1 //===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation implements the well known scalar replacement of
11 // aggregates transformation. This xform breaks up alloca instructions of
12 // aggregate type (structure or array) into individual alloca instructions for
13 // each member (if possible). Then, if possible, it transforms the individual
14 // alloca instructions into nice clean scalar SSA form.
16 // This combines a simple SRoA algorithm with the Mem2Reg algorithm because
17 // often interact, especially for C++ programs. As such, iterating between
18 // SRoA, then Mem2Reg until we run out of things to promote works well.
20 //===----------------------------------------------------------------------===//
22 #define DEBUG_TYPE "scalarrepl"
23 #include "llvm/Transforms/Scalar.h"
24 #include "llvm/Constants.h"
25 #include "llvm/DerivedTypes.h"
26 #include "llvm/Function.h"
27 #include "llvm/GlobalVariable.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/IntrinsicInst.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/Pass.h"
32 #include "llvm/Analysis/Dominators.h"
33 #include "llvm/Target/TargetData.h"
34 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
35 #include "llvm/Transforms/Utils/Local.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/ErrorHandling.h"
38 #include "llvm/Support/GetElementPtrTypeIterator.h"
39 #include "llvm/Support/IRBuilder.h"
40 #include "llvm/Support/MathExtras.h"
41 #include "llvm/Support/raw_ostream.h"
42 #include "llvm/ADT/SmallVector.h"
43 #include "llvm/ADT/Statistic.h"
46 STATISTIC(NumReplaced, "Number of allocas broken up");
47 STATISTIC(NumPromoted, "Number of allocas promoted");
48 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
49 STATISTIC(NumGlobals, "Number of allocas copied from constant global");
52 struct SROA : public FunctionPass {
53 static char ID; // Pass identification, replacement for typeid
54 explicit SROA(signed T = -1) : FunctionPass(&ID) {
61 bool runOnFunction(Function &F);
63 bool performScalarRepl(Function &F);
64 bool performPromotion(Function &F);
66 // getAnalysisUsage - This pass does not require any passes, but we know it
67 // will not alter the CFG, so say so.
68 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
69 AU.addRequired<DominatorTree>();
70 AU.addRequired<DominanceFrontier>();
77 /// DeadInsts - Keep track of instructions we have made dead, so that
78 /// we can remove them after we are done working.
79 SmallVector<Value*, 32> DeadInsts;
81 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
82 /// information about the uses. All these fields are initialized to false
83 /// and set to true when something is learned.
85 /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
88 /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
91 /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
95 : isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false) {}
100 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; }
102 bool isSafeAllocaToScalarRepl(AllocaInst *AI);
104 void isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
106 void isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t &Offset,
108 void isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize,
109 const Type *MemOpType, bool isStore, AllocaInfo &Info);
110 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size);
111 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset,
114 void DoScalarReplacement(AllocaInst *AI,
115 std::vector<AllocaInst*> &WorkList);
116 void DeleteDeadInstructions();
117 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocaInst *Base);
119 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
120 SmallVector<AllocaInst*, 32> &NewElts);
121 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
122 SmallVector<AllocaInst*, 32> &NewElts);
123 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
124 SmallVector<AllocaInst*, 32> &NewElts);
125 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
127 SmallVector<AllocaInst*, 32> &NewElts);
128 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
129 SmallVector<AllocaInst*, 32> &NewElts);
130 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
131 SmallVector<AllocaInst*, 32> &NewElts);
133 bool CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy,
134 bool &SawVec, uint64_t Offset, unsigned AllocaSize);
135 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset);
136 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType,
137 uint64_t Offset, IRBuilder<> &Builder);
138 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
139 uint64_t Offset, IRBuilder<> &Builder);
140 static Instruction *isOnlyCopiedFromConstantGlobal(AllocaInst *AI);
145 static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates");
147 // Public interface to the ScalarReplAggregates pass
148 FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) {
149 return new SROA(Threshold);
153 bool SROA::runOnFunction(Function &F) {
154 TD = getAnalysisIfAvailable<TargetData>();
156 bool Changed = performPromotion(F);
158 // FIXME: ScalarRepl currently depends on TargetData more than it
159 // theoretically needs to. It should be refactored in order to support
160 // target-independent IR. Until this is done, just skip the actual
161 // scalar-replacement portion of this pass.
162 if (!TD) return Changed;
165 bool LocalChange = performScalarRepl(F);
166 if (!LocalChange) break; // No need to repromote if no scalarrepl
168 LocalChange = performPromotion(F);
169 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
176 bool SROA::performPromotion(Function &F) {
177 std::vector<AllocaInst*> Allocas;
178 DominatorTree &DT = getAnalysis<DominatorTree>();
179 DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
181 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
183 bool Changed = false;
188 // Find allocas that are safe to promote, by looking at all instructions in
190 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
191 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
192 if (isAllocaPromotable(AI))
193 Allocas.push_back(AI);
195 if (Allocas.empty()) break;
197 PromoteMemToReg(Allocas, DT, DF);
198 NumPromoted += Allocas.size();
205 /// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for
206 /// SROA. It must be a struct or array type with a small number of elements.
207 static bool ShouldAttemptScalarRepl(AllocaInst *AI) {
208 const Type *T = AI->getAllocatedType();
209 // Do not promote any struct into more than 32 separate vars.
210 if (const StructType *ST = dyn_cast<StructType>(T))
211 return ST->getNumElements() <= 32;
212 // Arrays are much less likely to be safe for SROA; only consider
213 // them if they are very small.
214 if (const ArrayType *AT = dyn_cast<ArrayType>(T))
215 return AT->getNumElements() <= 8;
219 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
220 // which runs on all of the malloc/alloca instructions in the function, removing
221 // them if they are only used by getelementptr instructions.
223 bool SROA::performScalarRepl(Function &F) {
224 std::vector<AllocaInst*> WorkList;
226 // Scan the entry basic block, adding any alloca's and mallocs to the worklist
227 BasicBlock &BB = F.getEntryBlock();
228 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
229 if (AllocaInst *A = dyn_cast<AllocaInst>(I))
230 WorkList.push_back(A);
232 // Process the worklist
233 bool Changed = false;
234 while (!WorkList.empty()) {
235 AllocaInst *AI = WorkList.back();
238 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
239 // with unused elements.
240 if (AI->use_empty()) {
241 AI->eraseFromParent();
245 // If this alloca is impossible for us to promote, reject it early.
246 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
249 // Check to see if this allocation is only modified by a memcpy/memmove from
250 // a constant global. If this is the case, we can change all users to use
251 // the constant global instead. This is commonly produced by the CFE by
252 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
253 // is only subsequently read.
254 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) {
255 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n');
256 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n');
257 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2));
258 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
259 TheCopy->eraseFromParent(); // Don't mutate the global.
260 AI->eraseFromParent();
266 // Check to see if we can perform the core SROA transformation. We cannot
267 // transform the allocation instruction if it is an array allocation
268 // (allocations OF arrays are ok though), and an allocation of a scalar
269 // value cannot be decomposed at all.
270 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
272 // Do not promote [0 x %struct].
273 if (AllocaSize == 0) continue;
275 // If the alloca looks like a good candidate for scalar replacement, and if
276 // all its users can be transformed, then split up the aggregate into its
277 // separate elements.
278 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) {
279 DoScalarReplacement(AI, WorkList);
284 // Do not promote any struct whose size is too big.
285 if (AllocaSize > SRThreshold) continue;
287 // If we can turn this aggregate value (potentially with casts) into a
288 // simple scalar value that can be mem2reg'd into a register value.
289 // IsNotTrivial tracks whether this is something that mem2reg could have
290 // promoted itself. If so, we don't want to transform it needlessly. Note
291 // that we can't just check based on the type: the alloca may be of an i32
292 // but that has pointer arithmetic to set byte 3 of it or something.
293 bool IsNotTrivial = false;
294 const Type *VectorTy = 0;
295 bool HadAVector = false;
296 if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector,
297 0, unsigned(AllocaSize)) && IsNotTrivial) {
299 // If we were able to find a vector type that can handle this with
300 // insert/extract elements, and if there was at least one use that had
301 // a vector type, promote this to a vector. We don't want to promote
302 // random stuff that doesn't use vectors (e.g. <9 x double>) because then
303 // we just get a lot of insert/extracts. If at least one vector is
304 // involved, then we probably really do have a union of vector/array.
305 if (VectorTy && VectorTy->isVectorTy() && HadAVector) {
306 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = "
307 << *VectorTy << '\n');
309 // Create and insert the vector alloca.
310 NewAI = new AllocaInst(VectorTy, 0, "", AI->getParent()->begin());
311 ConvertUsesToScalar(AI, NewAI, 0);
313 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n");
315 // Create and insert the integer alloca.
316 const Type *NewTy = IntegerType::get(AI->getContext(), AllocaSize*8);
317 NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
318 ConvertUsesToScalar(AI, NewAI, 0);
321 AI->eraseFromParent();
327 // Otherwise, couldn't process this alloca.
333 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
334 /// predicate, do SROA now.
335 void SROA::DoScalarReplacement(AllocaInst *AI,
336 std::vector<AllocaInst*> &WorkList) {
337 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n');
338 SmallVector<AllocaInst*, 32> ElementAllocas;
339 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
340 ElementAllocas.reserve(ST->getNumContainedTypes());
341 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
342 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
344 AI->getName() + "." + Twine(i), AI);
345 ElementAllocas.push_back(NA);
346 WorkList.push_back(NA); // Add to worklist for recursive processing
349 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
350 ElementAllocas.reserve(AT->getNumElements());
351 const Type *ElTy = AT->getElementType();
352 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
353 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
354 AI->getName() + "." + Twine(i), AI);
355 ElementAllocas.push_back(NA);
356 WorkList.push_back(NA); // Add to worklist for recursive processing
360 // Now that we have created the new alloca instructions, rewrite all the
361 // uses of the old alloca.
362 RewriteForScalarRepl(AI, AI, 0, ElementAllocas);
364 // Now erase any instructions that were made dead while rewriting the alloca.
365 DeleteDeadInstructions();
366 AI->eraseFromParent();
371 /// DeleteDeadInstructions - Erase instructions on the DeadInstrs list,
372 /// recursively including all their operands that become trivially dead.
373 void SROA::DeleteDeadInstructions() {
374 while (!DeadInsts.empty()) {
375 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val());
377 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
378 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
379 // Zero out the operand and see if it becomes trivially dead.
380 // (But, don't add allocas to the dead instruction list -- they are
381 // already on the worklist and will be deleted separately.)
383 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U))
384 DeadInsts.push_back(U);
387 I->eraseFromParent();
391 /// isSafeForScalarRepl - Check if instruction I is a safe use with regard to
392 /// performing scalar replacement of alloca AI. The results are flagged in
393 /// the Info parameter. Offset indicates the position within AI that is
394 /// referenced by this instruction.
395 void SROA::isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
397 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
398 Instruction *User = cast<Instruction>(*UI);
400 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
401 isSafeForScalarRepl(BC, AI, Offset, Info);
402 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
403 uint64_t GEPOffset = Offset;
404 isSafeGEP(GEPI, AI, GEPOffset, Info);
406 isSafeForScalarRepl(GEPI, AI, GEPOffset, Info);
407 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) {
408 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
410 isSafeMemAccess(AI, Offset, Length->getZExtValue(), 0,
411 UI.getOperandNo() == 1, Info);
414 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
415 if (!LI->isVolatile()) {
416 const Type *LIType = LI->getType();
417 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(LIType),
418 LIType, false, Info);
421 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
422 // Store is ok if storing INTO the pointer, not storing the pointer
423 if (!SI->isVolatile() && SI->getOperand(0) != I) {
424 const Type *SIType = SI->getOperand(0)->getType();
425 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(SIType),
430 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n');
433 if (Info.isUnsafe) return;
437 /// isSafeGEP - Check if a GEP instruction can be handled for scalar
438 /// replacement. It is safe when all the indices are constant, in-bounds
439 /// references, and when the resulting offset corresponds to an element within
440 /// the alloca type. The results are flagged in the Info parameter. Upon
441 /// return, Offset is adjusted as specified by the GEP indices.
442 void SROA::isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI,
443 uint64_t &Offset, AllocaInfo &Info) {
444 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI);
448 // Walk through the GEP type indices, checking the types that this indexes
450 for (; GEPIt != E; ++GEPIt) {
451 // Ignore struct elements, no extra checking needed for these.
452 if ((*GEPIt)->isStructTy())
455 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand());
457 return MarkUnsafe(Info);
460 // Compute the offset due to this GEP and check if the alloca has a
461 // component element at that offset.
462 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
463 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(),
464 &Indices[0], Indices.size());
465 if (!TypeHasComponent(AI->getAllocatedType(), Offset, 0))
469 /// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI
470 /// alloca or has an offset and size that corresponds to a component element
471 /// within it. The offset checked here may have been formed from a GEP with a
472 /// pointer bitcasted to a different type.
473 void SROA::isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize,
474 const Type *MemOpType, bool isStore,
476 // Check if this is a load/store of the entire alloca.
477 if (Offset == 0 && MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) {
478 bool UsesAggregateType = (MemOpType == AI->getAllocatedType());
479 // This is safe for MemIntrinsics (where MemOpType is 0), integer types
480 // (which are essentially the same as the MemIntrinsics, especially with
481 // regard to copying padding between elements), or references using the
482 // aggregate type of the alloca.
483 if (!MemOpType || MemOpType->isIntegerTy() || UsesAggregateType) {
484 if (!UsesAggregateType) {
486 Info.isMemCpyDst = true;
488 Info.isMemCpySrc = true;
493 // Check if the offset/size correspond to a component within the alloca type.
494 const Type *T = AI->getAllocatedType();
495 if (TypeHasComponent(T, Offset, MemSize))
498 return MarkUnsafe(Info);
501 /// TypeHasComponent - Return true if T has a component type with the
502 /// specified offset and size. If Size is zero, do not check the size.
503 bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) {
506 if (const StructType *ST = dyn_cast<StructType>(T)) {
507 const StructLayout *Layout = TD->getStructLayout(ST);
508 unsigned EltIdx = Layout->getElementContainingOffset(Offset);
509 EltTy = ST->getContainedType(EltIdx);
510 EltSize = TD->getTypeAllocSize(EltTy);
511 Offset -= Layout->getElementOffset(EltIdx);
512 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) {
513 EltTy = AT->getElementType();
514 EltSize = TD->getTypeAllocSize(EltTy);
515 if (Offset >= AT->getNumElements() * EltSize)
521 if (Offset == 0 && (Size == 0 || EltSize == Size))
523 // Check if the component spans multiple elements.
524 if (Offset + Size > EltSize)
526 return TypeHasComponent(EltTy, Offset, Size);
529 /// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite
530 /// the instruction I, which references it, to use the separate elements.
531 /// Offset indicates the position within AI that is referenced by this
533 void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
534 SmallVector<AllocaInst*, 32> &NewElts) {
535 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
536 Instruction *User = cast<Instruction>(*UI);
538 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
539 RewriteBitCast(BC, AI, Offset, NewElts);
540 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
541 RewriteGEP(GEPI, AI, Offset, NewElts);
542 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
543 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
544 uint64_t MemSize = Length->getZExtValue();
546 MemSize == TD->getTypeAllocSize(AI->getAllocatedType()))
547 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts);
548 // Otherwise the intrinsic can only touch a single element and the
549 // address operand will be updated, so nothing else needs to be done.
550 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
551 const Type *LIType = LI->getType();
552 if (LIType == AI->getAllocatedType()) {
554 // %res = load { i32, i32 }* %alloc
556 // %load.0 = load i32* %alloc.0
557 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
558 // %load.1 = load i32* %alloc.1
559 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
560 // (Also works for arrays instead of structs)
561 Value *Insert = UndefValue::get(LIType);
562 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
563 Value *Load = new LoadInst(NewElts[i], "load", LI);
564 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI);
566 LI->replaceAllUsesWith(Insert);
567 DeadInsts.push_back(LI);
568 } else if (LIType->isIntegerTy() &&
569 TD->getTypeAllocSize(LIType) ==
570 TD->getTypeAllocSize(AI->getAllocatedType())) {
571 // If this is a load of the entire alloca to an integer, rewrite it.
572 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
574 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
575 Value *Val = SI->getOperand(0);
576 const Type *SIType = Val->getType();
577 if (SIType == AI->getAllocatedType()) {
579 // store { i32, i32 } %val, { i32, i32 }* %alloc
581 // %val.0 = extractvalue { i32, i32 } %val, 0
582 // store i32 %val.0, i32* %alloc.0
583 // %val.1 = extractvalue { i32, i32 } %val, 1
584 // store i32 %val.1, i32* %alloc.1
585 // (Also works for arrays instead of structs)
586 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
587 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI);
588 new StoreInst(Extract, NewElts[i], SI);
590 DeadInsts.push_back(SI);
591 } else if (SIType->isIntegerTy() &&
592 TD->getTypeAllocSize(SIType) ==
593 TD->getTypeAllocSize(AI->getAllocatedType())) {
594 // If this is a store of the entire alloca from an integer, rewrite it.
595 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
601 /// RewriteBitCast - Update a bitcast reference to the alloca being replaced
602 /// and recursively continue updating all of its uses.
603 void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
604 SmallVector<AllocaInst*, 32> &NewElts) {
605 RewriteForScalarRepl(BC, AI, Offset, NewElts);
606 if (BC->getOperand(0) != AI)
609 // The bitcast references the original alloca. Replace its uses with
610 // references to the first new element alloca.
611 Instruction *Val = NewElts[0];
612 if (Val->getType() != BC->getDestTy()) {
613 Val = new BitCastInst(Val, BC->getDestTy(), "", BC);
616 BC->replaceAllUsesWith(Val);
617 DeadInsts.push_back(BC);
620 /// FindElementAndOffset - Return the index of the element containing Offset
621 /// within the specified type, which must be either a struct or an array.
622 /// Sets T to the type of the element and Offset to the offset within that
623 /// element. IdxTy is set to the type of the index result to be used in a
625 uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset,
626 const Type *&IdxTy) {
628 if (const StructType *ST = dyn_cast<StructType>(T)) {
629 const StructLayout *Layout = TD->getStructLayout(ST);
630 Idx = Layout->getElementContainingOffset(Offset);
631 T = ST->getContainedType(Idx);
632 Offset -= Layout->getElementOffset(Idx);
633 IdxTy = Type::getInt32Ty(T->getContext());
636 const ArrayType *AT = cast<ArrayType>(T);
637 T = AT->getElementType();
638 uint64_t EltSize = TD->getTypeAllocSize(T);
639 Idx = Offset / EltSize;
640 Offset -= Idx * EltSize;
641 IdxTy = Type::getInt64Ty(T->getContext());
645 /// RewriteGEP - Check if this GEP instruction moves the pointer across
646 /// elements of the alloca that are being split apart, and if so, rewrite
647 /// the GEP to be relative to the new element.
648 void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
649 SmallVector<AllocaInst*, 32> &NewElts) {
650 uint64_t OldOffset = Offset;
651 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
652 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(),
653 &Indices[0], Indices.size());
655 RewriteForScalarRepl(GEPI, AI, Offset, NewElts);
657 const Type *T = AI->getAllocatedType();
659 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy);
660 if (GEPI->getOperand(0) == AI)
661 OldIdx = ~0ULL; // Force the GEP to be rewritten.
663 T = AI->getAllocatedType();
664 uint64_t EltOffset = Offset;
665 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy);
667 // If this GEP does not move the pointer across elements of the alloca
668 // being split, then it does not needs to be rewritten.
672 const Type *i32Ty = Type::getInt32Ty(AI->getContext());
673 SmallVector<Value*, 8> NewArgs;
674 NewArgs.push_back(Constant::getNullValue(i32Ty));
675 while (EltOffset != 0) {
676 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy);
677 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx));
679 Instruction *Val = NewElts[Idx];
680 if (NewArgs.size() > 1) {
681 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(),
682 NewArgs.end(), "", GEPI);
685 if (Val->getType() != GEPI->getType())
686 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI);
687 GEPI->replaceAllUsesWith(Val);
688 DeadInsts.push_back(GEPI);
691 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
692 /// Rewrite it to copy or set the elements of the scalarized memory.
693 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
695 SmallVector<AllocaInst*, 32> &NewElts) {
696 // If this is a memcpy/memmove, construct the other pointer as the
697 // appropriate type. The "Other" pointer is the pointer that goes to memory
698 // that doesn't have anything to do with the alloca that we are promoting. For
699 // memset, this Value* stays null.
701 LLVMContext &Context = MI->getContext();
702 unsigned MemAlignment = MI->getAlignment();
703 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy
704 if (Inst == MTI->getRawDest())
705 OtherPtr = MTI->getRawSource();
707 assert(Inst == MTI->getRawSource());
708 OtherPtr = MTI->getRawDest();
712 // If there is an other pointer, we want to convert it to the same pointer
713 // type as AI has, so we can GEP through it safely.
716 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an
717 // optimization, but it's also required to detect the corner case where
718 // both pointer operands are referencing the same memory, and where
719 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This
720 // function is only called for mem intrinsics that access the whole
721 // aggregate, so non-zero GEPs are not an issue here.)
723 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) {
724 OtherPtr = BC->getOperand(0);
727 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) {
728 // All zero GEPs are effectively bitcasts.
729 if (GEP->hasAllZeroIndices()) {
730 OtherPtr = GEP->getOperand(0);
736 // Copying the alloca to itself is a no-op: just delete it.
737 if (OtherPtr == AI || OtherPtr == NewElts[0]) {
738 // This code will run twice for a no-op memcpy -- once for each operand.
739 // Put only one reference to MI on the DeadInsts list.
740 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(),
741 E = DeadInsts.end(); I != E; ++I)
742 if (*I == MI) return;
743 DeadInsts.push_back(MI);
747 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr))
748 if (BCE->getOpcode() == Instruction::BitCast)
749 OtherPtr = BCE->getOperand(0);
751 // If the pointer is not the right type, insert a bitcast to the right
753 if (OtherPtr->getType() != AI->getType())
754 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(),
758 // Process each element of the aggregate.
759 Value *TheFn = MI->getOperand(0);
760 const Type *BytePtrTy = MI->getRawDest()->getType();
761 bool SROADest = MI->getRawDest() == Inst;
763 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext()));
765 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
766 // If this is a memcpy/memmove, emit a GEP of the other element address.
768 unsigned OtherEltAlign = MemAlignment;
771 Value *Idx[2] = { Zero,
772 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) };
773 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2,
774 OtherPtr->getName()+"."+Twine(i),
777 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
778 if (const StructType *ST =
779 dyn_cast<StructType>(OtherPtrTy->getElementType())) {
780 EltOffset = TD->getStructLayout(ST)->getElementOffset(i);
783 cast<SequentialType>(OtherPtr->getType())->getElementType();
784 EltOffset = TD->getTypeAllocSize(EltTy)*i;
787 // The alignment of the other pointer is the guaranteed alignment of the
788 // element, which is affected by both the known alignment of the whole
789 // mem intrinsic and the alignment of the element. If the alignment of
790 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
791 // known alignment is just 4 bytes.
792 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
795 Value *EltPtr = NewElts[i];
796 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
798 // If we got down to a scalar, insert a load or store as appropriate.
799 if (EltTy->isSingleValueType()) {
800 if (isa<MemTransferInst>(MI)) {
802 // From Other to Alloca.
803 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
804 new StoreInst(Elt, EltPtr, MI);
806 // From Alloca to Other.
807 Value *Elt = new LoadInst(EltPtr, "tmp", MI);
808 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI);
812 assert(isa<MemSetInst>(MI));
814 // If the stored element is zero (common case), just store a null
817 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) {
819 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
821 // If EltTy is a vector type, get the element type.
822 const Type *ValTy = EltTy->getScalarType();
824 // Construct an integer with the right value.
825 unsigned EltSize = TD->getTypeSizeInBits(ValTy);
826 APInt OneVal(EltSize, CI->getZExtValue());
827 APInt TotalVal(OneVal);
829 for (unsigned i = 0; 8*i < EltSize; ++i) {
830 TotalVal = TotalVal.shl(8);
834 // Convert the integer value to the appropriate type.
835 StoreVal = ConstantInt::get(Context, TotalVal);
836 if (ValTy->isPointerTy())
837 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
838 else if (ValTy->isFloatingPointTy())
839 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
840 assert(StoreVal->getType() == ValTy && "Type mismatch!");
842 // If the requested value was a vector constant, create it.
843 if (EltTy != ValTy) {
844 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements();
845 SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
846 StoreVal = ConstantVector::get(&Elts[0], NumElts);
849 new StoreInst(StoreVal, EltPtr, MI);
852 // Otherwise, if we're storing a byte variable, use a memset call for
856 // Cast the element pointer to BytePtrTy.
857 if (EltPtr->getType() != BytePtrTy)
858 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getName(), MI);
860 // Cast the other pointer (if we have one) to BytePtrTy.
861 if (OtherElt && OtherElt->getType() != BytePtrTy)
862 OtherElt = new BitCastInst(OtherElt, BytePtrTy, OtherElt->getName(), MI);
864 unsigned EltSize = TD->getTypeAllocSize(EltTy);
866 // Finally, insert the meminst for this element.
867 if (isa<MemTransferInst>(MI)) {
869 SROADest ? EltPtr : OtherElt, // Dest ptr
870 SROADest ? OtherElt : EltPtr, // Src ptr
871 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
873 ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign)
875 CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
877 assert(isa<MemSetInst>(MI));
879 EltPtr, MI->getOperand(2), // Dest, Value,
880 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
883 CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
886 DeadInsts.push_back(MI);
889 /// RewriteStoreUserOfWholeAlloca - We found a store of an integer that
890 /// overwrites the entire allocation. Extract out the pieces of the stored
891 /// integer and store them individually.
892 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
893 SmallVector<AllocaInst*, 32> &NewElts){
894 // Extract each element out of the integer according to its structure offset
895 // and store the element value to the individual alloca.
896 Value *SrcVal = SI->getOperand(0);
897 const Type *AllocaEltTy = AI->getAllocatedType();
898 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
900 // Handle tail padding by extending the operand
901 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
902 SrcVal = new ZExtInst(SrcVal,
903 IntegerType::get(SI->getContext(), AllocaSizeBits),
906 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI
909 // There are two forms here: AI could be an array or struct. Both cases
910 // have different ways to compute the element offset.
911 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
912 const StructLayout *Layout = TD->getStructLayout(EltSTy);
914 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
915 // Get the number of bits to shift SrcVal to get the value.
916 const Type *FieldTy = EltSTy->getElementType(i);
917 uint64_t Shift = Layout->getElementOffsetInBits(i);
919 if (TD->isBigEndian())
920 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy);
922 Value *EltVal = SrcVal;
924 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
925 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
926 "sroa.store.elt", SI);
929 // Truncate down to an integer of the right size.
930 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
932 // Ignore zero sized fields like {}, they obviously contain no data.
933 if (FieldSizeBits == 0) continue;
935 if (FieldSizeBits != AllocaSizeBits)
936 EltVal = new TruncInst(EltVal,
937 IntegerType::get(SI->getContext(), FieldSizeBits),
939 Value *DestField = NewElts[i];
940 if (EltVal->getType() == FieldTy) {
941 // Storing to an integer field of this size, just do it.
942 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) {
943 // Bitcast to the right element type (for fp/vector values).
944 EltVal = new BitCastInst(EltVal, FieldTy, "", SI);
946 // Otherwise, bitcast the dest pointer (for aggregates).
947 DestField = new BitCastInst(DestField,
948 PointerType::getUnqual(EltVal->getType()),
951 new StoreInst(EltVal, DestField, SI);
955 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
956 const Type *ArrayEltTy = ATy->getElementType();
957 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
958 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
962 if (TD->isBigEndian())
963 Shift = AllocaSizeBits-ElementOffset;
967 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
968 // Ignore zero sized fields like {}, they obviously contain no data.
969 if (ElementSizeBits == 0) continue;
971 Value *EltVal = SrcVal;
973 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
974 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
975 "sroa.store.elt", SI);
978 // Truncate down to an integer of the right size.
979 if (ElementSizeBits != AllocaSizeBits)
980 EltVal = new TruncInst(EltVal,
981 IntegerType::get(SI->getContext(),
982 ElementSizeBits),"",SI);
983 Value *DestField = NewElts[i];
984 if (EltVal->getType() == ArrayEltTy) {
985 // Storing to an integer field of this size, just do it.
986 } else if (ArrayEltTy->isFloatingPointTy() ||
987 ArrayEltTy->isVectorTy()) {
988 // Bitcast to the right element type (for fp/vector values).
989 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI);
991 // Otherwise, bitcast the dest pointer (for aggregates).
992 DestField = new BitCastInst(DestField,
993 PointerType::getUnqual(EltVal->getType()),
996 new StoreInst(EltVal, DestField, SI);
998 if (TD->isBigEndian())
999 Shift -= ElementOffset;
1001 Shift += ElementOffset;
1005 DeadInsts.push_back(SI);
1008 /// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to
1009 /// an integer. Load the individual pieces to form the aggregate value.
1010 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
1011 SmallVector<AllocaInst*, 32> &NewElts) {
1012 // Extract each element out of the NewElts according to its structure offset
1013 // and form the result value.
1014 const Type *AllocaEltTy = AI->getAllocatedType();
1015 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
1017 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI
1020 // There are two forms here: AI could be an array or struct. Both cases
1021 // have different ways to compute the element offset.
1022 const StructLayout *Layout = 0;
1023 uint64_t ArrayEltBitOffset = 0;
1024 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
1025 Layout = TD->getStructLayout(EltSTy);
1027 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
1028 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
1032 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits));
1034 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1035 // Load the value from the alloca. If the NewElt is an aggregate, cast
1036 // the pointer to an integer of the same size before doing the load.
1037 Value *SrcField = NewElts[i];
1038 const Type *FieldTy =
1039 cast<PointerType>(SrcField->getType())->getElementType();
1040 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
1042 // Ignore zero sized fields like {}, they obviously contain no data.
1043 if (FieldSizeBits == 0) continue;
1045 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(),
1047 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() &&
1048 !FieldTy->isVectorTy())
1049 SrcField = new BitCastInst(SrcField,
1050 PointerType::getUnqual(FieldIntTy),
1052 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
1054 // If SrcField is a fp or vector of the right size but that isn't an
1055 // integer type, bitcast to an integer so we can shift it.
1056 if (SrcField->getType() != FieldIntTy)
1057 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
1059 // Zero extend the field to be the same size as the final alloca so that
1060 // we can shift and insert it.
1061 if (SrcField->getType() != ResultVal->getType())
1062 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
1064 // Determine the number of bits to shift SrcField.
1066 if (Layout) // Struct case.
1067 Shift = Layout->getElementOffsetInBits(i);
1069 Shift = i*ArrayEltBitOffset;
1071 if (TD->isBigEndian())
1072 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
1075 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
1076 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
1079 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
1082 // Handle tail padding by truncating the result
1083 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
1084 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI);
1086 LI->replaceAllUsesWith(ResultVal);
1087 DeadInsts.push_back(LI);
1090 /// HasPadding - Return true if the specified type has any structure or
1091 /// alignment padding, false otherwise.
1092 static bool HasPadding(const Type *Ty, const TargetData &TD) {
1093 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1094 const StructLayout *SL = TD.getStructLayout(STy);
1095 unsigned PrevFieldBitOffset = 0;
1096 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1097 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
1099 // Padding in sub-elements?
1100 if (HasPadding(STy->getElementType(i), TD))
1103 // Check to see if there is any padding between this element and the
1106 unsigned PrevFieldEnd =
1107 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
1108 if (PrevFieldEnd < FieldBitOffset)
1112 PrevFieldBitOffset = FieldBitOffset;
1115 // Check for tail padding.
1116 if (unsigned EltCount = STy->getNumElements()) {
1117 unsigned PrevFieldEnd = PrevFieldBitOffset +
1118 TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
1119 if (PrevFieldEnd < SL->getSizeInBits())
1123 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1124 return HasPadding(ATy->getElementType(), TD);
1125 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1126 return HasPadding(VTy->getElementType(), TD);
1128 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
1131 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
1132 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
1133 /// or 1 if safe after canonicalization has been performed.
1134 bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) {
1135 // Loop over the use list of the alloca. We can only transform it if all of
1136 // the users are safe to transform.
1139 isSafeForScalarRepl(AI, AI, 0, Info);
1140 if (Info.isUnsafe) {
1141 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n');
1145 // Okay, we know all the users are promotable. If the aggregate is a memcpy
1146 // source and destination, we have to be careful. In particular, the memcpy
1147 // could be moving around elements that live in structure padding of the LLVM
1148 // types, but may actually be used. In these cases, we refuse to promote the
1150 if (Info.isMemCpySrc && Info.isMemCpyDst &&
1151 HasPadding(AI->getAllocatedType(), *TD))
1157 /// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at
1158 /// the offset specified by Offset (which is specified in bytes).
1160 /// There are two cases we handle here:
1161 /// 1) A union of vector types of the same size and potentially its elements.
1162 /// Here we turn element accesses into insert/extract element operations.
1163 /// This promotes a <4 x float> with a store of float to the third element
1164 /// into a <4 x float> that uses insert element.
1165 /// 2) A fully general blob of memory, which we turn into some (potentially
1166 /// large) integer type with extract and insert operations where the loads
1167 /// and stores would mutate the memory.
1168 static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy,
1169 unsigned AllocaSize, const TargetData &TD,
1170 LLVMContext &Context) {
1171 // If this could be contributing to a vector, analyze it.
1172 if (VecTy != Type::getVoidTy(Context)) { // either null or a vector type.
1174 // If the In type is a vector that is the same size as the alloca, see if it
1175 // matches the existing VecTy.
1176 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) {
1177 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) {
1178 // If we're storing/loading a vector of the right size, allow it as a
1179 // vector. If this the first vector we see, remember the type so that
1180 // we know the element size.
1185 } else if (In->isFloatTy() || In->isDoubleTy() ||
1186 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 &&
1187 isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
1188 // If we're accessing something that could be an element of a vector, see
1189 // if the implied vector agrees with what we already have and if Offset is
1190 // compatible with it.
1191 unsigned EltSize = In->getPrimitiveSizeInBits()/8;
1192 if (Offset % EltSize == 0 &&
1193 AllocaSize % EltSize == 0 &&
1195 cast<VectorType>(VecTy)->getElementType()
1196 ->getPrimitiveSizeInBits()/8 == EltSize)) {
1198 VecTy = VectorType::get(In, AllocaSize/EltSize);
1204 // Otherwise, we have a case that we can't handle with an optimized vector
1205 // form. We can still turn this into a large integer.
1206 VecTy = Type::getVoidTy(Context);
1209 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
1210 /// its accesses to a single vector type, return true and set VecTy to
1211 /// the new type. If we could convert the alloca into a single promotable
1212 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a
1213 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
1214 /// is the current offset from the base of the alloca being analyzed.
1216 /// If we see at least one access to the value that is as a vector type, set the
1218 bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy,
1219 bool &SawVec, uint64_t Offset,
1220 unsigned AllocaSize) {
1221 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1222 Instruction *User = cast<Instruction>(*UI);
1224 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1225 // Don't break volatile loads.
1226 if (LI->isVolatile())
1228 MergeInType(LI->getType(), Offset, VecTy,
1229 AllocaSize, *TD, V->getContext());
1230 SawVec |= LI->getType()->isVectorTy();
1234 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1235 // Storing the pointer, not into the value?
1236 if (SI->getOperand(0) == V || SI->isVolatile()) return 0;
1237 MergeInType(SI->getOperand(0)->getType(), Offset,
1238 VecTy, AllocaSize, *TD, V->getContext());
1239 SawVec |= SI->getOperand(0)->getType()->isVectorTy();
1243 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
1244 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset,
1247 IsNotTrivial = true;
1251 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1252 // If this is a GEP with a variable indices, we can't handle it.
1253 if (!GEP->hasAllConstantIndices())
1256 // Compute the offset that this GEP adds to the pointer.
1257 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
1258 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getPointerOperandType(),
1259 &Indices[0], Indices.size());
1260 // See if all uses can be converted.
1261 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset,
1264 IsNotTrivial = true;
1268 // If this is a constant sized memset of a constant value (e.g. 0) we can
1270 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
1271 // Store of constant value and constant size.
1272 if (isa<ConstantInt>(MSI->getValue()) &&
1273 isa<ConstantInt>(MSI->getLength())) {
1274 IsNotTrivial = true;
1279 // If this is a memcpy or memmove into or out of the whole allocation, we
1280 // can handle it like a load or store of the scalar type.
1281 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
1282 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()))
1283 if (Len->getZExtValue() == AllocaSize && Offset == 0) {
1284 IsNotTrivial = true;
1289 // Otherwise, we cannot handle this!
1296 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
1297 /// directly. This happens when we are converting an "integer union" to a
1298 /// single integer scalar, or when we are converting a "vector union" to a
1299 /// vector with insert/extractelement instructions.
1301 /// Offset is an offset from the original alloca, in bits that need to be
1302 /// shifted to the right. By the end of this, there should be no uses of Ptr.
1303 void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) {
1304 while (!Ptr->use_empty()) {
1305 Instruction *User = cast<Instruction>(Ptr->use_back());
1307 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
1308 ConvertUsesToScalar(CI, NewAI, Offset);
1309 CI->eraseFromParent();
1313 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1314 // Compute the offset that this GEP adds to the pointer.
1315 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
1316 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getPointerOperandType(),
1317 &Indices[0], Indices.size());
1318 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
1319 GEP->eraseFromParent();
1323 IRBuilder<> Builder(User->getParent(), User);
1325 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1326 // The load is a bit extract from NewAI shifted right by Offset bits.
1327 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp");
1329 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
1330 LI->replaceAllUsesWith(NewLoadVal);
1331 LI->eraseFromParent();
1335 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1336 assert(SI->getOperand(0) != Ptr && "Consistency error!");
1337 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
1338 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
1340 Builder.CreateStore(New, NewAI);
1341 SI->eraseFromParent();
1343 // If the load we just inserted is now dead, then the inserted store
1344 // overwrote the entire thing.
1345 if (Old->use_empty())
1346 Old->eraseFromParent();
1350 // If this is a constant sized memset of a constant value (e.g. 0) we can
1351 // transform it into a store of the expanded constant value.
1352 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
1353 assert(MSI->getRawDest() == Ptr && "Consistency error!");
1354 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
1355 if (NumBytes != 0) {
1356 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
1358 // Compute the value replicated the right number of times.
1359 APInt APVal(NumBytes*8, Val);
1361 // Splat the value if non-zero.
1363 for (unsigned i = 1; i != NumBytes; ++i)
1364 APVal |= APVal << 8;
1366 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
1367 Value *New = ConvertScalar_InsertValue(
1368 ConstantInt::get(User->getContext(), APVal),
1369 Old, Offset, Builder);
1370 Builder.CreateStore(New, NewAI);
1372 // If the load we just inserted is now dead, then the memset overwrote
1373 // the entire thing.
1374 if (Old->use_empty())
1375 Old->eraseFromParent();
1377 MSI->eraseFromParent();
1381 // If this is a memcpy or memmove into or out of the whole allocation, we
1382 // can handle it like a load or store of the scalar type.
1383 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
1384 assert(Offset == 0 && "must be store to start of alloca");
1386 // If the source and destination are both to the same alloca, then this is
1387 // a noop copy-to-self, just delete it. Otherwise, emit a load and store
1389 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject(0));
1391 if (MTI->getSource()->getUnderlyingObject(0) != OrigAI) {
1392 // Dest must be OrigAI, change this to be a load from the original
1393 // pointer (bitcasted), then a store to our new alloca.
1394 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
1395 Value *SrcPtr = MTI->getSource();
1396 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType());
1398 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
1399 SrcVal->setAlignment(MTI->getAlignment());
1400 Builder.CreateStore(SrcVal, NewAI);
1401 } else if (MTI->getDest()->getUnderlyingObject(0) != OrigAI) {
1402 // Src must be OrigAI, change this to be a load from NewAI then a store
1403 // through the original dest pointer (bitcasted).
1404 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
1405 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
1407 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType());
1408 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
1409 NewStore->setAlignment(MTI->getAlignment());
1411 // Noop transfer. Src == Dst
1414 MTI->eraseFromParent();
1418 llvm_unreachable("Unsupported operation!");
1422 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
1423 /// or vector value FromVal, extracting the bits from the offset specified by
1424 /// Offset. This returns the value, which is of type ToType.
1426 /// This happens when we are converting an "integer union" to a single
1427 /// integer scalar, or when we are converting a "vector union" to a vector with
1428 /// insert/extractelement instructions.
1430 /// Offset is an offset from the original alloca, in bits that need to be
1431 /// shifted to the right.
1432 Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType,
1433 uint64_t Offset, IRBuilder<> &Builder) {
1434 // If the load is of the whole new alloca, no conversion is needed.
1435 if (FromVal->getType() == ToType && Offset == 0)
1438 // If the result alloca is a vector type, this is either an element
1439 // access or a bitcast to another vector type of the same size.
1440 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) {
1441 if (ToType->isVectorTy())
1442 return Builder.CreateBitCast(FromVal, ToType, "tmp");
1444 // Otherwise it must be an element access.
1447 unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType());
1448 Elt = Offset/EltSize;
1449 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
1451 // Return the element extracted out of it.
1452 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get(
1453 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp");
1454 if (V->getType() != ToType)
1455 V = Builder.CreateBitCast(V, ToType, "tmp");
1459 // If ToType is a first class aggregate, extract out each of the pieces and
1460 // use insertvalue's to form the FCA.
1461 if (const StructType *ST = dyn_cast<StructType>(ToType)) {
1462 const StructLayout &Layout = *TD->getStructLayout(ST);
1463 Value *Res = UndefValue::get(ST);
1464 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
1465 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
1466 Offset+Layout.getElementOffsetInBits(i),
1468 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
1473 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
1474 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType());
1475 Value *Res = UndefValue::get(AT);
1476 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1477 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
1478 Offset+i*EltSize, Builder);
1479 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
1484 // Otherwise, this must be a union that was converted to an integer value.
1485 const IntegerType *NTy = cast<IntegerType>(FromVal->getType());
1487 // If this is a big-endian system and the load is narrower than the
1488 // full alloca type, we need to do a shift to get the right bits.
1490 if (TD->isBigEndian()) {
1491 // On big-endian machines, the lowest bit is stored at the bit offset
1492 // from the pointer given by getTypeStoreSizeInBits. This matters for
1493 // integers with a bitwidth that is not a multiple of 8.
1494 ShAmt = TD->getTypeStoreSizeInBits(NTy) -
1495 TD->getTypeStoreSizeInBits(ToType) - Offset;
1500 // Note: we support negative bitwidths (with shl) which are not defined.
1501 // We do this to support (f.e.) loads off the end of a structure where
1502 // only some bits are used.
1503 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
1504 FromVal = Builder.CreateLShr(FromVal,
1505 ConstantInt::get(FromVal->getType(),
1507 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
1508 FromVal = Builder.CreateShl(FromVal,
1509 ConstantInt::get(FromVal->getType(),
1512 // Finally, unconditionally truncate the integer to the right width.
1513 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType);
1514 if (LIBitWidth < NTy->getBitWidth())
1516 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(),
1517 LIBitWidth), "tmp");
1518 else if (LIBitWidth > NTy->getBitWidth())
1520 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(),
1521 LIBitWidth), "tmp");
1523 // If the result is an integer, this is a trunc or bitcast.
1524 if (ToType->isIntegerTy()) {
1526 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) {
1527 // Just do a bitcast, we know the sizes match up.
1528 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp");
1530 // Otherwise must be a pointer.
1531 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp");
1533 assert(FromVal->getType() == ToType && "Didn't convert right?");
1537 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
1538 /// or vector value "Old" at the offset specified by Offset.
1540 /// This happens when we are converting an "integer union" to a
1541 /// single integer scalar, or when we are converting a "vector union" to a
1542 /// vector with insert/extractelement instructions.
1544 /// Offset is an offset from the original alloca, in bits that need to be
1545 /// shifted to the right.
1546 Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old,
1547 uint64_t Offset, IRBuilder<> &Builder) {
1549 // Convert the stored type to the actual type, shift it left to insert
1550 // then 'or' into place.
1551 const Type *AllocaType = Old->getType();
1552 LLVMContext &Context = Old->getContext();
1554 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
1555 uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy);
1556 uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType());
1558 // Changing the whole vector with memset or with an access of a different
1560 if (ValSize == VecSize)
1561 return Builder.CreateBitCast(SV, AllocaType, "tmp");
1563 uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType());
1565 // Must be an element insertion.
1566 unsigned Elt = Offset/EltSize;
1568 if (SV->getType() != VTy->getElementType())
1569 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp");
1571 SV = Builder.CreateInsertElement(Old, SV,
1572 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt),
1577 // If SV is a first-class aggregate value, insert each value recursively.
1578 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) {
1579 const StructLayout &Layout = *TD->getStructLayout(ST);
1580 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
1581 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
1582 Old = ConvertScalar_InsertValue(Elt, Old,
1583 Offset+Layout.getElementOffsetInBits(i),
1589 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
1590 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType());
1591 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1592 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
1593 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
1598 // If SV is a float, convert it to the appropriate integer type.
1599 // If it is a pointer, do the same.
1600 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType());
1601 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType);
1602 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType());
1603 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType);
1604 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy())
1605 SV = Builder.CreateBitCast(SV,
1606 IntegerType::get(SV->getContext(),SrcWidth), "tmp");
1607 else if (SV->getType()->isPointerTy())
1608 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(SV->getContext()), "tmp");
1610 // Zero extend or truncate the value if needed.
1611 if (SV->getType() != AllocaType) {
1612 if (SV->getType()->getPrimitiveSizeInBits() <
1613 AllocaType->getPrimitiveSizeInBits())
1614 SV = Builder.CreateZExt(SV, AllocaType, "tmp");
1616 // Truncation may be needed if storing more than the alloca can hold
1617 // (undefined behavior).
1618 SV = Builder.CreateTrunc(SV, AllocaType, "tmp");
1619 SrcWidth = DestWidth;
1620 SrcStoreWidth = DestStoreWidth;
1624 // If this is a big-endian system and the store is narrower than the
1625 // full alloca type, we need to do a shift to get the right bits.
1627 if (TD->isBigEndian()) {
1628 // On big-endian machines, the lowest bit is stored at the bit offset
1629 // from the pointer given by getTypeStoreSizeInBits. This matters for
1630 // integers with a bitwidth that is not a multiple of 8.
1631 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
1636 // Note: we support negative bitwidths (with shr) which are not defined.
1637 // We do this to support (f.e.) stores off the end of a structure where
1638 // only some bits in the structure are set.
1639 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
1640 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
1641 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(),
1644 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
1645 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(),
1647 Mask = Mask.lshr(-ShAmt);
1650 // Mask out the bits we are about to insert from the old value, and or
1652 if (SrcWidth != DestWidth) {
1653 assert(DestWidth > SrcWidth);
1654 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask");
1655 SV = Builder.CreateOr(Old, SV, "ins");
1662 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
1663 /// some part of a constant global variable. This intentionally only accepts
1664 /// constant expressions because we don't can't rewrite arbitrary instructions.
1665 static bool PointsToConstantGlobal(Value *V) {
1666 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
1667 return GV->isConstant();
1668 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
1669 if (CE->getOpcode() == Instruction::BitCast ||
1670 CE->getOpcode() == Instruction::GetElementPtr)
1671 return PointsToConstantGlobal(CE->getOperand(0));
1675 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
1676 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
1677 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
1678 /// track of whether it moves the pointer (with isOffset) but otherwise traverse
1679 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
1680 /// the alloca, and if the source pointer is a pointer to a constant global, we
1681 /// can optimize this.
1682 static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy,
1684 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1685 if (LoadInst *LI = dyn_cast<LoadInst>(*UI))
1686 // Ignore non-volatile loads, they are always ok.
1687 if (!LI->isVolatile())
1690 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) {
1691 // If uses of the bitcast are ok, we are ok.
1692 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset))
1696 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
1697 // If the GEP has all zero indices, it doesn't offset the pointer. If it
1698 // doesn't, it does.
1699 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
1700 isOffset || !GEP->hasAllZeroIndices()))
1705 // If this is isn't our memcpy/memmove, reject it as something we can't
1707 if (!isa<MemTransferInst>(*UI))
1710 // If we already have seen a copy, reject the second one.
1711 if (TheCopy) return false;
1713 // If the pointer has been offset from the start of the alloca, we can't
1714 // safely handle this.
1715 if (isOffset) return false;
1717 // If the memintrinsic isn't using the alloca as the dest, reject it.
1718 if (UI.getOperandNo() != 1) return false;
1720 MemIntrinsic *MI = cast<MemIntrinsic>(*UI);
1722 // If the source of the memcpy/move is not a constant global, reject it.
1723 if (!PointsToConstantGlobal(MI->getOperand(2)))
1726 // Otherwise, the transform is safe. Remember the copy instruction.
1732 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
1733 /// modified by a copy from a constant global. If we can prove this, we can
1734 /// replace any uses of the alloca with uses of the global directly.
1735 Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) {
1736 Instruction *TheCopy = 0;
1737 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false))