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/Support/raw_ostream.h"
43 #include "llvm/ADT/SmallVector.h"
44 #include "llvm/ADT/Statistic.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>();
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 = getAnalysisIfAvailable<TargetData>();
155 bool Changed = performPromotion(F);
157 // FIXME: ScalarRepl currently depends on TargetData more than it
158 // theoretically needs to. It should be refactored in order to support
159 // target-independent IR. Until this is done, just skip the actual
160 // scalar-replacement portion of this pass.
161 if (!TD) return Changed;
164 bool LocalChange = performScalarRepl(F);
165 if (!LocalChange) break; // No need to repromote if no scalarrepl
167 LocalChange = performPromotion(F);
168 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
175 bool SROA::performPromotion(Function &F) {
176 std::vector<AllocaInst*> Allocas;
177 DominatorTree &DT = getAnalysis<DominatorTree>();
178 DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
180 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
182 bool Changed = false;
187 // Find allocas that are safe to promote, by looking at all instructions in
189 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
190 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
191 if (isAllocaPromotable(AI))
192 Allocas.push_back(AI);
194 if (Allocas.empty()) break;
196 PromoteMemToReg(Allocas, DT, DF, F.getContext());
197 NumPromoted += Allocas.size();
204 /// getNumSAElements - Return the number of elements in the specific struct or
206 static uint64_t getNumSAElements(const Type *T) {
207 if (const StructType *ST = dyn_cast<StructType>(T))
208 return ST->getNumElements();
209 return cast<ArrayType>(T)->getNumElements();
212 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
213 // which runs on all of the malloc/alloca instructions in the function, removing
214 // them if they are only used by getelementptr instructions.
216 bool SROA::performScalarRepl(Function &F) {
217 std::vector<AllocationInst*> WorkList;
219 // Scan the entry basic block, adding any alloca's and mallocs to the worklist
220 BasicBlock &BB = F.getEntryBlock();
221 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
222 if (AllocationInst *A = dyn_cast<AllocationInst>(I))
223 WorkList.push_back(A);
225 // Process the worklist
226 bool Changed = false;
227 while (!WorkList.empty()) {
228 AllocationInst *AI = WorkList.back();
231 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
232 // with unused elements.
233 if (AI->use_empty()) {
234 AI->eraseFromParent();
238 // If this alloca is impossible for us to promote, reject it early.
239 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
242 // Check to see if this allocation is only modified by a memcpy/memmove from
243 // a constant global. If this is the case, we can change all users to use
244 // the constant global instead. This is commonly produced by the CFE by
245 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
246 // is only subsequently read.
247 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) {
248 DEBUG(errs() << "Found alloca equal to global: " << *AI);
249 DEBUG(errs() << " memcpy = " << *TheCopy);
250 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2));
251 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
252 TheCopy->eraseFromParent(); // Don't mutate the global.
253 AI->eraseFromParent();
259 // Check to see if we can perform the core SROA transformation. We cannot
260 // transform the allocation instruction if it is an array allocation
261 // (allocations OF arrays are ok though), and an allocation of a scalar
262 // value cannot be decomposed at all.
263 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
265 // Do not promote [0 x %struct].
266 if (AllocaSize == 0) continue;
268 // Do not promote any struct whose size is too big.
269 if (AllocaSize > SRThreshold) continue;
271 if ((isa<StructType>(AI->getAllocatedType()) ||
272 isa<ArrayType>(AI->getAllocatedType())) &&
273 // Do not promote any struct into more than "32" separate vars.
274 getNumSAElements(AI->getAllocatedType()) <= SRThreshold/4) {
275 // Check that all of the users of the allocation are capable of being
277 switch (isSafeAllocaToScalarRepl(AI)) {
278 default: llvm_unreachable("Unexpected value!");
279 case 0: // Not safe to scalar replace.
281 case 1: // Safe, but requires cleanup/canonicalizations first
282 CleanupAllocaUsers(AI);
284 case 3: // Safe to scalar replace.
285 DoScalarReplacement(AI, WorkList);
291 // If we can turn this aggregate value (potentially with casts) into a
292 // simple scalar value that can be mem2reg'd into a register value.
293 // IsNotTrivial tracks whether this is something that mem2reg could have
294 // promoted itself. If so, we don't want to transform it needlessly. Note
295 // that we can't just check based on the type: the alloca may be of an i32
296 // but that has pointer arithmetic to set byte 3 of it or something.
297 bool IsNotTrivial = false;
298 const Type *VectorTy = 0;
299 bool HadAVector = false;
300 if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector,
301 0, unsigned(AllocaSize)) && IsNotTrivial) {
303 // If we were able to find a vector type that can handle this with
304 // insert/extract elements, and if there was at least one use that had
305 // a vector type, promote this to a vector. We don't want to promote
306 // random stuff that doesn't use vectors (e.g. <9 x double>) because then
307 // we just get a lot of insert/extracts. If at least one vector is
308 // involved, then we probably really do have a union of vector/array.
309 if (VectorTy && isa<VectorType>(VectorTy) && HadAVector) {
310 DEBUG(errs() << "CONVERT TO VECTOR: " << *AI << " TYPE = "
311 << *VectorTy << '\n');
313 // Create and insert the vector alloca.
314 NewAI = new AllocaInst(VectorTy, 0, "", AI->getParent()->begin());
315 ConvertUsesToScalar(AI, NewAI, 0);
317 DEBUG(errs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n");
319 // Create and insert the integer alloca.
320 const Type *NewTy = IntegerType::get(AI->getContext(), AllocaSize*8);
321 NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
322 ConvertUsesToScalar(AI, NewAI, 0);
325 AI->eraseFromParent();
331 // Otherwise, couldn't process this alloca.
337 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
338 /// predicate, do SROA now.
339 void SROA::DoScalarReplacement(AllocationInst *AI,
340 std::vector<AllocationInst*> &WorkList) {
341 DEBUG(errs() << "Found inst to SROA: " << *AI);
342 SmallVector<AllocaInst*, 32> ElementAllocas;
343 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
344 ElementAllocas.reserve(ST->getNumContainedTypes());
345 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
346 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
348 AI->getName() + "." + Twine(i), AI);
349 ElementAllocas.push_back(NA);
350 WorkList.push_back(NA); // Add to worklist for recursive processing
353 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
354 ElementAllocas.reserve(AT->getNumElements());
355 const Type *ElTy = AT->getElementType();
356 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
357 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
358 AI->getName() + "." + Twine(i), AI);
359 ElementAllocas.push_back(NA);
360 WorkList.push_back(NA); // Add to worklist for recursive processing
364 // Now that we have created the alloca instructions that we want to use,
365 // expand the getelementptr instructions to use them.
367 while (!AI->use_empty()) {
368 Instruction *User = cast<Instruction>(AI->use_back());
369 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) {
370 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas);
371 BCInst->eraseFromParent();
376 // %res = load { i32, i32 }* %alloc
378 // %load.0 = load i32* %alloc.0
379 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
380 // %load.1 = load i32* %alloc.1
381 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
382 // (Also works for arrays instead of structs)
383 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
384 Value *Insert = UndefValue::get(LI->getType());
385 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) {
386 Value *Load = new LoadInst(ElementAllocas[i], "load", LI);
387 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI);
389 LI->replaceAllUsesWith(Insert);
390 LI->eraseFromParent();
395 // store { i32, i32 } %val, { i32, i32 }* %alloc
397 // %val.0 = extractvalue { i32, i32 } %val, 0
398 // store i32 %val.0, i32* %alloc.0
399 // %val.1 = extractvalue { i32, i32 } %val, 1
400 // store i32 %val.1, i32* %alloc.1
401 // (Also works for arrays instead of structs)
402 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
403 Value *Val = SI->getOperand(0);
404 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) {
405 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI);
406 new StoreInst(Extract, ElementAllocas[i], SI);
408 SI->eraseFromParent();
412 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
413 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst>
415 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
417 assert(Idx < ElementAllocas.size() && "Index out of range?");
418 AllocaInst *AllocaToUse = ElementAllocas[Idx];
421 if (GEPI->getNumOperands() == 3) {
422 // Do not insert a new getelementptr instruction with zero indices, only
423 // to have it optimized out later.
424 RepValue = AllocaToUse;
426 // We are indexing deeply into the structure, so we still need a
427 // getelement ptr instruction to finish the indexing. This may be
428 // expanded itself once the worklist is rerun.
430 SmallVector<Value*, 8> NewArgs;
431 NewArgs.push_back(Constant::getNullValue(
432 Type::getInt32Ty(AI->getContext())));
433 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end());
434 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(),
435 NewArgs.end(), "", GEPI);
436 RepValue->takeName(GEPI);
439 // If this GEP is to the start of the aggregate, check for memcpys.
440 if (Idx == 0 && GEPI->hasAllZeroIndices())
441 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas);
443 // Move all of the users over to the new GEP.
444 GEPI->replaceAllUsesWith(RepValue);
445 // Delete the old GEP
446 GEPI->eraseFromParent();
449 // Finally, delete the Alloca instruction
450 AI->eraseFromParent();
455 /// isSafeElementUse - Check to see if this use is an allowed use for a
456 /// getelementptr instruction of an array aggregate allocation. isFirstElt
457 /// indicates whether Ptr is known to the start of the aggregate.
459 void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI,
461 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
463 Instruction *User = cast<Instruction>(*I);
464 switch (User->getOpcode()) {
465 case Instruction::Load: break;
466 case Instruction::Store:
467 // Store is ok if storing INTO the pointer, not storing the pointer
468 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info);
470 case Instruction::GetElementPtr: {
471 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User);
472 bool AreAllZeroIndices = isFirstElt;
473 if (GEP->getNumOperands() > 1) {
474 if (!isa<ConstantInt>(GEP->getOperand(1)) ||
475 !cast<ConstantInt>(GEP->getOperand(1))->isZero())
476 // Using pointer arithmetic to navigate the array.
477 return MarkUnsafe(Info);
479 if (AreAllZeroIndices)
480 AreAllZeroIndices = GEP->hasAllZeroIndices();
482 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info);
483 if (Info.isUnsafe) return;
486 case Instruction::BitCast:
488 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info);
489 if (Info.isUnsafe) return;
492 DEBUG(errs() << " Transformation preventing inst: " << *User);
493 return MarkUnsafe(Info);
494 case Instruction::Call:
495 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
497 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info);
498 if (Info.isUnsafe) return;
502 DEBUG(errs() << " Transformation preventing inst: " << *User);
503 return MarkUnsafe(Info);
505 DEBUG(errs() << " Transformation preventing inst: " << *User);
506 return MarkUnsafe(Info);
509 return; // All users look ok :)
512 /// AllUsersAreLoads - Return true if all users of this value are loads.
513 static bool AllUsersAreLoads(Value *Ptr) {
514 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
516 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load)
521 /// isSafeUseOfAllocation - Check to see if this user is an allowed use for an
522 /// aggregate allocation.
524 void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI,
526 if (BitCastInst *C = dyn_cast<BitCastInst>(User))
527 return isSafeUseOfBitCastedAllocation(C, AI, Info);
529 if (LoadInst *LI = dyn_cast<LoadInst>(User))
530 if (!LI->isVolatile())
531 return;// Loads (returning a first class aggregrate) are always rewritable
533 if (StoreInst *SI = dyn_cast<StoreInst>(User))
534 if (!SI->isVolatile() && SI->getOperand(0) != AI)
535 return;// Store is ok if storing INTO the pointer, not storing the pointer
537 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User);
539 return MarkUnsafe(Info);
541 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI);
543 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>".
545 I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) {
546 return MarkUnsafe(Info);
550 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices??
552 bool IsAllZeroIndices = true;
554 // If the first index is a non-constant index into an array, see if we can
555 // handle it as a special case.
556 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
557 if (!isa<ConstantInt>(I.getOperand())) {
558 IsAllZeroIndices = 0;
559 uint64_t NumElements = AT->getNumElements();
561 // If this is an array index and the index is not constant, we cannot
562 // promote... that is unless the array has exactly one or two elements in
563 // it, in which case we CAN promote it, but we have to canonicalize this
564 // out if this is the only problem.
565 if ((NumElements == 1 || NumElements == 2) &&
566 AllUsersAreLoads(GEPI)) {
567 Info.needsCleanup = true;
568 return; // Canonicalization required!
570 return MarkUnsafe(Info);
574 // Walk through the GEP type indices, checking the types that this indexes
576 for (; I != E; ++I) {
577 // Ignore struct elements, no extra checking needed for these.
578 if (isa<StructType>(*I))
581 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand());
582 if (!IdxVal) return MarkUnsafe(Info);
584 // Are all indices still zero?
585 IsAllZeroIndices &= IdxVal->isZero();
587 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
588 // This GEP indexes an array. Verify that this is an in-range constant
589 // integer. Specifically, consider A[0][i]. We cannot know that the user
590 // isn't doing invalid things like allowing i to index an out-of-range
591 // subscript that accesses A[1]. Because of this, we have to reject SROA
592 // of any accesses into structs where any of the components are variables.
593 if (IdxVal->getZExtValue() >= AT->getNumElements())
594 return MarkUnsafe(Info);
595 } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) {
596 if (IdxVal->getZExtValue() >= VT->getNumElements())
597 return MarkUnsafe(Info);
601 // If there are any non-simple uses of this getelementptr, make sure to reject
603 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info);
606 /// isSafeMemIntrinsicOnAllocation - Return true if the specified memory
607 /// intrinsic can be promoted by SROA. At this point, we know that the operand
608 /// of the memintrinsic is a pointer to the beginning of the allocation.
609 void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI,
610 unsigned OpNo, AllocaInfo &Info) {
611 // If not constant length, give up.
612 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
613 if (!Length) return MarkUnsafe(Info);
615 // If not the whole aggregate, give up.
616 if (Length->getZExtValue() !=
617 TD->getTypeAllocSize(AI->getType()->getElementType()))
618 return MarkUnsafe(Info);
620 // We only know about memcpy/memset/memmove.
621 if (!isa<MemIntrinsic>(MI))
622 return MarkUnsafe(Info);
624 // Otherwise, we can transform it. Determine whether this is a memcpy/set
625 // into or out of the aggregate.
627 Info.isMemCpyDst = true;
630 Info.isMemCpySrc = true;
634 /// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast
636 void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI,
638 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end();
640 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) {
641 isSafeUseOfBitCastedAllocation(BCU, AI, Info);
642 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) {
643 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info);
644 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
645 if (SI->isVolatile())
646 return MarkUnsafe(Info);
648 // If storing the entire alloca in one chunk through a bitcasted pointer
649 // to integer, we can transform it. This happens (for example) when you
650 // cast a {i32,i32}* to i64* and store through it. This is similar to the
651 // memcpy case and occurs in various "byval" cases and emulated memcpys.
652 if (isa<IntegerType>(SI->getOperand(0)->getType()) &&
653 TD->getTypeAllocSize(SI->getOperand(0)->getType()) ==
654 TD->getTypeAllocSize(AI->getType()->getElementType())) {
655 Info.isMemCpyDst = true;
658 return MarkUnsafe(Info);
659 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) {
660 if (LI->isVolatile())
661 return MarkUnsafe(Info);
663 // If loading the entire alloca in one chunk through a bitcasted pointer
664 // to integer, we can transform it. This happens (for example) when you
665 // cast a {i32,i32}* to i64* and load through it. This is similar to the
666 // memcpy case and occurs in various "byval" cases and emulated memcpys.
667 if (isa<IntegerType>(LI->getType()) &&
668 TD->getTypeAllocSize(LI->getType()) ==
669 TD->getTypeAllocSize(AI->getType()->getElementType())) {
670 Info.isMemCpySrc = true;
673 return MarkUnsafe(Info);
674 } else if (isa<DbgInfoIntrinsic>(UI)) {
675 // If one user is DbgInfoIntrinsic then check if all users are
676 // DbgInfoIntrinsics.
677 if (OnlyUsedByDbgInfoIntrinsics(BC)) {
678 Info.needsCleanup = true;
685 return MarkUnsafe(Info);
687 if (Info.isUnsafe) return;
691 /// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes
692 /// to its first element. Transform users of the cast to use the new values
694 void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI,
695 SmallVector<AllocaInst*, 32> &NewElts) {
696 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end();
698 Instruction *User = cast<Instruction>(*UI++);
699 if (BitCastInst *BCU = dyn_cast<BitCastInst>(User)) {
700 RewriteBitCastUserOfAlloca(BCU, AI, NewElts);
701 if (BCU->use_empty()) BCU->eraseFromParent();
705 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
706 // This must be memcpy/memmove/memset of the entire aggregate.
707 // Split into one per element.
708 RewriteMemIntrinUserOfAlloca(MI, BCInst, AI, NewElts);
712 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
713 // If this is a store of the entire alloca from an integer, rewrite it.
714 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
718 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
719 // If this is a load of the entire alloca to an integer, rewrite it.
720 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
724 // Otherwise it must be some other user of a gep of the first pointer. Just
725 // leave these alone.
730 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
731 /// Rewrite it to copy or set the elements of the scalarized memory.
732 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst,
734 SmallVector<AllocaInst*, 32> &NewElts) {
736 // If this is a memcpy/memmove, construct the other pointer as the
737 // appropriate type. The "Other" pointer is the pointer that goes to memory
738 // that doesn't have anything to do with the alloca that we are promoting. For
739 // memset, this Value* stays null.
741 LLVMContext &Context = MI->getContext();
742 unsigned MemAlignment = MI->getAlignment();
743 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy
744 if (BCInst == MTI->getRawDest())
745 OtherPtr = MTI->getRawSource();
747 assert(BCInst == MTI->getRawSource());
748 OtherPtr = MTI->getRawDest();
752 // If there is an other pointer, we want to convert it to the same pointer
753 // type as AI has, so we can GEP through it safely.
755 // It is likely that OtherPtr is a bitcast, if so, remove it.
756 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr))
757 OtherPtr = BC->getOperand(0);
758 // All zero GEPs are effectively bitcasts.
759 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr))
760 if (GEP->hasAllZeroIndices())
761 OtherPtr = GEP->getOperand(0);
763 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr))
764 if (BCE->getOpcode() == Instruction::BitCast)
765 OtherPtr = BCE->getOperand(0);
767 // If the pointer is not the right type, insert a bitcast to the right
769 if (OtherPtr->getType() != AI->getType())
770 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(),
774 // Process each element of the aggregate.
775 Value *TheFn = MI->getOperand(0);
776 const Type *BytePtrTy = MI->getRawDest()->getType();
777 bool SROADest = MI->getRawDest() == BCInst;
779 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext()));
781 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
782 // If this is a memcpy/memmove, emit a GEP of the other element address.
784 unsigned OtherEltAlign = MemAlignment;
787 Value *Idx[2] = { Zero,
788 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) };
789 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2,
790 OtherPtr->getNameStr()+"."+Twine(i),
793 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
794 if (const StructType *ST =
795 dyn_cast<StructType>(OtherPtrTy->getElementType())) {
796 EltOffset = TD->getStructLayout(ST)->getElementOffset(i);
799 cast<SequentialType>(OtherPtr->getType())->getElementType();
800 EltOffset = TD->getTypeAllocSize(EltTy)*i;
803 // The alignment of the other pointer is the guaranteed alignment of the
804 // element, which is affected by both the known alignment of the whole
805 // mem intrinsic and the alignment of the element. If the alignment of
806 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
807 // known alignment is just 4 bytes.
808 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
811 Value *EltPtr = NewElts[i];
812 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
814 // If we got down to a scalar, insert a load or store as appropriate.
815 if (EltTy->isSingleValueType()) {
816 if (isa<MemTransferInst>(MI)) {
818 // From Other to Alloca.
819 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
820 new StoreInst(Elt, EltPtr, MI);
822 // From Alloca to Other.
823 Value *Elt = new LoadInst(EltPtr, "tmp", MI);
824 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI);
828 assert(isa<MemSetInst>(MI));
830 // If the stored element is zero (common case), just store a null
833 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) {
835 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
837 // If EltTy is a vector type, get the element type.
838 const Type *ValTy = EltTy->getScalarType();
840 // Construct an integer with the right value.
841 unsigned EltSize = TD->getTypeSizeInBits(ValTy);
842 APInt OneVal(EltSize, CI->getZExtValue());
843 APInt TotalVal(OneVal);
845 for (unsigned i = 0; 8*i < EltSize; ++i) {
846 TotalVal = TotalVal.shl(8);
850 // Convert the integer value to the appropriate type.
851 StoreVal = ConstantInt::get(Context, TotalVal);
852 if (isa<PointerType>(ValTy))
853 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
854 else if (ValTy->isFloatingPoint())
855 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
856 assert(StoreVal->getType() == ValTy && "Type mismatch!");
858 // If the requested value was a vector constant, create it.
859 if (EltTy != ValTy) {
860 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements();
861 SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
862 StoreVal = ConstantVector::get(&Elts[0], NumElts);
865 new StoreInst(StoreVal, EltPtr, MI);
868 // Otherwise, if we're storing a byte variable, use a memset call for
872 // Cast the element pointer to BytePtrTy.
873 if (EltPtr->getType() != BytePtrTy)
874 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI);
876 // Cast the other pointer (if we have one) to BytePtrTy.
877 if (OtherElt && OtherElt->getType() != BytePtrTy)
878 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(),
881 unsigned EltSize = TD->getTypeAllocSize(EltTy);
883 // Finally, insert the meminst for this element.
884 if (isa<MemTransferInst>(MI)) {
886 SROADest ? EltPtr : OtherElt, // Dest ptr
887 SROADest ? OtherElt : EltPtr, // Src ptr
888 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
890 ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign)
892 CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
894 assert(isa<MemSetInst>(MI));
896 EltPtr, MI->getOperand(2), // Dest, Value,
897 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
900 CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
903 MI->eraseFromParent();
906 /// RewriteStoreUserOfWholeAlloca - We found an store of an integer that
907 /// overwrites the entire allocation. Extract out the pieces of the stored
908 /// integer and store them individually.
909 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI,
911 SmallVector<AllocaInst*, 32> &NewElts){
912 // Extract each element out of the integer according to its structure offset
913 // and store the element value to the individual alloca.
914 Value *SrcVal = SI->getOperand(0);
915 const Type *AllocaEltTy = AI->getType()->getElementType();
916 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
918 // If this isn't a store of an integer to the whole alloca, it may be a store
919 // to the first element. Just ignore the store in this case and normal SROA
921 if (!isa<IntegerType>(SrcVal->getType()) ||
922 TD->getTypeAllocSizeInBits(SrcVal->getType()) != AllocaSizeBits)
924 // Handle tail padding by extending the operand
925 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
926 SrcVal = new ZExtInst(SrcVal,
927 IntegerType::get(SI->getContext(), AllocaSizeBits),
930 DEBUG(errs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << *SI);
932 // There are two forms here: AI could be an array or struct. Both cases
933 // have different ways to compute the element offset.
934 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
935 const StructLayout *Layout = TD->getStructLayout(EltSTy);
937 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
938 // Get the number of bits to shift SrcVal to get the value.
939 const Type *FieldTy = EltSTy->getElementType(i);
940 uint64_t Shift = Layout->getElementOffsetInBits(i);
942 if (TD->isBigEndian())
943 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy);
945 Value *EltVal = SrcVal;
947 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
948 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
949 "sroa.store.elt", SI);
952 // Truncate down to an integer of the right size.
953 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
955 // Ignore zero sized fields like {}, they obviously contain no data.
956 if (FieldSizeBits == 0) continue;
958 if (FieldSizeBits != AllocaSizeBits)
959 EltVal = new TruncInst(EltVal,
960 IntegerType::get(SI->getContext(), FieldSizeBits),
962 Value *DestField = NewElts[i];
963 if (EltVal->getType() == FieldTy) {
964 // Storing to an integer field of this size, just do it.
965 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) {
966 // Bitcast to the right element type (for fp/vector values).
967 EltVal = new BitCastInst(EltVal, FieldTy, "", SI);
969 // Otherwise, bitcast the dest pointer (for aggregates).
970 DestField = new BitCastInst(DestField,
971 PointerType::getUnqual(EltVal->getType()),
974 new StoreInst(EltVal, DestField, SI);
978 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
979 const Type *ArrayEltTy = ATy->getElementType();
980 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
981 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
985 if (TD->isBigEndian())
986 Shift = AllocaSizeBits-ElementOffset;
990 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
991 // Ignore zero sized fields like {}, they obviously contain no data.
992 if (ElementSizeBits == 0) continue;
994 Value *EltVal = SrcVal;
996 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
997 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
998 "sroa.store.elt", SI);
1001 // Truncate down to an integer of the right size.
1002 if (ElementSizeBits != AllocaSizeBits)
1003 EltVal = new TruncInst(EltVal,
1004 IntegerType::get(SI->getContext(),
1005 ElementSizeBits),"",SI);
1006 Value *DestField = NewElts[i];
1007 if (EltVal->getType() == ArrayEltTy) {
1008 // Storing to an integer field of this size, just do it.
1009 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) {
1010 // Bitcast to the right element type (for fp/vector values).
1011 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI);
1013 // Otherwise, bitcast the dest pointer (for aggregates).
1014 DestField = new BitCastInst(DestField,
1015 PointerType::getUnqual(EltVal->getType()),
1018 new StoreInst(EltVal, DestField, SI);
1020 if (TD->isBigEndian())
1021 Shift -= ElementOffset;
1023 Shift += ElementOffset;
1027 SI->eraseFromParent();
1030 /// RewriteLoadUserOfWholeAlloca - We found an load of the entire allocation to
1031 /// an integer. Load the individual pieces to form the aggregate value.
1032 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI,
1033 SmallVector<AllocaInst*, 32> &NewElts) {
1034 // Extract each element out of the NewElts according to its structure offset
1035 // and form the result value.
1036 const Type *AllocaEltTy = AI->getType()->getElementType();
1037 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
1039 // If this isn't a load of the whole alloca to an integer, it may be a load
1040 // of the first element. Just ignore the load in this case and normal SROA
1042 if (!isa<IntegerType>(LI->getType()) ||
1043 TD->getTypeAllocSizeInBits(LI->getType()) != AllocaSizeBits)
1046 DEBUG(errs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << *LI);
1048 // There are two forms here: AI could be an array or struct. Both cases
1049 // have different ways to compute the element offset.
1050 const StructLayout *Layout = 0;
1051 uint64_t ArrayEltBitOffset = 0;
1052 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
1053 Layout = TD->getStructLayout(EltSTy);
1055 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
1056 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
1060 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits));
1062 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1063 // Load the value from the alloca. If the NewElt is an aggregate, cast
1064 // the pointer to an integer of the same size before doing the load.
1065 Value *SrcField = NewElts[i];
1066 const Type *FieldTy =
1067 cast<PointerType>(SrcField->getType())->getElementType();
1068 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
1070 // Ignore zero sized fields like {}, they obviously contain no data.
1071 if (FieldSizeBits == 0) continue;
1073 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(),
1075 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() &&
1076 !isa<VectorType>(FieldTy))
1077 SrcField = new BitCastInst(SrcField,
1078 PointerType::getUnqual(FieldIntTy),
1080 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
1082 // If SrcField is a fp or vector of the right size but that isn't an
1083 // integer type, bitcast to an integer so we can shift it.
1084 if (SrcField->getType() != FieldIntTy)
1085 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
1087 // Zero extend the field to be the same size as the final alloca so that
1088 // we can shift and insert it.
1089 if (SrcField->getType() != ResultVal->getType())
1090 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
1092 // Determine the number of bits to shift SrcField.
1094 if (Layout) // Struct case.
1095 Shift = Layout->getElementOffsetInBits(i);
1097 Shift = i*ArrayEltBitOffset;
1099 if (TD->isBigEndian())
1100 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
1103 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
1104 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
1107 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
1110 // Handle tail padding by truncating the result
1111 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
1112 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI);
1114 LI->replaceAllUsesWith(ResultVal);
1115 LI->eraseFromParent();
1119 /// HasPadding - Return true if the specified type has any structure or
1120 /// alignment padding, false otherwise.
1121 static bool HasPadding(const Type *Ty, const TargetData &TD) {
1122 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1123 const StructLayout *SL = TD.getStructLayout(STy);
1124 unsigned PrevFieldBitOffset = 0;
1125 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1126 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
1128 // Padding in sub-elements?
1129 if (HasPadding(STy->getElementType(i), TD))
1132 // Check to see if there is any padding between this element and the
1135 unsigned PrevFieldEnd =
1136 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
1137 if (PrevFieldEnd < FieldBitOffset)
1141 PrevFieldBitOffset = FieldBitOffset;
1144 // Check for tail padding.
1145 if (unsigned EltCount = STy->getNumElements()) {
1146 unsigned PrevFieldEnd = PrevFieldBitOffset +
1147 TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
1148 if (PrevFieldEnd < SL->getSizeInBits())
1152 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1153 return HasPadding(ATy->getElementType(), TD);
1154 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1155 return HasPadding(VTy->getElementType(), TD);
1157 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
1160 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
1161 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
1162 /// or 1 if safe after canonicalization has been performed.
1164 int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) {
1165 // Loop over the use list of the alloca. We can only transform it if all of
1166 // the users are safe to transform.
1169 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end();
1171 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info);
1172 if (Info.isUnsafe) {
1173 DEBUG(errs() << "Cannot transform: " << *AI << " due to user: " << **I);
1178 // Okay, we know all the users are promotable. If the aggregate is a memcpy
1179 // source and destination, we have to be careful. In particular, the memcpy
1180 // could be moving around elements that live in structure padding of the LLVM
1181 // types, but may actually be used. In these cases, we refuse to promote the
1183 if (Info.isMemCpySrc && Info.isMemCpyDst &&
1184 HasPadding(AI->getType()->getElementType(), *TD))
1187 // If we require cleanup, return 1, otherwise return 3.
1188 return Info.needsCleanup ? 1 : 3;
1191 /// CleanupGEP - GEP is used by an Alloca, which can be prompted after the GEP
1192 /// is canonicalized here.
1193 void SROA::CleanupGEP(GetElementPtrInst *GEPI) {
1194 gep_type_iterator I = gep_type_begin(GEPI);
1197 const ArrayType *AT = dyn_cast<ArrayType>(*I);
1201 uint64_t NumElements = AT->getNumElements();
1203 if (isa<ConstantInt>(I.getOperand()))
1206 if (NumElements == 1) {
1208 Constant::getNullValue(Type::getInt32Ty(GEPI->getContext())));
1212 assert(NumElements == 2 && "Unhandled case!");
1213 // All users of the GEP must be loads. At each use of the GEP, insert
1214 // two loads of the appropriate indexed GEP and select between them.
1215 Value *IsOne = new ICmpInst(GEPI, ICmpInst::ICMP_NE, I.getOperand(),
1216 Constant::getNullValue(I.getOperand()->getType()),
1218 // Insert the new GEP instructions, which are properly indexed.
1219 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end());
1220 Indices[1] = Constant::getNullValue(Type::getInt32Ty(GEPI->getContext()));
1221 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
1224 GEPI->getName()+".0", GEPI);
1225 Indices[1] = ConstantInt::get(Type::getInt32Ty(GEPI->getContext()), 1);
1226 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
1229 GEPI->getName()+".1", GEPI);
1230 // Replace all loads of the variable index GEP with loads from both
1231 // indexes and a select.
1232 while (!GEPI->use_empty()) {
1233 LoadInst *LI = cast<LoadInst>(GEPI->use_back());
1234 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI);
1235 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI);
1236 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI);
1237 LI->replaceAllUsesWith(R);
1238 LI->eraseFromParent();
1240 GEPI->eraseFromParent();
1244 /// CleanupAllocaUsers - If SROA reported that it can promote the specified
1245 /// allocation, but only if cleaned up, perform the cleanups required.
1246 void SROA::CleanupAllocaUsers(AllocationInst *AI) {
1247 // At this point, we know that the end result will be SROA'd and promoted, so
1248 // we can insert ugly code if required so long as sroa+mem2reg will clean it
1250 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
1253 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U))
1256 Instruction *I = cast<Instruction>(U);
1257 SmallVector<DbgInfoIntrinsic *, 2> DbgInUses;
1258 if (!isa<StoreInst>(I) && OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) {
1259 // Safe to remove debug info uses.
1260 while (!DbgInUses.empty()) {
1261 DbgInfoIntrinsic *DI = DbgInUses.back(); DbgInUses.pop_back();
1262 DI->eraseFromParent();
1264 I->eraseFromParent();
1270 /// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at
1271 /// the offset specified by Offset (which is specified in bytes).
1273 /// There are two cases we handle here:
1274 /// 1) A union of vector types of the same size and potentially its elements.
1275 /// Here we turn element accesses into insert/extract element operations.
1276 /// This promotes a <4 x float> with a store of float to the third element
1277 /// into a <4 x float> that uses insert element.
1278 /// 2) A fully general blob of memory, which we turn into some (potentially
1279 /// large) integer type with extract and insert operations where the loads
1280 /// and stores would mutate the memory.
1281 static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy,
1282 unsigned AllocaSize, const TargetData &TD,
1283 LLVMContext &Context) {
1284 // If this could be contributing to a vector, analyze it.
1285 if (VecTy != Type::getVoidTy(Context)) { // either null or a vector type.
1287 // If the In type is a vector that is the same size as the alloca, see if it
1288 // matches the existing VecTy.
1289 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) {
1290 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) {
1291 // If we're storing/loading a vector of the right size, allow it as a
1292 // vector. If this the first vector we see, remember the type so that
1293 // we know the element size.
1298 } else if (In == Type::getFloatTy(Context) ||
1299 In == Type::getDoubleTy(Context) ||
1300 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 &&
1301 isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
1302 // If we're accessing something that could be an element of a vector, see
1303 // if the implied vector agrees with what we already have and if Offset is
1304 // compatible with it.
1305 unsigned EltSize = In->getPrimitiveSizeInBits()/8;
1306 if (Offset % EltSize == 0 &&
1307 AllocaSize % EltSize == 0 &&
1309 cast<VectorType>(VecTy)->getElementType()
1310 ->getPrimitiveSizeInBits()/8 == EltSize)) {
1312 VecTy = VectorType::get(In, AllocaSize/EltSize);
1318 // Otherwise, we have a case that we can't handle with an optimized vector
1319 // form. We can still turn this into a large integer.
1320 VecTy = Type::getVoidTy(Context);
1323 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
1324 /// its accesses to use a to single vector type, return true, and set VecTy to
1325 /// the new type. If we could convert the alloca into a single promotable
1326 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a
1327 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
1328 /// is the current offset from the base of the alloca being analyzed.
1330 /// If we see at least one access to the value that is as a vector type, set the
1333 bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy,
1334 bool &SawVec, uint64_t Offset,
1335 unsigned AllocaSize) {
1336 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1337 Instruction *User = cast<Instruction>(*UI);
1339 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1340 // Don't break volatile loads.
1341 if (LI->isVolatile())
1343 MergeInType(LI->getType(), Offset, VecTy,
1344 AllocaSize, *TD, V->getContext());
1345 SawVec |= isa<VectorType>(LI->getType());
1349 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1350 // Storing the pointer, not into the value?
1351 if (SI->getOperand(0) == V || SI->isVolatile()) return 0;
1352 MergeInType(SI->getOperand(0)->getType(), Offset,
1353 VecTy, AllocaSize, *TD, V->getContext());
1354 SawVec |= isa<VectorType>(SI->getOperand(0)->getType());
1358 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
1359 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset,
1362 IsNotTrivial = true;
1366 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1367 // If this is a GEP with a variable indices, we can't handle it.
1368 if (!GEP->hasAllConstantIndices())
1371 // Compute the offset that this GEP adds to the pointer.
1372 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
1373 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(),
1374 &Indices[0], Indices.size());
1375 // See if all uses can be converted.
1376 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset,
1379 IsNotTrivial = true;
1383 // If this is a constant sized memset of a constant value (e.g. 0) we can
1385 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
1386 // Store of constant value and constant size.
1387 if (isa<ConstantInt>(MSI->getValue()) &&
1388 isa<ConstantInt>(MSI->getLength())) {
1389 IsNotTrivial = true;
1394 // If this is a memcpy or memmove into or out of the whole allocation, we
1395 // can handle it like a load or store of the scalar type.
1396 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
1397 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()))
1398 if (Len->getZExtValue() == AllocaSize && Offset == 0) {
1399 IsNotTrivial = true;
1404 // Ignore dbg intrinsic.
1405 if (isa<DbgInfoIntrinsic>(User))
1408 // Otherwise, we cannot handle this!
1416 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
1417 /// directly. This happens when we are converting an "integer union" to a
1418 /// single integer scalar, or when we are converting a "vector union" to a
1419 /// vector with insert/extractelement instructions.
1421 /// Offset is an offset from the original alloca, in bits that need to be
1422 /// shifted to the right. By the end of this, there should be no uses of Ptr.
1423 void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) {
1424 while (!Ptr->use_empty()) {
1425 Instruction *User = cast<Instruction>(Ptr->use_back());
1427 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
1428 ConvertUsesToScalar(CI, NewAI, Offset);
1429 CI->eraseFromParent();
1433 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1434 // Compute the offset that this GEP adds to the pointer.
1435 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
1436 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(),
1437 &Indices[0], Indices.size());
1438 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
1439 GEP->eraseFromParent();
1443 IRBuilder<> Builder(User->getParent(), User);
1445 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1446 // The load is a bit extract from NewAI shifted right by Offset bits.
1447 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp");
1449 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
1450 LI->replaceAllUsesWith(NewLoadVal);
1451 LI->eraseFromParent();
1455 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1456 assert(SI->getOperand(0) != Ptr && "Consistency error!");
1457 // FIXME: Remove once builder has Twine API.
1458 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").str().c_str());
1459 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
1461 Builder.CreateStore(New, NewAI);
1462 SI->eraseFromParent();
1466 // If this is a constant sized memset of a constant value (e.g. 0) we can
1467 // transform it into a store of the expanded constant value.
1468 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
1469 assert(MSI->getRawDest() == Ptr && "Consistency error!");
1470 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
1471 if (NumBytes != 0) {
1472 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
1474 // Compute the value replicated the right number of times.
1475 APInt APVal(NumBytes*8, Val);
1477 // Splat the value if non-zero.
1479 for (unsigned i = 1; i != NumBytes; ++i)
1480 APVal |= APVal << 8;
1482 // FIXME: Remove once builder has Twine API.
1483 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").str().c_str());
1484 Value *New = ConvertScalar_InsertValue(
1485 ConstantInt::get(User->getContext(), APVal),
1486 Old, Offset, Builder);
1487 Builder.CreateStore(New, NewAI);
1489 MSI->eraseFromParent();
1493 // If this is a memcpy or memmove into or out of the whole allocation, we
1494 // can handle it like a load or store of the scalar type.
1495 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
1496 assert(Offset == 0 && "must be store to start of alloca");
1498 // If the source and destination are both to the same alloca, then this is
1499 // a noop copy-to-self, just delete it. Otherwise, emit a load and store
1501 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject());
1503 if (MTI->getSource()->getUnderlyingObject() != OrigAI) {
1504 // Dest must be OrigAI, change this to be a load from the original
1505 // pointer (bitcasted), then a store to our new alloca.
1506 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
1507 Value *SrcPtr = MTI->getSource();
1508 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType());
1510 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
1511 SrcVal->setAlignment(MTI->getAlignment());
1512 Builder.CreateStore(SrcVal, NewAI);
1513 } else if (MTI->getDest()->getUnderlyingObject() != OrigAI) {
1514 // Src must be OrigAI, change this to be a load from NewAI then a store
1515 // through the original dest pointer (bitcasted).
1516 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
1517 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
1519 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType());
1520 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
1521 NewStore->setAlignment(MTI->getAlignment());
1523 // Noop transfer. Src == Dst
1527 MTI->eraseFromParent();
1531 // If user is a dbg info intrinsic then it is safe to remove it.
1532 if (isa<DbgInfoIntrinsic>(User)) {
1533 User->eraseFromParent();
1537 llvm_unreachable("Unsupported operation!");
1541 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
1542 /// or vector value FromVal, extracting the bits from the offset specified by
1543 /// Offset. This returns the value, which is of type ToType.
1545 /// This happens when we are converting an "integer union" to a single
1546 /// integer scalar, or when we are converting a "vector union" to a vector with
1547 /// insert/extractelement instructions.
1549 /// Offset is an offset from the original alloca, in bits that need to be
1550 /// shifted to the right.
1551 Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType,
1552 uint64_t Offset, IRBuilder<> &Builder) {
1553 // If the load is of the whole new alloca, no conversion is needed.
1554 if (FromVal->getType() == ToType && Offset == 0)
1557 // If the result alloca is a vector type, this is either an element
1558 // access or a bitcast to another vector type of the same size.
1559 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) {
1560 if (isa<VectorType>(ToType))
1561 return Builder.CreateBitCast(FromVal, ToType, "tmp");
1563 // Otherwise it must be an element access.
1566 unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType());
1567 Elt = Offset/EltSize;
1568 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
1570 // Return the element extracted out of it.
1571 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get(
1572 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp");
1573 if (V->getType() != ToType)
1574 V = Builder.CreateBitCast(V, ToType, "tmp");
1578 // If ToType is a first class aggregate, extract out each of the pieces and
1579 // use insertvalue's to form the FCA.
1580 if (const StructType *ST = dyn_cast<StructType>(ToType)) {
1581 const StructLayout &Layout = *TD->getStructLayout(ST);
1582 Value *Res = UndefValue::get(ST);
1583 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
1584 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
1585 Offset+Layout.getElementOffsetInBits(i),
1587 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
1592 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
1593 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType());
1594 Value *Res = UndefValue::get(AT);
1595 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1596 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
1597 Offset+i*EltSize, Builder);
1598 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
1603 // Otherwise, this must be a union that was converted to an integer value.
1604 const IntegerType *NTy = cast<IntegerType>(FromVal->getType());
1606 // If this is a big-endian system and the load is narrower than the
1607 // full alloca type, we need to do a shift to get the right bits.
1609 if (TD->isBigEndian()) {
1610 // On big-endian machines, the lowest bit is stored at the bit offset
1611 // from the pointer given by getTypeStoreSizeInBits. This matters for
1612 // integers with a bitwidth that is not a multiple of 8.
1613 ShAmt = TD->getTypeStoreSizeInBits(NTy) -
1614 TD->getTypeStoreSizeInBits(ToType) - Offset;
1619 // Note: we support negative bitwidths (with shl) which are not defined.
1620 // We do this to support (f.e.) loads off the end of a structure where
1621 // only some bits are used.
1622 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
1623 FromVal = Builder.CreateLShr(FromVal,
1624 ConstantInt::get(FromVal->getType(),
1626 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
1627 FromVal = Builder.CreateShl(FromVal,
1628 ConstantInt::get(FromVal->getType(),
1631 // Finally, unconditionally truncate the integer to the right width.
1632 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType);
1633 if (LIBitWidth < NTy->getBitWidth())
1635 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(),
1636 LIBitWidth), "tmp");
1637 else if (LIBitWidth > NTy->getBitWidth())
1639 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(),
1640 LIBitWidth), "tmp");
1642 // If the result is an integer, this is a trunc or bitcast.
1643 if (isa<IntegerType>(ToType)) {
1645 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) {
1646 // Just do a bitcast, we know the sizes match up.
1647 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp");
1649 // Otherwise must be a pointer.
1650 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp");
1652 assert(FromVal->getType() == ToType && "Didn't convert right?");
1657 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
1658 /// or vector value "Old" at the offset specified by Offset.
1660 /// This happens when we are converting an "integer union" to a
1661 /// single integer scalar, or when we are converting a "vector union" to a
1662 /// vector with insert/extractelement instructions.
1664 /// Offset is an offset from the original alloca, in bits that need to be
1665 /// shifted to the right.
1666 Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old,
1667 uint64_t Offset, IRBuilder<> &Builder) {
1669 // Convert the stored type to the actual type, shift it left to insert
1670 // then 'or' into place.
1671 const Type *AllocaType = Old->getType();
1672 LLVMContext &Context = Old->getContext();
1674 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
1675 uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy);
1676 uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType());
1678 // Changing the whole vector with memset or with an access of a different
1680 if (ValSize == VecSize)
1681 return Builder.CreateBitCast(SV, AllocaType, "tmp");
1683 uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType());
1685 // Must be an element insertion.
1686 unsigned Elt = Offset/EltSize;
1688 if (SV->getType() != VTy->getElementType())
1689 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp");
1691 SV = Builder.CreateInsertElement(Old, SV,
1692 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt),
1697 // If SV is a first-class aggregate value, insert each value recursively.
1698 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) {
1699 const StructLayout &Layout = *TD->getStructLayout(ST);
1700 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
1701 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
1702 Old = ConvertScalar_InsertValue(Elt, Old,
1703 Offset+Layout.getElementOffsetInBits(i),
1709 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
1710 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType());
1711 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1712 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
1713 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
1718 // If SV is a float, convert it to the appropriate integer type.
1719 // If it is a pointer, do the same.
1720 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType());
1721 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType);
1722 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType());
1723 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType);
1724 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType()))
1725 SV = Builder.CreateBitCast(SV,
1726 IntegerType::get(SV->getContext(),SrcWidth), "tmp");
1727 else if (isa<PointerType>(SV->getType()))
1728 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(SV->getContext()), "tmp");
1730 // Zero extend or truncate the value if needed.
1731 if (SV->getType() != AllocaType) {
1732 if (SV->getType()->getPrimitiveSizeInBits() <
1733 AllocaType->getPrimitiveSizeInBits())
1734 SV = Builder.CreateZExt(SV, AllocaType, "tmp");
1736 // Truncation may be needed if storing more than the alloca can hold
1737 // (undefined behavior).
1738 SV = Builder.CreateTrunc(SV, AllocaType, "tmp");
1739 SrcWidth = DestWidth;
1740 SrcStoreWidth = DestStoreWidth;
1744 // If this is a big-endian system and the store is narrower than the
1745 // full alloca type, we need to do a shift to get the right bits.
1747 if (TD->isBigEndian()) {
1748 // On big-endian machines, the lowest bit is stored at the bit offset
1749 // from the pointer given by getTypeStoreSizeInBits. This matters for
1750 // integers with a bitwidth that is not a multiple of 8.
1751 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
1756 // Note: we support negative bitwidths (with shr) which are not defined.
1757 // We do this to support (f.e.) stores off the end of a structure where
1758 // only some bits in the structure are set.
1759 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
1760 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
1761 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(),
1764 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
1765 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(),
1767 Mask = Mask.lshr(-ShAmt);
1770 // Mask out the bits we are about to insert from the old value, and or
1772 if (SrcWidth != DestWidth) {
1773 assert(DestWidth > SrcWidth);
1774 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask");
1775 SV = Builder.CreateOr(Old, SV, "ins");
1782 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
1783 /// some part of a constant global variable. This intentionally only accepts
1784 /// constant expressions because we don't can't rewrite arbitrary instructions.
1785 static bool PointsToConstantGlobal(Value *V) {
1786 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
1787 return GV->isConstant();
1788 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
1789 if (CE->getOpcode() == Instruction::BitCast ||
1790 CE->getOpcode() == Instruction::GetElementPtr)
1791 return PointsToConstantGlobal(CE->getOperand(0));
1795 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
1796 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
1797 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
1798 /// track of whether it moves the pointer (with isOffset) but otherwise traverse
1799 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
1800 /// the alloca, and if the source pointer is a pointer to a constant global, we
1801 /// can optimize this.
1802 static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy,
1804 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1805 if (LoadInst *LI = dyn_cast<LoadInst>(*UI))
1806 // Ignore non-volatile loads, they are always ok.
1807 if (!LI->isVolatile())
1810 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) {
1811 // If uses of the bitcast are ok, we are ok.
1812 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset))
1816 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
1817 // If the GEP has all zero indices, it doesn't offset the pointer. If it
1818 // doesn't, it does.
1819 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
1820 isOffset || !GEP->hasAllZeroIndices()))
1825 // If this is isn't our memcpy/memmove, reject it as something we can't
1827 if (!isa<MemTransferInst>(*UI))
1830 // If we already have seen a copy, reject the second one.
1831 if (TheCopy) return false;
1833 // If the pointer has been offset from the start of the alloca, we can't
1834 // safely handle this.
1835 if (isOffset) return false;
1837 // If the memintrinsic isn't using the alloca as the dest, reject it.
1838 if (UI.getOperandNo() != 1) return false;
1840 MemIntrinsic *MI = cast<MemIntrinsic>(*UI);
1842 // If the source of the memcpy/move is not a constant global, reject it.
1843 if (!PointsToConstantGlobal(MI->getOperand(2)))
1846 // Otherwise, the transform is safe. Remember the copy instruction.
1852 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
1853 /// modified by a copy from a constant global. If we can prove this, we can
1854 /// replace any uses of the alloca with uses of the global directly.
1855 Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) {
1856 Instruction *TheCopy = 0;
1857 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false))