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/Pass.h"
31 #include "llvm/Analysis/Dominators.h"
32 #include "llvm/Target/TargetData.h"
33 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
34 #include "llvm/Transforms/Utils/Local.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/GetElementPtrTypeIterator.h"
37 #include "llvm/Support/IRBuilder.h"
38 #include "llvm/Support/MathExtras.h"
39 #include "llvm/Support/Compiler.h"
40 #include "llvm/ADT/SmallVector.h"
41 #include "llvm/ADT/Statistic.h"
42 #include "llvm/ADT/StringExtras.h"
45 STATISTIC(NumReplaced, "Number of allocas broken up");
46 STATISTIC(NumPromoted, "Number of allocas promoted");
47 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
48 STATISTIC(NumGlobals, "Number of allocas copied from constant global");
51 struct VISIBILITY_HIDDEN SROA : public FunctionPass {
52 static char ID; // Pass identification, replacement for typeid
53 explicit SROA(signed T = -1) : FunctionPass(&ID) {
60 bool runOnFunction(Function &F);
62 bool performScalarRepl(Function &F);
63 bool performPromotion(Function &F);
65 // getAnalysisUsage - This pass does not require any passes, but we know it
66 // will not alter the CFG, so say so.
67 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
68 AU.addRequired<DominatorTree>();
69 AU.addRequired<DominanceFrontier>();
70 AU.addRequired<TargetData>();
77 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
78 /// information about the uses. All these fields are initialized to false
79 /// and set to true when something is learned.
81 /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
84 /// needsCleanup - This is set to true if there is some use of the alloca
85 /// that requires cleanup.
86 bool needsCleanup : 1;
88 /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
91 /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
95 : isUnsafe(false), needsCleanup(false),
96 isMemCpySrc(false), isMemCpyDst(false) {}
101 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; }
103 int isSafeAllocaToScalarRepl(AllocationInst *AI);
105 void isSafeUseOfAllocation(Instruction *User, AllocationInst *AI,
107 void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI,
109 void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI,
110 unsigned OpNo, AllocaInfo &Info);
111 void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI,
114 void DoScalarReplacement(AllocationInst *AI,
115 std::vector<AllocationInst*> &WorkList);
116 void CleanupGEP(GetElementPtrInst *GEP);
117 void CleanupAllocaUsers(AllocationInst *AI);
118 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base);
120 void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI,
121 SmallVector<AllocaInst*, 32> &NewElts);
123 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst,
125 SmallVector<AllocaInst*, 32> &NewElts);
126 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocationInst *AI,
127 SmallVector<AllocaInst*, 32> &NewElts);
128 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI,
129 SmallVector<AllocaInst*, 32> &NewElts);
131 bool CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy,
132 bool &SawVec, uint64_t Offset, unsigned AllocaSize);
133 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset);
134 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType,
135 uint64_t Offset, IRBuilder<> &Builder);
136 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
137 uint64_t Offset, IRBuilder<> &Builder);
138 static Instruction *isOnlyCopiedFromConstantGlobal(AllocationInst *AI);
143 static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates");
145 // Public interface to the ScalarReplAggregates pass
146 FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) {
147 return new SROA(Threshold);
151 bool SROA::runOnFunction(Function &F) {
152 TD = &getAnalysis<TargetData>();
154 bool Changed = performPromotion(F);
156 bool LocalChange = performScalarRepl(F);
157 if (!LocalChange) break; // No need to repromote if no scalarrepl
159 LocalChange = performPromotion(F);
160 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
167 bool SROA::performPromotion(Function &F) {
168 std::vector<AllocaInst*> Allocas;
169 DominatorTree &DT = getAnalysis<DominatorTree>();
170 DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
172 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
174 bool Changed = false;
179 // Find allocas that are safe to promote, by looking at all instructions in
181 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
182 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
183 if (isAllocaPromotable(AI))
184 Allocas.push_back(AI);
186 if (Allocas.empty()) break;
188 PromoteMemToReg(Allocas, DT, DF);
189 NumPromoted += Allocas.size();
196 /// getNumSAElements - Return the number of elements in the specific struct or
198 static uint64_t getNumSAElements(const Type *T) {
199 if (const StructType *ST = dyn_cast<StructType>(T))
200 return ST->getNumElements();
201 return cast<ArrayType>(T)->getNumElements();
204 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
205 // which runs on all of the malloc/alloca instructions in the function, removing
206 // them if they are only used by getelementptr instructions.
208 bool SROA::performScalarRepl(Function &F) {
209 std::vector<AllocationInst*> WorkList;
211 // Scan the entry basic block, adding any alloca's and mallocs to the worklist
212 BasicBlock &BB = F.getEntryBlock();
213 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
214 if (AllocationInst *A = dyn_cast<AllocationInst>(I))
215 WorkList.push_back(A);
217 // Process the worklist
218 bool Changed = false;
219 while (!WorkList.empty()) {
220 AllocationInst *AI = WorkList.back();
223 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
224 // with unused elements.
225 if (AI->use_empty()) {
226 AI->eraseFromParent();
230 // If this alloca is impossible for us to promote, reject it early.
231 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
234 // Check to see if this allocation is only modified by a memcpy/memmove from
235 // a constant global. If this is the case, we can change all users to use
236 // the constant global instead. This is commonly produced by the CFE by
237 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
238 // is only subsequently read.
239 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) {
240 DOUT << "Found alloca equal to global: " << *AI;
241 DOUT << " memcpy = " << *TheCopy;
242 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2));
243 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
244 TheCopy->eraseFromParent(); // Don't mutate the global.
245 AI->eraseFromParent();
251 // Check to see if we can perform the core SROA transformation. We cannot
252 // transform the allocation instruction if it is an array allocation
253 // (allocations OF arrays are ok though), and an allocation of a scalar
254 // value cannot be decomposed at all.
255 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
257 // Do not promote any struct whose size is too big.
258 if (AllocaSize > SRThreshold) continue;
260 if ((isa<StructType>(AI->getAllocatedType()) ||
261 isa<ArrayType>(AI->getAllocatedType())) &&
262 // Do not promote any struct into more than "32" separate vars.
263 getNumSAElements(AI->getAllocatedType()) <= SRThreshold/4) {
264 // Check that all of the users of the allocation are capable of being
266 switch (isSafeAllocaToScalarRepl(AI)) {
267 default: assert(0 && "Unexpected value!");
268 case 0: // Not safe to scalar replace.
270 case 1: // Safe, but requires cleanup/canonicalizations first
271 CleanupAllocaUsers(AI);
273 case 3: // Safe to scalar replace.
274 DoScalarReplacement(AI, WorkList);
280 // If we can turn this aggregate value (potentially with casts) into a
281 // simple scalar value that can be mem2reg'd into a register value.
282 // IsNotTrivial tracks whether this is something that mem2reg could have
283 // promoted itself. If so, we don't want to transform it needlessly. Note
284 // that we can't just check based on the type: the alloca may be of an i32
285 // but that has pointer arithmetic to set byte 3 of it or something.
286 bool IsNotTrivial = false;
287 const Type *VectorTy = 0;
288 bool HadAVector = false;
289 if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector,
290 0, unsigned(AllocaSize)) && IsNotTrivial) {
292 // If we were able to find a vector type that can handle this with
293 // insert/extract elements, and if there was at least one use that had
294 // a vector type, promote this to a vector. We don't want to promote
295 // random stuff that doesn't use vectors (e.g. <9 x double>) because then
296 // we just get a lot of insert/extracts. If at least one vector is
297 // involved, then we probably really do have a union of vector/array.
298 if (VectorTy && isa<VectorType>(VectorTy) && HadAVector) {
299 DOUT << "CONVERT TO VECTOR: " << *AI << " TYPE = " << *VectorTy <<"\n";
301 // Create and insert the vector alloca.
302 NewAI = new AllocaInst(VectorTy, 0, "", AI->getParent()->begin());
303 ConvertUsesToScalar(AI, NewAI, 0);
305 DOUT << "CONVERT TO SCALAR INTEGER: " << *AI << "\n";
307 // Create and insert the integer alloca.
308 const Type *NewTy = IntegerType::get(AllocaSize*8);
309 NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
310 ConvertUsesToScalar(AI, NewAI, 0);
313 AI->eraseFromParent();
319 // Otherwise, couldn't process this alloca.
325 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
326 /// predicate, do SROA now.
327 void SROA::DoScalarReplacement(AllocationInst *AI,
328 std::vector<AllocationInst*> &WorkList) {
329 DOUT << "Found inst to SROA: " << *AI;
330 SmallVector<AllocaInst*, 32> ElementAllocas;
331 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
332 ElementAllocas.reserve(ST->getNumContainedTypes());
333 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
334 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
336 AI->getName() + "." + utostr(i), AI);
337 ElementAllocas.push_back(NA);
338 WorkList.push_back(NA); // Add to worklist for recursive processing
341 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
342 ElementAllocas.reserve(AT->getNumElements());
343 const Type *ElTy = AT->getElementType();
344 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
345 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
346 AI->getName() + "." + utostr(i), AI);
347 ElementAllocas.push_back(NA);
348 WorkList.push_back(NA); // Add to worklist for recursive processing
352 // Now that we have created the alloca instructions that we want to use,
353 // expand the getelementptr instructions to use them.
355 while (!AI->use_empty()) {
356 Instruction *User = cast<Instruction>(AI->use_back());
357 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) {
358 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas);
359 BCInst->eraseFromParent();
364 // %res = load { i32, i32 }* %alloc
366 // %load.0 = load i32* %alloc.0
367 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
368 // %load.1 = load i32* %alloc.1
369 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
370 // (Also works for arrays instead of structs)
371 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
372 Value *Insert = UndefValue::get(LI->getType());
373 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) {
374 Value *Load = new LoadInst(ElementAllocas[i], "load", LI);
375 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI);
377 LI->replaceAllUsesWith(Insert);
378 LI->eraseFromParent();
383 // store { i32, i32 } %val, { i32, i32 }* %alloc
385 // %val.0 = extractvalue { i32, i32 } %val, 0
386 // store i32 %val.0, i32* %alloc.0
387 // %val.1 = extractvalue { i32, i32 } %val, 1
388 // store i32 %val.1, i32* %alloc.1
389 // (Also works for arrays instead of structs)
390 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
391 Value *Val = SI->getOperand(0);
392 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) {
393 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI);
394 new StoreInst(Extract, ElementAllocas[i], SI);
396 SI->eraseFromParent();
400 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
401 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst>
403 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
405 assert(Idx < ElementAllocas.size() && "Index out of range?");
406 AllocaInst *AllocaToUse = ElementAllocas[Idx];
409 if (GEPI->getNumOperands() == 3) {
410 // Do not insert a new getelementptr instruction with zero indices, only
411 // to have it optimized out later.
412 RepValue = AllocaToUse;
414 // We are indexing deeply into the structure, so we still need a
415 // getelement ptr instruction to finish the indexing. This may be
416 // expanded itself once the worklist is rerun.
418 SmallVector<Value*, 8> NewArgs;
419 NewArgs.push_back(Constant::getNullValue(Type::Int32Ty));
420 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end());
421 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(),
422 NewArgs.end(), "", GEPI);
423 RepValue->takeName(GEPI);
426 // If this GEP is to the start of the aggregate, check for memcpys.
427 if (Idx == 0 && GEPI->hasAllZeroIndices())
428 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas);
430 // Move all of the users over to the new GEP.
431 GEPI->replaceAllUsesWith(RepValue);
432 // Delete the old GEP
433 GEPI->eraseFromParent();
436 // Finally, delete the Alloca instruction
437 AI->eraseFromParent();
442 /// isSafeElementUse - Check to see if this use is an allowed use for a
443 /// getelementptr instruction of an array aggregate allocation. isFirstElt
444 /// indicates whether Ptr is known to the start of the aggregate.
446 void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI,
448 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
450 Instruction *User = cast<Instruction>(*I);
451 switch (User->getOpcode()) {
452 case Instruction::Load: break;
453 case Instruction::Store:
454 // Store is ok if storing INTO the pointer, not storing the pointer
455 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info);
457 case Instruction::GetElementPtr: {
458 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User);
459 bool AreAllZeroIndices = isFirstElt;
460 if (GEP->getNumOperands() > 1) {
461 if (!isa<ConstantInt>(GEP->getOperand(1)) ||
462 !cast<ConstantInt>(GEP->getOperand(1))->isZero())
463 // Using pointer arithmetic to navigate the array.
464 return MarkUnsafe(Info);
466 if (AreAllZeroIndices)
467 AreAllZeroIndices = GEP->hasAllZeroIndices();
469 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info);
470 if (Info.isUnsafe) return;
473 case Instruction::BitCast:
475 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info);
476 if (Info.isUnsafe) return;
479 DOUT << " Transformation preventing inst: " << *User;
480 return MarkUnsafe(Info);
481 case Instruction::Call:
482 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
484 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info);
485 if (Info.isUnsafe) return;
489 DOUT << " Transformation preventing inst: " << *User;
490 return MarkUnsafe(Info);
492 DOUT << " Transformation preventing inst: " << *User;
493 return MarkUnsafe(Info);
496 return; // All users look ok :)
499 /// AllUsersAreLoads - Return true if all users of this value are loads.
500 static bool AllUsersAreLoads(Value *Ptr) {
501 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
503 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load)
508 /// isSafeUseOfAllocation - Check to see if this user is an allowed use for an
509 /// aggregate allocation.
511 void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI,
513 if (BitCastInst *C = dyn_cast<BitCastInst>(User))
514 return isSafeUseOfBitCastedAllocation(C, AI, Info);
516 if (LoadInst *LI = dyn_cast<LoadInst>(User))
517 if (!LI->isVolatile())
518 return;// Loads (returning a first class aggregrate) are always rewritable
520 if (StoreInst *SI = dyn_cast<StoreInst>(User))
521 if (!SI->isVolatile() && SI->getOperand(0) != AI)
522 return;// Store is ok if storing INTO the pointer, not storing the pointer
524 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User);
526 return MarkUnsafe(Info);
528 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI);
530 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>".
532 I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) {
533 return MarkUnsafe(Info);
537 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices??
539 bool IsAllZeroIndices = true;
541 // If the first index is a non-constant index into an array, see if we can
542 // handle it as a special case.
543 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
544 if (!isa<ConstantInt>(I.getOperand())) {
545 IsAllZeroIndices = 0;
546 uint64_t NumElements = AT->getNumElements();
548 // If this is an array index and the index is not constant, we cannot
549 // promote... that is unless the array has exactly one or two elements in
550 // it, in which case we CAN promote it, but we have to canonicalize this
551 // out if this is the only problem.
552 if ((NumElements == 1 || NumElements == 2) &&
553 AllUsersAreLoads(GEPI)) {
554 Info.needsCleanup = true;
555 return; // Canonicalization required!
557 return MarkUnsafe(Info);
561 // Walk through the GEP type indices, checking the types that this indexes
563 for (; I != E; ++I) {
564 // Ignore struct elements, no extra checking needed for these.
565 if (isa<StructType>(*I))
568 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand());
569 if (!IdxVal) return MarkUnsafe(Info);
571 // Are all indices still zero?
572 IsAllZeroIndices &= IdxVal->isZero();
574 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
575 // This GEP indexes an array. Verify that this is an in-range constant
576 // integer. Specifically, consider A[0][i]. We cannot know that the user
577 // isn't doing invalid things like allowing i to index an out-of-range
578 // subscript that accesses A[1]. Because of this, we have to reject SROA
579 // of any accesses into structs where any of the components are variables.
580 if (IdxVal->getZExtValue() >= AT->getNumElements())
581 return MarkUnsafe(Info);
582 } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) {
583 if (IdxVal->getZExtValue() >= VT->getNumElements())
584 return MarkUnsafe(Info);
588 // If there are any non-simple uses of this getelementptr, make sure to reject
590 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info);
593 /// isSafeMemIntrinsicOnAllocation - Return true if the specified memory
594 /// intrinsic can be promoted by SROA. At this point, we know that the operand
595 /// of the memintrinsic is a pointer to the beginning of the allocation.
596 void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI,
597 unsigned OpNo, AllocaInfo &Info) {
598 // If not constant length, give up.
599 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
600 if (!Length) return MarkUnsafe(Info);
602 // If not the whole aggregate, give up.
603 if (Length->getZExtValue() !=
604 TD->getTypeAllocSize(AI->getType()->getElementType()))
605 return MarkUnsafe(Info);
607 // We only know about memcpy/memset/memmove.
608 if (!isa<MemIntrinsic>(MI))
609 return MarkUnsafe(Info);
611 // Otherwise, we can transform it. Determine whether this is a memcpy/set
612 // into or out of the aggregate.
614 Info.isMemCpyDst = true;
617 Info.isMemCpySrc = true;
621 /// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast
623 void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI,
625 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end();
627 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) {
628 isSafeUseOfBitCastedAllocation(BCU, AI, Info);
629 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) {
630 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info);
631 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
632 if (SI->isVolatile())
633 return MarkUnsafe(Info);
635 // If storing the entire alloca in one chunk through a bitcasted pointer
636 // to integer, we can transform it. This happens (for example) when you
637 // cast a {i32,i32}* to i64* and store through it. This is similar to the
638 // memcpy case and occurs in various "byval" cases and emulated memcpys.
639 if (isa<IntegerType>(SI->getOperand(0)->getType()) &&
640 TD->getTypeAllocSize(SI->getOperand(0)->getType()) ==
641 TD->getTypeAllocSize(AI->getType()->getElementType())) {
642 Info.isMemCpyDst = true;
645 return MarkUnsafe(Info);
646 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) {
647 if (LI->isVolatile())
648 return MarkUnsafe(Info);
650 // If loading the entire alloca in one chunk through a bitcasted pointer
651 // to integer, we can transform it. This happens (for example) when you
652 // cast a {i32,i32}* to i64* and load through it. This is similar to the
653 // memcpy case and occurs in various "byval" cases and emulated memcpys.
654 if (isa<IntegerType>(LI->getType()) &&
655 TD->getTypeAllocSize(LI->getType()) ==
656 TD->getTypeAllocSize(AI->getType()->getElementType())) {
657 Info.isMemCpySrc = true;
660 return MarkUnsafe(Info);
661 } else if (isa<DbgInfoIntrinsic>(UI)) {
662 // If one user is DbgInfoIntrinsic then check if all users are
663 // DbgInfoIntrinsics.
664 if (OnlyUsedByDbgInfoIntrinsics(BC)) {
665 Info.needsCleanup = true;
672 return MarkUnsafe(Info);
674 if (Info.isUnsafe) return;
678 /// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes
679 /// to its first element. Transform users of the cast to use the new values
681 void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI,
682 SmallVector<AllocaInst*, 32> &NewElts) {
683 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end();
685 Instruction *User = cast<Instruction>(*UI++);
686 if (BitCastInst *BCU = dyn_cast<BitCastInst>(User)) {
687 RewriteBitCastUserOfAlloca(BCU, AI, NewElts);
688 if (BCU->use_empty()) BCU->eraseFromParent();
692 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
693 // This must be memcpy/memmove/memset of the entire aggregate.
694 // Split into one per element.
695 RewriteMemIntrinUserOfAlloca(MI, BCInst, AI, NewElts);
699 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
700 // If this is a store of the entire alloca from an integer, rewrite it.
701 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
705 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
706 // If this is a load of the entire alloca to an integer, rewrite it.
707 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
711 // Otherwise it must be some other user of a gep of the first pointer. Just
712 // leave these alone.
717 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
718 /// Rewrite it to copy or set the elements of the scalarized memory.
719 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst,
721 SmallVector<AllocaInst*, 32> &NewElts) {
723 // If this is a memcpy/memmove, construct the other pointer as the
724 // appropriate type. The "Other" pointer is the pointer that goes to memory
725 // that doesn't have anything to do with the alloca that we are promoting. For
726 // memset, this Value* stays null.
728 unsigned MemAlignment = MI->getAlignment();
729 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy
730 if (BCInst == MTI->getRawDest())
731 OtherPtr = MTI->getRawSource();
733 assert(BCInst == MTI->getRawSource());
734 OtherPtr = MTI->getRawDest();
738 // If there is an other pointer, we want to convert it to the same pointer
739 // type as AI has, so we can GEP through it safely.
741 // It is likely that OtherPtr is a bitcast, if so, remove it.
742 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr))
743 OtherPtr = BC->getOperand(0);
744 // All zero GEPs are effectively bitcasts.
745 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr))
746 if (GEP->hasAllZeroIndices())
747 OtherPtr = GEP->getOperand(0);
749 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr))
750 if (BCE->getOpcode() == Instruction::BitCast)
751 OtherPtr = BCE->getOperand(0);
753 // If the pointer is not the right type, insert a bitcast to the right
755 if (OtherPtr->getType() != AI->getType())
756 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(),
760 // Process each element of the aggregate.
761 Value *TheFn = MI->getOperand(0);
762 const Type *BytePtrTy = MI->getRawDest()->getType();
763 bool SROADest = MI->getRawDest() == BCInst;
765 Constant *Zero = Constant::getNullValue(Type::Int32Ty);
767 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
768 // If this is a memcpy/memmove, emit a GEP of the other element address.
770 unsigned OtherEltAlign = MemAlignment;
773 Value *Idx[2] = { Zero, ConstantInt::get(Type::Int32Ty, i) };
774 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2,
775 OtherPtr->getNameStr()+"."+utostr(i),
778 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
779 if (const StructType *ST =
780 dyn_cast<StructType>(OtherPtrTy->getElementType())) {
781 EltOffset = TD->getStructLayout(ST)->getElementOffset(i);
784 cast<SequentialType>(OtherPtr->getType())->getElementType();
785 EltOffset = TD->getTypeAllocSize(EltTy)*i;
788 // The alignment of the other pointer is the guaranteed alignment of the
789 // element, which is affected by both the known alignment of the whole
790 // mem intrinsic and the alignment of the element. If the alignment of
791 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
792 // known alignment is just 4 bytes.
793 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
796 Value *EltPtr = NewElts[i];
797 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
799 // If we got down to a scalar, insert a load or store as appropriate.
800 if (EltTy->isSingleValueType()) {
801 if (isa<MemTransferInst>(MI)) {
803 // From Other to Alloca.
804 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
805 new StoreInst(Elt, EltPtr, MI);
807 // From Alloca to Other.
808 Value *Elt = new LoadInst(EltPtr, "tmp", MI);
809 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI);
813 assert(isa<MemSetInst>(MI));
815 // If the stored element is zero (common case), just store a null
818 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) {
820 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
822 // If EltTy is a vector type, get the element type.
823 const Type *ValTy = EltTy;
824 if (const VectorType *VTy = dyn_cast<VectorType>(ValTy))
825 ValTy = VTy->getElementType();
827 // Construct an integer with the right value.
828 unsigned EltSize = TD->getTypeSizeInBits(ValTy);
829 APInt OneVal(EltSize, CI->getZExtValue());
830 APInt TotalVal(OneVal);
832 for (unsigned i = 0; 8*i < EltSize; ++i) {
833 TotalVal = TotalVal.shl(8);
837 // Convert the integer value to the appropriate type.
838 StoreVal = ConstantInt::get(TotalVal);
839 if (isa<PointerType>(ValTy))
840 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
841 else if (ValTy->isFloatingPoint())
842 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
843 assert(StoreVal->getType() == ValTy && "Type mismatch!");
845 // If the requested value was a vector constant, create it.
846 if (EltTy != ValTy) {
847 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements();
848 SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
849 StoreVal = ConstantVector::get(&Elts[0], NumElts);
852 new StoreInst(StoreVal, EltPtr, MI);
855 // Otherwise, if we're storing a byte variable, use a memset call for
859 // Cast the element pointer to BytePtrTy.
860 if (EltPtr->getType() != BytePtrTy)
861 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI);
863 // Cast the other pointer (if we have one) to BytePtrTy.
864 if (OtherElt && OtherElt->getType() != BytePtrTy)
865 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(),
868 unsigned EltSize = TD->getTypeAllocSize(EltTy);
870 // Finally, insert the meminst for this element.
871 if (isa<MemTransferInst>(MI)) {
873 SROADest ? EltPtr : OtherElt, // Dest ptr
874 SROADest ? OtherElt : EltPtr, // Src ptr
875 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
876 ConstantInt::get(Type::Int32Ty, OtherEltAlign) // Align
878 CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
880 assert(isa<MemSetInst>(MI));
882 EltPtr, MI->getOperand(2), // Dest, Value,
883 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
886 CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
889 MI->eraseFromParent();
892 /// RewriteStoreUserOfWholeAlloca - We found an store of an integer that
893 /// overwrites the entire allocation. Extract out the pieces of the stored
894 /// integer and store them individually.
895 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI,
897 SmallVector<AllocaInst*, 32> &NewElts){
898 // Extract each element out of the integer according to its structure offset
899 // and store the element value to the individual alloca.
900 Value *SrcVal = SI->getOperand(0);
901 const Type *AllocaEltTy = AI->getType()->getElementType();
902 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
904 // If this isn't a store of an integer to the whole alloca, it may be a store
905 // to the first element. Just ignore the store in this case and normal SROA
906 // will handle it. We don't handle types here that have tail padding, like
907 // an alloca of type {i1}.
908 if (!isa<IntegerType>(SrcVal->getType()) ||
909 TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
912 DOUT << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << *SI;
914 // There are two forms here: AI could be an array or struct. Both cases
915 // have different ways to compute the element offset.
916 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
917 const StructLayout *Layout = TD->getStructLayout(EltSTy);
919 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
920 // Get the number of bits to shift SrcVal to get the value.
921 const Type *FieldTy = EltSTy->getElementType(i);
922 uint64_t Shift = Layout->getElementOffsetInBits(i);
924 if (TD->isBigEndian())
925 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy);
927 Value *EltVal = SrcVal;
929 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
930 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
931 "sroa.store.elt", SI);
934 // Truncate down to an integer of the right size.
935 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
937 // Ignore zero sized fields like {}, they obviously contain no data.
938 if (FieldSizeBits == 0) continue;
940 if (FieldSizeBits != AllocaSizeBits)
941 EltVal = new TruncInst(EltVal, IntegerType::get(FieldSizeBits), "", SI);
942 Value *DestField = NewElts[i];
943 if (EltVal->getType() == FieldTy) {
944 // Storing to an integer field of this size, just do it.
945 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) {
946 // Bitcast to the right element type (for fp/vector values).
947 EltVal = new BitCastInst(EltVal, FieldTy, "", SI);
949 // Otherwise, bitcast the dest pointer (for aggregates).
950 DestField = new BitCastInst(DestField,
951 PointerType::getUnqual(EltVal->getType()),
954 new StoreInst(EltVal, DestField, SI);
958 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
959 const Type *ArrayEltTy = ATy->getElementType();
960 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
961 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
965 if (TD->isBigEndian())
966 Shift = AllocaSizeBits-ElementOffset;
970 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
971 // Ignore zero sized fields like {}, they obviously contain no data.
972 if (ElementSizeBits == 0) continue;
974 Value *EltVal = SrcVal;
976 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
977 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
978 "sroa.store.elt", SI);
981 // Truncate down to an integer of the right size.
982 if (ElementSizeBits != AllocaSizeBits)
983 EltVal = new TruncInst(EltVal, IntegerType::get(ElementSizeBits),"",SI);
984 Value *DestField = NewElts[i];
985 if (EltVal->getType() == ArrayEltTy) {
986 // Storing to an integer field of this size, just do it.
987 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) {
988 // Bitcast to the right element type (for fp/vector values).
989 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI);
991 // Otherwise, bitcast the dest pointer (for aggregates).
992 DestField = new BitCastInst(DestField,
993 PointerType::getUnqual(EltVal->getType()),
996 new StoreInst(EltVal, DestField, SI);
998 if (TD->isBigEndian())
999 Shift -= ElementOffset;
1001 Shift += ElementOffset;
1005 SI->eraseFromParent();
1008 /// RewriteLoadUserOfWholeAlloca - We found an load of the entire allocation to
1009 /// an integer. Load the individual pieces to form the aggregate value.
1010 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI,
1011 SmallVector<AllocaInst*, 32> &NewElts) {
1012 // Extract each element out of the NewElts according to its structure offset
1013 // and form the result value.
1014 const Type *AllocaEltTy = AI->getType()->getElementType();
1015 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
1017 // If this isn't a load of the whole alloca to an integer, it may be a load
1018 // of the first element. Just ignore the load in this case and normal SROA
1019 // will handle it. We don't handle types here that have tail padding, like
1020 // an alloca of type {i1}.
1021 if (!isa<IntegerType>(LI->getType()) ||
1022 TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
1025 DOUT << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << *LI;
1027 // There are two forms here: AI could be an array or struct. Both cases
1028 // have different ways to compute the element offset.
1029 const StructLayout *Layout = 0;
1030 uint64_t ArrayEltBitOffset = 0;
1031 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
1032 Layout = TD->getStructLayout(EltSTy);
1034 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
1035 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
1038 Value *ResultVal = Constant::getNullValue(LI->getType());
1040 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1041 // Load the value from the alloca. If the NewElt is an aggregate, cast
1042 // the pointer to an integer of the same size before doing the load.
1043 Value *SrcField = NewElts[i];
1044 const Type *FieldTy =
1045 cast<PointerType>(SrcField->getType())->getElementType();
1046 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
1048 // Ignore zero sized fields like {}, they obviously contain no data.
1049 if (FieldSizeBits == 0) continue;
1051 const IntegerType *FieldIntTy = IntegerType::get(FieldSizeBits);
1052 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() &&
1053 !isa<VectorType>(FieldTy))
1054 SrcField = new BitCastInst(SrcField, PointerType::getUnqual(FieldIntTy),
1056 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
1058 // If SrcField is a fp or vector of the right size but that isn't an
1059 // integer type, bitcast to an integer so we can shift it.
1060 if (SrcField->getType() != FieldIntTy)
1061 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
1063 // Zero extend the field to be the same size as the final alloca so that
1064 // we can shift and insert it.
1065 if (SrcField->getType() != ResultVal->getType())
1066 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
1068 // Determine the number of bits to shift SrcField.
1070 if (Layout) // Struct case.
1071 Shift = Layout->getElementOffsetInBits(i);
1073 Shift = i*ArrayEltBitOffset;
1075 if (TD->isBigEndian())
1076 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
1079 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
1080 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
1083 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
1086 LI->replaceAllUsesWith(ResultVal);
1087 LI->eraseFromParent();
1091 /// HasPadding - Return true if the specified type has any structure or
1092 /// alignment padding, false otherwise.
1093 static bool HasPadding(const Type *Ty, const TargetData &TD) {
1094 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1095 const StructLayout *SL = TD.getStructLayout(STy);
1096 unsigned PrevFieldBitOffset = 0;
1097 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1098 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
1100 // Padding in sub-elements?
1101 if (HasPadding(STy->getElementType(i), TD))
1104 // Check to see if there is any padding between this element and the
1107 unsigned PrevFieldEnd =
1108 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
1109 if (PrevFieldEnd < FieldBitOffset)
1113 PrevFieldBitOffset = FieldBitOffset;
1116 // Check for tail padding.
1117 if (unsigned EltCount = STy->getNumElements()) {
1118 unsigned PrevFieldEnd = PrevFieldBitOffset +
1119 TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
1120 if (PrevFieldEnd < SL->getSizeInBits())
1124 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1125 return HasPadding(ATy->getElementType(), TD);
1126 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1127 return HasPadding(VTy->getElementType(), TD);
1129 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
1132 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
1133 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
1134 /// or 1 if safe after canonicalization has been performed.
1136 int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) {
1137 // Loop over the use list of the alloca. We can only transform it if all of
1138 // the users are safe to transform.
1141 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end();
1143 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info);
1144 if (Info.isUnsafe) {
1145 DOUT << "Cannot transform: " << *AI << " due to user: " << **I;
1150 // Okay, we know all the users are promotable. If the aggregate is a memcpy
1151 // source and destination, we have to be careful. In particular, the memcpy
1152 // could be moving around elements that live in structure padding of the LLVM
1153 // types, but may actually be used. In these cases, we refuse to promote the
1155 if (Info.isMemCpySrc && Info.isMemCpyDst &&
1156 HasPadding(AI->getType()->getElementType(), *TD))
1159 // If we require cleanup, return 1, otherwise return 3.
1160 return Info.needsCleanup ? 1 : 3;
1163 /// CleanupGEP - GEP is used by an Alloca, which can be prompted after the GEP
1164 /// is canonicalized here.
1165 void SROA::CleanupGEP(GetElementPtrInst *GEPI) {
1166 gep_type_iterator I = gep_type_begin(GEPI);
1169 const ArrayType *AT = dyn_cast<ArrayType>(*I);
1173 uint64_t NumElements = AT->getNumElements();
1175 if (isa<ConstantInt>(I.getOperand()))
1178 if (NumElements == 1) {
1179 GEPI->setOperand(2, Constant::getNullValue(Type::Int32Ty));
1183 assert(NumElements == 2 && "Unhandled case!");
1184 // All users of the GEP must be loads. At each use of the GEP, insert
1185 // two loads of the appropriate indexed GEP and select between them.
1186 Value *IsOne = new ICmpInst(ICmpInst::ICMP_NE, I.getOperand(),
1187 Constant::getNullValue(I.getOperand()->getType()),
1189 // Insert the new GEP instructions, which are properly indexed.
1190 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end());
1191 Indices[1] = Constant::getNullValue(Type::Int32Ty);
1192 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
1195 GEPI->getName()+".0", GEPI);
1196 Indices[1] = ConstantInt::get(Type::Int32Ty, 1);
1197 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
1200 GEPI->getName()+".1", GEPI);
1201 // Replace all loads of the variable index GEP with loads from both
1202 // indexes and a select.
1203 while (!GEPI->use_empty()) {
1204 LoadInst *LI = cast<LoadInst>(GEPI->use_back());
1205 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI);
1206 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI);
1207 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI);
1208 LI->replaceAllUsesWith(R);
1209 LI->eraseFromParent();
1211 GEPI->eraseFromParent();
1215 /// CleanupAllocaUsers - If SROA reported that it can promote the specified
1216 /// allocation, but only if cleaned up, perform the cleanups required.
1217 void SROA::CleanupAllocaUsers(AllocationInst *AI) {
1218 // At this point, we know that the end result will be SROA'd and promoted, so
1219 // we can insert ugly code if required so long as sroa+mem2reg will clean it
1221 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
1224 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U))
1226 else if (Instruction *I = dyn_cast<Instruction>(U)) {
1227 SmallVector<DbgInfoIntrinsic *, 2> DbgInUses;
1228 if (!isa<StoreInst>(I) && OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) {
1229 // Safe to remove debug info uses.
1230 while (!DbgInUses.empty()) {
1231 DbgInfoIntrinsic *DI = DbgInUses.back(); DbgInUses.pop_back();
1232 DI->eraseFromParent();
1234 I->eraseFromParent();
1240 /// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at
1241 /// the offset specified by Offset (which is specified in bytes).
1243 /// There are two cases we handle here:
1244 /// 1) A union of vector types of the same size and potentially its elements.
1245 /// Here we turn element accesses into insert/extract element operations.
1246 /// This promotes a <4 x float> with a store of float to the third element
1247 /// into a <4 x float> that uses insert element.
1248 /// 2) A fully general blob of memory, which we turn into some (potentially
1249 /// large) integer type with extract and insert operations where the loads
1250 /// and stores would mutate the memory.
1251 static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy,
1252 unsigned AllocaSize, const TargetData &TD) {
1253 // If this could be contributing to a vector, analyze it.
1254 if (VecTy != Type::VoidTy) { // either null or a vector type.
1256 // If the In type is a vector that is the same size as the alloca, see if it
1257 // matches the existing VecTy.
1258 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) {
1259 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) {
1260 // If we're storing/loading a vector of the right size, allow it as a
1261 // vector. If this the first vector we see, remember the type so that
1262 // we know the element size.
1267 } else if (In == Type::FloatTy || In == Type::DoubleTy ||
1268 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 &&
1269 isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
1270 // If we're accessing something that could be an element of a vector, see
1271 // if the implied vector agrees with what we already have and if Offset is
1272 // compatible with it.
1273 unsigned EltSize = In->getPrimitiveSizeInBits()/8;
1274 if (Offset % EltSize == 0 &&
1275 AllocaSize % EltSize == 0 &&
1277 cast<VectorType>(VecTy)->getElementType()
1278 ->getPrimitiveSizeInBits()/8 == EltSize)) {
1280 VecTy = VectorType::get(In, AllocaSize/EltSize);
1286 // Otherwise, we have a case that we can't handle with an optimized vector
1287 // form. We can still turn this into a large integer.
1288 VecTy = Type::VoidTy;
1291 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
1292 /// its accesses to use a to single vector type, return true, and set VecTy to
1293 /// the new type. If we could convert the alloca into a single promotable
1294 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a
1295 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
1296 /// is the current offset from the base of the alloca being analyzed.
1298 /// If we see at least one access to the value that is as a vector type, set the
1301 bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy,
1302 bool &SawVec, uint64_t Offset,
1303 unsigned AllocaSize) {
1304 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1305 Instruction *User = cast<Instruction>(*UI);
1307 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1308 // Don't break volatile loads.
1309 if (LI->isVolatile())
1311 MergeInType(LI->getType(), Offset, VecTy, AllocaSize, *TD);
1312 SawVec |= isa<VectorType>(LI->getType());
1316 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1317 // Storing the pointer, not into the value?
1318 if (SI->getOperand(0) == V || SI->isVolatile()) return 0;
1319 MergeInType(SI->getOperand(0)->getType(), Offset, VecTy, AllocaSize, *TD);
1320 SawVec |= isa<VectorType>(SI->getOperand(0)->getType());
1324 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
1325 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset,
1328 IsNotTrivial = true;
1332 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1333 // If this is a GEP with a variable indices, we can't handle it.
1334 if (!GEP->hasAllConstantIndices())
1337 // Compute the offset that this GEP adds to the pointer.
1338 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
1339 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(),
1340 &Indices[0], Indices.size());
1341 // See if all uses can be converted.
1342 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset,
1345 IsNotTrivial = true;
1349 // If this is a constant sized memset of a constant value (e.g. 0) we can
1351 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
1352 // Store of constant value and constant size.
1353 if (isa<ConstantInt>(MSI->getValue()) &&
1354 isa<ConstantInt>(MSI->getLength())) {
1355 IsNotTrivial = true;
1360 // If this is a memcpy or memmove into or out of the whole allocation, we
1361 // can handle it like a load or store of the scalar type.
1362 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
1363 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()))
1364 if (Len->getZExtValue() == AllocaSize && Offset == 0) {
1365 IsNotTrivial = true;
1370 // Ignore dbg intrinsic.
1371 if (isa<DbgInfoIntrinsic>(User))
1374 // Otherwise, we cannot handle this!
1382 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
1383 /// directly. This happens when we are converting an "integer union" to a
1384 /// single integer scalar, or when we are converting a "vector union" to a
1385 /// vector with insert/extractelement instructions.
1387 /// Offset is an offset from the original alloca, in bits that need to be
1388 /// shifted to the right. By the end of this, there should be no uses of Ptr.
1389 void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) {
1390 while (!Ptr->use_empty()) {
1391 Instruction *User = cast<Instruction>(Ptr->use_back());
1393 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
1394 ConvertUsesToScalar(CI, NewAI, Offset);
1395 CI->eraseFromParent();
1399 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1400 // Compute the offset that this GEP adds to the pointer.
1401 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
1402 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(),
1403 &Indices[0], Indices.size());
1404 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
1405 GEP->eraseFromParent();
1409 IRBuilder<> Builder(User->getParent(), User);
1411 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1412 // The load is a bit extract from NewAI shifted right by Offset bits.
1413 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp");
1415 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
1416 LI->replaceAllUsesWith(NewLoadVal);
1417 LI->eraseFromParent();
1421 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1422 assert(SI->getOperand(0) != Ptr && "Consistency error!");
1423 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").c_str());
1424 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
1426 Builder.CreateStore(New, NewAI);
1427 SI->eraseFromParent();
1431 // If this is a constant sized memset of a constant value (e.g. 0) we can
1432 // transform it into a store of the expanded constant value.
1433 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
1434 assert(MSI->getRawDest() == Ptr && "Consistency error!");
1435 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
1436 if (NumBytes != 0) {
1437 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
1439 // Compute the value replicated the right number of times.
1440 APInt APVal(NumBytes*8, Val);
1442 // Splat the value if non-zero.
1444 for (unsigned i = 1; i != NumBytes; ++i)
1445 APVal |= APVal << 8;
1447 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").c_str());
1448 Value *New = ConvertScalar_InsertValue(ConstantInt::get(APVal), Old,
1450 Builder.CreateStore(New, NewAI);
1452 MSI->eraseFromParent();
1456 // If this is a memcpy or memmove into or out of the whole allocation, we
1457 // can handle it like a load or store of the scalar type.
1458 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
1459 assert(Offset == 0 && "must be store to start of alloca");
1461 // If the source and destination are both to the same alloca, then this is
1462 // a noop copy-to-self, just delete it. Otherwise, emit a load and store
1464 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject());
1466 if (MTI->getSource()->getUnderlyingObject() != OrigAI) {
1467 // Dest must be OrigAI, change this to be a load from the original
1468 // pointer (bitcasted), then a store to our new alloca.
1469 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
1470 Value *SrcPtr = MTI->getSource();
1471 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType());
1473 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
1474 SrcVal->setAlignment(MTI->getAlignment());
1475 Builder.CreateStore(SrcVal, NewAI);
1476 } else if (MTI->getDest()->getUnderlyingObject() != OrigAI) {
1477 // Src must be OrigAI, change this to be a load from NewAI then a store
1478 // through the original dest pointer (bitcasted).
1479 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
1480 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
1482 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType());
1483 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
1484 NewStore->setAlignment(MTI->getAlignment());
1486 // Noop transfer. Src == Dst
1490 MTI->eraseFromParent();
1494 // If user is a dbg info intrinsic then it is safe to remove it.
1495 if (isa<DbgInfoIntrinsic>(User)) {
1496 User->eraseFromParent();
1500 assert(0 && "Unsupported operation!");
1505 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
1506 /// or vector value FromVal, extracting the bits from the offset specified by
1507 /// Offset. This returns the value, which is of type ToType.
1509 /// This happens when we are converting an "integer union" to a single
1510 /// integer scalar, or when we are converting a "vector union" to a vector with
1511 /// insert/extractelement instructions.
1513 /// Offset is an offset from the original alloca, in bits that need to be
1514 /// shifted to the right.
1515 Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType,
1516 uint64_t Offset, IRBuilder<> &Builder) {
1517 // If the load is of the whole new alloca, no conversion is needed.
1518 if (FromVal->getType() == ToType && Offset == 0)
1521 // If the result alloca is a vector type, this is either an element
1522 // access or a bitcast to another vector type of the same size.
1523 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) {
1524 if (isa<VectorType>(ToType))
1525 return Builder.CreateBitCast(FromVal, ToType, "tmp");
1527 // Otherwise it must be an element access.
1530 unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType());
1531 Elt = Offset/EltSize;
1532 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
1534 // Return the element extracted out of it.
1535 Value *V = Builder.CreateExtractElement(FromVal,
1536 ConstantInt::get(Type::Int32Ty,Elt),
1538 if (V->getType() != ToType)
1539 V = Builder.CreateBitCast(V, ToType, "tmp");
1543 // If ToType is a first class aggregate, extract out each of the pieces and
1544 // use insertvalue's to form the FCA.
1545 if (const StructType *ST = dyn_cast<StructType>(ToType)) {
1546 const StructLayout &Layout = *TD->getStructLayout(ST);
1547 Value *Res = UndefValue::get(ST);
1548 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
1549 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
1550 Offset+Layout.getElementOffsetInBits(i),
1552 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
1557 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
1558 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType());
1559 Value *Res = UndefValue::get(AT);
1560 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1561 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
1562 Offset+i*EltSize, Builder);
1563 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
1568 // Otherwise, this must be a union that was converted to an integer value.
1569 const IntegerType *NTy = cast<IntegerType>(FromVal->getType());
1571 // If this is a big-endian system and the load is narrower than the
1572 // full alloca type, we need to do a shift to get the right bits.
1574 if (TD->isBigEndian()) {
1575 // On big-endian machines, the lowest bit is stored at the bit offset
1576 // from the pointer given by getTypeStoreSizeInBits. This matters for
1577 // integers with a bitwidth that is not a multiple of 8.
1578 ShAmt = TD->getTypeStoreSizeInBits(NTy) -
1579 TD->getTypeStoreSizeInBits(ToType) - Offset;
1584 // Note: we support negative bitwidths (with shl) which are not defined.
1585 // We do this to support (f.e.) loads off the end of a structure where
1586 // only some bits are used.
1587 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
1588 FromVal = Builder.CreateLShr(FromVal, ConstantInt::get(FromVal->getType(),
1590 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
1591 FromVal = Builder.CreateShl(FromVal, ConstantInt::get(FromVal->getType(),
1594 // Finally, unconditionally truncate the integer to the right width.
1595 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType);
1596 if (LIBitWidth < NTy->getBitWidth())
1597 FromVal = Builder.CreateTrunc(FromVal, IntegerType::get(LIBitWidth), "tmp");
1598 else if (LIBitWidth > NTy->getBitWidth())
1599 FromVal = Builder.CreateZExt(FromVal, IntegerType::get(LIBitWidth), "tmp");
1601 // If the result is an integer, this is a trunc or bitcast.
1602 if (isa<IntegerType>(ToType)) {
1604 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) {
1605 // Just do a bitcast, we know the sizes match up.
1606 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp");
1608 // Otherwise must be a pointer.
1609 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp");
1611 assert(FromVal->getType() == ToType && "Didn't convert right?");
1616 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
1617 /// or vector value "Old" at the offset specified by Offset.
1619 /// This happens when we are converting an "integer union" to a
1620 /// single integer scalar, or when we are converting a "vector union" to a
1621 /// vector with insert/extractelement instructions.
1623 /// Offset is an offset from the original alloca, in bits that need to be
1624 /// shifted to the right.
1625 Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old,
1626 uint64_t Offset, IRBuilder<> &Builder) {
1628 // Convert the stored type to the actual type, shift it left to insert
1629 // then 'or' into place.
1630 const Type *AllocaType = Old->getType();
1632 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
1633 uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy);
1634 uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType());
1636 // Changing the whole vector with memset or with an access of a different
1638 if (ValSize == VecSize)
1639 return Builder.CreateBitCast(SV, AllocaType, "tmp");
1641 uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType());
1643 // Must be an element insertion.
1644 unsigned Elt = Offset/EltSize;
1646 if (SV->getType() != VTy->getElementType())
1647 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp");
1649 SV = Builder.CreateInsertElement(Old, SV,
1650 ConstantInt::get(Type::Int32Ty, Elt),
1655 // If SV is a first-class aggregate value, insert each value recursively.
1656 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) {
1657 const StructLayout &Layout = *TD->getStructLayout(ST);
1658 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
1659 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
1660 Old = ConvertScalar_InsertValue(Elt, Old,
1661 Offset+Layout.getElementOffsetInBits(i),
1667 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
1668 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType());
1669 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1670 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
1671 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
1676 // If SV is a float, convert it to the appropriate integer type.
1677 // If it is a pointer, do the same.
1678 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType());
1679 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType);
1680 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType());
1681 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType);
1682 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType()))
1683 SV = Builder.CreateBitCast(SV, IntegerType::get(SrcWidth), "tmp");
1684 else if (isa<PointerType>(SV->getType()))
1685 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(), "tmp");
1687 // Zero extend or truncate the value if needed.
1688 if (SV->getType() != AllocaType) {
1689 if (SV->getType()->getPrimitiveSizeInBits() <
1690 AllocaType->getPrimitiveSizeInBits())
1691 SV = Builder.CreateZExt(SV, AllocaType, "tmp");
1693 // Truncation may be needed if storing more than the alloca can hold
1694 // (undefined behavior).
1695 SV = Builder.CreateTrunc(SV, AllocaType, "tmp");
1696 SrcWidth = DestWidth;
1697 SrcStoreWidth = DestStoreWidth;
1701 // If this is a big-endian system and the store is narrower than the
1702 // full alloca type, we need to do a shift to get the right bits.
1704 if (TD->isBigEndian()) {
1705 // On big-endian machines, the lowest bit is stored at the bit offset
1706 // from the pointer given by getTypeStoreSizeInBits. This matters for
1707 // integers with a bitwidth that is not a multiple of 8.
1708 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
1713 // Note: we support negative bitwidths (with shr) which are not defined.
1714 // We do this to support (f.e.) stores off the end of a structure where
1715 // only some bits in the structure are set.
1716 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
1717 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
1718 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt), "tmp");
1720 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
1721 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt), "tmp");
1722 Mask = Mask.lshr(-ShAmt);
1725 // Mask out the bits we are about to insert from the old value, and or
1727 if (SrcWidth != DestWidth) {
1728 assert(DestWidth > SrcWidth);
1729 Old = Builder.CreateAnd(Old, ConstantInt::get(~Mask), "mask");
1730 SV = Builder.CreateOr(Old, SV, "ins");
1737 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
1738 /// some part of a constant global variable. This intentionally only accepts
1739 /// constant expressions because we don't can't rewrite arbitrary instructions.
1740 static bool PointsToConstantGlobal(Value *V) {
1741 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
1742 return GV->isConstant();
1743 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
1744 if (CE->getOpcode() == Instruction::BitCast ||
1745 CE->getOpcode() == Instruction::GetElementPtr)
1746 return PointsToConstantGlobal(CE->getOperand(0));
1750 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
1751 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
1752 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
1753 /// track of whether it moves the pointer (with isOffset) but otherwise traverse
1754 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
1755 /// the alloca, and if the source pointer is a pointer to a constant global, we
1756 /// can optimize this.
1757 static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy,
1759 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1760 if (LoadInst *LI = dyn_cast<LoadInst>(*UI))
1761 // Ignore non-volatile loads, they are always ok.
1762 if (!LI->isVolatile())
1765 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) {
1766 // If uses of the bitcast are ok, we are ok.
1767 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset))
1771 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
1772 // If the GEP has all zero indices, it doesn't offset the pointer. If it
1773 // doesn't, it does.
1774 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
1775 isOffset || !GEP->hasAllZeroIndices()))
1780 // If this is isn't our memcpy/memmove, reject it as something we can't
1782 if (!isa<MemTransferInst>(*UI))
1785 // If we already have seen a copy, reject the second one.
1786 if (TheCopy) return false;
1788 // If the pointer has been offset from the start of the alloca, we can't
1789 // safely handle this.
1790 if (isOffset) return false;
1792 // If the memintrinsic isn't using the alloca as the dest, reject it.
1793 if (UI.getOperandNo() != 1) return false;
1795 MemIntrinsic *MI = cast<MemIntrinsic>(*UI);
1797 // If the source of the memcpy/move is not a constant global, reject it.
1798 if (!PointsToConstantGlobal(MI->getOperand(2)))
1801 // Otherwise, the transform is safe. Remember the copy instruction.
1807 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
1808 /// modified by a copy from a constant global. If we can prove this, we can
1809 /// replace any uses of the alloca with uses of the global directly.
1810 Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) {
1811 Instruction *TheCopy = 0;
1812 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false))