1 //===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation implements the well known scalar replacement of
11 // aggregates transformation. This xform breaks up alloca instructions of
12 // aggregate type (structure or array) into individual alloca instructions for
13 // each member (if possible). Then, if possible, it transforms the individual
14 // alloca instructions into nice clean scalar SSA form.
16 // This combines a simple SRoA algorithm with the Mem2Reg algorithm because
17 // often interact, especially for C++ programs. As such, iterating between
18 // SRoA, then Mem2Reg until we run out of things to promote works well.
20 //===----------------------------------------------------------------------===//
22 #define DEBUG_TYPE "scalarrepl"
23 #include "llvm/Transforms/Scalar.h"
24 #include "llvm/Constants.h"
25 #include "llvm/DerivedTypes.h"
26 #include "llvm/Function.h"
27 #include "llvm/GlobalVariable.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/IntrinsicInst.h"
30 #include "llvm/Pass.h"
31 #include "llvm/Analysis/Dominators.h"
32 #include "llvm/Target/TargetData.h"
33 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
34 #include "llvm/Transforms/Utils/Local.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/GetElementPtrTypeIterator.h"
37 #include "llvm/Support/IRBuilder.h"
38 #include "llvm/Support/MathExtras.h"
39 #include "llvm/Support/Compiler.h"
40 #include "llvm/ADT/SmallVector.h"
41 #include "llvm/ADT/Statistic.h"
42 #include "llvm/ADT/StringExtras.h"
45 STATISTIC(NumReplaced, "Number of allocas broken up");
46 STATISTIC(NumPromoted, "Number of allocas promoted");
47 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
48 STATISTIC(NumGlobals, "Number of allocas copied from constant global");
51 struct VISIBILITY_HIDDEN SROA : public FunctionPass {
52 static char ID; // Pass identification, replacement for typeid
53 explicit SROA(signed T = -1) : FunctionPass(&ID) {
60 bool runOnFunction(Function &F);
62 bool performScalarRepl(Function &F);
63 bool performPromotion(Function &F);
65 // getAnalysisUsage - This pass does not require any passes, but we know it
66 // will not alter the CFG, so say so.
67 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
68 AU.addRequired<DominatorTree>();
69 AU.addRequired<DominanceFrontier>();
70 AU.addRequired<TargetData>();
77 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
78 /// information about the uses. All these fields are initialized to false
79 /// and set to true when something is learned.
81 /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
84 /// needsCleanup - This is set to true if there is some use of the alloca
85 /// that requires cleanup.
86 bool needsCleanup : 1;
88 /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
91 /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
95 : isUnsafe(false), needsCleanup(false),
96 isMemCpySrc(false), isMemCpyDst(false) {}
101 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; }
103 int isSafeAllocaToScalarRepl(AllocationInst *AI);
105 void isSafeUseOfAllocation(Instruction *User, AllocationInst *AI,
107 void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI,
109 void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI,
110 unsigned OpNo, AllocaInfo &Info);
111 void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI,
114 void DoScalarReplacement(AllocationInst *AI,
115 std::vector<AllocationInst*> &WorkList);
116 void CleanupGEP(GetElementPtrInst *GEP);
117 void CleanupAllocaUsers(AllocationInst *AI);
118 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base);
120 void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI,
121 SmallVector<AllocaInst*, 32> &NewElts);
123 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst,
125 SmallVector<AllocaInst*, 32> &NewElts);
126 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocationInst *AI,
127 SmallVector<AllocaInst*, 32> &NewElts);
128 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI,
129 SmallVector<AllocaInst*, 32> &NewElts);
131 bool CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy,
132 bool &SawVec, uint64_t Offset, unsigned AllocaSize);
133 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset);
134 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType,
135 uint64_t Offset, IRBuilder<> &Builder);
136 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
137 uint64_t Offset, IRBuilder<> &Builder);
138 static Instruction *isOnlyCopiedFromConstantGlobal(AllocationInst *AI);
143 static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates");
145 // Public interface to the ScalarReplAggregates pass
146 FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) {
147 return new SROA(Threshold);
151 bool SROA::runOnFunction(Function &F) {
152 TD = &getAnalysis<TargetData>();
154 bool Changed = performPromotion(F);
156 bool LocalChange = performScalarRepl(F);
157 if (!LocalChange) break; // No need to repromote if no scalarrepl
159 LocalChange = performPromotion(F);
160 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
167 bool SROA::performPromotion(Function &F) {
168 std::vector<AllocaInst*> Allocas;
169 DominatorTree &DT = getAnalysis<DominatorTree>();
170 DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
172 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
174 bool Changed = false;
179 // Find allocas that are safe to promote, by looking at all instructions in
181 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
182 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
183 if (isAllocaPromotable(AI))
184 Allocas.push_back(AI);
186 if (Allocas.empty()) break;
188 PromoteMemToReg(Allocas, DT, DF);
189 NumPromoted += Allocas.size();
196 /// getNumSAElements - Return the number of elements in the specific struct or
198 static uint64_t getNumSAElements(const Type *T) {
199 if (const StructType *ST = dyn_cast<StructType>(T))
200 return ST->getNumElements();
201 return cast<ArrayType>(T)->getNumElements();
204 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
205 // which runs on all of the malloc/alloca instructions in the function, removing
206 // them if they are only used by getelementptr instructions.
208 bool SROA::performScalarRepl(Function &F) {
209 std::vector<AllocationInst*> WorkList;
211 // Scan the entry basic block, adding any alloca's and mallocs to the worklist
212 BasicBlock &BB = F.getEntryBlock();
213 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
214 if (AllocationInst *A = dyn_cast<AllocationInst>(I))
215 WorkList.push_back(A);
217 // Process the worklist
218 bool Changed = false;
219 while (!WorkList.empty()) {
220 AllocationInst *AI = WorkList.back();
223 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
224 // with unused elements.
225 if (AI->use_empty()) {
226 AI->eraseFromParent();
230 // If this alloca is impossible for us to promote, reject it early.
231 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
234 // Check to see if this allocation is only modified by a memcpy/memmove from
235 // a constant global. If this is the case, we can change all users to use
236 // the constant global instead. This is commonly produced by the CFE by
237 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
238 // is only subsequently read.
239 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) {
240 DOUT << "Found alloca equal to global: " << *AI;
241 DOUT << " memcpy = " << *TheCopy;
242 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2));
243 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
244 TheCopy->eraseFromParent(); // Don't mutate the global.
245 AI->eraseFromParent();
251 // Check to see if we can perform the core SROA transformation. We cannot
252 // transform the allocation instruction if it is an array allocation
253 // (allocations OF arrays are ok though), and an allocation of a scalar
254 // value cannot be decomposed at all.
255 uint64_t AllocaSize = TD->getTypePaddedSize(AI->getAllocatedType());
257 // Do not promote any struct whose size is too big.
258 if (AllocaSize > SRThreshold) continue;
260 if ((isa<StructType>(AI->getAllocatedType()) ||
261 isa<ArrayType>(AI->getAllocatedType())) &&
262 // Do not promote any struct into more than "32" separate vars.
263 getNumSAElements(AI->getAllocatedType()) < SRThreshold/4) {
264 // Check that all of the users of the allocation are capable of being
266 switch (isSafeAllocaToScalarRepl(AI)) {
267 default: assert(0 && "Unexpected value!");
268 case 0: // Not safe to scalar replace.
270 case 1: // Safe, but requires cleanup/canonicalizations first
271 CleanupAllocaUsers(AI);
273 case 3: // Safe to scalar replace.
274 DoScalarReplacement(AI, WorkList);
280 // If we can turn this aggregate value (potentially with casts) into a
281 // simple scalar value that can be mem2reg'd into a register value.
282 // IsNotTrivial tracks whether this is something that mem2reg could have
283 // promoted itself. If so, we don't want to transform it needlessly. Note
284 // that we can't just check based on the type: the alloca may be of an i32
285 // but that has pointer arithmetic to set byte 3 of it or something.
286 bool IsNotTrivial = false;
287 const Type *VectorTy = 0;
288 bool HadAVector = false;
289 if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector,
290 0, unsigned(AllocaSize)) && IsNotTrivial &&
293 // If we were able to find a vector type that can handle this with
294 // insert/extract elements, and if there was at least one use that had
295 // a vector type, promote this to a vector. We don't want to promote
296 // random stuff that doesn't use vectors (e.g. <9 x double>) because then
297 // we just get a lot of insert/extracts. If at least one vector is
298 // involved, then we probably really do have a union of vector/array.
299 if (VectorTy && isa<VectorType>(VectorTy) && HadAVector) {
300 DOUT << "CONVERT TO VECTOR: " << *AI << " TYPE = " << *VectorTy <<"\n";
302 // Create and insert the vector alloca.
303 NewAI = new AllocaInst(VectorTy, 0, "", AI->getParent()->begin());
304 ConvertUsesToScalar(AI, NewAI, 0);
306 DOUT << "CONVERT TO SCALAR INTEGER: " << *AI << "\n";
308 // Create and insert the integer alloca.
309 const Type *NewTy = IntegerType::get(AllocaSize*8);
310 NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
311 ConvertUsesToScalar(AI, NewAI, 0);
314 AI->eraseFromParent();
320 // Otherwise, couldn't process this alloca.
326 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
327 /// predicate, do SROA now.
328 void SROA::DoScalarReplacement(AllocationInst *AI,
329 std::vector<AllocationInst*> &WorkList) {
330 DOUT << "Found inst to SROA: " << *AI;
331 SmallVector<AllocaInst*, 32> ElementAllocas;
332 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
333 ElementAllocas.reserve(ST->getNumContainedTypes());
334 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
335 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
337 AI->getName() + "." + utostr(i), AI);
338 ElementAllocas.push_back(NA);
339 WorkList.push_back(NA); // Add to worklist for recursive processing
342 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
343 ElementAllocas.reserve(AT->getNumElements());
344 const Type *ElTy = AT->getElementType();
345 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
346 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
347 AI->getName() + "." + utostr(i), AI);
348 ElementAllocas.push_back(NA);
349 WorkList.push_back(NA); // Add to worklist for recursive processing
353 // Now that we have created the alloca instructions that we want to use,
354 // expand the getelementptr instructions to use them.
356 while (!AI->use_empty()) {
357 Instruction *User = cast<Instruction>(AI->use_back());
358 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) {
359 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas);
360 BCInst->eraseFromParent();
365 // %res = load { i32, i32 }* %alloc
367 // %load.0 = load i32* %alloc.0
368 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
369 // %load.1 = load i32* %alloc.1
370 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
371 // (Also works for arrays instead of structs)
372 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
373 Value *Insert = UndefValue::get(LI->getType());
374 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) {
375 Value *Load = new LoadInst(ElementAllocas[i], "load", LI);
376 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI);
378 LI->replaceAllUsesWith(Insert);
379 LI->eraseFromParent();
384 // store { i32, i32 } %val, { i32, i32 }* %alloc
386 // %val.0 = extractvalue { i32, i32 } %val, 0
387 // store i32 %val.0, i32* %alloc.0
388 // %val.1 = extractvalue { i32, i32 } %val, 1
389 // store i32 %val.1, i32* %alloc.1
390 // (Also works for arrays instead of structs)
391 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
392 Value *Val = SI->getOperand(0);
393 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) {
394 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI);
395 new StoreInst(Extract, ElementAllocas[i], SI);
397 SI->eraseFromParent();
401 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
402 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst>
404 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
406 assert(Idx < ElementAllocas.size() && "Index out of range?");
407 AllocaInst *AllocaToUse = ElementAllocas[Idx];
410 if (GEPI->getNumOperands() == 3) {
411 // Do not insert a new getelementptr instruction with zero indices, only
412 // to have it optimized out later.
413 RepValue = AllocaToUse;
415 // We are indexing deeply into the structure, so we still need a
416 // getelement ptr instruction to finish the indexing. This may be
417 // expanded itself once the worklist is rerun.
419 SmallVector<Value*, 8> NewArgs;
420 NewArgs.push_back(Constant::getNullValue(Type::Int32Ty));
421 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end());
422 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(),
423 NewArgs.end(), "", GEPI);
424 RepValue->takeName(GEPI);
427 // If this GEP is to the start of the aggregate, check for memcpys.
428 if (Idx == 0 && GEPI->hasAllZeroIndices())
429 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas);
431 // Move all of the users over to the new GEP.
432 GEPI->replaceAllUsesWith(RepValue);
433 // Delete the old GEP
434 GEPI->eraseFromParent();
437 // Finally, delete the Alloca instruction
438 AI->eraseFromParent();
443 /// isSafeElementUse - Check to see if this use is an allowed use for a
444 /// getelementptr instruction of an array aggregate allocation. isFirstElt
445 /// indicates whether Ptr is known to the start of the aggregate.
447 void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI,
449 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
451 Instruction *User = cast<Instruction>(*I);
452 switch (User->getOpcode()) {
453 case Instruction::Load: break;
454 case Instruction::Store:
455 // Store is ok if storing INTO the pointer, not storing the pointer
456 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info);
458 case Instruction::GetElementPtr: {
459 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User);
460 bool AreAllZeroIndices = isFirstElt;
461 if (GEP->getNumOperands() > 1) {
462 if (!isa<ConstantInt>(GEP->getOperand(1)) ||
463 !cast<ConstantInt>(GEP->getOperand(1))->isZero())
464 // Using pointer arithmetic to navigate the array.
465 return MarkUnsafe(Info);
467 if (AreAllZeroIndices)
468 AreAllZeroIndices = GEP->hasAllZeroIndices();
470 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info);
471 if (Info.isUnsafe) return;
474 case Instruction::BitCast:
476 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info);
477 if (Info.isUnsafe) return;
480 DOUT << " Transformation preventing inst: " << *User;
481 return MarkUnsafe(Info);
482 case Instruction::Call:
483 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
485 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info);
486 if (Info.isUnsafe) return;
490 DOUT << " Transformation preventing inst: " << *User;
491 return MarkUnsafe(Info);
493 DOUT << " Transformation preventing inst: " << *User;
494 return MarkUnsafe(Info);
497 return; // All users look ok :)
500 /// AllUsersAreLoads - Return true if all users of this value are loads.
501 static bool AllUsersAreLoads(Value *Ptr) {
502 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
504 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load)
509 /// isSafeUseOfAllocation - Check to see if this user is an allowed use for an
510 /// aggregate allocation.
512 void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI,
514 if (BitCastInst *C = dyn_cast<BitCastInst>(User))
515 return isSafeUseOfBitCastedAllocation(C, AI, Info);
517 if (LoadInst *LI = dyn_cast<LoadInst>(User))
518 if (!LI->isVolatile())
519 return;// Loads (returning a first class aggregrate) are always rewritable
521 if (StoreInst *SI = dyn_cast<StoreInst>(User))
522 if (!SI->isVolatile() && SI->getOperand(0) != AI)
523 return;// Store is ok if storing INTO the pointer, not storing the pointer
525 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User);
527 return MarkUnsafe(Info);
529 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI);
531 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>".
533 I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) {
534 return MarkUnsafe(Info);
538 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices??
540 bool IsAllZeroIndices = true;
542 // If the first index is a non-constant index into an array, see if we can
543 // handle it as a special case.
544 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
545 if (!isa<ConstantInt>(I.getOperand())) {
546 IsAllZeroIndices = 0;
547 uint64_t NumElements = AT->getNumElements();
549 // If this is an array index and the index is not constant, we cannot
550 // promote... that is unless the array has exactly one or two elements in
551 // it, in which case we CAN promote it, but we have to canonicalize this
552 // out if this is the only problem.
553 if ((NumElements == 1 || NumElements == 2) &&
554 AllUsersAreLoads(GEPI)) {
555 Info.needsCleanup = true;
556 return; // Canonicalization required!
558 return MarkUnsafe(Info);
562 // Walk through the GEP type indices, checking the types that this indexes
564 for (; I != E; ++I) {
565 // Ignore struct elements, no extra checking needed for these.
566 if (isa<StructType>(*I))
569 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand());
570 if (!IdxVal) return MarkUnsafe(Info);
572 // Are all indices still zero?
573 IsAllZeroIndices &= IdxVal->isZero();
575 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
576 // This GEP indexes an array. Verify that this is an in-range constant
577 // integer. Specifically, consider A[0][i]. We cannot know that the user
578 // isn't doing invalid things like allowing i to index an out-of-range
579 // subscript that accesses A[1]. Because of this, we have to reject SROA
580 // of any accesses into structs where any of the components are variables.
581 if (IdxVal->getZExtValue() >= AT->getNumElements())
582 return MarkUnsafe(Info);
583 } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) {
584 if (IdxVal->getZExtValue() >= VT->getNumElements())
585 return MarkUnsafe(Info);
589 // If there are any non-simple uses of this getelementptr, make sure to reject
591 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info);
594 /// isSafeMemIntrinsicOnAllocation - Return true if the specified memory
595 /// intrinsic can be promoted by SROA. At this point, we know that the operand
596 /// of the memintrinsic is a pointer to the beginning of the allocation.
597 void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI,
598 unsigned OpNo, AllocaInfo &Info) {
599 // If not constant length, give up.
600 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
601 if (!Length) return MarkUnsafe(Info);
603 // If not the whole aggregate, give up.
604 if (Length->getZExtValue() !=
605 TD->getTypePaddedSize(AI->getType()->getElementType()))
606 return MarkUnsafe(Info);
608 // We only know about memcpy/memset/memmove.
609 if (!isa<MemCpyInst>(MI) && !isa<MemSetInst>(MI) && !isa<MemMoveInst>(MI))
610 return MarkUnsafe(Info);
612 // Otherwise, we can transform it. Determine whether this is a memcpy/set
613 // into or out of the aggregate.
615 Info.isMemCpyDst = true;
618 Info.isMemCpySrc = true;
622 /// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast
624 void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI,
626 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end();
628 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) {
629 isSafeUseOfBitCastedAllocation(BCU, AI, Info);
630 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) {
631 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info);
632 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
633 if (SI->isVolatile())
634 return MarkUnsafe(Info);
636 // If storing the entire alloca in one chunk through a bitcasted pointer
637 // to integer, we can transform it. This happens (for example) when you
638 // cast a {i32,i32}* to i64* and store through it. This is similar to the
639 // memcpy case and occurs in various "byval" cases and emulated memcpys.
640 if (isa<IntegerType>(SI->getOperand(0)->getType()) &&
641 TD->getTypePaddedSize(SI->getOperand(0)->getType()) ==
642 TD->getTypePaddedSize(AI->getType()->getElementType())) {
643 Info.isMemCpyDst = true;
646 return MarkUnsafe(Info);
647 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) {
648 if (LI->isVolatile())
649 return MarkUnsafe(Info);
651 // If loading the entire alloca in one chunk through a bitcasted pointer
652 // to integer, we can transform it. This happens (for example) when you
653 // cast a {i32,i32}* to i64* and load through it. This is similar to the
654 // memcpy case and occurs in various "byval" cases and emulated memcpys.
655 if (isa<IntegerType>(LI->getType()) &&
656 TD->getTypePaddedSize(LI->getType()) ==
657 TD->getTypePaddedSize(AI->getType()->getElementType())) {
658 Info.isMemCpySrc = true;
661 return MarkUnsafe(Info);
662 } else if (isa<DbgInfoIntrinsic>(UI)) {
663 // If one user is DbgInfoIntrinsic then check if all users are
664 // DbgInfoIntrinsics.
665 if (OnlyUsedByDbgInfoIntrinsics(BC)) {
666 Info.needsCleanup = true;
673 return MarkUnsafe(Info);
675 if (Info.isUnsafe) return;
679 /// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes
680 /// to its first element. Transform users of the cast to use the new values
682 void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI,
683 SmallVector<AllocaInst*, 32> &NewElts) {
684 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end();
686 Instruction *User = cast<Instruction>(*UI++);
687 if (BitCastInst *BCU = dyn_cast<BitCastInst>(User)) {
688 RewriteBitCastUserOfAlloca(BCU, AI, NewElts);
689 if (BCU->use_empty()) BCU->eraseFromParent();
693 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
694 // This must be memcpy/memmove/memset of the entire aggregate.
695 // Split into one per element.
696 RewriteMemIntrinUserOfAlloca(MI, BCInst, AI, NewElts);
700 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
701 // If this is a store of the entire alloca from an integer, rewrite it.
702 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
706 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
707 // If this is a load of the entire alloca to an integer, rewrite it.
708 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
712 // Otherwise it must be some other user of a gep of the first pointer. Just
713 // leave these alone.
718 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
719 /// Rewrite it to copy or set the elements of the scalarized memory.
720 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst,
722 SmallVector<AllocaInst*, 32> &NewElts) {
724 // If this is a memcpy/memmove, construct the other pointer as the
725 // appropriate type. The "Other" pointer is the pointer that goes to
727 unsigned MemAlignment = MI->getAlignment()->getZExtValue();
728 if (MemCpyInst *MCI = dyn_cast<MemCpyInst>(MI)) {
729 if (BCInst == MCI->getRawDest())
730 OtherPtr = MCI->getRawSource();
732 assert(BCInst == MCI->getRawSource());
733 OtherPtr = MCI->getRawDest();
735 } else if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
736 if (BCInst == MMI->getRawDest())
737 OtherPtr = MMI->getRawSource();
739 assert(BCInst == MMI->getRawSource());
740 OtherPtr = MMI->getRawDest();
744 // If there is an other pointer, we want to convert it to the same pointer
745 // type as AI has, so we can GEP through it safely.
747 // It is likely that OtherPtr is a bitcast, if so, remove it.
748 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr))
749 OtherPtr = BC->getOperand(0);
750 // All zero GEPs are effectively bitcasts.
751 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr))
752 if (GEP->hasAllZeroIndices())
753 OtherPtr = GEP->getOperand(0);
755 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr))
756 if (BCE->getOpcode() == Instruction::BitCast)
757 OtherPtr = BCE->getOperand(0);
759 // If the pointer is not the right type, insert a bitcast to the right
761 if (OtherPtr->getType() != AI->getType())
762 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(),
766 // Process each element of the aggregate.
767 Value *TheFn = MI->getOperand(0);
768 const Type *BytePtrTy = MI->getRawDest()->getType();
769 bool SROADest = MI->getRawDest() == BCInst;
771 Constant *Zero = Constant::getNullValue(Type::Int32Ty);
773 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
774 // If this is a memcpy/memmove, emit a GEP of the other element address.
776 unsigned OtherEltAlign = MemAlignment;
779 Value *Idx[2] = { Zero, ConstantInt::get(Type::Int32Ty, i) };
780 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2,
781 OtherPtr->getNameStr()+"."+utostr(i),
784 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
785 if (const StructType *ST =
786 dyn_cast<StructType>(OtherPtrTy->getElementType())) {
787 EltOffset = TD->getStructLayout(ST)->getElementOffset(i);
790 cast<SequentialType>(OtherPtr->getType())->getElementType();
791 EltOffset = TD->getTypePaddedSize(EltTy)*i;
794 // The alignment of the other pointer is the guaranteed alignment of the
795 // element, which is affected by both the known alignment of the whole
796 // mem intrinsic and the alignment of the element. If the alignment of
797 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
798 // known alignment is just 4 bytes.
799 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
802 Value *EltPtr = NewElts[i];
803 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
805 // If we got down to a scalar, insert a load or store as appropriate.
806 if (EltTy->isSingleValueType()) {
807 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
809 // From Other to Alloca.
810 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
811 new StoreInst(Elt, EltPtr, MI);
813 // From Alloca to Other.
814 Value *Elt = new LoadInst(EltPtr, "tmp", MI);
815 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI);
819 assert(isa<MemSetInst>(MI));
821 // If the stored element is zero (common case), just store a null
824 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) {
826 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
828 // If EltTy is a vector type, get the element type.
829 const Type *ValTy = EltTy;
830 if (const VectorType *VTy = dyn_cast<VectorType>(ValTy))
831 ValTy = VTy->getElementType();
833 // Construct an integer with the right value.
834 unsigned EltSize = TD->getTypeSizeInBits(ValTy);
835 APInt OneVal(EltSize, CI->getZExtValue());
836 APInt TotalVal(OneVal);
838 for (unsigned i = 0; 8*i < EltSize; ++i) {
839 TotalVal = TotalVal.shl(8);
843 // Convert the integer value to the appropriate type.
844 StoreVal = ConstantInt::get(TotalVal);
845 if (isa<PointerType>(ValTy))
846 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
847 else if (ValTy->isFloatingPoint())
848 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
849 assert(StoreVal->getType() == ValTy && "Type mismatch!");
851 // If the requested value was a vector constant, create it.
852 if (EltTy != ValTy) {
853 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements();
854 SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
855 StoreVal = ConstantVector::get(&Elts[0], NumElts);
858 new StoreInst(StoreVal, EltPtr, MI);
861 // Otherwise, if we're storing a byte variable, use a memset call for
865 // Cast the element pointer to BytePtrTy.
866 if (EltPtr->getType() != BytePtrTy)
867 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI);
869 // Cast the other pointer (if we have one) to BytePtrTy.
870 if (OtherElt && OtherElt->getType() != BytePtrTy)
871 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(),
874 unsigned EltSize = TD->getTypePaddedSize(EltTy);
876 // Finally, insert the meminst for this element.
877 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
879 SROADest ? EltPtr : OtherElt, // Dest ptr
880 SROADest ? OtherElt : EltPtr, // Src ptr
881 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
882 ConstantInt::get(Type::Int32Ty, OtherEltAlign) // Align
884 CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
886 assert(isa<MemSetInst>(MI));
888 EltPtr, MI->getOperand(2), // Dest, Value,
889 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
892 CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
895 MI->eraseFromParent();
898 /// RewriteStoreUserOfWholeAlloca - We found an store of an integer that
899 /// overwrites the entire allocation. Extract out the pieces of the stored
900 /// integer and store them individually.
901 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI,
903 SmallVector<AllocaInst*, 32> &NewElts){
904 // Extract each element out of the integer according to its structure offset
905 // and store the element value to the individual alloca.
906 Value *SrcVal = SI->getOperand(0);
907 const Type *AllocaEltTy = AI->getType()->getElementType();
908 uint64_t AllocaSizeBits = TD->getTypePaddedSizeInBits(AllocaEltTy);
910 // If this isn't a store of an integer to the whole alloca, it may be a store
911 // to the first element. Just ignore the store in this case and normal SROA
913 if (!isa<IntegerType>(SrcVal->getType()) ||
914 TD->getTypePaddedSizeInBits(SrcVal->getType()) != AllocaSizeBits)
917 DOUT << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << *SI;
919 // There are two forms here: AI could be an array or struct. Both cases
920 // have different ways to compute the element offset.
921 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
922 const StructLayout *Layout = TD->getStructLayout(EltSTy);
924 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
925 // Get the number of bits to shift SrcVal to get the value.
926 const Type *FieldTy = EltSTy->getElementType(i);
927 uint64_t Shift = Layout->getElementOffsetInBits(i);
929 if (TD->isBigEndian())
930 Shift = AllocaSizeBits-Shift-TD->getTypePaddedSizeInBits(FieldTy);
932 Value *EltVal = SrcVal;
934 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
935 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
936 "sroa.store.elt", SI);
939 // Truncate down to an integer of the right size.
940 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
942 // Ignore zero sized fields like {}, they obviously contain no data.
943 if (FieldSizeBits == 0) continue;
945 if (FieldSizeBits != AllocaSizeBits)
946 EltVal = new TruncInst(EltVal, IntegerType::get(FieldSizeBits), "", SI);
947 Value *DestField = NewElts[i];
948 if (EltVal->getType() == FieldTy) {
949 // Storing to an integer field of this size, just do it.
950 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) {
951 // Bitcast to the right element type (for fp/vector values).
952 EltVal = new BitCastInst(EltVal, FieldTy, "", SI);
954 // Otherwise, bitcast the dest pointer (for aggregates).
955 DestField = new BitCastInst(DestField,
956 PointerType::getUnqual(EltVal->getType()),
959 new StoreInst(EltVal, DestField, SI);
963 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
964 const Type *ArrayEltTy = ATy->getElementType();
965 uint64_t ElementOffset = TD->getTypePaddedSizeInBits(ArrayEltTy);
966 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
970 if (TD->isBigEndian())
971 Shift = AllocaSizeBits-ElementOffset;
975 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
976 // Ignore zero sized fields like {}, they obviously contain no data.
977 if (ElementSizeBits == 0) continue;
979 Value *EltVal = SrcVal;
981 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
982 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
983 "sroa.store.elt", SI);
986 // Truncate down to an integer of the right size.
987 if (ElementSizeBits != AllocaSizeBits)
988 EltVal = new TruncInst(EltVal, IntegerType::get(ElementSizeBits),"",SI);
989 Value *DestField = NewElts[i];
990 if (EltVal->getType() == ArrayEltTy) {
991 // Storing to an integer field of this size, just do it.
992 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) {
993 // Bitcast to the right element type (for fp/vector values).
994 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI);
996 // Otherwise, bitcast the dest pointer (for aggregates).
997 DestField = new BitCastInst(DestField,
998 PointerType::getUnqual(EltVal->getType()),
1001 new StoreInst(EltVal, DestField, SI);
1003 if (TD->isBigEndian())
1004 Shift -= ElementOffset;
1006 Shift += ElementOffset;
1010 SI->eraseFromParent();
1013 /// RewriteLoadUserOfWholeAlloca - We found an load of the entire allocation to
1014 /// an integer. Load the individual pieces to form the aggregate value.
1015 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI,
1016 SmallVector<AllocaInst*, 32> &NewElts) {
1017 // Extract each element out of the NewElts according to its structure offset
1018 // and form the result value.
1019 const Type *AllocaEltTy = AI->getType()->getElementType();
1020 uint64_t AllocaSizeBits = TD->getTypePaddedSizeInBits(AllocaEltTy);
1022 // If this isn't a load of the whole alloca to an integer, it may be a load
1023 // of the first element. Just ignore the load in this case and normal SROA
1025 if (!isa<IntegerType>(LI->getType()) ||
1026 TD->getTypePaddedSizeInBits(LI->getType()) != AllocaSizeBits)
1029 DOUT << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << *LI;
1031 // There are two forms here: AI could be an array or struct. Both cases
1032 // have different ways to compute the element offset.
1033 const StructLayout *Layout = 0;
1034 uint64_t ArrayEltBitOffset = 0;
1035 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
1036 Layout = TD->getStructLayout(EltSTy);
1038 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
1039 ArrayEltBitOffset = TD->getTypePaddedSizeInBits(ArrayEltTy);
1042 Value *ResultVal = Constant::getNullValue(LI->getType());
1044 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1045 // Load the value from the alloca. If the NewElt is an aggregate, cast
1046 // the pointer to an integer of the same size before doing the load.
1047 Value *SrcField = NewElts[i];
1048 const Type *FieldTy =
1049 cast<PointerType>(SrcField->getType())->getElementType();
1050 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
1052 // Ignore zero sized fields like {}, they obviously contain no data.
1053 if (FieldSizeBits == 0) continue;
1055 const IntegerType *FieldIntTy = IntegerType::get(FieldSizeBits);
1056 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() &&
1057 !isa<VectorType>(FieldTy))
1058 SrcField = new BitCastInst(SrcField, PointerType::getUnqual(FieldIntTy),
1060 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
1062 // If SrcField is a fp or vector of the right size but that isn't an
1063 // integer type, bitcast to an integer so we can shift it.
1064 if (SrcField->getType() != FieldIntTy)
1065 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
1067 // Zero extend the field to be the same size as the final alloca so that
1068 // we can shift and insert it.
1069 if (SrcField->getType() != ResultVal->getType())
1070 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
1072 // Determine the number of bits to shift SrcField.
1074 if (Layout) // Struct case.
1075 Shift = Layout->getElementOffsetInBits(i);
1077 Shift = i*ArrayEltBitOffset;
1079 if (TD->isBigEndian())
1080 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
1083 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
1084 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
1087 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
1090 LI->replaceAllUsesWith(ResultVal);
1091 LI->eraseFromParent();
1095 /// HasPadding - Return true if the specified type has any structure or
1096 /// alignment padding, false otherwise.
1097 static bool HasPadding(const Type *Ty, const TargetData &TD) {
1098 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1099 const StructLayout *SL = TD.getStructLayout(STy);
1100 unsigned PrevFieldBitOffset = 0;
1101 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1102 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
1104 // Padding in sub-elements?
1105 if (HasPadding(STy->getElementType(i), TD))
1108 // Check to see if there is any padding between this element and the
1111 unsigned PrevFieldEnd =
1112 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
1113 if (PrevFieldEnd < FieldBitOffset)
1117 PrevFieldBitOffset = FieldBitOffset;
1120 // Check for tail padding.
1121 if (unsigned EltCount = STy->getNumElements()) {
1122 unsigned PrevFieldEnd = PrevFieldBitOffset +
1123 TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
1124 if (PrevFieldEnd < SL->getSizeInBits())
1128 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1129 return HasPadding(ATy->getElementType(), TD);
1130 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1131 return HasPadding(VTy->getElementType(), TD);
1133 return TD.getTypeSizeInBits(Ty) != TD.getTypePaddedSizeInBits(Ty);
1136 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
1137 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
1138 /// or 1 if safe after canonicalization has been performed.
1140 int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) {
1141 // Loop over the use list of the alloca. We can only transform it if all of
1142 // the users are safe to transform.
1145 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end();
1147 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info);
1148 if (Info.isUnsafe) {
1149 DOUT << "Cannot transform: " << *AI << " due to user: " << **I;
1154 // Okay, we know all the users are promotable. If the aggregate is a memcpy
1155 // source and destination, we have to be careful. In particular, the memcpy
1156 // could be moving around elements that live in structure padding of the LLVM
1157 // types, but may actually be used. In these cases, we refuse to promote the
1159 if (Info.isMemCpySrc && Info.isMemCpyDst &&
1160 HasPadding(AI->getType()->getElementType(), *TD))
1163 // If we require cleanup, return 1, otherwise return 3.
1164 return Info.needsCleanup ? 1 : 3;
1167 /// CleanupGEP - GEP is used by an Alloca, which can be prompted after the GEP
1168 /// is canonicalized here.
1169 void SROA::CleanupGEP(GetElementPtrInst *GEPI) {
1170 gep_type_iterator I = gep_type_begin(GEPI);
1173 const ArrayType *AT = dyn_cast<ArrayType>(*I);
1177 uint64_t NumElements = AT->getNumElements();
1179 if (isa<ConstantInt>(I.getOperand()))
1182 if (NumElements == 1) {
1183 GEPI->setOperand(2, Constant::getNullValue(Type::Int32Ty));
1187 assert(NumElements == 2 && "Unhandled case!");
1188 // All users of the GEP must be loads. At each use of the GEP, insert
1189 // two loads of the appropriate indexed GEP and select between them.
1190 Value *IsOne = new ICmpInst(ICmpInst::ICMP_NE, I.getOperand(),
1191 Constant::getNullValue(I.getOperand()->getType()),
1193 // Insert the new GEP instructions, which are properly indexed.
1194 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end());
1195 Indices[1] = Constant::getNullValue(Type::Int32Ty);
1196 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
1199 GEPI->getName()+".0", GEPI);
1200 Indices[1] = ConstantInt::get(Type::Int32Ty, 1);
1201 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
1204 GEPI->getName()+".1", GEPI);
1205 // Replace all loads of the variable index GEP with loads from both
1206 // indexes and a select.
1207 while (!GEPI->use_empty()) {
1208 LoadInst *LI = cast<LoadInst>(GEPI->use_back());
1209 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI);
1210 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI);
1211 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI);
1212 LI->replaceAllUsesWith(R);
1213 LI->eraseFromParent();
1215 GEPI->eraseFromParent();
1219 /// CleanupAllocaUsers - If SROA reported that it can promote the specified
1220 /// allocation, but only if cleaned up, perform the cleanups required.
1221 void SROA::CleanupAllocaUsers(AllocationInst *AI) {
1222 // At this point, we know that the end result will be SROA'd and promoted, so
1223 // we can insert ugly code if required so long as sroa+mem2reg will clean it
1225 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
1228 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U))
1230 else if (Instruction *I = dyn_cast<Instruction>(U)) {
1231 SmallVector<DbgInfoIntrinsic *, 2> DbgInUses;
1232 if (OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) {
1233 // Safe to remove debug info uses.
1234 while (!DbgInUses.empty()) {
1235 DbgInfoIntrinsic *DI = DbgInUses.back(); DbgInUses.pop_back();
1236 DI->eraseFromParent();
1238 I->eraseFromParent();
1244 /// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at
1245 /// the offset specified by Offset (which is specified in bytes).
1247 /// There are two cases we handle here:
1248 /// 1) A union of vector types of the same size and potentially its elements.
1249 /// Here we turn element accesses into insert/extract element operations.
1250 /// This promotes a <4 x float> with a store of float to the third element
1251 /// into a <4 x float> that uses insert element.
1252 /// 2) A fully general blob of memory, which we turn into some (potentially
1253 /// large) integer type with extract and insert operations where the loads
1254 /// and stores would mutate the memory.
1255 static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy,
1256 unsigned AllocaSize, const TargetData &TD) {
1257 // If this could be contributing to a vector, analyze it.
1258 if (VecTy != Type::VoidTy) { // either null or a vector type.
1260 // If the In type is a vector that is the same size as the alloca, see if it
1261 // matches the existing VecTy.
1262 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) {
1263 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) {
1264 // If we're storing/loading a vector of the right size, allow it as a
1265 // vector. If this the first vector we see, remember the type so that
1266 // we know the element size.
1271 } else if (In == Type::FloatTy || In == Type::DoubleTy ||
1272 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 &&
1273 isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
1274 // If we're accessing something that could be an element of a vector, see
1275 // if the implied vector agrees with what we already have and if Offset is
1276 // compatible with it.
1277 unsigned EltSize = In->getPrimitiveSizeInBits()/8;
1278 if (Offset % EltSize == 0 &&
1279 AllocaSize % EltSize == 0 &&
1281 cast<VectorType>(VecTy)->getElementType()
1282 ->getPrimitiveSizeInBits()/8 == EltSize)) {
1284 VecTy = VectorType::get(In, AllocaSize/EltSize);
1290 // Otherwise, we have a case that we can't handle with an optimized vector
1291 // form. We can still turn this into a large integer.
1292 VecTy = Type::VoidTy;
1295 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
1296 /// its accesses to use a to single vector type, return true, and set VecTy to
1297 /// the new type. If we could convert the alloca into a single promotable
1298 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a
1299 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
1300 /// is the current offset from the base of the alloca being analyzed.
1302 /// If we see at least one access to the value that is as a vector type, set the
1305 bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy,
1306 bool &SawVec, uint64_t Offset,
1307 unsigned AllocaSize) {
1308 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1309 Instruction *User = cast<Instruction>(*UI);
1311 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1312 // Don't break volatile loads.
1313 if (LI->isVolatile())
1315 MergeInType(LI->getType(), Offset, VecTy, AllocaSize, *TD);
1316 SawVec |= isa<VectorType>(LI->getType());
1320 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1321 // Storing the pointer, not into the value?
1322 if (SI->getOperand(0) == V || SI->isVolatile()) return 0;
1323 MergeInType(SI->getOperand(0)->getType(), Offset, VecTy, AllocaSize, *TD);
1324 SawVec |= isa<VectorType>(SI->getOperand(0)->getType());
1328 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
1329 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset,
1332 IsNotTrivial = true;
1336 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1337 // If this is a GEP with a variable indices, we can't handle it.
1338 if (!GEP->hasAllConstantIndices())
1341 // Compute the offset that this GEP adds to the pointer.
1342 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
1343 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(),
1344 &Indices[0], Indices.size());
1345 // See if all uses can be converted.
1346 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset,
1349 IsNotTrivial = true;
1353 // If this is a constant sized memset of a constant value (e.g. 0) we can
1355 if (isa<MemSetInst>(User) &&
1356 // Store of constant value.
1357 isa<ConstantInt>(User->getOperand(2)) &&
1358 // Store with constant size.
1359 isa<ConstantInt>(User->getOperand(3))) {
1360 VecTy = Type::VoidTy;
1361 IsNotTrivial = true;
1365 // Otherwise, we cannot handle this!
1373 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
1374 /// directly. This happens when we are converting an "integer union" to a
1375 /// single integer scalar, or when we are converting a "vector union" to a
1376 /// vector with insert/extractelement instructions.
1378 /// Offset is an offset from the original alloca, in bits that need to be
1379 /// shifted to the right. By the end of this, there should be no uses of Ptr.
1380 void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) {
1381 while (!Ptr->use_empty()) {
1382 Instruction *User = cast<Instruction>(Ptr->use_back());
1384 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
1385 ConvertUsesToScalar(CI, NewAI, Offset);
1386 CI->eraseFromParent();
1390 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1391 // Compute the offset that this GEP adds to the pointer.
1392 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
1393 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(),
1394 &Indices[0], Indices.size());
1395 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
1396 GEP->eraseFromParent();
1400 IRBuilder<> Builder(User->getParent(), User);
1402 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1403 // The load is a bit extract from NewAI shifted right by Offset bits.
1404 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp");
1406 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
1407 LI->replaceAllUsesWith(NewLoadVal);
1408 LI->eraseFromParent();
1412 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1413 assert(SI->getOperand(0) != Ptr && "Consistency error!");
1414 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").c_str());
1415 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
1417 Builder.CreateStore(New, NewAI);
1418 SI->eraseFromParent();
1422 // If this is a constant sized memset of a constant value (e.g. 0) we can
1423 // transform it into a store of the expanded constant value.
1424 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
1425 assert(MSI->getRawDest() == Ptr && "Consistency error!");
1426 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
1427 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
1429 // Compute the value replicated the right number of times.
1430 APInt APVal(NumBytes*8, Val);
1432 // Splat the value if non-zero.
1434 for (unsigned i = 1; i != NumBytes; ++i)
1435 APVal |= APVal << 8;
1437 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").c_str());
1438 Value *New = ConvertScalar_InsertValue(ConstantInt::get(APVal), Old,
1440 Builder.CreateStore(New, NewAI);
1441 MSI->eraseFromParent();
1446 assert(0 && "Unsupported operation!");
1451 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
1452 /// or vector value FromVal, extracting the bits from the offset specified by
1453 /// Offset. This returns the value, which is of type ToType.
1455 /// This happens when we are converting an "integer union" to a single
1456 /// integer scalar, or when we are converting a "vector union" to a vector with
1457 /// insert/extractelement instructions.
1459 /// Offset is an offset from the original alloca, in bits that need to be
1460 /// shifted to the right.
1461 Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType,
1462 uint64_t Offset, IRBuilder<> &Builder) {
1463 // If the load is of the whole new alloca, no conversion is needed.
1464 if (FromVal->getType() == ToType && Offset == 0)
1467 // If the result alloca is a vector type, this is either an element
1468 // access or a bitcast to another vector type of the same size.
1469 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) {
1470 if (isa<VectorType>(ToType))
1471 return Builder.CreateBitCast(FromVal, ToType, "tmp");
1473 // Otherwise it must be an element access.
1476 unsigned EltSize = TD->getTypePaddedSizeInBits(VTy->getElementType());
1477 Elt = Offset/EltSize;
1478 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
1480 // Return the element extracted out of it.
1481 Value *V = Builder.CreateExtractElement(FromVal,
1482 ConstantInt::get(Type::Int32Ty,Elt),
1484 if (V->getType() != ToType)
1485 V = Builder.CreateBitCast(V, ToType, "tmp");
1489 // If ToType is a first class aggregate, extract out each of the pieces and
1490 // use insertvalue's to form the FCA.
1491 if (const StructType *ST = dyn_cast<StructType>(ToType)) {
1492 const StructLayout &Layout = *TD->getStructLayout(ST);
1493 Value *Res = UndefValue::get(ST);
1494 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
1495 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
1496 Offset+Layout.getElementOffsetInBits(i),
1498 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
1503 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
1504 uint64_t EltSize = TD->getTypePaddedSizeInBits(AT->getElementType());
1505 Value *Res = UndefValue::get(AT);
1506 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1507 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
1508 Offset+i*EltSize, Builder);
1509 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
1514 // Otherwise, this must be a union that was converted to an integer value.
1515 const IntegerType *NTy = cast<IntegerType>(FromVal->getType());
1517 // If this is a big-endian system and the load is narrower than the
1518 // full alloca type, we need to do a shift to get the right bits.
1520 if (TD->isBigEndian()) {
1521 // On big-endian machines, the lowest bit is stored at the bit offset
1522 // from the pointer given by getTypeStoreSizeInBits. This matters for
1523 // integers with a bitwidth that is not a multiple of 8.
1524 ShAmt = TD->getTypeStoreSizeInBits(NTy) -
1525 TD->getTypeStoreSizeInBits(ToType) - Offset;
1530 // Note: we support negative bitwidths (with shl) which are not defined.
1531 // We do this to support (f.e.) loads off the end of a structure where
1532 // only some bits are used.
1533 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
1534 FromVal = Builder.CreateLShr(FromVal, ConstantInt::get(FromVal->getType(),
1536 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
1537 FromVal = Builder.CreateShl(FromVal, ConstantInt::get(FromVal->getType(),
1540 // Finally, unconditionally truncate the integer to the right width.
1541 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType);
1542 if (LIBitWidth < NTy->getBitWidth())
1543 FromVal = Builder.CreateTrunc(FromVal, IntegerType::get(LIBitWidth), "tmp");
1544 else if (LIBitWidth > NTy->getBitWidth())
1545 FromVal = Builder.CreateZExt(FromVal, IntegerType::get(LIBitWidth), "tmp");
1547 // If the result is an integer, this is a trunc or bitcast.
1548 if (isa<IntegerType>(ToType)) {
1550 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) {
1551 // Just do a bitcast, we know the sizes match up.
1552 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp");
1554 // Otherwise must be a pointer.
1555 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp");
1557 assert(FromVal->getType() == ToType && "Didn't convert right?");
1562 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
1563 /// or vector value "Old" at the offset specified by Offset.
1565 /// This happens when we are converting an "integer union" to a
1566 /// single integer scalar, or when we are converting a "vector union" to a
1567 /// vector with insert/extractelement instructions.
1569 /// Offset is an offset from the original alloca, in bits that need to be
1570 /// shifted to the right.
1571 Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old,
1572 uint64_t Offset, IRBuilder<> &Builder) {
1574 // Convert the stored type to the actual type, shift it left to insert
1575 // then 'or' into place.
1576 const Type *AllocaType = Old->getType();
1578 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
1579 // If the result alloca is a vector type, this is either an element
1580 // access or a bitcast to another vector type.
1581 if (isa<VectorType>(SV->getType())) {
1582 SV = Builder.CreateBitCast(SV, AllocaType, "tmp");
1584 // Must be an element insertion.
1585 unsigned Elt = Offset/TD->getTypePaddedSizeInBits(VTy->getElementType());
1587 if (SV->getType() != VTy->getElementType())
1588 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp");
1590 SV = Builder.CreateInsertElement(Old, SV,
1591 ConstantInt::get(Type::Int32Ty, Elt),
1597 // If SV is a first-class aggregate value, insert each value recursively.
1598 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) {
1599 const StructLayout &Layout = *TD->getStructLayout(ST);
1600 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
1601 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
1602 Old = ConvertScalar_InsertValue(Elt, Old,
1603 Offset+Layout.getElementOffsetInBits(i),
1609 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
1610 uint64_t EltSize = TD->getTypePaddedSizeInBits(AT->getElementType());
1611 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1612 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
1613 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
1618 // If SV is a float, convert it to the appropriate integer type.
1619 // If it is a pointer, do the same.
1620 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType());
1621 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType);
1622 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType());
1623 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType);
1624 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType()))
1625 SV = Builder.CreateBitCast(SV, IntegerType::get(SrcWidth), "tmp");
1626 else if (isa<PointerType>(SV->getType()))
1627 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(), "tmp");
1629 // Zero extend or truncate the value if needed.
1630 if (SV->getType() != AllocaType) {
1631 if (SV->getType()->getPrimitiveSizeInBits() <
1632 AllocaType->getPrimitiveSizeInBits())
1633 SV = Builder.CreateZExt(SV, AllocaType, "tmp");
1635 // Truncation may be needed if storing more than the alloca can hold
1636 // (undefined behavior).
1637 SV = Builder.CreateTrunc(SV, AllocaType, "tmp");
1638 SrcWidth = DestWidth;
1639 SrcStoreWidth = DestStoreWidth;
1643 // If this is a big-endian system and the store is narrower than the
1644 // full alloca type, we need to do a shift to get the right bits.
1646 if (TD->isBigEndian()) {
1647 // On big-endian machines, the lowest bit is stored at the bit offset
1648 // from the pointer given by getTypeStoreSizeInBits. This matters for
1649 // integers with a bitwidth that is not a multiple of 8.
1650 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
1655 // Note: we support negative bitwidths (with shr) which are not defined.
1656 // We do this to support (f.e.) stores off the end of a structure where
1657 // only some bits in the structure are set.
1658 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
1659 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
1660 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt), "tmp");
1662 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
1663 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt), "tmp");
1664 Mask = Mask.lshr(-ShAmt);
1667 // Mask out the bits we are about to insert from the old value, and or
1669 if (SrcWidth != DestWidth) {
1670 assert(DestWidth > SrcWidth);
1671 Old = Builder.CreateAnd(Old, ConstantInt::get(~Mask), "mask");
1672 SV = Builder.CreateOr(Old, SV, "ins");
1679 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
1680 /// some part of a constant global variable. This intentionally only accepts
1681 /// constant expressions because we don't can't rewrite arbitrary instructions.
1682 static bool PointsToConstantGlobal(Value *V) {
1683 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
1684 return GV->isConstant();
1685 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
1686 if (CE->getOpcode() == Instruction::BitCast ||
1687 CE->getOpcode() == Instruction::GetElementPtr)
1688 return PointsToConstantGlobal(CE->getOperand(0));
1692 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
1693 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
1694 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
1695 /// track of whether it moves the pointer (with isOffset) but otherwise traverse
1696 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
1697 /// the alloca, and if the source pointer is a pointer to a constant global, we
1698 /// can optimize this.
1699 static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy,
1701 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1702 if (LoadInst *LI = dyn_cast<LoadInst>(*UI))
1703 // Ignore non-volatile loads, they are always ok.
1704 if (!LI->isVolatile())
1707 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) {
1708 // If uses of the bitcast are ok, we are ok.
1709 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset))
1713 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
1714 // If the GEP has all zero indices, it doesn't offset the pointer. If it
1715 // doesn't, it does.
1716 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
1717 isOffset || !GEP->hasAllZeroIndices()))
1722 // If this is isn't our memcpy/memmove, reject it as something we can't
1724 if (!isa<MemCpyInst>(*UI) && !isa<MemMoveInst>(*UI))
1727 // If we already have seen a copy, reject the second one.
1728 if (TheCopy) return false;
1730 // If the pointer has been offset from the start of the alloca, we can't
1731 // safely handle this.
1732 if (isOffset) return false;
1734 // If the memintrinsic isn't using the alloca as the dest, reject it.
1735 if (UI.getOperandNo() != 1) return false;
1737 MemIntrinsic *MI = cast<MemIntrinsic>(*UI);
1739 // If the source of the memcpy/move is not a constant global, reject it.
1740 if (!PointsToConstantGlobal(MI->getOperand(2)))
1743 // Otherwise, the transform is safe. Remember the copy instruction.
1749 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
1750 /// modified by a copy from a constant global. If we can prove this, we can
1751 /// replace any uses of the alloca with uses of the global directly.
1752 Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) {
1753 Instruction *TheCopy = 0;
1754 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false))