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/Compiler.h"
42 #include "llvm/ADT/SmallVector.h"
43 #include "llvm/ADT/Statistic.h"
44 #include "llvm/ADT/StringExtras.h"
47 STATISTIC(NumReplaced, "Number of allocas broken up");
48 STATISTIC(NumPromoted, "Number of allocas promoted");
49 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
50 STATISTIC(NumGlobals, "Number of allocas copied from constant global");
53 struct VISIBILITY_HIDDEN SROA : public FunctionPass {
54 static char ID; // Pass identification, replacement for typeid
55 explicit SROA(signed T = -1) : FunctionPass(&ID) {
62 bool runOnFunction(Function &F);
64 bool performScalarRepl(Function &F);
65 bool performPromotion(Function &F);
67 // getAnalysisUsage - This pass does not require any passes, but we know it
68 // will not alter the CFG, so say so.
69 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
70 AU.addRequired<DominatorTree>();
71 AU.addRequired<DominanceFrontier>();
72 AU.addRequired<TargetData>();
79 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
80 /// information about the uses. All these fields are initialized to false
81 /// and set to true when something is learned.
83 /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
86 /// needsCleanup - This is set to true if there is some use of the alloca
87 /// that requires cleanup.
88 bool needsCleanup : 1;
90 /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
93 /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
97 : isUnsafe(false), needsCleanup(false),
98 isMemCpySrc(false), isMemCpyDst(false) {}
101 unsigned SRThreshold;
103 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; }
105 int isSafeAllocaToScalarRepl(AllocationInst *AI);
107 void isSafeUseOfAllocation(Instruction *User, AllocationInst *AI,
109 void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI,
111 void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI,
112 unsigned OpNo, AllocaInfo &Info);
113 void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI,
116 void DoScalarReplacement(AllocationInst *AI,
117 std::vector<AllocationInst*> &WorkList);
118 void CleanupGEP(GetElementPtrInst *GEP);
119 void CleanupAllocaUsers(AllocationInst *AI);
120 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base);
122 void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI,
123 SmallVector<AllocaInst*, 32> &NewElts);
125 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst,
127 SmallVector<AllocaInst*, 32> &NewElts);
128 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocationInst *AI,
129 SmallVector<AllocaInst*, 32> &NewElts);
130 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI,
131 SmallVector<AllocaInst*, 32> &NewElts);
133 bool CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy,
134 bool &SawVec, uint64_t Offset, unsigned AllocaSize);
135 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset);
136 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType,
137 uint64_t Offset, IRBuilder<> &Builder);
138 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
139 uint64_t Offset, IRBuilder<> &Builder);
140 static Instruction *isOnlyCopiedFromConstantGlobal(AllocationInst *AI);
145 static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates");
147 // Public interface to the ScalarReplAggregates pass
148 FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) {
149 return new SROA(Threshold);
153 bool SROA::runOnFunction(Function &F) {
154 TD = &getAnalysis<TargetData>();
156 bool Changed = performPromotion(F);
158 bool LocalChange = performScalarRepl(F);
159 if (!LocalChange) break; // No need to repromote if no scalarrepl
161 LocalChange = performPromotion(F);
162 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
169 bool SROA::performPromotion(Function &F) {
170 std::vector<AllocaInst*> Allocas;
171 DominatorTree &DT = getAnalysis<DominatorTree>();
172 DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
174 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
176 bool Changed = false;
181 // Find allocas that are safe to promote, by looking at all instructions in
183 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
184 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
185 if (isAllocaPromotable(AI))
186 Allocas.push_back(AI);
188 if (Allocas.empty()) break;
190 PromoteMemToReg(Allocas, DT, DF, F.getContext());
191 NumPromoted += Allocas.size();
198 /// getNumSAElements - Return the number of elements in the specific struct or
200 static uint64_t getNumSAElements(const Type *T) {
201 if (const StructType *ST = dyn_cast<StructType>(T))
202 return ST->getNumElements();
203 return cast<ArrayType>(T)->getNumElements();
206 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
207 // which runs on all of the malloc/alloca instructions in the function, removing
208 // them if they are only used by getelementptr instructions.
210 bool SROA::performScalarRepl(Function &F) {
211 std::vector<AllocationInst*> WorkList;
213 // Scan the entry basic block, adding any alloca's and mallocs to the worklist
214 BasicBlock &BB = F.getEntryBlock();
215 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
216 if (AllocationInst *A = dyn_cast<AllocationInst>(I))
217 WorkList.push_back(A);
219 // Process the worklist
220 bool Changed = false;
221 while (!WorkList.empty()) {
222 AllocationInst *AI = WorkList.back();
225 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
226 // with unused elements.
227 if (AI->use_empty()) {
228 AI->eraseFromParent();
232 // If this alloca is impossible for us to promote, reject it early.
233 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
236 // Check to see if this allocation is only modified by a memcpy/memmove from
237 // a constant global. If this is the case, we can change all users to use
238 // the constant global instead. This is commonly produced by the CFE by
239 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
240 // is only subsequently read.
241 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) {
242 DOUT << "Found alloca equal to global: " << *AI;
243 DOUT << " memcpy = " << *TheCopy;
244 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2));
245 AI->replaceAllUsesWith(
246 F.getContext().getConstantExprBitCast(TheSrc, AI->getType()));
247 TheCopy->eraseFromParent(); // Don't mutate the global.
248 AI->eraseFromParent();
254 // Check to see if we can perform the core SROA transformation. We cannot
255 // transform the allocation instruction if it is an array allocation
256 // (allocations OF arrays are ok though), and an allocation of a scalar
257 // value cannot be decomposed at all.
258 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
260 // Do not promote any struct whose size is too big.
261 if (AllocaSize > SRThreshold) continue;
263 if ((isa<StructType>(AI->getAllocatedType()) ||
264 isa<ArrayType>(AI->getAllocatedType())) &&
265 // Do not promote any struct into more than "32" separate vars.
266 getNumSAElements(AI->getAllocatedType()) <= SRThreshold/4) {
267 // Check that all of the users of the allocation are capable of being
269 switch (isSafeAllocaToScalarRepl(AI)) {
270 default: llvm_unreachable("Unexpected value!");
271 case 0: // Not safe to scalar replace.
273 case 1: // Safe, but requires cleanup/canonicalizations first
274 CleanupAllocaUsers(AI);
276 case 3: // Safe to scalar replace.
277 DoScalarReplacement(AI, WorkList);
283 // If we can turn this aggregate value (potentially with casts) into a
284 // simple scalar value that can be mem2reg'd into a register value.
285 // IsNotTrivial tracks whether this is something that mem2reg could have
286 // promoted itself. If so, we don't want to transform it needlessly. Note
287 // that we can't just check based on the type: the alloca may be of an i32
288 // but that has pointer arithmetic to set byte 3 of it or something.
289 bool IsNotTrivial = false;
290 const Type *VectorTy = 0;
291 bool HadAVector = false;
292 if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector,
293 0, unsigned(AllocaSize)) && IsNotTrivial) {
295 // If we were able to find a vector type that can handle this with
296 // insert/extract elements, and if there was at least one use that had
297 // a vector type, promote this to a vector. We don't want to promote
298 // random stuff that doesn't use vectors (e.g. <9 x double>) because then
299 // we just get a lot of insert/extracts. If at least one vector is
300 // involved, then we probably really do have a union of vector/array.
301 if (VectorTy && isa<VectorType>(VectorTy) && HadAVector) {
302 DOUT << "CONVERT TO VECTOR: " << *AI << " TYPE = " << *VectorTy <<"\n";
304 // Create and insert the vector alloca.
305 NewAI = new AllocaInst(VectorTy, 0, "", AI->getParent()->begin());
306 ConvertUsesToScalar(AI, NewAI, 0);
308 DOUT << "CONVERT TO SCALAR INTEGER: " << *AI << "\n";
310 // Create and insert the integer alloca.
311 const Type *NewTy = F.getContext().getIntegerType(AllocaSize*8);
312 NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
313 ConvertUsesToScalar(AI, NewAI, 0);
316 AI->eraseFromParent();
322 // Otherwise, couldn't process this alloca.
328 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
329 /// predicate, do SROA now.
330 void SROA::DoScalarReplacement(AllocationInst *AI,
331 std::vector<AllocationInst*> &WorkList) {
332 DOUT << "Found inst to SROA: " << *AI;
333 SmallVector<AllocaInst*, 32> ElementAllocas;
334 LLVMContext &Context = AI->getContext();
335 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
336 ElementAllocas.reserve(ST->getNumContainedTypes());
337 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
338 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
340 AI->getName() + "." + utostr(i), AI);
341 ElementAllocas.push_back(NA);
342 WorkList.push_back(NA); // Add to worklist for recursive processing
345 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
346 ElementAllocas.reserve(AT->getNumElements());
347 const Type *ElTy = AT->getElementType();
348 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
349 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
350 AI->getName() + "." + utostr(i), AI);
351 ElementAllocas.push_back(NA);
352 WorkList.push_back(NA); // Add to worklist for recursive processing
356 // Now that we have created the alloca instructions that we want to use,
357 // expand the getelementptr instructions to use them.
359 while (!AI->use_empty()) {
360 Instruction *User = cast<Instruction>(AI->use_back());
361 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) {
362 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas);
363 BCInst->eraseFromParent();
368 // %res = load { i32, i32 }* %alloc
370 // %load.0 = load i32* %alloc.0
371 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
372 // %load.1 = load i32* %alloc.1
373 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
374 // (Also works for arrays instead of structs)
375 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
376 Value *Insert = Context.getUndef(LI->getType());
377 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) {
378 Value *Load = new LoadInst(ElementAllocas[i], "load", LI);
379 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI);
381 LI->replaceAllUsesWith(Insert);
382 LI->eraseFromParent();
387 // store { i32, i32 } %val, { i32, i32 }* %alloc
389 // %val.0 = extractvalue { i32, i32 } %val, 0
390 // store i32 %val.0, i32* %alloc.0
391 // %val.1 = extractvalue { i32, i32 } %val, 1
392 // store i32 %val.1, i32* %alloc.1
393 // (Also works for arrays instead of structs)
394 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
395 Value *Val = SI->getOperand(0);
396 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) {
397 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI);
398 new StoreInst(Extract, ElementAllocas[i], SI);
400 SI->eraseFromParent();
404 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
405 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst>
407 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
409 assert(Idx < ElementAllocas.size() && "Index out of range?");
410 AllocaInst *AllocaToUse = ElementAllocas[Idx];
413 if (GEPI->getNumOperands() == 3) {
414 // Do not insert a new getelementptr instruction with zero indices, only
415 // to have it optimized out later.
416 RepValue = AllocaToUse;
418 // We are indexing deeply into the structure, so we still need a
419 // getelement ptr instruction to finish the indexing. This may be
420 // expanded itself once the worklist is rerun.
422 SmallVector<Value*, 8> NewArgs;
423 NewArgs.push_back(Context.getNullValue(Type::Int32Ty));
424 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end());
425 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(),
426 NewArgs.end(), "", GEPI);
427 RepValue->takeName(GEPI);
430 // If this GEP is to the start of the aggregate, check for memcpys.
431 if (Idx == 0 && GEPI->hasAllZeroIndices())
432 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas);
434 // Move all of the users over to the new GEP.
435 GEPI->replaceAllUsesWith(RepValue);
436 // Delete the old GEP
437 GEPI->eraseFromParent();
440 // Finally, delete the Alloca instruction
441 AI->eraseFromParent();
446 /// isSafeElementUse - Check to see if this use is an allowed use for a
447 /// getelementptr instruction of an array aggregate allocation. isFirstElt
448 /// indicates whether Ptr is known to the start of the aggregate.
450 void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI,
452 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
454 Instruction *User = cast<Instruction>(*I);
455 switch (User->getOpcode()) {
456 case Instruction::Load: break;
457 case Instruction::Store:
458 // Store is ok if storing INTO the pointer, not storing the pointer
459 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info);
461 case Instruction::GetElementPtr: {
462 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User);
463 bool AreAllZeroIndices = isFirstElt;
464 if (GEP->getNumOperands() > 1) {
465 if (!isa<ConstantInt>(GEP->getOperand(1)) ||
466 !cast<ConstantInt>(GEP->getOperand(1))->isZero())
467 // Using pointer arithmetic to navigate the array.
468 return MarkUnsafe(Info);
470 if (AreAllZeroIndices)
471 AreAllZeroIndices = GEP->hasAllZeroIndices();
473 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info);
474 if (Info.isUnsafe) return;
477 case Instruction::BitCast:
479 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info);
480 if (Info.isUnsafe) return;
483 DOUT << " Transformation preventing inst: " << *User;
484 return MarkUnsafe(Info);
485 case Instruction::Call:
486 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
488 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info);
489 if (Info.isUnsafe) return;
493 DOUT << " Transformation preventing inst: " << *User;
494 return MarkUnsafe(Info);
496 DOUT << " Transformation preventing inst: " << *User;
497 return MarkUnsafe(Info);
500 return; // All users look ok :)
503 /// AllUsersAreLoads - Return true if all users of this value are loads.
504 static bool AllUsersAreLoads(Value *Ptr) {
505 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
507 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load)
512 /// isSafeUseOfAllocation - Check to see if this user is an allowed use for an
513 /// aggregate allocation.
515 void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI,
517 LLVMContext &Context = User->getContext();
518 if (BitCastInst *C = dyn_cast<BitCastInst>(User))
519 return isSafeUseOfBitCastedAllocation(C, AI, Info);
521 if (LoadInst *LI = dyn_cast<LoadInst>(User))
522 if (!LI->isVolatile())
523 return;// Loads (returning a first class aggregrate) are always rewritable
525 if (StoreInst *SI = dyn_cast<StoreInst>(User))
526 if (!SI->isVolatile() && SI->getOperand(0) != AI)
527 return;// Store is ok if storing INTO the pointer, not storing the pointer
529 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User);
531 return MarkUnsafe(Info);
533 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI);
535 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>".
537 I.getOperand() != Context.getNullValue(I.getOperand()->getType())) {
538 return MarkUnsafe(Info);
542 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices??
544 bool IsAllZeroIndices = true;
546 // If the first index is a non-constant index into an array, see if we can
547 // handle it as a special case.
548 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
549 if (!isa<ConstantInt>(I.getOperand())) {
550 IsAllZeroIndices = 0;
551 uint64_t NumElements = AT->getNumElements();
553 // If this is an array index and the index is not constant, we cannot
554 // promote... that is unless the array has exactly one or two elements in
555 // it, in which case we CAN promote it, but we have to canonicalize this
556 // out if this is the only problem.
557 if ((NumElements == 1 || NumElements == 2) &&
558 AllUsersAreLoads(GEPI)) {
559 Info.needsCleanup = true;
560 return; // Canonicalization required!
562 return MarkUnsafe(Info);
566 // Walk through the GEP type indices, checking the types that this indexes
568 for (; I != E; ++I) {
569 // Ignore struct elements, no extra checking needed for these.
570 if (isa<StructType>(*I))
573 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand());
574 if (!IdxVal) return MarkUnsafe(Info);
576 // Are all indices still zero?
577 IsAllZeroIndices &= IdxVal->isZero();
579 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
580 // This GEP indexes an array. Verify that this is an in-range constant
581 // integer. Specifically, consider A[0][i]. We cannot know that the user
582 // isn't doing invalid things like allowing i to index an out-of-range
583 // subscript that accesses A[1]. Because of this, we have to reject SROA
584 // of any accesses into structs where any of the components are variables.
585 if (IdxVal->getZExtValue() >= AT->getNumElements())
586 return MarkUnsafe(Info);
587 } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) {
588 if (IdxVal->getZExtValue() >= VT->getNumElements())
589 return MarkUnsafe(Info);
593 // If there are any non-simple uses of this getelementptr, make sure to reject
595 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info);
598 /// isSafeMemIntrinsicOnAllocation - Return true if the specified memory
599 /// intrinsic can be promoted by SROA. At this point, we know that the operand
600 /// of the memintrinsic is a pointer to the beginning of the allocation.
601 void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI,
602 unsigned OpNo, AllocaInfo &Info) {
603 // If not constant length, give up.
604 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
605 if (!Length) return MarkUnsafe(Info);
607 // If not the whole aggregate, give up.
608 if (Length->getZExtValue() !=
609 TD->getTypeAllocSize(AI->getType()->getElementType()))
610 return MarkUnsafe(Info);
612 // We only know about memcpy/memset/memmove.
613 if (!isa<MemIntrinsic>(MI))
614 return MarkUnsafe(Info);
616 // Otherwise, we can transform it. Determine whether this is a memcpy/set
617 // into or out of the aggregate.
619 Info.isMemCpyDst = true;
622 Info.isMemCpySrc = true;
626 /// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast
628 void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI,
630 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end();
632 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) {
633 isSafeUseOfBitCastedAllocation(BCU, AI, Info);
634 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) {
635 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info);
636 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
637 if (SI->isVolatile())
638 return MarkUnsafe(Info);
640 // If storing the entire alloca in one chunk through a bitcasted pointer
641 // to integer, we can transform it. This happens (for example) when you
642 // cast a {i32,i32}* to i64* and store through it. This is similar to the
643 // memcpy case and occurs in various "byval" cases and emulated memcpys.
644 if (isa<IntegerType>(SI->getOperand(0)->getType()) &&
645 TD->getTypeAllocSize(SI->getOperand(0)->getType()) ==
646 TD->getTypeAllocSize(AI->getType()->getElementType())) {
647 Info.isMemCpyDst = true;
650 return MarkUnsafe(Info);
651 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) {
652 if (LI->isVolatile())
653 return MarkUnsafe(Info);
655 // If loading the entire alloca in one chunk through a bitcasted pointer
656 // to integer, we can transform it. This happens (for example) when you
657 // cast a {i32,i32}* to i64* and load through it. This is similar to the
658 // memcpy case and occurs in various "byval" cases and emulated memcpys.
659 if (isa<IntegerType>(LI->getType()) &&
660 TD->getTypeAllocSize(LI->getType()) ==
661 TD->getTypeAllocSize(AI->getType()->getElementType())) {
662 Info.isMemCpySrc = true;
665 return MarkUnsafe(Info);
666 } else if (isa<DbgInfoIntrinsic>(UI)) {
667 // If one user is DbgInfoIntrinsic then check if all users are
668 // DbgInfoIntrinsics.
669 if (OnlyUsedByDbgInfoIntrinsics(BC)) {
670 Info.needsCleanup = true;
677 return MarkUnsafe(Info);
679 if (Info.isUnsafe) return;
683 /// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes
684 /// to its first element. Transform users of the cast to use the new values
686 void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI,
687 SmallVector<AllocaInst*, 32> &NewElts) {
688 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end();
690 Instruction *User = cast<Instruction>(*UI++);
691 if (BitCastInst *BCU = dyn_cast<BitCastInst>(User)) {
692 RewriteBitCastUserOfAlloca(BCU, AI, NewElts);
693 if (BCU->use_empty()) BCU->eraseFromParent();
697 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
698 // This must be memcpy/memmove/memset of the entire aggregate.
699 // Split into one per element.
700 RewriteMemIntrinUserOfAlloca(MI, BCInst, AI, NewElts);
704 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
705 // If this is a store of the entire alloca from an integer, rewrite it.
706 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
710 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
711 // If this is a load of the entire alloca to an integer, rewrite it.
712 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
716 // Otherwise it must be some other user of a gep of the first pointer. Just
717 // leave these alone.
722 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
723 /// Rewrite it to copy or set the elements of the scalarized memory.
724 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst,
726 SmallVector<AllocaInst*, 32> &NewElts) {
728 // If this is a memcpy/memmove, construct the other pointer as the
729 // appropriate type. The "Other" pointer is the pointer that goes to memory
730 // that doesn't have anything to do with the alloca that we are promoting. For
731 // memset, this Value* stays null.
733 LLVMContext &Context = MI->getContext();
734 unsigned MemAlignment = MI->getAlignment();
735 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy
736 if (BCInst == MTI->getRawDest())
737 OtherPtr = MTI->getRawSource();
739 assert(BCInst == MTI->getRawSource());
740 OtherPtr = MTI->getRawDest();
744 // If there is an other pointer, we want to convert it to the same pointer
745 // type as AI has, so we can GEP through it safely.
747 // It is likely that OtherPtr is a bitcast, if so, remove it.
748 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr))
749 OtherPtr = BC->getOperand(0);
750 // All zero GEPs are effectively bitcasts.
751 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr))
752 if (GEP->hasAllZeroIndices())
753 OtherPtr = GEP->getOperand(0);
755 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr))
756 if (BCE->getOpcode() == Instruction::BitCast)
757 OtherPtr = BCE->getOperand(0);
759 // If the pointer is not the right type, insert a bitcast to the right
761 if (OtherPtr->getType() != AI->getType())
762 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(),
766 // Process each element of the aggregate.
767 Value *TheFn = MI->getOperand(0);
768 const Type *BytePtrTy = MI->getRawDest()->getType();
769 bool SROADest = MI->getRawDest() == BCInst;
771 Constant *Zero = Context.getNullValue(Type::Int32Ty);
773 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
774 // If this is a memcpy/memmove, emit a GEP of the other element address.
776 unsigned OtherEltAlign = MemAlignment;
779 Value *Idx[2] = { Zero, ConstantInt::get(Type::Int32Ty, i) };
780 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2,
781 OtherPtr->getNameStr()+"."+utostr(i),
784 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
785 if (const StructType *ST =
786 dyn_cast<StructType>(OtherPtrTy->getElementType())) {
787 EltOffset = TD->getStructLayout(ST)->getElementOffset(i);
790 cast<SequentialType>(OtherPtr->getType())->getElementType();
791 EltOffset = TD->getTypeAllocSize(EltTy)*i;
794 // The alignment of the other pointer is the guaranteed alignment of the
795 // element, which is affected by both the known alignment of the whole
796 // mem intrinsic and the alignment of the element. If the alignment of
797 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
798 // known alignment is just 4 bytes.
799 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
802 Value *EltPtr = NewElts[i];
803 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
805 // If we got down to a scalar, insert a load or store as appropriate.
806 if (EltTy->isSingleValueType()) {
807 if (isa<MemTransferInst>(MI)) {
809 // From Other to Alloca.
810 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
811 new StoreInst(Elt, EltPtr, MI);
813 // From Alloca to Other.
814 Value *Elt = new LoadInst(EltPtr, "tmp", MI);
815 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI);
819 assert(isa<MemSetInst>(MI));
821 // If the stored element is zero (common case), just store a null
824 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) {
826 StoreVal = Context.getNullValue(EltTy); // 0.0, null, 0, <0,0>
828 // If EltTy is a vector type, get the element type.
829 const Type *ValTy = EltTy->getScalarType();
831 // Construct an integer with the right value.
832 unsigned EltSize = TD->getTypeSizeInBits(ValTy);
833 APInt OneVal(EltSize, CI->getZExtValue());
834 APInt TotalVal(OneVal);
836 for (unsigned i = 0; 8*i < EltSize; ++i) {
837 TotalVal = TotalVal.shl(8);
841 // Convert the integer value to the appropriate type.
842 StoreVal = ConstantInt::get(Context, TotalVal);
843 if (isa<PointerType>(ValTy))
844 StoreVal = Context.getConstantExprIntToPtr(StoreVal, ValTy);
845 else if (ValTy->isFloatingPoint())
846 StoreVal = Context.getConstantExprBitCast(StoreVal, ValTy);
847 assert(StoreVal->getType() == ValTy && "Type mismatch!");
849 // If the requested value was a vector constant, create it.
850 if (EltTy != ValTy) {
851 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements();
852 SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
853 StoreVal = Context.getConstantVector(&Elts[0], NumElts);
856 new StoreInst(StoreVal, EltPtr, MI);
859 // Otherwise, if we're storing a byte variable, use a memset call for
863 // Cast the element pointer to BytePtrTy.
864 if (EltPtr->getType() != BytePtrTy)
865 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI);
867 // Cast the other pointer (if we have one) to BytePtrTy.
868 if (OtherElt && OtherElt->getType() != BytePtrTy)
869 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(),
872 unsigned EltSize = TD->getTypeAllocSize(EltTy);
874 // Finally, insert the meminst for this element.
875 if (isa<MemTransferInst>(MI)) {
877 SROADest ? EltPtr : OtherElt, // Dest ptr
878 SROADest ? OtherElt : EltPtr, // Src ptr
879 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
880 ConstantInt::get(Type::Int32Ty, OtherEltAlign) // Align
882 CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
884 assert(isa<MemSetInst>(MI));
886 EltPtr, MI->getOperand(2), // Dest, Value,
887 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
890 CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
893 MI->eraseFromParent();
896 /// RewriteStoreUserOfWholeAlloca - We found an store of an integer that
897 /// overwrites the entire allocation. Extract out the pieces of the stored
898 /// integer and store them individually.
899 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI,
901 SmallVector<AllocaInst*, 32> &NewElts){
902 // Extract each element out of the integer according to its structure offset
903 // and store the element value to the individual alloca.
904 LLVMContext &Context = SI->getContext();
905 Value *SrcVal = SI->getOperand(0);
906 const Type *AllocaEltTy = AI->getType()->getElementType();
907 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
909 // If this isn't a store of an integer to the whole alloca, it may be a store
910 // to the first element. Just ignore the store in this case and normal SROA
912 if (!isa<IntegerType>(SrcVal->getType()) ||
913 TD->getTypeAllocSizeInBits(SrcVal->getType()) != AllocaSizeBits)
915 // Handle tail padding by extending the operand
916 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
917 SrcVal = new ZExtInst(SrcVal,
918 Context.getIntegerType(AllocaSizeBits), "", SI);
920 DOUT << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << *SI;
922 // There are two forms here: AI could be an array or struct. Both cases
923 // have different ways to compute the element offset.
924 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
925 const StructLayout *Layout = TD->getStructLayout(EltSTy);
927 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
928 // Get the number of bits to shift SrcVal to get the value.
929 const Type *FieldTy = EltSTy->getElementType(i);
930 uint64_t Shift = Layout->getElementOffsetInBits(i);
932 if (TD->isBigEndian())
933 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy);
935 Value *EltVal = SrcVal;
937 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
938 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
939 "sroa.store.elt", SI);
942 // Truncate down to an integer of the right size.
943 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
945 // Ignore zero sized fields like {}, they obviously contain no data.
946 if (FieldSizeBits == 0) continue;
948 if (FieldSizeBits != AllocaSizeBits)
949 EltVal = new TruncInst(EltVal,
950 Context.getIntegerType(FieldSizeBits), "", SI);
951 Value *DestField = NewElts[i];
952 if (EltVal->getType() == FieldTy) {
953 // Storing to an integer field of this size, just do it.
954 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) {
955 // Bitcast to the right element type (for fp/vector values).
956 EltVal = new BitCastInst(EltVal, FieldTy, "", SI);
958 // Otherwise, bitcast the dest pointer (for aggregates).
959 DestField = new BitCastInst(DestField,
960 Context.getPointerTypeUnqual(EltVal->getType()),
963 new StoreInst(EltVal, DestField, SI);
967 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
968 const Type *ArrayEltTy = ATy->getElementType();
969 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
970 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
974 if (TD->isBigEndian())
975 Shift = AllocaSizeBits-ElementOffset;
979 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
980 // Ignore zero sized fields like {}, they obviously contain no data.
981 if (ElementSizeBits == 0) continue;
983 Value *EltVal = SrcVal;
985 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
986 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
987 "sroa.store.elt", SI);
990 // Truncate down to an integer of the right size.
991 if (ElementSizeBits != AllocaSizeBits)
992 EltVal = new TruncInst(EltVal,
993 Context.getIntegerType(ElementSizeBits),"",SI);
994 Value *DestField = NewElts[i];
995 if (EltVal->getType() == ArrayEltTy) {
996 // Storing to an integer field of this size, just do it.
997 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) {
998 // Bitcast to the right element type (for fp/vector values).
999 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI);
1001 // Otherwise, bitcast the dest pointer (for aggregates).
1002 DestField = new BitCastInst(DestField,
1003 Context.getPointerTypeUnqual(EltVal->getType()),
1006 new StoreInst(EltVal, DestField, SI);
1008 if (TD->isBigEndian())
1009 Shift -= ElementOffset;
1011 Shift += ElementOffset;
1015 SI->eraseFromParent();
1018 /// RewriteLoadUserOfWholeAlloca - We found an load of the entire allocation to
1019 /// an integer. Load the individual pieces to form the aggregate value.
1020 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI,
1021 SmallVector<AllocaInst*, 32> &NewElts) {
1022 // Extract each element out of the NewElts according to its structure offset
1023 // and form the result value.
1024 const Type *AllocaEltTy = AI->getType()->getElementType();
1025 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
1027 // If this isn't a load of the whole alloca to an integer, it may be a load
1028 // of the first element. Just ignore the load in this case and normal SROA
1030 if (!isa<IntegerType>(LI->getType()) ||
1031 TD->getTypeAllocSizeInBits(LI->getType()) != AllocaSizeBits)
1034 DOUT << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << *LI;
1036 // There are two forms here: AI could be an array or struct. Both cases
1037 // have different ways to compute the element offset.
1038 const StructLayout *Layout = 0;
1039 uint64_t ArrayEltBitOffset = 0;
1040 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
1041 Layout = TD->getStructLayout(EltSTy);
1043 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
1044 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
1047 LLVMContext &Context = LI->getContext();
1050 Context.getNullValue(Context.getIntegerType(AllocaSizeBits));
1052 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1053 // Load the value from the alloca. If the NewElt is an aggregate, cast
1054 // the pointer to an integer of the same size before doing the load.
1055 Value *SrcField = NewElts[i];
1056 const Type *FieldTy =
1057 cast<PointerType>(SrcField->getType())->getElementType();
1058 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
1060 // Ignore zero sized fields like {}, they obviously contain no data.
1061 if (FieldSizeBits == 0) continue;
1063 const IntegerType *FieldIntTy = Context.getIntegerType(FieldSizeBits);
1064 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() &&
1065 !isa<VectorType>(FieldTy))
1066 SrcField = new BitCastInst(SrcField,
1067 Context.getPointerTypeUnqual(FieldIntTy),
1069 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
1071 // If SrcField is a fp or vector of the right size but that isn't an
1072 // integer type, bitcast to an integer so we can shift it.
1073 if (SrcField->getType() != FieldIntTy)
1074 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
1076 // Zero extend the field to be the same size as the final alloca so that
1077 // we can shift and insert it.
1078 if (SrcField->getType() != ResultVal->getType())
1079 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
1081 // Determine the number of bits to shift SrcField.
1083 if (Layout) // Struct case.
1084 Shift = Layout->getElementOffsetInBits(i);
1086 Shift = i*ArrayEltBitOffset;
1088 if (TD->isBigEndian())
1089 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
1092 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
1093 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
1096 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
1099 // Handle tail padding by truncating the result
1100 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
1101 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI);
1103 LI->replaceAllUsesWith(ResultVal);
1104 LI->eraseFromParent();
1108 /// HasPadding - Return true if the specified type has any structure or
1109 /// alignment padding, false otherwise.
1110 static bool HasPadding(const Type *Ty, const TargetData &TD) {
1111 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1112 const StructLayout *SL = TD.getStructLayout(STy);
1113 unsigned PrevFieldBitOffset = 0;
1114 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1115 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
1117 // Padding in sub-elements?
1118 if (HasPadding(STy->getElementType(i), TD))
1121 // Check to see if there is any padding between this element and the
1124 unsigned PrevFieldEnd =
1125 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
1126 if (PrevFieldEnd < FieldBitOffset)
1130 PrevFieldBitOffset = FieldBitOffset;
1133 // Check for tail padding.
1134 if (unsigned EltCount = STy->getNumElements()) {
1135 unsigned PrevFieldEnd = PrevFieldBitOffset +
1136 TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
1137 if (PrevFieldEnd < SL->getSizeInBits())
1141 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1142 return HasPadding(ATy->getElementType(), TD);
1143 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1144 return HasPadding(VTy->getElementType(), TD);
1146 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
1149 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
1150 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
1151 /// or 1 if safe after canonicalization has been performed.
1153 int SROA::isSafeAllocaToScalarRepl(AllocationInst *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 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end();
1160 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info);
1161 if (Info.isUnsafe) {
1162 DOUT << "Cannot transform: " << *AI << " due to user: " << **I;
1167 // Okay, we know all the users are promotable. If the aggregate is a memcpy
1168 // source and destination, we have to be careful. In particular, the memcpy
1169 // could be moving around elements that live in structure padding of the LLVM
1170 // types, but may actually be used. In these cases, we refuse to promote the
1172 if (Info.isMemCpySrc && Info.isMemCpyDst &&
1173 HasPadding(AI->getType()->getElementType(), *TD))
1176 // If we require cleanup, return 1, otherwise return 3.
1177 return Info.needsCleanup ? 1 : 3;
1180 /// CleanupGEP - GEP is used by an Alloca, which can be prompted after the GEP
1181 /// is canonicalized here.
1182 void SROA::CleanupGEP(GetElementPtrInst *GEPI) {
1183 gep_type_iterator I = gep_type_begin(GEPI);
1186 const ArrayType *AT = dyn_cast<ArrayType>(*I);
1190 uint64_t NumElements = AT->getNumElements();
1192 if (isa<ConstantInt>(I.getOperand()))
1195 LLVMContext &Context = GEPI->getContext();
1197 if (NumElements == 1) {
1198 GEPI->setOperand(2, Context.getNullValue(Type::Int32Ty));
1202 assert(NumElements == 2 && "Unhandled case!");
1203 // All users of the GEP must be loads. At each use of the GEP, insert
1204 // two loads of the appropriate indexed GEP and select between them.
1205 Value *IsOne = new ICmpInst(GEPI, ICmpInst::ICMP_NE, I.getOperand(),
1206 Context.getNullValue(I.getOperand()->getType()),
1208 // Insert the new GEP instructions, which are properly indexed.
1209 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end());
1210 Indices[1] = Context.getNullValue(Type::Int32Ty);
1211 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
1214 GEPI->getName()+".0", GEPI);
1215 Indices[1] = ConstantInt::get(Type::Int32Ty, 1);
1216 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
1219 GEPI->getName()+".1", GEPI);
1220 // Replace all loads of the variable index GEP with loads from both
1221 // indexes and a select.
1222 while (!GEPI->use_empty()) {
1223 LoadInst *LI = cast<LoadInst>(GEPI->use_back());
1224 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI);
1225 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI);
1226 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI);
1227 LI->replaceAllUsesWith(R);
1228 LI->eraseFromParent();
1230 GEPI->eraseFromParent();
1234 /// CleanupAllocaUsers - If SROA reported that it can promote the specified
1235 /// allocation, but only if cleaned up, perform the cleanups required.
1236 void SROA::CleanupAllocaUsers(AllocationInst *AI) {
1237 // At this point, we know that the end result will be SROA'd and promoted, so
1238 // we can insert ugly code if required so long as sroa+mem2reg will clean it
1240 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
1243 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U))
1246 Instruction *I = cast<Instruction>(U);
1247 SmallVector<DbgInfoIntrinsic *, 2> DbgInUses;
1248 if (!isa<StoreInst>(I) && OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) {
1249 // Safe to remove debug info uses.
1250 while (!DbgInUses.empty()) {
1251 DbgInfoIntrinsic *DI = DbgInUses.back(); DbgInUses.pop_back();
1252 DI->eraseFromParent();
1254 I->eraseFromParent();
1260 /// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at
1261 /// the offset specified by Offset (which is specified in bytes).
1263 /// There are two cases we handle here:
1264 /// 1) A union of vector types of the same size and potentially its elements.
1265 /// Here we turn element accesses into insert/extract element operations.
1266 /// This promotes a <4 x float> with a store of float to the third element
1267 /// into a <4 x float> that uses insert element.
1268 /// 2) A fully general blob of memory, which we turn into some (potentially
1269 /// large) integer type with extract and insert operations where the loads
1270 /// and stores would mutate the memory.
1271 static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy,
1272 unsigned AllocaSize, const TargetData &TD,
1273 LLVMContext &Context) {
1274 // If this could be contributing to a vector, analyze it.
1275 if (VecTy != Type::VoidTy) { // either null or a vector type.
1277 // If the In type is a vector that is the same size as the alloca, see if it
1278 // matches the existing VecTy.
1279 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) {
1280 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) {
1281 // If we're storing/loading a vector of the right size, allow it as a
1282 // vector. If this the first vector we see, remember the type so that
1283 // we know the element size.
1288 } else if (In == Type::FloatTy || In == Type::DoubleTy ||
1289 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 &&
1290 isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
1291 // If we're accessing something that could be an element of a vector, see
1292 // if the implied vector agrees with what we already have and if Offset is
1293 // compatible with it.
1294 unsigned EltSize = In->getPrimitiveSizeInBits()/8;
1295 if (Offset % EltSize == 0 &&
1296 AllocaSize % EltSize == 0 &&
1298 cast<VectorType>(VecTy)->getElementType()
1299 ->getPrimitiveSizeInBits()/8 == EltSize)) {
1301 VecTy = In->getContext().getVectorType(In, AllocaSize/EltSize);
1307 // Otherwise, we have a case that we can't handle with an optimized vector
1308 // form. We can still turn this into a large integer.
1309 VecTy = Type::VoidTy;
1312 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
1313 /// its accesses to use a to single vector type, return true, and set VecTy to
1314 /// the new type. If we could convert the alloca into a single promotable
1315 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a
1316 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
1317 /// is the current offset from the base of the alloca being analyzed.
1319 /// If we see at least one access to the value that is as a vector type, set the
1322 bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy,
1323 bool &SawVec, uint64_t Offset,
1324 unsigned AllocaSize) {
1325 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1326 Instruction *User = cast<Instruction>(*UI);
1328 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1329 // Don't break volatile loads.
1330 if (LI->isVolatile())
1332 MergeInType(LI->getType(), Offset, VecTy,
1333 AllocaSize, *TD, V->getContext());
1334 SawVec |= isa<VectorType>(LI->getType());
1338 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1339 // Storing the pointer, not into the value?
1340 if (SI->getOperand(0) == V || SI->isVolatile()) return 0;
1341 MergeInType(SI->getOperand(0)->getType(), Offset,
1342 VecTy, AllocaSize, *TD, V->getContext());
1343 SawVec |= isa<VectorType>(SI->getOperand(0)->getType());
1347 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
1348 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset,
1351 IsNotTrivial = true;
1355 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1356 // If this is a GEP with a variable indices, we can't handle it.
1357 if (!GEP->hasAllConstantIndices())
1360 // Compute the offset that this GEP adds to the pointer.
1361 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
1362 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(),
1363 &Indices[0], Indices.size());
1364 // See if all uses can be converted.
1365 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset,
1368 IsNotTrivial = true;
1372 // If this is a constant sized memset of a constant value (e.g. 0) we can
1374 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
1375 // Store of constant value and constant size.
1376 if (isa<ConstantInt>(MSI->getValue()) &&
1377 isa<ConstantInt>(MSI->getLength())) {
1378 IsNotTrivial = true;
1383 // If this is a memcpy or memmove into or out of the whole allocation, we
1384 // can handle it like a load or store of the scalar type.
1385 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
1386 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()))
1387 if (Len->getZExtValue() == AllocaSize && Offset == 0) {
1388 IsNotTrivial = true;
1393 // Ignore dbg intrinsic.
1394 if (isa<DbgInfoIntrinsic>(User))
1397 // Otherwise, we cannot handle this!
1405 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
1406 /// directly. This happens when we are converting an "integer union" to a
1407 /// single integer scalar, or when we are converting a "vector union" to a
1408 /// vector with insert/extractelement instructions.
1410 /// Offset is an offset from the original alloca, in bits that need to be
1411 /// shifted to the right. By the end of this, there should be no uses of Ptr.
1412 void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) {
1413 while (!Ptr->use_empty()) {
1414 Instruction *User = cast<Instruction>(Ptr->use_back());
1416 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
1417 ConvertUsesToScalar(CI, NewAI, Offset);
1418 CI->eraseFromParent();
1422 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1423 // Compute the offset that this GEP adds to the pointer.
1424 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
1425 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(),
1426 &Indices[0], Indices.size());
1427 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
1428 GEP->eraseFromParent();
1432 IRBuilder<> Builder(User->getParent(), User);
1434 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1435 // The load is a bit extract from NewAI shifted right by Offset bits.
1436 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp");
1438 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
1439 LI->replaceAllUsesWith(NewLoadVal);
1440 LI->eraseFromParent();
1444 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1445 assert(SI->getOperand(0) != Ptr && "Consistency error!");
1446 // FIXME: Remove once builder has Twine API.
1447 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").str().c_str());
1448 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
1450 Builder.CreateStore(New, NewAI);
1451 SI->eraseFromParent();
1455 // If this is a constant sized memset of a constant value (e.g. 0) we can
1456 // transform it into a store of the expanded constant value.
1457 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
1458 assert(MSI->getRawDest() == Ptr && "Consistency error!");
1459 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
1460 if (NumBytes != 0) {
1461 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
1463 // Compute the value replicated the right number of times.
1464 APInt APVal(NumBytes*8, Val);
1466 // Splat the value if non-zero.
1468 for (unsigned i = 1; i != NumBytes; ++i)
1469 APVal |= APVal << 8;
1471 // FIXME: Remove once builder has Twine API.
1472 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").str().c_str());
1473 Value *New = ConvertScalar_InsertValue(
1474 ConstantInt::get(User->getContext(), APVal),
1475 Old, Offset, Builder);
1476 Builder.CreateStore(New, NewAI);
1478 MSI->eraseFromParent();
1482 // If this is a memcpy or memmove into or out of the whole allocation, we
1483 // can handle it like a load or store of the scalar type.
1484 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
1485 assert(Offset == 0 && "must be store to start of alloca");
1487 // If the source and destination are both to the same alloca, then this is
1488 // a noop copy-to-self, just delete it. Otherwise, emit a load and store
1490 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject());
1492 if (MTI->getSource()->getUnderlyingObject() != OrigAI) {
1493 // Dest must be OrigAI, change this to be a load from the original
1494 // pointer (bitcasted), then a store to our new alloca.
1495 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
1496 Value *SrcPtr = MTI->getSource();
1497 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType());
1499 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
1500 SrcVal->setAlignment(MTI->getAlignment());
1501 Builder.CreateStore(SrcVal, NewAI);
1502 } else if (MTI->getDest()->getUnderlyingObject() != OrigAI) {
1503 // Src must be OrigAI, change this to be a load from NewAI then a store
1504 // through the original dest pointer (bitcasted).
1505 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
1506 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
1508 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType());
1509 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
1510 NewStore->setAlignment(MTI->getAlignment());
1512 // Noop transfer. Src == Dst
1516 MTI->eraseFromParent();
1520 // If user is a dbg info intrinsic then it is safe to remove it.
1521 if (isa<DbgInfoIntrinsic>(User)) {
1522 User->eraseFromParent();
1526 llvm_unreachable("Unsupported operation!");
1530 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
1531 /// or vector value FromVal, extracting the bits from the offset specified by
1532 /// Offset. This returns the value, which is of type ToType.
1534 /// This happens when we are converting an "integer union" to a single
1535 /// integer scalar, or when we are converting a "vector union" to a vector with
1536 /// insert/extractelement instructions.
1538 /// Offset is an offset from the original alloca, in bits that need to be
1539 /// shifted to the right.
1540 Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType,
1541 uint64_t Offset, IRBuilder<> &Builder) {
1542 // If the load is of the whole new alloca, no conversion is needed.
1543 if (FromVal->getType() == ToType && Offset == 0)
1546 LLVMContext &Context = FromVal->getContext();
1548 // If the result alloca is a vector type, this is either an element
1549 // access or a bitcast to another vector type of the same size.
1550 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) {
1551 if (isa<VectorType>(ToType))
1552 return Builder.CreateBitCast(FromVal, ToType, "tmp");
1554 // Otherwise it must be an element access.
1557 unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType());
1558 Elt = Offset/EltSize;
1559 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
1561 // Return the element extracted out of it.
1562 Value *V = Builder.CreateExtractElement(FromVal,
1563 ConstantInt::get(Type::Int32Ty,Elt),
1565 if (V->getType() != ToType)
1566 V = Builder.CreateBitCast(V, ToType, "tmp");
1570 // If ToType is a first class aggregate, extract out each of the pieces and
1571 // use insertvalue's to form the FCA.
1572 if (const StructType *ST = dyn_cast<StructType>(ToType)) {
1573 const StructLayout &Layout = *TD->getStructLayout(ST);
1574 Value *Res = Context.getUndef(ST);
1575 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
1576 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
1577 Offset+Layout.getElementOffsetInBits(i),
1579 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
1584 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
1585 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType());
1586 Value *Res = Context.getUndef(AT);
1587 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1588 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
1589 Offset+i*EltSize, Builder);
1590 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
1595 // Otherwise, this must be a union that was converted to an integer value.
1596 const IntegerType *NTy = cast<IntegerType>(FromVal->getType());
1598 // If this is a big-endian system and the load is narrower than the
1599 // full alloca type, we need to do a shift to get the right bits.
1601 if (TD->isBigEndian()) {
1602 // On big-endian machines, the lowest bit is stored at the bit offset
1603 // from the pointer given by getTypeStoreSizeInBits. This matters for
1604 // integers with a bitwidth that is not a multiple of 8.
1605 ShAmt = TD->getTypeStoreSizeInBits(NTy) -
1606 TD->getTypeStoreSizeInBits(ToType) - Offset;
1611 // Note: we support negative bitwidths (with shl) which are not defined.
1612 // We do this to support (f.e.) loads off the end of a structure where
1613 // only some bits are used.
1614 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
1615 FromVal = Builder.CreateLShr(FromVal,
1616 ConstantInt::get(FromVal->getType(),
1618 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
1619 FromVal = Builder.CreateShl(FromVal,
1620 ConstantInt::get(FromVal->getType(),
1623 // Finally, unconditionally truncate the integer to the right width.
1624 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType);
1625 if (LIBitWidth < NTy->getBitWidth())
1627 Builder.CreateTrunc(FromVal, Context.getIntegerType(LIBitWidth), "tmp");
1628 else if (LIBitWidth > NTy->getBitWidth())
1630 Builder.CreateZExt(FromVal, Context.getIntegerType(LIBitWidth), "tmp");
1632 // If the result is an integer, this is a trunc or bitcast.
1633 if (isa<IntegerType>(ToType)) {
1635 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) {
1636 // Just do a bitcast, we know the sizes match up.
1637 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp");
1639 // Otherwise must be a pointer.
1640 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp");
1642 assert(FromVal->getType() == ToType && "Didn't convert right?");
1647 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
1648 /// or vector value "Old" at the offset specified by Offset.
1650 /// This happens when we are converting an "integer union" to a
1651 /// single integer scalar, or when we are converting a "vector union" to a
1652 /// vector with insert/extractelement instructions.
1654 /// Offset is an offset from the original alloca, in bits that need to be
1655 /// shifted to the right.
1656 Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old,
1657 uint64_t Offset, IRBuilder<> &Builder) {
1659 // Convert the stored type to the actual type, shift it left to insert
1660 // then 'or' into place.
1661 const Type *AllocaType = Old->getType();
1662 LLVMContext &Context = Old->getContext();
1664 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
1665 uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy);
1666 uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType());
1668 // Changing the whole vector with memset or with an access of a different
1670 if (ValSize == VecSize)
1671 return Builder.CreateBitCast(SV, AllocaType, "tmp");
1673 uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType());
1675 // Must be an element insertion.
1676 unsigned Elt = Offset/EltSize;
1678 if (SV->getType() != VTy->getElementType())
1679 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp");
1681 SV = Builder.CreateInsertElement(Old, SV,
1682 ConstantInt::get(Type::Int32Ty, Elt),
1687 // If SV is a first-class aggregate value, insert each value recursively.
1688 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) {
1689 const StructLayout &Layout = *TD->getStructLayout(ST);
1690 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
1691 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
1692 Old = ConvertScalar_InsertValue(Elt, Old,
1693 Offset+Layout.getElementOffsetInBits(i),
1699 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
1700 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType());
1701 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1702 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
1703 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
1708 // If SV is a float, convert it to the appropriate integer type.
1709 // If it is a pointer, do the same.
1710 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType());
1711 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType);
1712 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType());
1713 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType);
1714 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType()))
1715 SV = Builder.CreateBitCast(SV, Context.getIntegerType(SrcWidth), "tmp");
1716 else if (isa<PointerType>(SV->getType()))
1717 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(), "tmp");
1719 // Zero extend or truncate the value if needed.
1720 if (SV->getType() != AllocaType) {
1721 if (SV->getType()->getPrimitiveSizeInBits() <
1722 AllocaType->getPrimitiveSizeInBits())
1723 SV = Builder.CreateZExt(SV, AllocaType, "tmp");
1725 // Truncation may be needed if storing more than the alloca can hold
1726 // (undefined behavior).
1727 SV = Builder.CreateTrunc(SV, AllocaType, "tmp");
1728 SrcWidth = DestWidth;
1729 SrcStoreWidth = DestStoreWidth;
1733 // If this is a big-endian system and the store is narrower than the
1734 // full alloca type, we need to do a shift to get the right bits.
1736 if (TD->isBigEndian()) {
1737 // On big-endian machines, the lowest bit is stored at the bit offset
1738 // from the pointer given by getTypeStoreSizeInBits. This matters for
1739 // integers with a bitwidth that is not a multiple of 8.
1740 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
1745 // Note: we support negative bitwidths (with shr) which are not defined.
1746 // We do this to support (f.e.) stores off the end of a structure where
1747 // only some bits in the structure are set.
1748 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
1749 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
1750 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(),
1753 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
1754 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(),
1756 Mask = Mask.lshr(-ShAmt);
1759 // Mask out the bits we are about to insert from the old value, and or
1761 if (SrcWidth != DestWidth) {
1762 assert(DestWidth > SrcWidth);
1763 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask");
1764 SV = Builder.CreateOr(Old, SV, "ins");
1771 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
1772 /// some part of a constant global variable. This intentionally only accepts
1773 /// constant expressions because we don't can't rewrite arbitrary instructions.
1774 static bool PointsToConstantGlobal(Value *V) {
1775 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
1776 return GV->isConstant();
1777 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
1778 if (CE->getOpcode() == Instruction::BitCast ||
1779 CE->getOpcode() == Instruction::GetElementPtr)
1780 return PointsToConstantGlobal(CE->getOperand(0));
1784 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
1785 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
1786 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
1787 /// track of whether it moves the pointer (with isOffset) but otherwise traverse
1788 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
1789 /// the alloca, and if the source pointer is a pointer to a constant global, we
1790 /// can optimize this.
1791 static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy,
1793 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1794 if (LoadInst *LI = dyn_cast<LoadInst>(*UI))
1795 // Ignore non-volatile loads, they are always ok.
1796 if (!LI->isVolatile())
1799 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) {
1800 // If uses of the bitcast are ok, we are ok.
1801 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset))
1805 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
1806 // If the GEP has all zero indices, it doesn't offset the pointer. If it
1807 // doesn't, it does.
1808 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
1809 isOffset || !GEP->hasAllZeroIndices()))
1814 // If this is isn't our memcpy/memmove, reject it as something we can't
1816 if (!isa<MemTransferInst>(*UI))
1819 // If we already have seen a copy, reject the second one.
1820 if (TheCopy) return false;
1822 // If the pointer has been offset from the start of the alloca, we can't
1823 // safely handle this.
1824 if (isOffset) return false;
1826 // If the memintrinsic isn't using the alloca as the dest, reject it.
1827 if (UI.getOperandNo() != 1) return false;
1829 MemIntrinsic *MI = cast<MemIntrinsic>(*UI);
1831 // If the source of the memcpy/move is not a constant global, reject it.
1832 if (!PointsToConstantGlobal(MI->getOperand(2)))
1835 // Otherwise, the transform is safe. Remember the copy instruction.
1841 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
1842 /// modified by a copy from a constant global. If we can prove this, we can
1843 /// replace any uses of the alloca with uses of the global directly.
1844 Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) {
1845 Instruction *TheCopy = 0;
1846 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false))