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/Support/Debug.h"
35 #include "llvm/Support/GetElementPtrTypeIterator.h"
36 #include "llvm/Support/MathExtras.h"
37 #include "llvm/Support/Compiler.h"
38 #include "llvm/ADT/SmallVector.h"
39 #include "llvm/ADT/Statistic.h"
40 #include "llvm/ADT/StringExtras.h"
43 STATISTIC(NumReplaced, "Number of allocas broken up");
44 STATISTIC(NumPromoted, "Number of allocas promoted");
45 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
46 STATISTIC(NumGlobals, "Number of allocas copied from constant global");
49 struct VISIBILITY_HIDDEN SROA : public FunctionPass {
50 static char ID; // Pass identification, replacement for typeid
51 explicit SROA(signed T = -1) : FunctionPass(&ID) {
58 bool runOnFunction(Function &F);
60 bool performScalarRepl(Function &F);
61 bool performPromotion(Function &F);
63 // getAnalysisUsage - This pass does not require any passes, but we know it
64 // will not alter the CFG, so say so.
65 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
66 AU.addRequired<DominatorTree>();
67 AU.addRequired<DominanceFrontier>();
68 AU.addRequired<TargetData>();
75 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
76 /// information about the uses. All these fields are initialized to false
77 /// and set to true when something is learned.
79 /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
82 /// needsCanon - This is set to true if there is some use of the alloca
83 /// that requires canonicalization.
86 /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
89 /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
93 : isUnsafe(false), needsCanon(false),
94 isMemCpySrc(false), isMemCpyDst(false) {}
99 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; }
101 int isSafeAllocaToScalarRepl(AllocationInst *AI);
103 void isSafeUseOfAllocation(Instruction *User, AllocationInst *AI,
105 void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI,
107 void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI,
108 unsigned OpNo, AllocaInfo &Info);
109 void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI,
112 void DoScalarReplacement(AllocationInst *AI,
113 std::vector<AllocationInst*> &WorkList);
114 void CanonicalizeAllocaUsers(AllocationInst *AI);
115 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base);
117 void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI,
118 SmallVector<AllocaInst*, 32> &NewElts);
120 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst,
122 SmallVector<AllocaInst*, 32> &NewElts);
123 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocationInst *AI,
124 SmallVector<AllocaInst*, 32> &NewElts);
125 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI,
126 SmallVector<AllocaInst*, 32> &NewElts);
128 const Type *CanConvertToScalar(Value *V, bool &IsNotTrivial);
129 void ConvertToScalar(AllocationInst *AI, const Type *Ty);
130 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset);
131 Value *ConvertUsesOfLoadToScalar(LoadInst *LI, AllocaInst *NewAI,
133 Value *ConvertUsesOfStoreToScalar(StoreInst *SI, AllocaInst *NewAI,
135 static Instruction *isOnlyCopiedFromConstantGlobal(AllocationInst *AI);
140 static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates");
142 // Public interface to the ScalarReplAggregates pass
143 FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) {
144 return new SROA(Threshold);
148 bool SROA::runOnFunction(Function &F) {
149 TD = &getAnalysis<TargetData>();
151 bool Changed = performPromotion(F);
153 bool LocalChange = performScalarRepl(F);
154 if (!LocalChange) break; // No need to repromote if no scalarrepl
156 LocalChange = performPromotion(F);
157 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
164 bool SROA::performPromotion(Function &F) {
165 std::vector<AllocaInst*> Allocas;
166 DominatorTree &DT = getAnalysis<DominatorTree>();
167 DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
169 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
171 bool Changed = false;
176 // Find allocas that are safe to promote, by looking at all instructions in
178 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
179 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
180 if (isAllocaPromotable(AI))
181 Allocas.push_back(AI);
183 if (Allocas.empty()) break;
185 PromoteMemToReg(Allocas, DT, DF);
186 NumPromoted += Allocas.size();
193 /// getNumSAElements - Return the number of elements in the specific struct or
195 static uint64_t getNumSAElements(const Type *T) {
196 if (const StructType *ST = dyn_cast<StructType>(T))
197 return ST->getNumElements();
198 return cast<ArrayType>(T)->getNumElements();
201 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
202 // which runs on all of the malloc/alloca instructions in the function, removing
203 // them if they are only used by getelementptr instructions.
205 bool SROA::performScalarRepl(Function &F) {
206 std::vector<AllocationInst*> WorkList;
208 // Scan the entry basic block, adding any alloca's and mallocs to the worklist
209 BasicBlock &BB = F.getEntryBlock();
210 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
211 if (AllocationInst *A = dyn_cast<AllocationInst>(I))
212 WorkList.push_back(A);
214 // Process the worklist
215 bool Changed = false;
216 while (!WorkList.empty()) {
217 AllocationInst *AI = WorkList.back();
220 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
221 // with unused elements.
222 if (AI->use_empty()) {
223 AI->eraseFromParent();
227 // Check to see if we can perform the core SROA transformation. We cannot
228 // transform the allocation instruction if it is an array allocation
229 // (allocations OF arrays are ok though), and an allocation of a scalar
230 // value cannot be decomposed at all.
231 if (!AI->isArrayAllocation() &&
232 (isa<StructType>(AI->getAllocatedType()) ||
233 isa<ArrayType>(AI->getAllocatedType())) &&
234 AI->getAllocatedType()->isSized() &&
235 // Do not promote any struct whose size is larger than "128" bytes.
236 TD->getTypePaddedSize(AI->getAllocatedType()) < SRThreshold &&
237 // Do not promote any struct into more than "32" separate vars.
238 getNumSAElements(AI->getAllocatedType()) < SRThreshold/4) {
239 // Check that all of the users of the allocation are capable of being
241 switch (isSafeAllocaToScalarRepl(AI)) {
242 default: assert(0 && "Unexpected value!");
243 case 0: // Not safe to scalar replace.
245 case 1: // Safe, but requires cleanup/canonicalizations first
246 CanonicalizeAllocaUsers(AI);
248 case 3: // Safe to scalar replace.
249 DoScalarReplacement(AI, WorkList);
255 // Check to see if this allocation is only modified by a memcpy/memmove from
256 // a constant global. If this is the case, we can change all users to use
257 // the constant global instead. This is commonly produced by the CFE by
258 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
259 // is only subsequently read.
260 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) {
261 DOUT << "Found alloca equal to global: " << *AI;
262 DOUT << " memcpy = " << *TheCopy;
263 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2));
264 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
265 TheCopy->eraseFromParent(); // Don't mutate the global.
266 AI->eraseFromParent();
272 // If we can turn this aggregate value (potentially with casts) into a
273 // simple scalar value that can be mem2reg'd into a register value.
274 bool IsNotTrivial = false;
275 if (const Type *ActualType = CanConvertToScalar(AI, IsNotTrivial))
276 if (IsNotTrivial && ActualType != Type::VoidTy) {
277 ConvertToScalar(AI, ActualType);
282 // Otherwise, couldn't process this.
288 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
289 /// predicate, do SROA now.
290 void SROA::DoScalarReplacement(AllocationInst *AI,
291 std::vector<AllocationInst*> &WorkList) {
292 DOUT << "Found inst to SROA: " << *AI;
293 SmallVector<AllocaInst*, 32> ElementAllocas;
294 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
295 ElementAllocas.reserve(ST->getNumContainedTypes());
296 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
297 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
299 AI->getName() + "." + utostr(i), AI);
300 ElementAllocas.push_back(NA);
301 WorkList.push_back(NA); // Add to worklist for recursive processing
304 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
305 ElementAllocas.reserve(AT->getNumElements());
306 const Type *ElTy = AT->getElementType();
307 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
308 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
309 AI->getName() + "." + utostr(i), AI);
310 ElementAllocas.push_back(NA);
311 WorkList.push_back(NA); // Add to worklist for recursive processing
315 // Now that we have created the alloca instructions that we want to use,
316 // expand the getelementptr instructions to use them.
318 while (!AI->use_empty()) {
319 Instruction *User = cast<Instruction>(AI->use_back());
320 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) {
321 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas);
322 BCInst->eraseFromParent();
327 // %res = load { i32, i32 }* %alloc
329 // %load.0 = load i32* %alloc.0
330 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
331 // %load.1 = load i32* %alloc.1
332 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
333 // (Also works for arrays instead of structs)
334 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
335 Value *Insert = UndefValue::get(LI->getType());
336 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) {
337 Value *Load = new LoadInst(ElementAllocas[i], "load", LI);
338 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI);
340 LI->replaceAllUsesWith(Insert);
341 LI->eraseFromParent();
346 // store { i32, i32 } %val, { i32, i32 }* %alloc
348 // %val.0 = extractvalue { i32, i32 } %val, 0
349 // store i32 %val.0, i32* %alloc.0
350 // %val.1 = extractvalue { i32, i32 } %val, 1
351 // store i32 %val.1, i32* %alloc.1
352 // (Also works for arrays instead of structs)
353 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
354 Value *Val = SI->getOperand(0);
355 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) {
356 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI);
357 new StoreInst(Extract, ElementAllocas[i], SI);
359 SI->eraseFromParent();
363 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
364 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst>
366 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
368 assert(Idx < ElementAllocas.size() && "Index out of range?");
369 AllocaInst *AllocaToUse = ElementAllocas[Idx];
372 if (GEPI->getNumOperands() == 3) {
373 // Do not insert a new getelementptr instruction with zero indices, only
374 // to have it optimized out later.
375 RepValue = AllocaToUse;
377 // We are indexing deeply into the structure, so we still need a
378 // getelement ptr instruction to finish the indexing. This may be
379 // expanded itself once the worklist is rerun.
381 SmallVector<Value*, 8> NewArgs;
382 NewArgs.push_back(Constant::getNullValue(Type::Int32Ty));
383 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end());
384 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(),
385 NewArgs.end(), "", GEPI);
386 RepValue->takeName(GEPI);
389 // If this GEP is to the start of the aggregate, check for memcpys.
390 if (Idx == 0 && GEPI->hasAllZeroIndices())
391 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas);
393 // Move all of the users over to the new GEP.
394 GEPI->replaceAllUsesWith(RepValue);
395 // Delete the old GEP
396 GEPI->eraseFromParent();
399 // Finally, delete the Alloca instruction
400 AI->eraseFromParent();
405 /// isSafeElementUse - Check to see if this use is an allowed use for a
406 /// getelementptr instruction of an array aggregate allocation. isFirstElt
407 /// indicates whether Ptr is known to the start of the aggregate.
409 void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI,
411 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
413 Instruction *User = cast<Instruction>(*I);
414 switch (User->getOpcode()) {
415 case Instruction::Load: break;
416 case Instruction::Store:
417 // Store is ok if storing INTO the pointer, not storing the pointer
418 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info);
420 case Instruction::GetElementPtr: {
421 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User);
422 bool AreAllZeroIndices = isFirstElt;
423 if (GEP->getNumOperands() > 1) {
424 if (!isa<ConstantInt>(GEP->getOperand(1)) ||
425 !cast<ConstantInt>(GEP->getOperand(1))->isZero())
426 // Using pointer arithmetic to navigate the array.
427 return MarkUnsafe(Info);
429 if (AreAllZeroIndices)
430 AreAllZeroIndices = GEP->hasAllZeroIndices();
432 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info);
433 if (Info.isUnsafe) return;
436 case Instruction::BitCast:
438 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info);
439 if (Info.isUnsafe) return;
442 DOUT << " Transformation preventing inst: " << *User;
443 return MarkUnsafe(Info);
444 case Instruction::Call:
445 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
447 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info);
448 if (Info.isUnsafe) return;
452 DOUT << " Transformation preventing inst: " << *User;
453 return MarkUnsafe(Info);
455 DOUT << " Transformation preventing inst: " << *User;
456 return MarkUnsafe(Info);
459 return; // All users look ok :)
462 /// AllUsersAreLoads - Return true if all users of this value are loads.
463 static bool AllUsersAreLoads(Value *Ptr) {
464 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
466 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load)
471 /// isSafeUseOfAllocation - Check to see if this user is an allowed use for an
472 /// aggregate allocation.
474 void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI,
476 if (BitCastInst *C = dyn_cast<BitCastInst>(User))
477 return isSafeUseOfBitCastedAllocation(C, AI, Info);
479 if (LoadInst *LI = dyn_cast<LoadInst>(User))
480 if (!LI->isVolatile())
481 return;// Loads (returning a first class aggregrate) are always rewritable
483 if (StoreInst *SI = dyn_cast<StoreInst>(User))
484 if (!SI->isVolatile() && SI->getOperand(0) != AI)
485 return;// Store is ok if storing INTO the pointer, not storing the pointer
487 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User);
489 return MarkUnsafe(Info);
491 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI);
493 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>".
495 I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) {
496 return MarkUnsafe(Info);
500 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices??
502 bool IsAllZeroIndices = true;
504 // If the first index is a non-constant index into an array, see if we can
505 // handle it as a special case.
506 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
507 if (!isa<ConstantInt>(I.getOperand())) {
508 IsAllZeroIndices = 0;
509 uint64_t NumElements = AT->getNumElements();
511 // If this is an array index and the index is not constant, we cannot
512 // promote... that is unless the array has exactly one or two elements in
513 // it, in which case we CAN promote it, but we have to canonicalize this
514 // out if this is the only problem.
515 if ((NumElements == 1 || NumElements == 2) &&
516 AllUsersAreLoads(GEPI)) {
517 Info.needsCanon = true;
518 return; // Canonicalization required!
520 return MarkUnsafe(Info);
524 // Walk through the GEP type indices, checking the types that this indexes
526 for (; I != E; ++I) {
527 // Ignore struct elements, no extra checking needed for these.
528 if (isa<StructType>(*I))
531 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand());
532 if (!IdxVal) return MarkUnsafe(Info);
534 // Are all indices still zero?
535 IsAllZeroIndices &= IdxVal->isZero();
537 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
538 // This GEP indexes an array. Verify that this is an in-range constant
539 // integer. Specifically, consider A[0][i]. We cannot know that the user
540 // isn't doing invalid things like allowing i to index an out-of-range
541 // subscript that accesses A[1]. Because of this, we have to reject SROA
542 // of any accesses into structs where any of the components are variables.
543 if (IdxVal->getZExtValue() >= AT->getNumElements())
544 return MarkUnsafe(Info);
545 } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) {
546 if (IdxVal->getZExtValue() >= VT->getNumElements())
547 return MarkUnsafe(Info);
551 // If there are any non-simple uses of this getelementptr, make sure to reject
553 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info);
556 /// isSafeMemIntrinsicOnAllocation - Return true if the specified memory
557 /// intrinsic can be promoted by SROA. At this point, we know that the operand
558 /// of the memintrinsic is a pointer to the beginning of the allocation.
559 void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI,
560 unsigned OpNo, AllocaInfo &Info) {
561 // If not constant length, give up.
562 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
563 if (!Length) return MarkUnsafe(Info);
565 // If not the whole aggregate, give up.
566 if (Length->getZExtValue() !=
567 TD->getTypePaddedSize(AI->getType()->getElementType()))
568 return MarkUnsafe(Info);
570 // We only know about memcpy/memset/memmove.
571 if (!isa<MemCpyInst>(MI) && !isa<MemSetInst>(MI) && !isa<MemMoveInst>(MI))
572 return MarkUnsafe(Info);
574 // Otherwise, we can transform it. Determine whether this is a memcpy/set
575 // into or out of the aggregate.
577 Info.isMemCpyDst = true;
580 Info.isMemCpySrc = true;
584 /// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast
586 void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI,
588 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end();
590 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) {
591 isSafeUseOfBitCastedAllocation(BCU, AI, Info);
592 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) {
593 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info);
594 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
595 if (SI->isVolatile())
596 return MarkUnsafe(Info);
598 // If storing the entire alloca in one chunk through a bitcasted pointer
599 // to integer, we can transform it. This happens (for example) when you
600 // cast a {i32,i32}* to i64* and store through it. This is similar to the
601 // memcpy case and occurs in various "byval" cases and emulated memcpys.
602 if (isa<IntegerType>(SI->getOperand(0)->getType()) &&
603 TD->getTypePaddedSize(SI->getOperand(0)->getType()) ==
604 TD->getTypePaddedSize(AI->getType()->getElementType())) {
605 Info.isMemCpyDst = true;
608 return MarkUnsafe(Info);
609 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) {
610 if (LI->isVolatile())
611 return MarkUnsafe(Info);
613 // If loading the entire alloca in one chunk through a bitcasted pointer
614 // to integer, we can transform it. This happens (for example) when you
615 // cast a {i32,i32}* to i64* and load through it. This is similar to the
616 // memcpy case and occurs in various "byval" cases and emulated memcpys.
617 if (isa<IntegerType>(LI->getType()) &&
618 TD->getTypePaddedSize(LI->getType()) ==
619 TD->getTypePaddedSize(AI->getType()->getElementType())) {
620 Info.isMemCpySrc = true;
623 return MarkUnsafe(Info);
625 return MarkUnsafe(Info);
627 if (Info.isUnsafe) return;
631 /// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes
632 /// to its first element. Transform users of the cast to use the new values
634 void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI,
635 SmallVector<AllocaInst*, 32> &NewElts) {
636 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end();
638 Instruction *User = cast<Instruction>(*UI++);
639 if (BitCastInst *BCU = dyn_cast<BitCastInst>(User)) {
640 RewriteBitCastUserOfAlloca(BCU, AI, NewElts);
641 if (BCU->use_empty()) BCU->eraseFromParent();
645 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
646 // This must be memcpy/memmove/memset of the entire aggregate.
647 // Split into one per element.
648 RewriteMemIntrinUserOfAlloca(MI, BCInst, AI, NewElts);
652 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
653 // If this is a store of the entire alloca from an integer, rewrite it.
654 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
658 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
659 // If this is a load of the entire alloca to an integer, rewrite it.
660 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
664 // Otherwise it must be some other user of a gep of the first pointer. Just
665 // leave these alone.
670 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
671 /// Rewrite it to copy or set the elements of the scalarized memory.
672 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst,
674 SmallVector<AllocaInst*, 32> &NewElts) {
676 // If this is a memcpy/memmove, construct the other pointer as the
679 if (MemCpyInst *MCI = dyn_cast<MemCpyInst>(MI)) {
680 if (BCInst == MCI->getRawDest())
681 OtherPtr = MCI->getRawSource();
683 assert(BCInst == MCI->getRawSource());
684 OtherPtr = MCI->getRawDest();
686 } else if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
687 if (BCInst == MMI->getRawDest())
688 OtherPtr = MMI->getRawSource();
690 assert(BCInst == MMI->getRawSource());
691 OtherPtr = MMI->getRawDest();
695 // If there is an other pointer, we want to convert it to the same pointer
696 // type as AI has, so we can GEP through it safely.
698 // It is likely that OtherPtr is a bitcast, if so, remove it.
699 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr))
700 OtherPtr = BC->getOperand(0);
701 // All zero GEPs are effectively bitcasts.
702 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr))
703 if (GEP->hasAllZeroIndices())
704 OtherPtr = GEP->getOperand(0);
706 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr))
707 if (BCE->getOpcode() == Instruction::BitCast)
708 OtherPtr = BCE->getOperand(0);
710 // If the pointer is not the right type, insert a bitcast to the right
712 if (OtherPtr->getType() != AI->getType())
713 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(),
717 // Process each element of the aggregate.
718 Value *TheFn = MI->getOperand(0);
719 const Type *BytePtrTy = MI->getRawDest()->getType();
720 bool SROADest = MI->getRawDest() == BCInst;
722 Constant *Zero = Constant::getNullValue(Type::Int32Ty);
724 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
725 // If this is a memcpy/memmove, emit a GEP of the other element address.
728 Value *Idx[2] = { Zero, ConstantInt::get(Type::Int32Ty, i) };
729 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2,
730 OtherPtr->getNameStr()+"."+utostr(i),
734 Value *EltPtr = NewElts[i];
735 const Type *EltTy =cast<PointerType>(EltPtr->getType())->getElementType();
737 // If we got down to a scalar, insert a load or store as appropriate.
738 if (EltTy->isSingleValueType()) {
739 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
740 Value *Elt = new LoadInst(SROADest ? OtherElt : EltPtr, "tmp",
742 new StoreInst(Elt, SROADest ? EltPtr : OtherElt, MI);
745 assert(isa<MemSetInst>(MI));
747 // If the stored element is zero (common case), just store a null
750 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) {
752 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
754 // If EltTy is a vector type, get the element type.
755 const Type *ValTy = EltTy;
756 if (const VectorType *VTy = dyn_cast<VectorType>(ValTy))
757 ValTy = VTy->getElementType();
759 // Construct an integer with the right value.
760 unsigned EltSize = TD->getTypeSizeInBits(ValTy);
761 APInt OneVal(EltSize, CI->getZExtValue());
762 APInt TotalVal(OneVal);
764 for (unsigned i = 0; 8*i < EltSize; ++i) {
765 TotalVal = TotalVal.shl(8);
769 // Convert the integer value to the appropriate type.
770 StoreVal = ConstantInt::get(TotalVal);
771 if (isa<PointerType>(ValTy))
772 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
773 else if (ValTy->isFloatingPoint())
774 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
775 assert(StoreVal->getType() == ValTy && "Type mismatch!");
777 // If the requested value was a vector constant, create it.
778 if (EltTy != ValTy) {
779 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements();
780 SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
781 StoreVal = ConstantVector::get(&Elts[0], NumElts);
784 new StoreInst(StoreVal, EltPtr, MI);
787 // Otherwise, if we're storing a byte variable, use a memset call for
791 // Cast the element pointer to BytePtrTy.
792 if (EltPtr->getType() != BytePtrTy)
793 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI);
795 // Cast the other pointer (if we have one) to BytePtrTy.
796 if (OtherElt && OtherElt->getType() != BytePtrTy)
797 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(),
800 unsigned EltSize = TD->getTypePaddedSize(EltTy);
802 // Finally, insert the meminst for this element.
803 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
805 SROADest ? EltPtr : OtherElt, // Dest ptr
806 SROADest ? OtherElt : EltPtr, // Src ptr
807 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
810 CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
812 assert(isa<MemSetInst>(MI));
814 EltPtr, MI->getOperand(2), // Dest, Value,
815 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
818 CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
821 MI->eraseFromParent();
824 /// RewriteStoreUserOfWholeAlloca - We found an store of an integer that
825 /// overwrites the entire allocation. Extract out the pieces of the stored
826 /// integer and store them individually.
827 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI,
829 SmallVector<AllocaInst*, 32> &NewElts){
830 // Extract each element out of the integer according to its structure offset
831 // and store the element value to the individual alloca.
832 Value *SrcVal = SI->getOperand(0);
833 const Type *AllocaEltTy = AI->getType()->getElementType();
834 uint64_t AllocaSizeBits = TD->getTypePaddedSizeInBits(AllocaEltTy);
836 // If this isn't a store of an integer to the whole alloca, it may be a store
837 // to the first element. Just ignore the store in this case and normal SROA
839 if (!isa<IntegerType>(SrcVal->getType()) ||
840 TD->getTypePaddedSizeInBits(SrcVal->getType()) != AllocaSizeBits)
843 DOUT << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << *SI;
845 // There are two forms here: AI could be an array or struct. Both cases
846 // have different ways to compute the element offset.
847 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
848 const StructLayout *Layout = TD->getStructLayout(EltSTy);
850 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
851 // Get the number of bits to shift SrcVal to get the value.
852 const Type *FieldTy = EltSTy->getElementType(i);
853 uint64_t Shift = Layout->getElementOffsetInBits(i);
855 if (TD->isBigEndian())
856 Shift = AllocaSizeBits-Shift-TD->getTypePaddedSizeInBits(FieldTy);
858 Value *EltVal = SrcVal;
860 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
861 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
862 "sroa.store.elt", SI);
865 // Truncate down to an integer of the right size.
866 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
868 // Ignore zero sized fields like {}, they obviously contain no data.
869 if (FieldSizeBits == 0) continue;
871 if (FieldSizeBits != AllocaSizeBits)
872 EltVal = new TruncInst(EltVal, IntegerType::get(FieldSizeBits), "", SI);
873 Value *DestField = NewElts[i];
874 if (EltVal->getType() == FieldTy) {
875 // Storing to an integer field of this size, just do it.
876 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) {
877 // Bitcast to the right element type (for fp/vector values).
878 EltVal = new BitCastInst(EltVal, FieldTy, "", SI);
880 // Otherwise, bitcast the dest pointer (for aggregates).
881 DestField = new BitCastInst(DestField,
882 PointerType::getUnqual(EltVal->getType()),
885 new StoreInst(EltVal, DestField, SI);
889 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
890 const Type *ArrayEltTy = ATy->getElementType();
891 uint64_t ElementOffset = TD->getTypePaddedSizeInBits(ArrayEltTy);
892 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
896 if (TD->isBigEndian())
897 Shift = AllocaSizeBits-ElementOffset;
901 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
902 // Ignore zero sized fields like {}, they obviously contain no data.
903 if (ElementSizeBits == 0) continue;
905 Value *EltVal = SrcVal;
907 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
908 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
909 "sroa.store.elt", SI);
912 // Truncate down to an integer of the right size.
913 if (ElementSizeBits != AllocaSizeBits)
914 EltVal = new TruncInst(EltVal, IntegerType::get(ElementSizeBits),"",SI);
915 Value *DestField = NewElts[i];
916 if (EltVal->getType() == ArrayEltTy) {
917 // Storing to an integer field of this size, just do it.
918 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) {
919 // Bitcast to the right element type (for fp/vector values).
920 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI);
922 // Otherwise, bitcast the dest pointer (for aggregates).
923 DestField = new BitCastInst(DestField,
924 PointerType::getUnqual(EltVal->getType()),
927 new StoreInst(EltVal, DestField, SI);
929 if (TD->isBigEndian())
930 Shift -= ElementOffset;
932 Shift += ElementOffset;
936 SI->eraseFromParent();
939 /// RewriteLoadUserOfWholeAlloca - We found an load of the entire allocation to
940 /// an integer. Load the individual pieces to form the aggregate value.
941 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI,
942 SmallVector<AllocaInst*, 32> &NewElts) {
943 // Extract each element out of the NewElts according to its structure offset
944 // and form the result value.
945 const Type *AllocaEltTy = AI->getType()->getElementType();
946 uint64_t AllocaSizeBits = TD->getTypePaddedSizeInBits(AllocaEltTy);
948 // If this isn't a load of the whole alloca to an integer, it may be a load
949 // of the first element. Just ignore the load in this case and normal SROA
951 if (!isa<IntegerType>(LI->getType()) ||
952 TD->getTypePaddedSizeInBits(LI->getType()) != AllocaSizeBits)
955 DOUT << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << *LI;
957 // There are two forms here: AI could be an array or struct. Both cases
958 // have different ways to compute the element offset.
959 const StructLayout *Layout = 0;
960 uint64_t ArrayEltBitOffset = 0;
961 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
962 Layout = TD->getStructLayout(EltSTy);
964 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
965 ArrayEltBitOffset = TD->getTypePaddedSizeInBits(ArrayEltTy);
968 Value *ResultVal = Constant::getNullValue(LI->getType());
970 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
971 // Load the value from the alloca. If the NewElt is an aggregate, cast
972 // the pointer to an integer of the same size before doing the load.
973 Value *SrcField = NewElts[i];
974 const Type *FieldTy =
975 cast<PointerType>(SrcField->getType())->getElementType();
976 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
978 // Ignore zero sized fields like {}, they obviously contain no data.
979 if (FieldSizeBits == 0) continue;
981 const IntegerType *FieldIntTy = IntegerType::get(FieldSizeBits);
982 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() &&
983 !isa<VectorType>(FieldTy))
984 SrcField = new BitCastInst(SrcField, PointerType::getUnqual(FieldIntTy),
986 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
988 // If SrcField is a fp or vector of the right size but that isn't an
989 // integer type, bitcast to an integer so we can shift it.
990 if (SrcField->getType() != FieldIntTy)
991 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
993 // Zero extend the field to be the same size as the final alloca so that
994 // we can shift and insert it.
995 if (SrcField->getType() != ResultVal->getType())
996 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
998 // Determine the number of bits to shift SrcField.
1000 if (Layout) // Struct case.
1001 Shift = Layout->getElementOffsetInBits(i);
1003 Shift = i*ArrayEltBitOffset;
1005 if (TD->isBigEndian())
1006 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
1009 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
1010 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
1013 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
1016 LI->replaceAllUsesWith(ResultVal);
1017 LI->eraseFromParent();
1021 /// HasPadding - Return true if the specified type has any structure or
1022 /// alignment padding, false otherwise.
1023 static bool HasPadding(const Type *Ty, const TargetData &TD) {
1024 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1025 const StructLayout *SL = TD.getStructLayout(STy);
1026 unsigned PrevFieldBitOffset = 0;
1027 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1028 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
1030 // Padding in sub-elements?
1031 if (HasPadding(STy->getElementType(i), TD))
1034 // Check to see if there is any padding between this element and the
1037 unsigned PrevFieldEnd =
1038 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
1039 if (PrevFieldEnd < FieldBitOffset)
1043 PrevFieldBitOffset = FieldBitOffset;
1046 // Check for tail padding.
1047 if (unsigned EltCount = STy->getNumElements()) {
1048 unsigned PrevFieldEnd = PrevFieldBitOffset +
1049 TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
1050 if (PrevFieldEnd < SL->getSizeInBits())
1054 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1055 return HasPadding(ATy->getElementType(), TD);
1056 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1057 return HasPadding(VTy->getElementType(), TD);
1059 return TD.getTypeSizeInBits(Ty) != TD.getTypePaddedSizeInBits(Ty);
1062 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
1063 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
1064 /// or 1 if safe after canonicalization has been performed.
1066 int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) {
1067 // Loop over the use list of the alloca. We can only transform it if all of
1068 // the users are safe to transform.
1071 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end();
1073 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info);
1074 if (Info.isUnsafe) {
1075 DOUT << "Cannot transform: " << *AI << " due to user: " << **I;
1080 // Okay, we know all the users are promotable. If the aggregate is a memcpy
1081 // source and destination, we have to be careful. In particular, the memcpy
1082 // could be moving around elements that live in structure padding of the LLVM
1083 // types, but may actually be used. In these cases, we refuse to promote the
1085 if (Info.isMemCpySrc && Info.isMemCpyDst &&
1086 HasPadding(AI->getType()->getElementType(), *TD))
1089 // If we require cleanup, return 1, otherwise return 3.
1090 return Info.needsCanon ? 1 : 3;
1093 /// CanonicalizeAllocaUsers - If SROA reported that it can promote the specified
1094 /// allocation, but only if cleaned up, perform the cleanups required.
1095 void SROA::CanonicalizeAllocaUsers(AllocationInst *AI) {
1096 // At this point, we know that the end result will be SROA'd and promoted, so
1097 // we can insert ugly code if required so long as sroa+mem2reg will clean it
1099 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
1101 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI++);
1102 if (!GEPI) continue;
1103 gep_type_iterator I = gep_type_begin(GEPI);
1106 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
1107 uint64_t NumElements = AT->getNumElements();
1109 if (!isa<ConstantInt>(I.getOperand())) {
1110 if (NumElements == 1) {
1111 GEPI->setOperand(2, Constant::getNullValue(Type::Int32Ty));
1113 assert(NumElements == 2 && "Unhandled case!");
1114 // All users of the GEP must be loads. At each use of the GEP, insert
1115 // two loads of the appropriate indexed GEP and select between them.
1116 Value *IsOne = new ICmpInst(ICmpInst::ICMP_NE, I.getOperand(),
1117 Constant::getNullValue(I.getOperand()->getType()),
1119 // Insert the new GEP instructions, which are properly indexed.
1120 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end());
1121 Indices[1] = Constant::getNullValue(Type::Int32Ty);
1122 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
1125 GEPI->getName()+".0", GEPI);
1126 Indices[1] = ConstantInt::get(Type::Int32Ty, 1);
1127 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
1130 GEPI->getName()+".1", GEPI);
1131 // Replace all loads of the variable index GEP with loads from both
1132 // indexes and a select.
1133 while (!GEPI->use_empty()) {
1134 LoadInst *LI = cast<LoadInst>(GEPI->use_back());
1135 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI);
1136 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI);
1137 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI);
1138 LI->replaceAllUsesWith(R);
1139 LI->eraseFromParent();
1141 GEPI->eraseFromParent();
1148 /// MergeInType - Add the 'In' type to the accumulated type so far. If the
1149 /// types are incompatible, return true, otherwise update Accum and return
1152 /// There are three cases we handle here:
1153 /// 1) An effectively-integer union, where the pieces are stored into as
1154 /// smaller integers (common with byte swap and other idioms).
1155 /// 2) A union of vector types of the same size and potentially its elements.
1156 /// Here we turn element accesses into insert/extract element operations.
1157 /// 3) A union of scalar types, such as int/float or int/pointer. Here we
1158 /// merge together into integers, allowing the xform to work with #1 as
1160 static bool MergeInType(const Type *In, const Type *&Accum,
1161 const TargetData &TD) {
1162 // If this is our first type, just use it.
1163 const VectorType *PTy;
1164 if (Accum == Type::VoidTy || In == Accum) {
1166 } else if (In == Type::VoidTy) {
1168 } else if (In->isInteger() && Accum->isInteger()) { // integer union.
1169 // Otherwise pick whichever type is larger.
1170 if (cast<IntegerType>(In)->getBitWidth() >
1171 cast<IntegerType>(Accum)->getBitWidth())
1173 } else if (isa<PointerType>(In) && isa<PointerType>(Accum)) {
1174 // Pointer unions just stay as one of the pointers.
1175 } else if (isa<VectorType>(In) || isa<VectorType>(Accum)) {
1176 if ((PTy = dyn_cast<VectorType>(Accum)) &&
1177 PTy->getElementType() == In) {
1178 // Accum is a vector, and we are accessing an element: ok.
1179 } else if ((PTy = dyn_cast<VectorType>(In)) &&
1180 PTy->getElementType() == Accum) {
1181 // In is a vector, and accum is an element: ok, remember In.
1183 } else if ((PTy = dyn_cast<VectorType>(In)) && isa<VectorType>(Accum) &&
1184 PTy->getBitWidth() == cast<VectorType>(Accum)->getBitWidth()) {
1185 // Two vectors of the same size: keep Accum.
1187 // Cannot insert an short into a <4 x int> or handle
1188 // <2 x int> -> <4 x int>
1192 // Pointer/FP/Integer unions merge together as integers.
1193 switch (Accum->getTypeID()) {
1194 case Type::PointerTyID: Accum = TD.getIntPtrType(); break;
1195 case Type::FloatTyID: Accum = Type::Int32Ty; break;
1196 case Type::DoubleTyID: Accum = Type::Int64Ty; break;
1197 case Type::X86_FP80TyID: return true;
1198 case Type::FP128TyID: return true;
1199 case Type::PPC_FP128TyID: return true;
1201 assert(Accum->isInteger() && "Unknown FP type!");
1205 switch (In->getTypeID()) {
1206 case Type::PointerTyID: In = TD.getIntPtrType(); break;
1207 case Type::FloatTyID: In = Type::Int32Ty; break;
1208 case Type::DoubleTyID: In = Type::Int64Ty; break;
1209 case Type::X86_FP80TyID: return true;
1210 case Type::FP128TyID: return true;
1211 case Type::PPC_FP128TyID: return true;
1213 assert(In->isInteger() && "Unknown FP type!");
1216 return MergeInType(In, Accum, TD);
1221 /// getIntAtLeastAsBigAs - Return an integer type that is at least as big as the
1222 /// specified type. If there is no suitable type, this returns null.
1223 const Type *getIntAtLeastAsBigAs(unsigned NumBits) {
1224 if (NumBits > 64) return 0;
1225 if (NumBits > 32) return Type::Int64Ty;
1226 if (NumBits > 16) return Type::Int32Ty;
1227 if (NumBits > 8) return Type::Int16Ty;
1228 return Type::Int8Ty;
1231 /// CanConvertToScalar - V is a pointer. If we can convert the pointee to a
1232 /// single scalar integer type, return that type. Further, if the use is not
1233 /// a completely trivial use that mem2reg could promote, set IsNotTrivial. If
1234 /// there are no uses of this pointer, return Type::VoidTy to differentiate from
1237 const Type *SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial) {
1238 const Type *UsedType = Type::VoidTy; // No uses, no forced type.
1239 const PointerType *PTy = cast<PointerType>(V->getType());
1241 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1242 Instruction *User = cast<Instruction>(*UI);
1244 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1245 if (LI->isVolatile())
1248 // FIXME: Loads of a first class aggregrate value could be converted to a
1249 // series of loads and insertvalues
1250 if (!LI->getType()->isSingleValueType())
1253 if (MergeInType(LI->getType(), UsedType, *TD))
1258 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1259 // Storing the pointer, not into the value?
1260 if (SI->getOperand(0) == V || SI->isVolatile()) return 0;
1262 // FIXME: Stores of a first class aggregrate value could be converted to a
1263 // series of extractvalues and stores
1264 if (!SI->getOperand(0)->getType()->isSingleValueType())
1267 // NOTE: We could handle storing of FP imms into integers here!
1269 if (MergeInType(SI->getOperand(0)->getType(), UsedType, *TD))
1273 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
1274 IsNotTrivial = true;
1275 const Type *SubTy = CanConvertToScalar(CI, IsNotTrivial);
1276 if (!SubTy || MergeInType(SubTy, UsedType, *TD)) return 0;
1280 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1281 // Check to see if this is stepping over an element: GEP Ptr, int C
1282 if (GEP->getNumOperands() == 2 && isa<ConstantInt>(GEP->getOperand(1))) {
1283 unsigned Idx = cast<ConstantInt>(GEP->getOperand(1))->getZExtValue();
1284 unsigned ElSize = TD->getTypePaddedSize(PTy->getElementType());
1285 unsigned BitOffset = Idx*ElSize*8;
1286 if (BitOffset > 64 || !isPowerOf2_32(ElSize)) return 0;
1288 IsNotTrivial = true;
1289 const Type *SubElt = CanConvertToScalar(GEP, IsNotTrivial);
1290 if (SubElt == 0) return 0;
1291 if (SubElt != Type::VoidTy && SubElt->isInteger()) {
1293 getIntAtLeastAsBigAs(TD->getTypePaddedSizeInBits(SubElt)+BitOffset);
1294 if (NewTy == 0 || MergeInType(NewTy, UsedType, *TD)) return 0;
1297 // Cannot handle this!
1301 if (GEP->getNumOperands() == 3 &&
1302 isa<ConstantInt>(GEP->getOperand(1)) &&
1303 isa<ConstantInt>(GEP->getOperand(2)) &&
1304 cast<ConstantInt>(GEP->getOperand(1))->isZero()) {
1305 // We are stepping into an element, e.g. a structure or an array:
1306 // GEP Ptr, i32 0, i32 Cst
1307 const Type *AggTy = PTy->getElementType();
1308 unsigned Idx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
1310 if (const ArrayType *ATy = dyn_cast<ArrayType>(AggTy)) {
1311 if (Idx >= ATy->getNumElements()) return 0; // Out of range.
1312 } else if (const VectorType *VectorTy = dyn_cast<VectorType>(AggTy)) {
1313 // Getting an element of the vector.
1314 if (Idx >= VectorTy->getNumElements()) return 0; // Out of range.
1316 // Merge in the vector type.
1317 if (MergeInType(VectorTy, UsedType, *TD)) return 0;
1319 const Type *SubTy = CanConvertToScalar(GEP, IsNotTrivial);
1320 if (SubTy == 0) return 0;
1322 if (SubTy != Type::VoidTy && MergeInType(SubTy, UsedType, *TD))
1325 // We'll need to change this to an insert/extract element operation.
1326 IsNotTrivial = true;
1327 continue; // Everything looks ok
1329 } else if (isa<StructType>(AggTy)) {
1330 // Structs are always ok.
1335 getIntAtLeastAsBigAs(TD->getTypePaddedSizeInBits(AggTy));
1336 if (NTy == 0 || MergeInType(NTy, UsedType, *TD)) return 0;
1337 const Type *SubTy = CanConvertToScalar(GEP, IsNotTrivial);
1338 if (SubTy == 0) return 0;
1339 if (SubTy != Type::VoidTy && MergeInType(SubTy, UsedType, *TD))
1341 continue; // Everything looks ok
1346 // Cannot handle this!
1353 /// ConvertToScalar - The specified alloca passes the CanConvertToScalar
1354 /// predicate and is non-trivial. Convert it to something that can be trivially
1355 /// promoted into a register by mem2reg.
1356 void SROA::ConvertToScalar(AllocationInst *AI, const Type *ActualTy) {
1357 DOUT << "CONVERT TO SCALAR: " << *AI << " TYPE = "
1358 << *ActualTy << "\n";
1361 BasicBlock *EntryBlock = AI->getParent();
1362 assert(EntryBlock == &EntryBlock->getParent()->getEntryBlock() &&
1363 "Not in the entry block!");
1364 EntryBlock->getInstList().remove(AI); // Take the alloca out of the program.
1366 // Create and insert the alloca.
1367 AllocaInst *NewAI = new AllocaInst(ActualTy, 0, AI->getName(),
1368 EntryBlock->begin());
1369 ConvertUsesToScalar(AI, NewAI, 0);
1374 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
1375 /// directly. This happens when we are converting an "integer union" to a
1376 /// single integer scalar, or when we are converting a "vector union" to a
1377 /// vector with insert/extractelement instructions.
1379 /// Offset is an offset from the original alloca, in bits that need to be
1380 /// shifted to the right. By the end of this, there should be no uses of Ptr.
1381 void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset) {
1382 while (!Ptr->use_empty()) {
1383 Instruction *User = cast<Instruction>(Ptr->use_back());
1385 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1386 Value *NV = ConvertUsesOfLoadToScalar(LI, NewAI, Offset);
1387 LI->replaceAllUsesWith(NV);
1388 LI->eraseFromParent();
1392 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1393 assert(SI->getOperand(0) != Ptr && "Consistency error!");
1395 Value *SV = ConvertUsesOfStoreToScalar(SI, NewAI, Offset);
1396 new StoreInst(SV, NewAI, SI);
1397 SI->eraseFromParent();
1401 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
1402 ConvertUsesToScalar(CI, NewAI, Offset);
1403 CI->eraseFromParent();
1407 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1408 const PointerType *AggPtrTy =
1409 cast<PointerType>(GEP->getOperand(0)->getType());
1410 unsigned AggSizeInBits =
1411 TD->getTypePaddedSizeInBits(AggPtrTy->getElementType());
1413 // Check to see if this is stepping over an element: GEP Ptr, int C
1414 unsigned NewOffset = Offset;
1415 if (GEP->getNumOperands() == 2) {
1416 unsigned Idx = cast<ConstantInt>(GEP->getOperand(1))->getZExtValue();
1417 unsigned BitOffset = Idx*AggSizeInBits;
1419 NewOffset += BitOffset;
1420 ConvertUsesToScalar(GEP, NewAI, NewOffset);
1421 GEP->eraseFromParent();
1425 assert(GEP->getNumOperands() == 3 && "Unsupported operation");
1427 // We know that operand #2 is zero.
1428 unsigned Idx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
1429 const Type *AggTy = AggPtrTy->getElementType();
1430 if (const SequentialType *SeqTy = dyn_cast<SequentialType>(AggTy)) {
1431 unsigned ElSizeBits =
1432 TD->getTypePaddedSizeInBits(SeqTy->getElementType());
1434 NewOffset += ElSizeBits*Idx;
1436 const StructType *STy = cast<StructType>(AggTy);
1437 unsigned EltBitOffset =
1438 TD->getStructLayout(STy)->getElementOffsetInBits(Idx);
1440 NewOffset += EltBitOffset;
1442 ConvertUsesToScalar(GEP, NewAI, NewOffset);
1443 GEP->eraseFromParent();
1447 assert(0 && "Unsupported operation!");
1452 /// ConvertUsesOfLoadToScalar - Convert all of the users the specified load to
1453 /// use the new alloca directly, returning the value that should replace the
1454 /// load. This happens when we are converting an "integer union" to a
1455 /// single integer scalar, or when we are converting a "vector union" to a
1456 /// vector with insert/extractelement instructions.
1458 /// Offset is an offset from the original alloca, in bits that need to be
1459 /// shifted to the right. By the end of this, there should be no uses of Ptr.
1460 Value *SROA::ConvertUsesOfLoadToScalar(LoadInst *LI, AllocaInst *NewAI,
1462 // The load is a bit extract from NewAI shifted right by Offset bits.
1463 Value *NV = new LoadInst(NewAI, LI->getName(), LI);
1465 if (NV->getType() == LI->getType() && Offset == 0) {
1466 // We win, no conversion needed.
1470 // If the result type of the 'union' is a pointer, then this must be ptr->ptr
1471 // cast. Anything else would result in NV being an integer.
1472 if (isa<PointerType>(NV->getType())) {
1473 assert(isa<PointerType>(LI->getType()));
1474 return new BitCastInst(NV, LI->getType(), LI->getName(), LI);
1477 if (const VectorType *VTy = dyn_cast<VectorType>(NV->getType())) {
1478 // If the result alloca is a vector type, this is either an element
1479 // access or a bitcast to another vector type.
1480 if (isa<VectorType>(LI->getType()))
1481 return new BitCastInst(NV, LI->getType(), LI->getName(), LI);
1483 // Otherwise it must be an element access.
1486 unsigned EltSize = TD->getTypePaddedSizeInBits(VTy->getElementType());
1487 Elt = Offset/EltSize;
1488 Offset -= EltSize*Elt;
1490 NV = new ExtractElementInst(NV, ConstantInt::get(Type::Int32Ty, Elt),
1493 // If we're done, return this element.
1494 if (NV->getType() == LI->getType() && Offset == 0)
1498 const IntegerType *NTy = cast<IntegerType>(NV->getType());
1500 // If this is a big-endian system and the load is narrower than the
1501 // full alloca type, we need to do a shift to get the right bits.
1503 if (TD->isBigEndian()) {
1504 // On big-endian machines, the lowest bit is stored at the bit offset
1505 // from the pointer given by getTypeStoreSizeInBits. This matters for
1506 // integers with a bitwidth that is not a multiple of 8.
1507 ShAmt = TD->getTypeStoreSizeInBits(NTy) -
1508 TD->getTypeStoreSizeInBits(LI->getType()) - Offset;
1513 // Note: we support negative bitwidths (with shl) which are not defined.
1514 // We do this to support (f.e.) loads off the end of a structure where
1515 // only some bits are used.
1516 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
1517 NV = BinaryOperator::CreateLShr(NV,
1518 ConstantInt::get(NV->getType(),ShAmt),
1520 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
1521 NV = BinaryOperator::CreateShl(NV,
1522 ConstantInt::get(NV->getType(),-ShAmt),
1525 // Finally, unconditionally truncate the integer to the right width.
1526 unsigned LIBitWidth = TD->getTypeSizeInBits(LI->getType());
1527 if (LIBitWidth < NTy->getBitWidth())
1528 NV = new TruncInst(NV, IntegerType::get(LIBitWidth),
1531 // If the result is an integer, this is a trunc or bitcast.
1532 if (isa<IntegerType>(LI->getType())) {
1534 } else if (LI->getType()->isFloatingPoint()) {
1535 // Just do a bitcast, we know the sizes match up.
1536 NV = new BitCastInst(NV, LI->getType(), LI->getName(), LI);
1538 // Otherwise must be a pointer.
1539 NV = new IntToPtrInst(NV, LI->getType(), LI->getName(), LI);
1541 assert(NV->getType() == LI->getType() && "Didn't convert right?");
1546 /// ConvertUsesOfStoreToScalar - Convert the specified store to a load+store
1547 /// pair of the new alloca directly, returning the value that should be stored
1548 /// to the alloca. This happens when we are converting an "integer union" to a
1549 /// single integer scalar, or when we are converting a "vector union" to a
1550 /// vector with insert/extractelement instructions.
1552 /// Offset is an offset from the original alloca, in bits that need to be
1553 /// shifted to the right. By the end of this, there should be no uses of Ptr.
1554 Value *SROA::ConvertUsesOfStoreToScalar(StoreInst *SI, AllocaInst *NewAI,
1557 // Convert the stored type to the actual type, shift it left to insert
1558 // then 'or' into place.
1559 Value *SV = SI->getOperand(0);
1560 const Type *AllocaType = NewAI->getType()->getElementType();
1561 if (SV->getType() == AllocaType && Offset == 0) {
1563 } else if (const VectorType *PTy = dyn_cast<VectorType>(AllocaType)) {
1564 Value *Old = new LoadInst(NewAI, NewAI->getName()+".in", SI);
1566 // If the result alloca is a vector type, this is either an element
1567 // access or a bitcast to another vector type.
1568 if (isa<VectorType>(SV->getType())) {
1569 SV = new BitCastInst(SV, AllocaType, SV->getName(), SI);
1571 // Must be an element insertion.
1572 unsigned Elt = Offset/TD->getTypePaddedSizeInBits(PTy->getElementType());
1573 SV = InsertElementInst::Create(Old, SV,
1574 ConstantInt::get(Type::Int32Ty, Elt),
1577 } else if (isa<PointerType>(AllocaType)) {
1578 // If the alloca type is a pointer, then all the elements must be
1580 if (SV->getType() != AllocaType)
1581 SV = new BitCastInst(SV, AllocaType, SV->getName(), SI);
1583 Value *Old = new LoadInst(NewAI, NewAI->getName()+".in", SI);
1585 // If SV is a float, convert it to the appropriate integer type.
1586 // If it is a pointer, do the same, and also handle ptr->ptr casts
1588 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType());
1589 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType);
1590 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType());
1591 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType);
1592 if (SV->getType()->isFloatingPoint())
1593 SV = new BitCastInst(SV, IntegerType::get(SrcWidth),
1595 else if (isa<PointerType>(SV->getType()))
1596 SV = new PtrToIntInst(SV, TD->getIntPtrType(), SV->getName(), SI);
1598 // Always zero extend the value if needed.
1599 if (SV->getType() != AllocaType)
1600 SV = new ZExtInst(SV, AllocaType, SV->getName(), SI);
1602 // If this is a big-endian system and the store is narrower than the
1603 // full alloca type, we need to do a shift to get the right bits.
1605 if (TD->isBigEndian()) {
1606 // On big-endian machines, the lowest bit is stored at the bit offset
1607 // from the pointer given by getTypeStoreSizeInBits. This matters for
1608 // integers with a bitwidth that is not a multiple of 8.
1609 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
1614 // Note: we support negative bitwidths (with shr) which are not defined.
1615 // We do this to support (f.e.) stores off the end of a structure where
1616 // only some bits in the structure are set.
1617 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
1618 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
1619 SV = BinaryOperator::CreateShl(SV,
1620 ConstantInt::get(SV->getType(), ShAmt),
1623 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
1624 SV = BinaryOperator::CreateLShr(SV,
1625 ConstantInt::get(SV->getType(),-ShAmt),
1627 Mask = Mask.lshr(ShAmt);
1630 // Mask out the bits we are about to insert from the old value, and or
1632 if (SrcWidth != DestWidth) {
1633 assert(DestWidth > SrcWidth);
1634 Old = BinaryOperator::CreateAnd(Old, ConstantInt::get(~Mask),
1635 Old->getName()+".mask", SI);
1636 SV = BinaryOperator::CreateOr(Old, SV, SV->getName()+".ins", SI);
1644 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
1645 /// some part of a constant global variable. This intentionally only accepts
1646 /// constant expressions because we don't can't rewrite arbitrary instructions.
1647 static bool PointsToConstantGlobal(Value *V) {
1648 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
1649 return GV->isConstant();
1650 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
1651 if (CE->getOpcode() == Instruction::BitCast ||
1652 CE->getOpcode() == Instruction::GetElementPtr)
1653 return PointsToConstantGlobal(CE->getOperand(0));
1657 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
1658 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
1659 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
1660 /// track of whether it moves the pointer (with isOffset) but otherwise traverse
1661 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
1662 /// the alloca, and if the source pointer is a pointer to a constant global, we
1663 /// can optimize this.
1664 static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy,
1666 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1667 if (LoadInst *LI = dyn_cast<LoadInst>(*UI))
1668 // Ignore non-volatile loads, they are always ok.
1669 if (!LI->isVolatile())
1672 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) {
1673 // If uses of the bitcast are ok, we are ok.
1674 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset))
1678 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
1679 // If the GEP has all zero indices, it doesn't offset the pointer. If it
1680 // doesn't, it does.
1681 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
1682 isOffset || !GEP->hasAllZeroIndices()))
1687 // If this is isn't our memcpy/memmove, reject it as something we can't
1689 if (!isa<MemCpyInst>(*UI) && !isa<MemMoveInst>(*UI))
1692 // If we already have seen a copy, reject the second one.
1693 if (TheCopy) return false;
1695 // If the pointer has been offset from the start of the alloca, we can't
1696 // safely handle this.
1697 if (isOffset) return false;
1699 // If the memintrinsic isn't using the alloca as the dest, reject it.
1700 if (UI.getOperandNo() != 1) return false;
1702 MemIntrinsic *MI = cast<MemIntrinsic>(*UI);
1704 // If the source of the memcpy/move is not a constant global, reject it.
1705 if (!PointsToConstantGlobal(MI->getOperand(2)))
1708 // Otherwise, the transform is safe. Remember the copy instruction.
1714 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
1715 /// modified by a copy from a constant global. If we can prove this, we can
1716 /// replace any uses of the alloca with uses of the global directly.
1717 Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) {
1718 Instruction *TheCopy = 0;
1719 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false))