize)
545 return TypeHasComponent(EltTy, Offset, Size);
548 /// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite
549 /// the instruction I, which references it, to use the separate elements.
550 /// Offset indicates the position within AI that is referenced by this
552 void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
553 SmallVector<AllocaInst*, 32> &NewElts) {
554 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
555 Instruction *User = cast<Instruction>(*UI);
557 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
558 RewriteBitCast(BC, AI, Offset, NewElts);
559 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
560 RewriteGEP(GEPI, AI, Offset, NewElts);
561 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
562 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
563 uint64_t MemSize = Length->getZExtValue();
565 MemSize == TD->getTypeAllocSize(AI->getAllocatedType()))
566 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts);
567 // Otherwise the intrinsic can only touch a single element and the
568 // address operand will be updated, so nothing else needs to be done.
569 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
570 const Type *LIType = LI->getType();
571 if (LIType == AI->getAllocatedType()) {
573 // %res = load { i32, i32 }* %alloc
575 // %load.0 = load i32* %alloc.0
576 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
577 // %load.1 = load i32* %alloc.1
578 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
579 // (Also works for arrays instead of structs)
580 Value *Insert = UndefValue::get(LIType);
581 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
582 Value *Load = new LoadInst(NewElts[i], "load", LI);
583 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI);
585 LI->replaceAllUsesWith(Insert);
586 DeadInsts.push_back(LI);
587 } else if (isa<IntegerType>(LIType) &&
588 TD->getTypeAllocSize(LIType) ==
589 TD->getTypeAllocSize(AI->getAllocatedType())) {
590 // If this is a load of the entire alloca to an integer, rewrite it.
591 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
593 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
594 Value *Val = SI->getOperand(0);
595 const Type *SIType = Val->getType();
596 if (SIType == AI->getAllocatedType()) {
598 // store { i32, i32 } %val, { i32, i32 }* %alloc
600 // %val.0 = extractvalue { i32, i32 } %val, 0
601 // store i32 %val.0, i32* %alloc.0
602 // %val.1 = extractvalue { i32, i32 } %val, 1
603 // store i32 %val.1, i32* %alloc.1
604 // (Also works for arrays instead of structs)
605 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
606 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI);
607 new StoreInst(Extract, NewElts[i], SI);
609 DeadInsts.push_back(SI);
610 } else if (isa<IntegerType>(SIType) &&
611 TD->getTypeAllocSize(SIType) ==
612 TD->getTypeAllocSize(AI->getAllocatedType())) {
613 // If this is a store of the entire alloca from an integer, rewrite it.
614 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
620 /// RewriteBitCast - Update a bitcast reference to the alloca being replaced
621 /// and recursively continue updating all of its uses.
622 void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
623 SmallVector<AllocaInst*, 32> &NewElts) {
624 RewriteForScalarRepl(BC, AI, Offset, NewElts);
625 if (BC->getOperand(0) != AI)
628 // The bitcast references the original alloca. Replace its uses with
629 // references to the first new element alloca.
630 Instruction *Val = NewElts[0];
631 if (Val->getType() != BC->getDestTy()) {
632 Val = new BitCastInst(Val, BC->getDestTy(), "", BC);
635 BC->replaceAllUsesWith(Val);
636 DeadInsts.push_back(BC);
639 /// FindElementAndOffset - Return the index of the element containing Offset
640 /// within the specified type, which must be either a struct or an array.
641 /// Sets T to the type of the element and Offset to the offset within that
642 /// element. IdxTy is set to the type of the index result to be used in a
644 uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset,
645 const Type *&IdxTy) {
647 if (const StructType *ST = dyn_cast<StructType>(T)) {
648 const StructLayout *Layout = TD->getStructLayout(ST);
649 Idx = Layout->getElementContainingOffset(Offset);
650 T = ST->getContainedType(Idx);
651 Offset -= Layout->getElementOffset(Idx);
652 IdxTy = Type::getInt32Ty(T->getContext());
655 const ArrayType *AT = cast<ArrayType>(T);
656 T = AT->getElementType();
657 uint64_t EltSize = TD->getTypeAllocSize(T);
658 Idx = Offset / EltSize;
659 Offset -= Idx * EltSize;
660 IdxTy = Type::getInt64Ty(T->getContext());
664 /// RewriteGEP - Check if this GEP instruction moves the pointer across
665 /// elements of the alloca that are being split apart, and if so, rewrite
666 /// the GEP to be relative to the new element.
667 void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
668 SmallVector<AllocaInst*, 32> &NewElts) {
669 uint64_t OldOffset = Offset;
670 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
671 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(),
672 &Indices[0], Indices.size());
674 RewriteForScalarRepl(GEPI, AI, Offset, NewElts);
676 const Type *T = AI->getAllocatedType();
678 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy);
679 if (GEPI->getOperand(0) == AI)
680 OldIdx = ~0ULL; // Force the GEP to be rewritten.
682 T = AI->getAllocatedType();
683 uint64_t EltOffset = Offset;
684 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy);
686 // If this GEP does not move the pointer across elements of the alloca
687 // being split, then it does not needs to be rewritten.
691 const Type *i32Ty = Type::getInt32Ty(AI->getContext());
692 SmallVector<Value*, 8> NewArgs;
693 NewArgs.push_back(Constant::getNullValue(i32Ty));
694 while (EltOffset != 0) {
695 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy);
696 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx));
698 Instruction *Val = NewElts[Idx];
699 if (NewArgs.size() > 1) {
700 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(),
701 NewArgs.end(), "", GEPI);
704 if (Val->getType() != GEPI->getType())
705 Val = new BitCastInst(Val, GEPI->getType(), Val->getNameStr(), GEPI);
706 GEPI->replaceAllUsesWith(Val);
707 DeadInsts.push_back(GEPI);
710 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
711 /// Rewrite it to copy or set the elements of the scalarized memory.
712 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
714 SmallVector<AllocaInst*, 32> &NewElts) {
715 // If this is a memcpy/memmove, construct the other pointer as the
716 // appropriate type. The "Other" pointer is the pointer that goes to memory
717 // that doesn't have anything to do with the alloca that we are promoting. For
718 // memset, this Value* stays null.
720 LLVMContext &Context = MI->getContext();
721 unsigned MemAlignment = MI->getAlignment();
722 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy
723 if (Inst == MTI->getRawDest())
724 OtherPtr = MTI->getRawSource();
726 assert(Inst == MTI->getRawSource());
727 OtherPtr = MTI->getRawDest();
731 // If there is an other pointer, we want to convert it to the same pointer
732 // type as AI has, so we can GEP through it safely.
735 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an
736 // optimization, but it's also required to detect the corner case where
737 // both pointer operands are referencing the same memory, and where
738 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This
739 // function is only called for mem intrinsics that access the whole
740 // aggregate, so non-zero GEPs are not an issue here.)
742 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) {
743 OtherPtr = BC->getOperand(0);
746 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) {
747 // All zero GEPs are effectively bitcasts.
748 if (GEP->hasAllZeroIndices()) {
749 OtherPtr = GEP->getOperand(0);
755 // Copying the alloca to itself is a no-op: just delete it.
756 if (OtherPtr == AI || OtherPtr == NewElts[0]) {
757 // This code will run twice for a no-op memcpy -- once for each operand.
758 // Put only one reference to MI on the DeadInsts list.
759 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(),
760 E = DeadInsts.end(); I != E; ++I)
761 if (*I == MI) return;
762 DeadInsts.push_back(MI);
766 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr))
767 if (BCE->getOpcode() == Instruction::BitCast)
768 OtherPtr = BCE->getOperand(0);
770 // If the pointer is not the right type, insert a bitcast to the right
772 if (OtherPtr->getType() != AI->getType())
773 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(),
777 // Process each element of the aggregate.
778 Value *TheFn = MI->getOperand(0);
779 const Type *BytePtrTy = MI->getRawDest()->getType();
780 bool SROADest = MI->getRawDest() == Inst;
782 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext()));
784 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
785 // If this is a memcpy/memmove, emit a GEP of the other element address.
787 unsigned OtherEltAlign = MemAlignment;
790 Value *Idx[2] = { Zero,
791 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) };
792 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2,
793 OtherPtr->getNameStr()+"."+Twine(i),
796 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
797 if (const StructType *ST =
798 dyn_cast<StructType>(OtherPtrTy->getElementType())) {
799 EltOffset = TD->getStructLayout(ST)->getElementOffset(i);
802 cast<SequentialType>(OtherPtr->getType())->getElementType();
803 EltOffset = TD->getTypeAllocSize(EltTy)*i;
806 // The alignment of the other pointer is the guaranteed alignment of the
807 // element, which is affected by both the known alignment of the whole
808 // mem intrinsic and the alignment of the element. If the alignment of
809 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
810 // known alignment is just 4 bytes.
811 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
814 Value *EltPtr = NewElts[i];
815 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
817 // If we got down to a scalar, insert a load or store as appropriate.
818 if (EltTy->isSingleValueType()) {
819 if (isa<MemTransferInst>(MI)) {
821 // From Other to Alloca.
822 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
823 new StoreInst(Elt, EltPtr, MI);
825 // From Alloca to Other.
826 Value *Elt = new LoadInst(EltPtr, "tmp", MI);
827 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI);
831 assert(isa<MemSetInst>(MI));
833 // If the stored element is zero (common case), just store a null
836 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) {
838 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
840 // If EltTy is a vector type, get the element type.
841 const Type *ValTy = EltTy->getScalarType();
843 // Construct an integer with the right value.
844 unsigned EltSize = TD->getTypeSizeInBits(ValTy);
845 APInt OneVal(EltSize, CI->getZExtValue());
846 APInt TotalVal(OneVal);
848 for (unsigned i = 0; 8*i < EltSize; ++i) {
849 TotalVal = TotalVal.shl(8);
853 // Convert the integer value to the appropriate type.
854 StoreVal = ConstantInt::get(Context, TotalVal);
855 if (isa<PointerType>(ValTy))
856 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
857 else if (ValTy->isFloatingPoint())
858 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
859 assert(StoreVal->getType() == ValTy && "Type mismatch!");
861 // If the requested value was a vector constant, create it.
862 if (EltTy != ValTy) {
863 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements();
864 SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
865 StoreVal = ConstantVector::get(&Elts[0], NumElts);
868 new StoreInst(StoreVal, EltPtr, MI);
871 // Otherwise, if we're storing a byte variable, use a memset call for
875 // Cast the element pointer to BytePtrTy.
876 if (EltPtr->getType() != BytePtrTy)
877 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI);
879 // Cast the other pointer (if we have one) to BytePtrTy.
880 if (OtherElt && OtherElt->getType() != BytePtrTy)
881 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(),
884 unsigned EltSize = TD->getTypeAllocSize(EltTy);
886 // Finally, insert the meminst for this element.
887 if (isa<MemTransferInst>(MI)) {
889 SROADest ? EltPtr : OtherElt, // Dest ptr
890 SROADest ? OtherElt : EltPtr, // Src ptr
891 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
893 ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign)
895 CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
897 assert(isa<MemSetInst>(MI));
899 EltPtr, MI->getOperand(2), // Dest, Value,
900 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
903 CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
906 DeadInsts.push_back(MI);
909 /// RewriteStoreUserOfWholeAlloca - We found a store of an integer that
910 /// overwrites the entire allocation. Extract out the pieces of the stored
911 /// integer and store them individually.
912 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
913 SmallVector<AllocaInst*, 32> &NewElts){
914 // Extract each element out of the integer according to its structure offset
915 // and store the element value to the individual alloca.
916 Value *SrcVal = SI->getOperand(0);
917 const Type *AllocaEltTy = AI->getAllocatedType();
918 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
920 // Handle tail padding by extending the operand
921 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
922 SrcVal = new ZExtInst(SrcVal,
923 IntegerType::get(SI->getContext(), AllocaSizeBits),
926 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI
929 // There are two forms here: AI could be an array or struct. Both cases
930 // have different ways to compute the element offset.
931 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
932 const StructLayout *Layout = TD->getStructLayout(EltSTy);
934 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
935 // Get the number of bits to shift SrcVal to get the value.
936 const Type *FieldTy = EltSTy->getElementType(i);
937 uint64_t Shift = Layout->getElementOffsetInBits(i);
939 if (TD->isBigEndian())
940 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy);
942 Value *EltVal = SrcVal;
944 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
945 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
946 "sroa.store.elt", SI);
949 // Truncate down to an integer of the right size.
950 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
952 // Ignore zero sized fields like {}, they obviously contain no data.
953 if (FieldSizeBits == 0) continue;
955 if (FieldSizeBits != AllocaSizeBits)
956 EltVal = new TruncInst(EltVal,
957 IntegerType::get(SI->getContext(), FieldSizeBits),
959 Value *DestField = NewElts[i];
960 if (EltVal->getType() == FieldTy) {
961 // Storing to an integer field of this size, just do it.
962 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) {
963 // Bitcast to the right element type (for fp/vector values).
964 EltVal = new BitCastInst(EltVal, FieldTy, "", SI);
966 // Otherwise, bitcast the dest pointer (for aggregates).
967 DestField = new BitCastInst(DestField,
968 PointerType::getUnqual(EltVal->getType()),
971 new StoreInst(EltVal, DestField, SI);
975 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
976 const Type *ArrayEltTy = ATy->getElementType();
977 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
978 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
982 if (TD->isBigEndian())
983 Shift = AllocaSizeBits-ElementOffset;
987 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
988 // Ignore zero sized fields like {}, they obviously contain no data.
989 if (ElementSizeBits == 0) continue;
991 Value *EltVal = SrcVal;
993 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
994 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
995 "sroa.store.elt", SI);
998 // Truncate down to an integer of the right size.
999 if (ElementSizeBits != AllocaSizeBits)
1000 EltVal = new TruncInst(EltVal,
1001 IntegerType::get(SI->getContext(),
1002 ElementSizeBits),"",SI);
1003 Value *DestField = NewElts[i];
1004 if (EltVal->getType() == ArrayEltTy) {
1005 // Storing to an integer field of this size, just do it.
1006 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) {
1007 // Bitcast to the right element type (for fp/vector values).
1008 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI);
1010 // Otherwise, bitcast the dest pointer (for aggregates).
1011 DestField = new BitCastInst(DestField,
1012 PointerType::getUnqual(EltVal->getType()),
1015 new StoreInst(EltVal, DestField, SI);
1017 if (TD->isBigEndian())
1018 Shift -= ElementOffset;
1020 Shift += ElementOffset;
1024 DeadInsts.push_back(SI);
1027 /// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to
1028 /// an integer. Load the individual pieces to form the aggregate value.
1029 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
1030 SmallVector<AllocaInst*, 32> &NewElts) {
1031 // Extract each element out of the NewElts according to its structure offset
1032 // and form the result value.
1033 const Type *AllocaEltTy = AI->getAllocatedType();
1034 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
1036 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI
1039 // There are two forms here: AI could be an array or struct. Both cases
1040 // have different ways to compute the element offset.
1041 const StructLayout *Layout = 0;
1042 uint64_t ArrayEltBitOffset = 0;
1043 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
1044 Layout = TD->getStructLayout(EltSTy);
1046 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
1047 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
1051 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits));
1053 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1054 // Load the value from the alloca. If the NewElt is an aggregate, cast
1055 // the pointer to an integer of the same size before doing the load.
1056 Value *SrcField = NewElts[i];
1057 const Type *FieldTy =
1058 cast<PointerType>(SrcField->getType())->getElementType();
1059 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
1061 // Ignore zero sized fields like {}, they obviously contain no data.
1062 if (FieldSizeBits == 0) continue;
1064 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(),
1066 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() &&
1067 !isa<VectorType>(FieldTy))
1068 SrcField = new BitCastInst(SrcField,
1069 PointerType::getUnqual(FieldIntTy),
1071 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
1073 // If SrcField is a fp or vector of the right size but that isn't an
1074 // integer type, bitcast to an integer so we can shift it.
1075 if (SrcField->getType() != FieldIntTy)
1076 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
1078 // Zero extend the field to be the same size as the final alloca so that
1079 // we can shift and insert it.
1080 if (SrcField->getType() != ResultVal->getType())
1081 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
1083 // Determine the number of bits to shift SrcField.
1085 if (Layout) // Struct case.
1086 Shift = Layout->getElementOffsetInBits(i);
1088 Shift = i*ArrayEltBitOffset;
1090 if (TD->isBigEndian())
1091 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
1094 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
1095 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
1098 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
1101 // Handle tail padding by truncating the result
1102 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
1103 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI);
1105 LI->replaceAllUsesWith(ResultVal);
1106 DeadInsts.push_back(LI);
1109 /// HasPadding - Return true if the specified type has any structure or
1110 /// alignment padding, false otherwise.
1111 static bool HasPadding(const Type *Ty, const TargetData &TD) {
1112 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1113 const StructLayout *SL = TD.getStructLayout(STy);
1114 unsigned PrevFieldBitOffset = 0;
1115 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1116 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
1118 // Padding in sub-elements?
1119 if (HasPadding(STy->getElementType(i), TD))
1122 // Check to see if there is any padding between this element and the
1125 unsigned PrevFieldEnd =
1126 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
1127 if (PrevFieldEnd < FieldBitOffset)
1131 PrevFieldBitOffset = FieldBitOffset;
1134 // Check for tail padding.
1135 if (unsigned EltCount = STy->getNumElements()) {
1136 unsigned PrevFieldEnd = PrevFieldBitOffset +
1137 TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
1138 if (PrevFieldEnd < SL->getSizeInBits())
1142 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1143 return HasPadding(ATy->getElementType(), TD);
1144 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1145 return HasPadding(VTy->getElementType(), TD);
1147 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
1150 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
1151 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
1152 /// or 1 if safe after canonicalization has been performed.
1153 int SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) {
1154 // Loop over the use list of the alloca. We can only transform it if all of
1155 // the users are safe to transform.
1158 isSafeForScalarRepl(AI, AI, 0, Info);
1159 if (Info.isUnsafe) {
1160 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n');
1164 // Okay, we know all the users are promotable. If the aggregate is a memcpy
1165 // source and destination, we have to be careful. In particular, the memcpy
1166 // could be moving around elements that live in structure padding of the LLVM
1167 // types, but may actually be used. In these cases, we refuse to promote the
1169 if (Info.isMemCpySrc && Info.isMemCpyDst &&
1170 HasPadding(AI->getAllocatedType(), *TD))
1173 // If we require cleanup, return 1, otherwise return 3.
1174 return Info.needsCleanup ? 1 : 3;
1177 /// CleanupAllocaUsers - If SROA reported that it can promote the specified
1178 /// allocation, but only if cleaned up, perform the cleanups required.
1179 void SROA::CleanupAllocaUsers(Value *V) {
1180 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
1183 Instruction *I = cast<Instruction>(U);
1184 SmallVector<DbgInfoIntrinsic *, 2> DbgInUses;
1185 if (!isa<StoreInst>(I) && OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) {
1186 // Safe to remove debug info uses.
1187 while (!DbgInUses.empty()) {
1188 DbgInfoIntrinsic *DI = DbgInUses.pop_back_val();
1189 DI->eraseFromParent();
1191 I->eraseFromParent();
1196 /// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at
1197 /// the offset specified by Offset (which is specified in bytes).
1199 /// There are two cases we handle here:
1200 /// 1) A union of vector types of the same size and potentially its elements.
1201 /// Here we turn element accesses into insert/extract element operations.
1202 /// This promotes a <4 x float> with a store of float to the third element
1203 /// into a <4 x float> that uses insert element.
1204 /// 2) A fully general blob of memory, which we turn into some (potentially
1205 /// large) integer type with extract and insert operations where the loads
1206 /// and stores would mutate the memory.
1207 static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy,
1208 unsigned AllocaSize, const TargetData &TD,
1209 LLVMContext &Context) {
1210 // If this could be contributing to a vector, analyze it.
1211 if (VecTy != Type::getVoidTy(Context)) { // either null or a vector type.
1213 // If the In type is a vector that is the same size as the alloca, see if it
1214 // matches the existing VecTy.
1215 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) {
1216 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) {
1217 // If we're storing/loading a vector of the right size, allow it as a
1218 // vector. If this the first vector we see, remember the type so that
1219 // we know the element size.
1224 } else if (In->isFloatTy() || In->isDoubleTy() ||
1225 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 &&
1226 isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
1227 // If we're accessing something that could be an element of a vector, see
1228 // if the implied vector agrees with what we already have and if Offset is
1229 // compatible with it.
1230 unsigned EltSize = In->getPrimitiveSizeInBits()/8;
1231 if (Offset % EltSize == 0 &&
1232 AllocaSize % EltSize == 0 &&
1234 cast<VectorType>(VecTy)->getElementType()
1235 ->getPrimitiveSizeInBits()/8 == EltSize)) {
1237 VecTy = VectorType::get(In, AllocaSize/EltSize);
1243 // Otherwise, we have a case that we can't handle with an optimized vector
1244 // form. We can still turn this into a large integer.
1245 VecTy = Type::getVoidTy(Context);
1248 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
1249 /// its accesses to a single vector type, return true and set VecTy to
1250 /// the new type. If we could convert the alloca into a single promotable
1251 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a
1252 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
1253 /// is the current offset from the base of the alloca being analyzed.
1255 /// If we see at least one access to the value that is as a vector type, set the
1257 bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy,
1258 bool &SawVec, uint64_t Offset,
1259 unsigned AllocaSize) {
1260 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1261 Instruction *User = cast<Instruction>(*UI);
1263 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1264 // Don't break volatile loads.
1265 if (LI->isVolatile())
1267 MergeInType(LI->getType(), Offset, VecTy,
1268 AllocaSize, *TD, V->getContext());
1269 SawVec |= isa<VectorType>(LI->getType());
1273 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1274 // Storing the pointer, not into the value?
1275 if (SI->getOperand(0) == V || SI->isVolatile()) return 0;
1276 MergeInType(SI->getOperand(0)->getType(), Offset,
1277 VecTy, AllocaSize, *TD, V->getContext());
1278 SawVec |= isa<VectorType>(SI->getOperand(0)->getType());
1282 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
1283 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset,
1286 IsNotTrivial = true;
1290 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1291 // If this is a GEP with a variable indices, we can't handle it.
1292 if (!GEP->hasAllConstantIndices())
1295 // Compute the offset that this GEP adds to the pointer.
1296 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
1297 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getPointerOperandType(),
1298 &Indices[0], Indices.size());
1299 // See if all uses can be converted.
1300 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset,
1303 IsNotTrivial = true;
1307 // If this is a constant sized memset of a constant value (e.g. 0) we can
1309 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
1310 // Store of constant value and constant size.
1311 if (isa<ConstantInt>(MSI->getValue()) &&
1312 isa<ConstantInt>(MSI->getLength())) {
1313 IsNotTrivial = true;
1318 // If this is a memcpy or memmove into or out of the whole allocation, we
1319 // can handle it like a load or store of the scalar type.
1320 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
1321 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()))
1322 if (Len->getZExtValue() == AllocaSize && Offset == 0) {
1323 IsNotTrivial = true;
1328 // Ignore dbg intrinsic.
1329 if (isa<DbgInfoIntrinsic>(User))
1332 // Otherwise, we cannot handle this!
1339 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
1340 /// directly. This happens when we are converting an "integer union" to a
1341 /// single integer scalar, or when we are converting a "vector union" to a
1342 /// vector with insert/extractelement instructions.
1344 /// Offset is an offset from the original alloca, in bits that need to be
1345 /// shifted to the right. By the end of this, there should be no uses of Ptr.
1346 void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) {
1347 while (!Ptr->use_empty()) {
1348 Instruction *User = cast<Instruction>(Ptr->use_back());
1350 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
1351 ConvertUsesToScalar(CI, NewAI, Offset);
1352 CI->eraseFromParent();
1356 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1357 // Compute the offset that this GEP adds to the pointer.
1358 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
1359 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getPointerOperandType(),
1360 &Indices[0], Indices.size());
1361 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
1362 GEP->eraseFromParent();
1366 IRBuilder<> Builder(User->getParent(), User);
1368 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1369 // The load is a bit extract from NewAI shifted right by Offset bits.
1370 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp");
1372 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
1373 LI->replaceAllUsesWith(NewLoadVal);
1374 LI->eraseFromParent();
1378 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1379 assert(SI->getOperand(0) != Ptr && "Consistency error!");
1380 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
1381 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
1383 Builder.CreateStore(New, NewAI);
1384 SI->eraseFromParent();
1386 // If the load we just inserted is now dead, then the inserted store
1387 // overwrote the entire thing.
1388 if (Old->use_empty())
1389 Old->eraseFromParent();
1393 // If this is a constant sized memset of a constant value (e.g. 0) we can
1394 // transform it into a store of the expanded constant value.
1395 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
1396 assert(MSI->getRawDest() == Ptr && "Consistency error!");
1397 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
1398 if (NumBytes != 0) {
1399 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
1401 // Compute the value replicated the right number of times.
1402 APInt APVal(NumBytes*8, Val);
1404 // Splat the value if non-zero.
1406 for (unsigned i = 1; i != NumBytes; ++i)
1407 APVal |= APVal << 8;
1409 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
1410 Value *New = ConvertScalar_InsertValue(
1411 ConstantInt::get(User->getContext(), APVal),
1412 Old, Offset, Builder);
1413 Builder.CreateStore(New, NewAI);
1415 // If the load we just inserted is now dead, then the memset overwrote
1416 // the entire thing.
1417 if (Old->use_empty())
1418 Old->eraseFromParent();
1420 MSI->eraseFromParent();
1424 // If this is a memcpy or memmove into or out of the whole allocation, we
1425 // can handle it like a load or store of the scalar type.
1426 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
1427 assert(Offset == 0 && "must be store to start of alloca");
1429 // If the source and destination are both to the same alloca, then this is
1430 // a noop copy-to-self, just delete it. Otherwise, emit a load and store
1432 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject());
1434 if (MTI->getSource()->getUnderlyingObject() != OrigAI) {
1435 // Dest must be OrigAI, change this to be a load from the original
1436 // pointer (bitcasted), then a store to our new alloca.
1437 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
1438 Value *SrcPtr = MTI->getSource();
1439 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType());
1441 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
1442 SrcVal->setAlignment(MTI->getAlignment());
1443 Builder.CreateStore(SrcVal, NewAI);
1444 } else if (MTI->getDest()->getUnderlyingObject() != OrigAI) {
1445 // Src must be OrigAI, change this to be a load from NewAI then a store
1446 // through the original dest pointer (bitcasted).
1447 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
1448 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
1450 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType());
1451 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
1452 NewStore->setAlignment(MTI->getAlignment());
1454 // Noop transfer. Src == Dst
1458 MTI->eraseFromParent();
1462 // If user is a dbg info intrinsic then it is safe to remove it.
1463 if (isa<DbgInfoIntrinsic>(User)) {
1464 User->eraseFromParent();
1468 llvm_unreachable("Unsupported operation!");
1472 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
1473 /// or vector value FromVal, extracting the bits from the offset specified by
1474 /// Offset. This returns the value, which is of type ToType.
1476 /// This happens when we are converting an "integer union" to a single
1477 /// integer scalar, or when we are converting a "vector union" to a vector with
1478 /// insert/extractelement instructions.
1480 /// Offset is an offset from the original alloca, in bits that need to be
1481 /// shifted to the right.
1482 Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType,
1483 uint64_t Offset, IRBuilder<> &Builder) {
1484 // If the load is of the whole new alloca, no conversion is needed.
1485 if (FromVal->getType() == ToType && Offset == 0)
1488 // If the result alloca is a vector type, this is either an element
1489 // access or a bitcast to another vector type of the same size.
1490 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) {
1491 if (isa<VectorType>(ToType))
1492 return Builder.CreateBitCast(FromVal, ToType, "tmp");
1494 // Otherwise it must be an element access.
1497 unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType());
1498 Elt = Offset/EltSize;
1499 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
1501 // Return the element extracted out of it.
1502 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get(
1503 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp");
1504 if (V->getType() != ToType)
1505 V = Builder.CreateBitCast(V, ToType, "tmp");
1509 // If ToType is a first class aggregate, extract out each of the pieces and
1510 // use insertvalue's to form the FCA.
1511 if (const StructType *ST = dyn_cast<StructType>(ToType)) {
1512 const StructLayout &Layout = *TD->getStructLayout(ST);
1513 Value *Res = UndefValue::get(ST);
1514 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
1515 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
1516 Offset+Layout.getElementOffsetInBits(i),
1518 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
1523 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
1524 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType());
1525 Value *Res = UndefValue::get(AT);
1526 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1527 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
1528 Offset+i*EltSize, Builder);
1529 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
1534 // Otherwise, this must be a union that was converted to an integer value.
1535 const IntegerType *NTy = cast<IntegerType>(FromVal->getType());
1537 // If this is a big-endian system and the load is narrower than the
1538 // full alloca type, we need to do a shift to get the right bits.
1540 if (TD->isBigEndian()) {
1541 // On big-endian machines, the lowest bit is stored at the bit offset
1542 // from the pointer given by getTypeStoreSizeInBits. This matters for
1543 // integers with a bitwidth that is not a multiple of 8.
1544 ShAmt = TD->getTypeStoreSizeInBits(NTy) -
1545 TD->getTypeStoreSizeInBits(ToType) - Offset;
1550 // Note: we support negative bitwidths (with shl) which are not defined.
1551 // We do this to support (f.e.) loads off the end of a structure where
1552 // only some bits are used.
1553 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
1554 FromVal = Builder.CreateLShr(FromVal,
1555 ConstantInt::get(FromVal->getType(),
1557 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
1558 FromVal = Builder.CreateShl(FromVal,
1559 ConstantInt::get(FromVal->getType(),
1562 // Finally, unconditionally truncate the integer to the right width.
1563 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType);
1564 if (LIBitWidth < NTy->getBitWidth())
1566 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(),
1567 LIBitWidth), "tmp");
1568 else if (LIBitWidth > NTy->getBitWidth())
1570 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(),
1571 LIBitWidth), "tmp");
1573 // If the result is an integer, this is a trunc or bitcast.
1574 if (isa<IntegerType>(ToType)) {
1576 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) {
1577 // Just do a bitcast, we know the sizes match up.
1578 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp");
1580 // Otherwise must be a pointer.
1581 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp");
1583 assert(FromVal->getType() == ToType && "Didn't convert right?");
1587 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
1588 /// or vector value "Old" at the offset specified by Offset.
1590 /// This happens when we are converting an "integer union" to a
1591 /// single integer scalar, or when we are converting a "vector union" to a
1592 /// vector with insert/extractelement instructions.
1594 /// Offset is an offset from the original alloca, in bits that need to be
1595 /// shifted to the right.
1596 Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old,
1597 uint64_t Offset, IRBuilder<> &Builder) {
1599 // Convert the stored type to the actual type, shift it left to insert
1600 // then 'or' into place.
1601 const Type *AllocaType = Old->getType();
1602 LLVMContext &Context = Old->getContext();
1604 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
1605 uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy);
1606 uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType());
1608 // Changing the whole vector with memset or with an access of a different
1610 if (ValSize == VecSize)
1611 return Builder.CreateBitCast(SV, AllocaType, "tmp");
1613 uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType());
1615 // Must be an element insertion.
1616 unsigned Elt = Offset/EltSize;
1618 if (SV->getType() != VTy->getElementType())
1619 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp");
1621 SV = Builder.CreateInsertElement(Old, SV,
1622 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt),
1627 // If SV is a first-class aggregate value, insert each value recursively.
1628 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) {
1629 const StructLayout &Layout = *TD->getStructLayout(ST);
1630 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
1631 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
1632 Old = ConvertScalar_InsertValue(Elt, Old,
1633 Offset+Layout.getElementOffsetInBits(i),
1639 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
1640 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType());
1641 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1642 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
1643 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
1648 // If SV is a float, convert it to the appropriate integer type.
1649 // If it is a pointer, do the same.
1650 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType());
1651 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType);
1652 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType());
1653 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType);
1654 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType()))
1655 SV = Builder.CreateBitCast(SV,
1656 IntegerType::get(SV->getContext(),SrcWidth), "tmp");
1657 else if (isa<PointerType>(SV->getType()))
1658 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(SV->getContext()), "tmp");
1660 // Zero extend or truncate the value if needed.
1661 if (SV->getType() != AllocaType) {
1662 if (SV->getType()->getPrimitiveSizeInBits() <
1663 AllocaType->getPrimitiveSizeInBits())
1664 SV = Builder.CreateZExt(SV, AllocaType, "tmp");
1666 // Truncation may be needed if storing more than the alloca can hold
1667 // (undefined behavior).
1668 SV = Builder.CreateTrunc(SV, AllocaType, "tmp");
1669 SrcWidth = DestWidth;
1670 SrcStoreWidth = DestStoreWidth;
1674 // If this is a big-endian system and the store is narrower than the
1675 // full alloca type, we need to do a shift to get the right bits.
1677 if (TD->isBigEndian()) {
1678 // On big-endian machines, the lowest bit is stored at the bit offset
1679 // from the pointer given by getTypeStoreSizeInBits. This matters for
1680 // integers with a bitwidth that is not a multiple of 8.
1681 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
1686 // Note: we support negative bitwidths (with shr) which are not defined.
1687 // We do this to support (f.e.) stores off the end of a structure where
1688 // only some bits in the structure are set.
1689 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
1690 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
1691 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(),
1694 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
1695 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(),
1697 Mask = Mask.lshr(-ShAmt);
1700 // Mask out the bits we are about to insert from the old value, and or
1702 if (SrcWidth != DestWidth) {
1703 assert(DestWidth > SrcWidth);
1704 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask");
1705 SV = Builder.CreateOr(Old, SV, "ins");
1712 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
1713 /// some part of a constant global variable. This intentionally only accepts
1714 /// constant expressions because we don't can't rewrite arbitrary instructions.
1715 static bool PointsToConstantGlobal(Value *V) {
1716 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
1717 return GV->isConstant();
1718 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
1719 if (CE->getOpcode() == Instruction::BitCast ||
1720 CE->getOpcode() == Instruction::GetElementPtr)
1721 return PointsToConstantGlobal(CE->getOperand(0));
1725 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
1726 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
1727 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
1728 /// track of whether it moves the pointer (with isOffset) but otherwise traverse
1729 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
1730 /// the alloca, and if the source pointer is a pointer to a constant global, we
1731 /// can optimize this.
1732 static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy,
1734 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1735 if (LoadInst *LI = dyn_cast<LoadInst>(*UI))
1736 // Ignore non-volatile loads, they are always ok.
1737 if (!LI->isVolatile())
1740 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) {
1741 // If uses of the bitcast are ok, we are ok.
1742 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset))
1746 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
1747 // If the GEP has all zero indices, it doesn't offset the pointer. If it
1748 // doesn't, it does.
1749 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
1750 isOffset || !GEP->hasAllZeroIndices()))
1755 // If this is isn't our memcpy/memmove, reject it as something we can't
1757 if (!isa<MemTransferInst>(*UI))
1760 // If we already have seen a copy, reject the second one.
1761 if (TheCopy) return false;
1763 // If the pointer has been offset from the start of the alloca, we can't
1764 // safely handle this.
1765 if (isOffset) return false;
1767 // If the memintrinsic isn't using the alloca as the dest, reject it.
1768 if (UI.getOperandNo() != 1) return false;
1770 MemIntrinsic *MI = cast<MemIntrinsic>(*UI);
1772 // If the source of the memcpy/move is not a constant global, reject it.
1773 if (!PointsToConstantGlobal(MI->getOperand(2)))
1776 // Otherwise, the transform is safe. Remember the copy instruction.
1782 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
1783 /// modified by a copy from a constant global. If we can prove this, we can
1784 /// replace any uses of the alloca with uses of the global directly.
1785 Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) {
1786 Instruction *TheCopy = 0;
1787 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false))