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"
46 STATISTIC(NumReplaced, "Number of allocas broken up");
47 STATISTIC(NumPromoted, "Number of allocas promoted");
48 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
49 STATISTIC(NumGlobals, "Number of allocas copied from constant global");
52 struct VISIBILITY_HIDDEN SROA : public FunctionPass {
53 static char ID; // Pass identification, replacement for typeid
54 explicit SROA(signed T = -1) : FunctionPass(&ID) {
61 bool runOnFunction(Function &F);
63 bool performScalarRepl(Function &F);
64 bool performPromotion(Function &F);
66 // getAnalysisUsage - This pass does not require any passes, but we know it
67 // will not alter the CFG, so say so.
68 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
69 AU.addRequired<DominatorTree>();
70 AU.addRequired<DominanceFrontier>();
71 AU.addRequired<TargetData>();
78 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
79 /// information about the uses. All these fields are initialized to false
80 /// and set to true when something is learned.
82 /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
85 /// needsCleanup - This is set to true if there is some use of the alloca
86 /// that requires cleanup.
87 bool needsCleanup : 1;
89 /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
92 /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
96 : isUnsafe(false), needsCleanup(false),
97 isMemCpySrc(false), isMemCpyDst(false) {}
100 unsigned SRThreshold;
102 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; }
104 int isSafeAllocaToScalarRepl(AllocationInst *AI);
106 void isSafeUseOfAllocation(Instruction *User, AllocationInst *AI,
108 void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI,
110 void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI,
111 unsigned OpNo, AllocaInfo &Info);
112 void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI,
115 void DoScalarReplacement(AllocationInst *AI,
116 std::vector<AllocationInst*> &WorkList);
117 void CleanupGEP(GetElementPtrInst *GEP);
118 void CleanupAllocaUsers(AllocationInst *AI);
119 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base);
121 void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI,
122 SmallVector<AllocaInst*, 32> &NewElts);
124 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst,
126 SmallVector<AllocaInst*, 32> &NewElts);
127 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocationInst *AI,
128 SmallVector<AllocaInst*, 32> &NewElts);
129 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI,
130 SmallVector<AllocaInst*, 32> &NewElts);
132 bool CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy,
133 bool &SawVec, uint64_t Offset, unsigned AllocaSize);
134 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset);
135 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType,
136 uint64_t Offset, IRBuilder<> &Builder);
137 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
138 uint64_t Offset, IRBuilder<> &Builder);
139 static Instruction *isOnlyCopiedFromConstantGlobal(AllocationInst *AI);
144 static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates");
146 // Public interface to the ScalarReplAggregates pass
147 FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) {
148 return new SROA(Threshold);
152 bool SROA::runOnFunction(Function &F) {
153 TD = &getAnalysis<TargetData>();
155 bool Changed = performPromotion(F);
157 bool LocalChange = performScalarRepl(F);
158 if (!LocalChange) break; // No need to repromote if no scalarrepl
160 LocalChange = performPromotion(F);
161 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
168 bool SROA::performPromotion(Function &F) {
169 std::vector<AllocaInst*> Allocas;
170 DominatorTree &DT = getAnalysis<DominatorTree>();
171 DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
173 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
175 bool Changed = false;
180 // Find allocas that are safe to promote, by looking at all instructions in
182 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
183 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
184 if (isAllocaPromotable(AI))
185 Allocas.push_back(AI);
187 if (Allocas.empty()) break;
189 PromoteMemToReg(Allocas, DT, DF, F.getContext());
190 NumPromoted += Allocas.size();
197 /// getNumSAElements - Return the number of elements in the specific struct or
199 static uint64_t getNumSAElements(const Type *T) {
200 if (const StructType *ST = dyn_cast<StructType>(T))
201 return ST->getNumElements();
202 return cast<ArrayType>(T)->getNumElements();
205 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
206 // which runs on all of the malloc/alloca instructions in the function, removing
207 // them if they are only used by getelementptr instructions.
209 bool SROA::performScalarRepl(Function &F) {
210 std::vector<AllocationInst*> WorkList;
212 // Scan the entry basic block, adding any alloca's and mallocs to the worklist
213 BasicBlock &BB = F.getEntryBlock();
214 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
215 if (AllocationInst *A = dyn_cast<AllocationInst>(I))
216 WorkList.push_back(A);
218 // Process the worklist
219 bool Changed = false;
220 while (!WorkList.empty()) {
221 AllocationInst *AI = WorkList.back();
224 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
225 // with unused elements.
226 if (AI->use_empty()) {
227 AI->eraseFromParent();
231 // If this alloca is impossible for us to promote, reject it early.
232 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
235 // Check to see if this allocation is only modified by a memcpy/memmove from
236 // a constant global. If this is the case, we can change all users to use
237 // the constant global instead. This is commonly produced by the CFE by
238 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
239 // is only subsequently read.
240 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) {
241 DOUT << "Found alloca equal to global: " << *AI;
242 DOUT << " memcpy = " << *TheCopy;
243 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2));
244 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
245 TheCopy->eraseFromParent(); // Don't mutate the global.
246 AI->eraseFromParent();
252 // Check to see if we can perform the core SROA transformation. We cannot
253 // transform the allocation instruction if it is an array allocation
254 // (allocations OF arrays are ok though), and an allocation of a scalar
255 // value cannot be decomposed at all.
256 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
258 // Do not promote [0 x %struct].
259 if (AllocaSize == 0) continue;
261 // Do not promote any struct whose size is too big.
262 if (AllocaSize > SRThreshold) continue;
264 if ((isa<StructType>(AI->getAllocatedType()) ||
265 isa<ArrayType>(AI->getAllocatedType())) &&
266 // Do not promote any struct into more than "32" separate vars.
267 getNumSAElements(AI->getAllocatedType()) <= SRThreshold/4) {
268 // Check that all of the users of the allocation are capable of being
270 switch (isSafeAllocaToScalarRepl(AI)) {
271 default: llvm_unreachable("Unexpected value!");
272 case 0: // Not safe to scalar replace.
274 case 1: // Safe, but requires cleanup/canonicalizations first
275 CleanupAllocaUsers(AI);
277 case 3: // Safe to scalar replace.
278 DoScalarReplacement(AI, WorkList);
284 // If we can turn this aggregate value (potentially with casts) into a
285 // simple scalar value that can be mem2reg'd into a register value.
286 // IsNotTrivial tracks whether this is something that mem2reg could have
287 // promoted itself. If so, we don't want to transform it needlessly. Note
288 // that we can't just check based on the type: the alloca may be of an i32
289 // but that has pointer arithmetic to set byte 3 of it or something.
290 bool IsNotTrivial = false;
291 const Type *VectorTy = 0;
292 bool HadAVector = false;
293 if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector,
294 0, unsigned(AllocaSize)) && IsNotTrivial) {
296 // If we were able to find a vector type that can handle this with
297 // insert/extract elements, and if there was at least one use that had
298 // a vector type, promote this to a vector. We don't want to promote
299 // random stuff that doesn't use vectors (e.g. <9 x double>) because then
300 // we just get a lot of insert/extracts. If at least one vector is
301 // involved, then we probably really do have a union of vector/array.
302 if (VectorTy && isa<VectorType>(VectorTy) && HadAVector) {
303 DOUT << "CONVERT TO VECTOR: " << *AI << " TYPE = " << *VectorTy <<"\n";
305 // Create and insert the vector alloca.
306 NewAI = new AllocaInst(VectorTy, 0, "", AI->getParent()->begin());
307 ConvertUsesToScalar(AI, NewAI, 0);
309 DOUT << "CONVERT TO SCALAR INTEGER: " << *AI << "\n";
311 // Create and insert the integer alloca.
312 const Type *NewTy = IntegerType::get(AI->getContext(), AllocaSize*8);
313 NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
314 ConvertUsesToScalar(AI, NewAI, 0);
317 AI->eraseFromParent();
323 // Otherwise, couldn't process this alloca.
329 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
330 /// predicate, do SROA now.
331 void SROA::DoScalarReplacement(AllocationInst *AI,
332 std::vector<AllocationInst*> &WorkList) {
333 DOUT << "Found inst to SROA: " << *AI;
334 SmallVector<AllocaInst*, 32> ElementAllocas;
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() + "." + Twine(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() + "." + Twine(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 = UndefValue::get(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(Constant::getNullValue(
424 Type::getInt32Ty(AI->getContext())));
425 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end());
426 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(),
427 NewArgs.end(), "", GEPI);
428 RepValue->takeName(GEPI);
431 // If this GEP is to the start of the aggregate, check for memcpys.
432 if (Idx == 0 && GEPI->hasAllZeroIndices())
433 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas);
435 // Move all of the users over to the new GEP.
436 GEPI->replaceAllUsesWith(RepValue);
437 // Delete the old GEP
438 GEPI->eraseFromParent();
441 // Finally, delete the Alloca instruction
442 AI->eraseFromParent();
447 /// isSafeElementUse - Check to see if this use is an allowed use for a
448 /// getelementptr instruction of an array aggregate allocation. isFirstElt
449 /// indicates whether Ptr is known to the start of the aggregate.
451 void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI,
453 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
455 Instruction *User = cast<Instruction>(*I);
456 switch (User->getOpcode()) {
457 case Instruction::Load: break;
458 case Instruction::Store:
459 // Store is ok if storing INTO the pointer, not storing the pointer
460 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info);
462 case Instruction::GetElementPtr: {
463 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User);
464 bool AreAllZeroIndices = isFirstElt;
465 if (GEP->getNumOperands() > 1) {
466 if (!isa<ConstantInt>(GEP->getOperand(1)) ||
467 !cast<ConstantInt>(GEP->getOperand(1))->isZero())
468 // Using pointer arithmetic to navigate the array.
469 return MarkUnsafe(Info);
471 if (AreAllZeroIndices)
472 AreAllZeroIndices = GEP->hasAllZeroIndices();
474 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info);
475 if (Info.isUnsafe) return;
478 case Instruction::BitCast:
480 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info);
481 if (Info.isUnsafe) return;
484 DOUT << " Transformation preventing inst: " << *User;
485 return MarkUnsafe(Info);
486 case Instruction::Call:
487 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
489 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info);
490 if (Info.isUnsafe) return;
494 DOUT << " Transformation preventing inst: " << *User;
495 return MarkUnsafe(Info);
497 DOUT << " Transformation preventing inst: " << *User;
498 return MarkUnsafe(Info);
501 return; // All users look ok :)
504 /// AllUsersAreLoads - Return true if all users of this value are loads.
505 static bool AllUsersAreLoads(Value *Ptr) {
506 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
508 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load)
513 /// isSafeUseOfAllocation - Check to see if this user is an allowed use for an
514 /// aggregate allocation.
516 void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI,
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() != Constant::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 = Constant::getNullValue(Type::getInt32Ty(MI->getContext()));
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,
780 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) };
781 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2,
782 OtherPtr->getNameStr()+"."+Twine(i),
785 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
786 if (const StructType *ST =
787 dyn_cast<StructType>(OtherPtrTy->getElementType())) {
788 EltOffset = TD->getStructLayout(ST)->getElementOffset(i);
791 cast<SequentialType>(OtherPtr->getType())->getElementType();
792 EltOffset = TD->getTypeAllocSize(EltTy)*i;
795 // The alignment of the other pointer is the guaranteed alignment of the
796 // element, which is affected by both the known alignment of the whole
797 // mem intrinsic and the alignment of the element. If the alignment of
798 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
799 // known alignment is just 4 bytes.
800 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
803 Value *EltPtr = NewElts[i];
804 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
806 // If we got down to a scalar, insert a load or store as appropriate.
807 if (EltTy->isSingleValueType()) {
808 if (isa<MemTransferInst>(MI)) {
810 // From Other to Alloca.
811 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
812 new StoreInst(Elt, EltPtr, MI);
814 // From Alloca to Other.
815 Value *Elt = new LoadInst(EltPtr, "tmp", MI);
816 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI);
820 assert(isa<MemSetInst>(MI));
822 // If the stored element is zero (common case), just store a null
825 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) {
827 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
829 // If EltTy is a vector type, get the element type.
830 const Type *ValTy = EltTy->getScalarType();
832 // Construct an integer with the right value.
833 unsigned EltSize = TD->getTypeSizeInBits(ValTy);
834 APInt OneVal(EltSize, CI->getZExtValue());
835 APInt TotalVal(OneVal);
837 for (unsigned i = 0; 8*i < EltSize; ++i) {
838 TotalVal = TotalVal.shl(8);
842 // Convert the integer value to the appropriate type.
843 StoreVal = ConstantInt::get(Context, TotalVal);
844 if (isa<PointerType>(ValTy))
845 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
846 else if (ValTy->isFloatingPoint())
847 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
848 assert(StoreVal->getType() == ValTy && "Type mismatch!");
850 // If the requested value was a vector constant, create it.
851 if (EltTy != ValTy) {
852 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements();
853 SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
854 StoreVal = ConstantVector::get(&Elts[0], NumElts);
857 new StoreInst(StoreVal, EltPtr, MI);
860 // Otherwise, if we're storing a byte variable, use a memset call for
864 // Cast the element pointer to BytePtrTy.
865 if (EltPtr->getType() != BytePtrTy)
866 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI);
868 // Cast the other pointer (if we have one) to BytePtrTy.
869 if (OtherElt && OtherElt->getType() != BytePtrTy)
870 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(),
873 unsigned EltSize = TD->getTypeAllocSize(EltTy);
875 // Finally, insert the meminst for this element.
876 if (isa<MemTransferInst>(MI)) {
878 SROADest ? EltPtr : OtherElt, // Dest ptr
879 SROADest ? OtherElt : EltPtr, // Src ptr
880 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
882 ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign)
884 CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
886 assert(isa<MemSetInst>(MI));
888 EltPtr, MI->getOperand(2), // Dest, Value,
889 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
892 CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
895 MI->eraseFromParent();
898 /// RewriteStoreUserOfWholeAlloca - We found an store of an integer that
899 /// overwrites the entire allocation. Extract out the pieces of the stored
900 /// integer and store them individually.
901 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI,
903 SmallVector<AllocaInst*, 32> &NewElts){
904 // Extract each element out of the integer according to its structure offset
905 // and store the element value to the individual alloca.
906 Value *SrcVal = SI->getOperand(0);
907 const Type *AllocaEltTy = AI->getType()->getElementType();
908 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
910 // If this isn't a store of an integer to the whole alloca, it may be a store
911 // to the first element. Just ignore the store in this case and normal SROA
913 if (!isa<IntegerType>(SrcVal->getType()) ||
914 TD->getTypeAllocSizeInBits(SrcVal->getType()) != AllocaSizeBits)
916 // Handle tail padding by extending the operand
917 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
918 SrcVal = new ZExtInst(SrcVal,
919 IntegerType::get(SI->getContext(), AllocaSizeBits),
922 DOUT << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << *SI;
924 // There are two forms here: AI could be an array or struct. Both cases
925 // have different ways to compute the element offset.
926 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
927 const StructLayout *Layout = TD->getStructLayout(EltSTy);
929 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
930 // Get the number of bits to shift SrcVal to get the value.
931 const Type *FieldTy = EltSTy->getElementType(i);
932 uint64_t Shift = Layout->getElementOffsetInBits(i);
934 if (TD->isBigEndian())
935 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy);
937 Value *EltVal = SrcVal;
939 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
940 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
941 "sroa.store.elt", SI);
944 // Truncate down to an integer of the right size.
945 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
947 // Ignore zero sized fields like {}, they obviously contain no data.
948 if (FieldSizeBits == 0) continue;
950 if (FieldSizeBits != AllocaSizeBits)
951 EltVal = new TruncInst(EltVal,
952 IntegerType::get(SI->getContext(), FieldSizeBits),
954 Value *DestField = NewElts[i];
955 if (EltVal->getType() == FieldTy) {
956 // Storing to an integer field of this size, just do it.
957 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) {
958 // Bitcast to the right element type (for fp/vector values).
959 EltVal = new BitCastInst(EltVal, FieldTy, "", SI);
961 // Otherwise, bitcast the dest pointer (for aggregates).
962 DestField = new BitCastInst(DestField,
963 PointerType::getUnqual(EltVal->getType()),
966 new StoreInst(EltVal, DestField, SI);
970 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
971 const Type *ArrayEltTy = ATy->getElementType();
972 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
973 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
977 if (TD->isBigEndian())
978 Shift = AllocaSizeBits-ElementOffset;
982 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
983 // Ignore zero sized fields like {}, they obviously contain no data.
984 if (ElementSizeBits == 0) continue;
986 Value *EltVal = SrcVal;
988 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
989 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
990 "sroa.store.elt", SI);
993 // Truncate down to an integer of the right size.
994 if (ElementSizeBits != AllocaSizeBits)
995 EltVal = new TruncInst(EltVal,
996 IntegerType::get(SI->getContext(),
997 ElementSizeBits),"",SI);
998 Value *DestField = NewElts[i];
999 if (EltVal->getType() == ArrayEltTy) {
1000 // Storing to an integer field of this size, just do it.
1001 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) {
1002 // Bitcast to the right element type (for fp/vector values).
1003 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI);
1005 // Otherwise, bitcast the dest pointer (for aggregates).
1006 DestField = new BitCastInst(DestField,
1007 PointerType::getUnqual(EltVal->getType()),
1010 new StoreInst(EltVal, DestField, SI);
1012 if (TD->isBigEndian())
1013 Shift -= ElementOffset;
1015 Shift += ElementOffset;
1019 SI->eraseFromParent();
1022 /// RewriteLoadUserOfWholeAlloca - We found an load of the entire allocation to
1023 /// an integer. Load the individual pieces to form the aggregate value.
1024 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI,
1025 SmallVector<AllocaInst*, 32> &NewElts) {
1026 // Extract each element out of the NewElts according to its structure offset
1027 // and form the result value.
1028 const Type *AllocaEltTy = AI->getType()->getElementType();
1029 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
1031 // If this isn't a load of the whole alloca to an integer, it may be a load
1032 // of the first element. Just ignore the load in this case and normal SROA
1034 if (!isa<IntegerType>(LI->getType()) ||
1035 TD->getTypeAllocSizeInBits(LI->getType()) != AllocaSizeBits)
1038 DOUT << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << *LI;
1040 // There are two forms here: AI could be an array or struct. Both cases
1041 // have different ways to compute the element offset.
1042 const StructLayout *Layout = 0;
1043 uint64_t ArrayEltBitOffset = 0;
1044 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
1045 Layout = TD->getStructLayout(EltSTy);
1047 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
1048 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
1052 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits));
1054 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1055 // Load the value from the alloca. If the NewElt is an aggregate, cast
1056 // the pointer to an integer of the same size before doing the load.
1057 Value *SrcField = NewElts[i];
1058 const Type *FieldTy =
1059 cast<PointerType>(SrcField->getType())->getElementType();
1060 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
1062 // Ignore zero sized fields like {}, they obviously contain no data.
1063 if (FieldSizeBits == 0) continue;
1065 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(),
1067 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() &&
1068 !isa<VectorType>(FieldTy))
1069 SrcField = new BitCastInst(SrcField,
1070 PointerType::getUnqual(FieldIntTy),
1072 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
1074 // If SrcField is a fp or vector of the right size but that isn't an
1075 // integer type, bitcast to an integer so we can shift it.
1076 if (SrcField->getType() != FieldIntTy)
1077 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
1079 // Zero extend the field to be the same size as the final alloca so that
1080 // we can shift and insert it.
1081 if (SrcField->getType() != ResultVal->getType())
1082 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
1084 // Determine the number of bits to shift SrcField.
1086 if (Layout) // Struct case.
1087 Shift = Layout->getElementOffsetInBits(i);
1089 Shift = i*ArrayEltBitOffset;
1091 if (TD->isBigEndian())
1092 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
1095 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
1096 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
1099 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
1102 // Handle tail padding by truncating the result
1103 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
1104 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI);
1106 LI->replaceAllUsesWith(ResultVal);
1107 LI->eraseFromParent();
1111 /// HasPadding - Return true if the specified type has any structure or
1112 /// alignment padding, false otherwise.
1113 static bool HasPadding(const Type *Ty, const TargetData &TD) {
1114 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1115 const StructLayout *SL = TD.getStructLayout(STy);
1116 unsigned PrevFieldBitOffset = 0;
1117 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1118 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
1120 // Padding in sub-elements?
1121 if (HasPadding(STy->getElementType(i), TD))
1124 // Check to see if there is any padding between this element and the
1127 unsigned PrevFieldEnd =
1128 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
1129 if (PrevFieldEnd < FieldBitOffset)
1133 PrevFieldBitOffset = FieldBitOffset;
1136 // Check for tail padding.
1137 if (unsigned EltCount = STy->getNumElements()) {
1138 unsigned PrevFieldEnd = PrevFieldBitOffset +
1139 TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
1140 if (PrevFieldEnd < SL->getSizeInBits())
1144 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1145 return HasPadding(ATy->getElementType(), TD);
1146 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1147 return HasPadding(VTy->getElementType(), TD);
1149 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
1152 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
1153 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
1154 /// or 1 if safe after canonicalization has been performed.
1156 int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) {
1157 // Loop over the use list of the alloca. We can only transform it if all of
1158 // the users are safe to transform.
1161 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end();
1163 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info);
1164 if (Info.isUnsafe) {
1165 DOUT << "Cannot transform: " << *AI << " due to user: " << **I;
1170 // Okay, we know all the users are promotable. If the aggregate is a memcpy
1171 // source and destination, we have to be careful. In particular, the memcpy
1172 // could be moving around elements that live in structure padding of the LLVM
1173 // types, but may actually be used. In these cases, we refuse to promote the
1175 if (Info.isMemCpySrc && Info.isMemCpyDst &&
1176 HasPadding(AI->getType()->getElementType(), *TD))
1179 // If we require cleanup, return 1, otherwise return 3.
1180 return Info.needsCleanup ? 1 : 3;
1183 /// CleanupGEP - GEP is used by an Alloca, which can be prompted after the GEP
1184 /// is canonicalized here.
1185 void SROA::CleanupGEP(GetElementPtrInst *GEPI) {
1186 gep_type_iterator I = gep_type_begin(GEPI);
1189 const ArrayType *AT = dyn_cast<ArrayType>(*I);
1193 uint64_t NumElements = AT->getNumElements();
1195 if (isa<ConstantInt>(I.getOperand()))
1198 if (NumElements == 1) {
1200 Constant::getNullValue(Type::getInt32Ty(GEPI->getContext())));
1204 assert(NumElements == 2 && "Unhandled case!");
1205 // All users of the GEP must be loads. At each use of the GEP, insert
1206 // two loads of the appropriate indexed GEP and select between them.
1207 Value *IsOne = new ICmpInst(GEPI, ICmpInst::ICMP_NE, I.getOperand(),
1208 Constant::getNullValue(I.getOperand()->getType()),
1210 // Insert the new GEP instructions, which are properly indexed.
1211 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end());
1212 Indices[1] = Constant::getNullValue(Type::getInt32Ty(GEPI->getContext()));
1213 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
1216 GEPI->getName()+".0", GEPI);
1217 Indices[1] = ConstantInt::get(Type::getInt32Ty(GEPI->getContext()), 1);
1218 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
1221 GEPI->getName()+".1", GEPI);
1222 // Replace all loads of the variable index GEP with loads from both
1223 // indexes and a select.
1224 while (!GEPI->use_empty()) {
1225 LoadInst *LI = cast<LoadInst>(GEPI->use_back());
1226 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI);
1227 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI);
1228 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI);
1229 LI->replaceAllUsesWith(R);
1230 LI->eraseFromParent();
1232 GEPI->eraseFromParent();
1236 /// CleanupAllocaUsers - If SROA reported that it can promote the specified
1237 /// allocation, but only if cleaned up, perform the cleanups required.
1238 void SROA::CleanupAllocaUsers(AllocationInst *AI) {
1239 // At this point, we know that the end result will be SROA'd and promoted, so
1240 // we can insert ugly code if required so long as sroa+mem2reg will clean it
1242 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
1245 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U))
1248 Instruction *I = cast<Instruction>(U);
1249 SmallVector<DbgInfoIntrinsic *, 2> DbgInUses;
1250 if (!isa<StoreInst>(I) && OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) {
1251 // Safe to remove debug info uses.
1252 while (!DbgInUses.empty()) {
1253 DbgInfoIntrinsic *DI = DbgInUses.back(); DbgInUses.pop_back();
1254 DI->eraseFromParent();
1256 I->eraseFromParent();
1262 /// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at
1263 /// the offset specified by Offset (which is specified in bytes).
1265 /// There are two cases we handle here:
1266 /// 1) A union of vector types of the same size and potentially its elements.
1267 /// Here we turn element accesses into insert/extract element operations.
1268 /// This promotes a <4 x float> with a store of float to the third element
1269 /// into a <4 x float> that uses insert element.
1270 /// 2) A fully general blob of memory, which we turn into some (potentially
1271 /// large) integer type with extract and insert operations where the loads
1272 /// and stores would mutate the memory.
1273 static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy,
1274 unsigned AllocaSize, const TargetData &TD,
1275 LLVMContext &Context) {
1276 // If this could be contributing to a vector, analyze it.
1277 if (VecTy != Type::getVoidTy(Context)) { // either null or a vector type.
1279 // If the In type is a vector that is the same size as the alloca, see if it
1280 // matches the existing VecTy.
1281 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) {
1282 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) {
1283 // If we're storing/loading a vector of the right size, allow it as a
1284 // vector. If this the first vector we see, remember the type so that
1285 // we know the element size.
1290 } else if (In == Type::getFloatTy(Context) ||
1291 In == Type::getDoubleTy(Context) ||
1292 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 &&
1293 isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
1294 // If we're accessing something that could be an element of a vector, see
1295 // if the implied vector agrees with what we already have and if Offset is
1296 // compatible with it.
1297 unsigned EltSize = In->getPrimitiveSizeInBits()/8;
1298 if (Offset % EltSize == 0 &&
1299 AllocaSize % EltSize == 0 &&
1301 cast<VectorType>(VecTy)->getElementType()
1302 ->getPrimitiveSizeInBits()/8 == EltSize)) {
1304 VecTy = VectorType::get(In, AllocaSize/EltSize);
1310 // Otherwise, we have a case that we can't handle with an optimized vector
1311 // form. We can still turn this into a large integer.
1312 VecTy = Type::getVoidTy(Context);
1315 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
1316 /// its accesses to use a to single vector type, return true, and set VecTy to
1317 /// the new type. If we could convert the alloca into a single promotable
1318 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a
1319 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
1320 /// is the current offset from the base of the alloca being analyzed.
1322 /// If we see at least one access to the value that is as a vector type, set the
1325 bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy,
1326 bool &SawVec, uint64_t Offset,
1327 unsigned AllocaSize) {
1328 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1329 Instruction *User = cast<Instruction>(*UI);
1331 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1332 // Don't break volatile loads.
1333 if (LI->isVolatile())
1335 MergeInType(LI->getType(), Offset, VecTy,
1336 AllocaSize, *TD, V->getContext());
1337 SawVec |= isa<VectorType>(LI->getType());
1341 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1342 // Storing the pointer, not into the value?
1343 if (SI->getOperand(0) == V || SI->isVolatile()) return 0;
1344 MergeInType(SI->getOperand(0)->getType(), Offset,
1345 VecTy, AllocaSize, *TD, V->getContext());
1346 SawVec |= isa<VectorType>(SI->getOperand(0)->getType());
1350 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
1351 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset,
1354 IsNotTrivial = true;
1358 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1359 // If this is a GEP with a variable indices, we can't handle it.
1360 if (!GEP->hasAllConstantIndices())
1363 // Compute the offset that this GEP adds to the pointer.
1364 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
1365 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(),
1366 &Indices[0], Indices.size());
1367 // See if all uses can be converted.
1368 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset,
1371 IsNotTrivial = true;
1375 // If this is a constant sized memset of a constant value (e.g. 0) we can
1377 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
1378 // Store of constant value and constant size.
1379 if (isa<ConstantInt>(MSI->getValue()) &&
1380 isa<ConstantInt>(MSI->getLength())) {
1381 IsNotTrivial = true;
1386 // If this is a memcpy or memmove into or out of the whole allocation, we
1387 // can handle it like a load or store of the scalar type.
1388 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
1389 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()))
1390 if (Len->getZExtValue() == AllocaSize && Offset == 0) {
1391 IsNotTrivial = true;
1396 // Ignore dbg intrinsic.
1397 if (isa<DbgInfoIntrinsic>(User))
1400 // Otherwise, we cannot handle this!
1408 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
1409 /// directly. This happens when we are converting an "integer union" to a
1410 /// single integer scalar, or when we are converting a "vector union" to a
1411 /// vector with insert/extractelement instructions.
1413 /// Offset is an offset from the original alloca, in bits that need to be
1414 /// shifted to the right. By the end of this, there should be no uses of Ptr.
1415 void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) {
1416 while (!Ptr->use_empty()) {
1417 Instruction *User = cast<Instruction>(Ptr->use_back());
1419 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
1420 ConvertUsesToScalar(CI, NewAI, Offset);
1421 CI->eraseFromParent();
1425 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1426 // Compute the offset that this GEP adds to the pointer.
1427 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
1428 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(),
1429 &Indices[0], Indices.size());
1430 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
1431 GEP->eraseFromParent();
1435 IRBuilder<> Builder(User->getParent(), User);
1437 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1438 // The load is a bit extract from NewAI shifted right by Offset bits.
1439 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp");
1441 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
1442 LI->replaceAllUsesWith(NewLoadVal);
1443 LI->eraseFromParent();
1447 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1448 assert(SI->getOperand(0) != Ptr && "Consistency error!");
1449 // FIXME: Remove once builder has Twine API.
1450 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").str().c_str());
1451 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
1453 Builder.CreateStore(New, NewAI);
1454 SI->eraseFromParent();
1458 // If this is a constant sized memset of a constant value (e.g. 0) we can
1459 // transform it into a store of the expanded constant value.
1460 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
1461 assert(MSI->getRawDest() == Ptr && "Consistency error!");
1462 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
1463 if (NumBytes != 0) {
1464 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
1466 // Compute the value replicated the right number of times.
1467 APInt APVal(NumBytes*8, Val);
1469 // Splat the value if non-zero.
1471 for (unsigned i = 1; i != NumBytes; ++i)
1472 APVal |= APVal << 8;
1474 // FIXME: Remove once builder has Twine API.
1475 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").str().c_str());
1476 Value *New = ConvertScalar_InsertValue(
1477 ConstantInt::get(User->getContext(), APVal),
1478 Old, Offset, Builder);
1479 Builder.CreateStore(New, NewAI);
1481 MSI->eraseFromParent();
1485 // If this is a memcpy or memmove into or out of the whole allocation, we
1486 // can handle it like a load or store of the scalar type.
1487 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
1488 assert(Offset == 0 && "must be store to start of alloca");
1490 // If the source and destination are both to the same alloca, then this is
1491 // a noop copy-to-self, just delete it. Otherwise, emit a load and store
1493 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject());
1495 if (MTI->getSource()->getUnderlyingObject() != OrigAI) {
1496 // Dest must be OrigAI, change this to be a load from the original
1497 // pointer (bitcasted), then a store to our new alloca.
1498 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
1499 Value *SrcPtr = MTI->getSource();
1500 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType());
1502 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
1503 SrcVal->setAlignment(MTI->getAlignment());
1504 Builder.CreateStore(SrcVal, NewAI);
1505 } else if (MTI->getDest()->getUnderlyingObject() != OrigAI) {
1506 // Src must be OrigAI, change this to be a load from NewAI then a store
1507 // through the original dest pointer (bitcasted).
1508 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
1509 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
1511 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType());
1512 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
1513 NewStore->setAlignment(MTI->getAlignment());
1515 // Noop transfer. Src == Dst
1519 MTI->eraseFromParent();
1523 // If user is a dbg info intrinsic then it is safe to remove it.
1524 if (isa<DbgInfoIntrinsic>(User)) {
1525 User->eraseFromParent();
1529 llvm_unreachable("Unsupported operation!");
1533 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
1534 /// or vector value FromVal, extracting the bits from the offset specified by
1535 /// Offset. This returns the value, which is of type ToType.
1537 /// This happens when we are converting an "integer union" to a single
1538 /// integer scalar, or when we are converting a "vector union" to a vector with
1539 /// insert/extractelement instructions.
1541 /// Offset is an offset from the original alloca, in bits that need to be
1542 /// shifted to the right.
1543 Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType,
1544 uint64_t Offset, IRBuilder<> &Builder) {
1545 // If the load is of the whole new alloca, no conversion is needed.
1546 if (FromVal->getType() == ToType && Offset == 0)
1549 // If the result alloca is a vector type, this is either an element
1550 // access or a bitcast to another vector type of the same size.
1551 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) {
1552 if (isa<VectorType>(ToType))
1553 return Builder.CreateBitCast(FromVal, ToType, "tmp");
1555 // Otherwise it must be an element access.
1558 unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType());
1559 Elt = Offset/EltSize;
1560 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
1562 // Return the element extracted out of it.
1563 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get(
1564 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp");
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 = UndefValue::get(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 = UndefValue::get(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, IntegerType::get(FromVal->getContext(),
1628 LIBitWidth), "tmp");
1629 else if (LIBitWidth > NTy->getBitWidth())
1631 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(),
1632 LIBitWidth), "tmp");
1634 // If the result is an integer, this is a trunc or bitcast.
1635 if (isa<IntegerType>(ToType)) {
1637 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) {
1638 // Just do a bitcast, we know the sizes match up.
1639 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp");
1641 // Otherwise must be a pointer.
1642 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp");
1644 assert(FromVal->getType() == ToType && "Didn't convert right?");
1649 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
1650 /// or vector value "Old" at the offset specified by Offset.
1652 /// This happens when we are converting an "integer union" to a
1653 /// single integer scalar, or when we are converting a "vector union" to a
1654 /// vector with insert/extractelement instructions.
1656 /// Offset is an offset from the original alloca, in bits that need to be
1657 /// shifted to the right.
1658 Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old,
1659 uint64_t Offset, IRBuilder<> &Builder) {
1661 // Convert the stored type to the actual type, shift it left to insert
1662 // then 'or' into place.
1663 const Type *AllocaType = Old->getType();
1664 LLVMContext &Context = Old->getContext();
1666 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
1667 uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy);
1668 uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType());
1670 // Changing the whole vector with memset or with an access of a different
1672 if (ValSize == VecSize)
1673 return Builder.CreateBitCast(SV, AllocaType, "tmp");
1675 uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType());
1677 // Must be an element insertion.
1678 unsigned Elt = Offset/EltSize;
1680 if (SV->getType() != VTy->getElementType())
1681 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp");
1683 SV = Builder.CreateInsertElement(Old, SV,
1684 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt),
1689 // If SV is a first-class aggregate value, insert each value recursively.
1690 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) {
1691 const StructLayout &Layout = *TD->getStructLayout(ST);
1692 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
1693 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
1694 Old = ConvertScalar_InsertValue(Elt, Old,
1695 Offset+Layout.getElementOffsetInBits(i),
1701 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
1702 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType());
1703 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1704 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
1705 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
1710 // If SV is a float, convert it to the appropriate integer type.
1711 // If it is a pointer, do the same.
1712 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType());
1713 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType);
1714 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType());
1715 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType);
1716 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType()))
1717 SV = Builder.CreateBitCast(SV,
1718 IntegerType::get(SV->getContext(),SrcWidth), "tmp");
1719 else if (isa<PointerType>(SV->getType()))
1720 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(SV->getContext()), "tmp");
1722 // Zero extend or truncate the value if needed.
1723 if (SV->getType() != AllocaType) {
1724 if (SV->getType()->getPrimitiveSizeInBits() <
1725 AllocaType->getPrimitiveSizeInBits())
1726 SV = Builder.CreateZExt(SV, AllocaType, "tmp");
1728 // Truncation may be needed if storing more than the alloca can hold
1729 // (undefined behavior).
1730 SV = Builder.CreateTrunc(SV, AllocaType, "tmp");
1731 SrcWidth = DestWidth;
1732 SrcStoreWidth = DestStoreWidth;
1736 // If this is a big-endian system and the store is narrower than the
1737 // full alloca type, we need to do a shift to get the right bits.
1739 if (TD->isBigEndian()) {
1740 // On big-endian machines, the lowest bit is stored at the bit offset
1741 // from the pointer given by getTypeStoreSizeInBits. This matters for
1742 // integers with a bitwidth that is not a multiple of 8.
1743 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
1748 // Note: we support negative bitwidths (with shr) which are not defined.
1749 // We do this to support (f.e.) stores off the end of a structure where
1750 // only some bits in the structure are set.
1751 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
1752 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
1753 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(),
1756 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
1757 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(),
1759 Mask = Mask.lshr(-ShAmt);
1762 // Mask out the bits we are about to insert from the old value, and or
1764 if (SrcWidth != DestWidth) {
1765 assert(DestWidth > SrcWidth);
1766 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask");
1767 SV = Builder.CreateOr(Old, SV, "ins");
1774 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
1775 /// some part of a constant global variable. This intentionally only accepts
1776 /// constant expressions because we don't can't rewrite arbitrary instructions.
1777 static bool PointsToConstantGlobal(Value *V) {
1778 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
1779 return GV->isConstant();
1780 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
1781 if (CE->getOpcode() == Instruction::BitCast ||
1782 CE->getOpcode() == Instruction::GetElementPtr)
1783 return PointsToConstantGlobal(CE->getOperand(0));
1787 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
1788 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
1789 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
1790 /// track of whether it moves the pointer (with isOffset) but otherwise traverse
1791 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
1792 /// the alloca, and if the source pointer is a pointer to a constant global, we
1793 /// can optimize this.
1794 static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy,
1796 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1797 if (LoadInst *LI = dyn_cast<LoadInst>(*UI))
1798 // Ignore non-volatile loads, they are always ok.
1799 if (!LI->isVolatile())
1802 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) {
1803 // If uses of the bitcast are ok, we are ok.
1804 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset))
1808 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
1809 // If the GEP has all zero indices, it doesn't offset the pointer. If it
1810 // doesn't, it does.
1811 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
1812 isOffset || !GEP->hasAllZeroIndices()))
1817 // If this is isn't our memcpy/memmove, reject it as something we can't
1819 if (!isa<MemTransferInst>(*UI))
1822 // If we already have seen a copy, reject the second one.
1823 if (TheCopy) return false;
1825 // If the pointer has been offset from the start of the alloca, we can't
1826 // safely handle this.
1827 if (isOffset) return false;
1829 // If the memintrinsic isn't using the alloca as the dest, reject it.
1830 if (UI.getOperandNo() != 1) return false;
1832 MemIntrinsic *MI = cast<MemIntrinsic>(*UI);
1834 // If the source of the memcpy/move is not a constant global, reject it.
1835 if (!PointsToConstantGlobal(MI->getOperand(2)))
1838 // Otherwise, the transform is safe. Remember the copy instruction.
1844 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
1845 /// modified by a copy from a constant global. If we can prove this, we can
1846 /// replace any uses of the alloca with uses of the global directly.
1847 Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) {
1848 Instruction *TheCopy = 0;
1849 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false))