1 //===- SROA.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. It tries to identify promotable elements of an
12 /// aggregate alloca, and promote them to registers. It will also try to
13 /// convert uses of an element (or set of elements) of an alloca into a vector
14 /// or bitfield-style integer scalar if appropriate.
16 /// It works to do this with minimal slicing of the alloca so that regions
17 /// which are merely transferred in and out of external memory remain unchanged
18 /// and are not decomposed to scalar code.
20 /// Because this also performs alloca promotion, it can be thought of as also
21 /// serving the purpose of SSA formation. The algorithm iterates on the
22 /// function until all opportunities for promotion have been realized.
24 //===----------------------------------------------------------------------===//
26 #define DEBUG_TYPE "sroa"
27 #include "llvm/Transforms/Scalar.h"
28 #include "llvm/ADT/STLExtras.h"
29 #include "llvm/ADT/SetVector.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/Analysis/Loads.h"
33 #include "llvm/Analysis/PtrUseVisitor.h"
34 #include "llvm/Analysis/ValueTracking.h"
35 #include "llvm/DIBuilder.h"
36 #include "llvm/DebugInfo.h"
37 #include "llvm/IR/Constants.h"
38 #include "llvm/IR/DataLayout.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/Dominators.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/IRBuilder.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/IntrinsicInst.h"
45 #include "llvm/IR/LLVMContext.h"
46 #include "llvm/IR/Operator.h"
47 #include "llvm/InstVisitor.h"
48 #include "llvm/Pass.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Compiler.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/ErrorHandling.h"
53 #include "llvm/Support/MathExtras.h"
54 #include "llvm/Support/TimeValue.h"
55 #include "llvm/Support/raw_ostream.h"
56 #include "llvm/Transforms/Utils/Local.h"
57 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
58 #include "llvm/Transforms/Utils/SSAUpdater.h"
60 #if __cplusplus >= 201103L && !defined(NDEBUG)
61 // We only use this for a debug check in C++11
67 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
68 STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed");
69 STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions per alloca");
70 STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten");
71 STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition");
72 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
73 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
74 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
75 STATISTIC(NumDeleted, "Number of instructions deleted");
76 STATISTIC(NumVectorized, "Number of vectorized aggregates");
78 /// Hidden option to force the pass to not use DomTree and mem2reg, instead
79 /// forming SSA values through the SSAUpdater infrastructure.
81 ForceSSAUpdater("force-ssa-updater", cl::init(false), cl::Hidden);
83 /// Hidden option to enable randomly shuffling the slices to help uncover
84 /// instability in their order.
85 static cl::opt<bool> SROARandomShuffleSlices("sroa-random-shuffle-slices",
86 cl::init(false), cl::Hidden);
88 /// Hidden option to experiment with completely strict handling of inbounds
90 static cl::opt<bool> SROAStrictInbounds("sroa-strict-inbounds",
91 cl::init(false), cl::Hidden);
94 /// \brief A custom IRBuilder inserter which prefixes all names if they are
96 template <bool preserveNames = true>
97 class IRBuilderPrefixedInserter :
98 public IRBuilderDefaultInserter<preserveNames> {
102 void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
105 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
106 BasicBlock::iterator InsertPt) const {
107 IRBuilderDefaultInserter<preserveNames>::InsertHelper(
108 I, Name.isTriviallyEmpty() ? Name : Prefix + Name, BB, InsertPt);
112 // Specialization for not preserving the name is trivial.
114 class IRBuilderPrefixedInserter<false> :
115 public IRBuilderDefaultInserter<false> {
117 void SetNamePrefix(const Twine &P) {}
120 /// \brief Provide a typedef for IRBuilder that drops names in release builds.
122 typedef llvm::IRBuilder<true, ConstantFolder,
123 IRBuilderPrefixedInserter<true> > IRBuilderTy;
125 typedef llvm::IRBuilder<false, ConstantFolder,
126 IRBuilderPrefixedInserter<false> > IRBuilderTy;
131 /// \brief A used slice of an alloca.
133 /// This structure represents a slice of an alloca used by some instruction. It
134 /// stores both the begin and end offsets of this use, a pointer to the use
135 /// itself, and a flag indicating whether we can classify the use as splittable
136 /// or not when forming partitions of the alloca.
138 /// \brief The beginning offset of the range.
139 uint64_t BeginOffset;
141 /// \brief The ending offset, not included in the range.
144 /// \brief Storage for both the use of this slice and whether it can be
146 PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
149 Slice() : BeginOffset(), EndOffset() {}
150 Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
151 : BeginOffset(BeginOffset), EndOffset(EndOffset),
152 UseAndIsSplittable(U, IsSplittable) {}
154 uint64_t beginOffset() const { return BeginOffset; }
155 uint64_t endOffset() const { return EndOffset; }
157 bool isSplittable() const { return UseAndIsSplittable.getInt(); }
158 void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
160 Use *getUse() const { return UseAndIsSplittable.getPointer(); }
162 bool isDead() const { return getUse() == 0; }
163 void kill() { UseAndIsSplittable.setPointer(0); }
165 /// \brief Support for ordering ranges.
167 /// This provides an ordering over ranges such that start offsets are
168 /// always increasing, and within equal start offsets, the end offsets are
169 /// decreasing. Thus the spanning range comes first in a cluster with the
170 /// same start position.
171 bool operator<(const Slice &RHS) const {
172 if (beginOffset() < RHS.beginOffset()) return true;
173 if (beginOffset() > RHS.beginOffset()) return false;
174 if (isSplittable() != RHS.isSplittable()) return !isSplittable();
175 if (endOffset() > RHS.endOffset()) return true;
179 /// \brief Support comparison with a single offset to allow binary searches.
180 friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Slice &LHS,
181 uint64_t RHSOffset) {
182 return LHS.beginOffset() < RHSOffset;
184 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
186 return LHSOffset < RHS.beginOffset();
189 bool operator==(const Slice &RHS) const {
190 return isSplittable() == RHS.isSplittable() &&
191 beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
193 bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
195 } // end anonymous namespace
198 template <typename T> struct isPodLike;
199 template <> struct isPodLike<Slice> {
200 static const bool value = true;
205 /// \brief Representation of the alloca slices.
207 /// This class represents the slices of an alloca which are formed by its
208 /// various uses. If a pointer escapes, we can't fully build a representation
209 /// for the slices used and we reflect that in this structure. The uses are
210 /// stored, sorted by increasing beginning offset and with unsplittable slices
211 /// starting at a particular offset before splittable slices.
214 /// \brief Construct the slices of a particular alloca.
215 AllocaSlices(const DataLayout &DL, AllocaInst &AI);
217 /// \brief Test whether a pointer to the allocation escapes our analysis.
219 /// If this is true, the slices are never fully built and should be
221 bool isEscaped() const { return PointerEscapingInstr; }
223 /// \brief Support for iterating over the slices.
225 typedef SmallVectorImpl<Slice>::iterator iterator;
226 iterator begin() { return Slices.begin(); }
227 iterator end() { return Slices.end(); }
229 typedef SmallVectorImpl<Slice>::const_iterator const_iterator;
230 const_iterator begin() const { return Slices.begin(); }
231 const_iterator end() const { return Slices.end(); }
234 /// \brief Allow iterating the dead users for this alloca.
236 /// These are instructions which will never actually use the alloca as they
237 /// are outside the allocated range. They are safe to replace with undef and
240 typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator;
241 dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); }
242 dead_user_iterator dead_user_end() const { return DeadUsers.end(); }
245 /// \brief Allow iterating the dead expressions referring to this alloca.
247 /// These are operands which have cannot actually be used to refer to the
248 /// alloca as they are outside its range and the user doesn't correct for
249 /// that. These mostly consist of PHI node inputs and the like which we just
250 /// need to replace with undef.
252 typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator;
253 dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); }
254 dead_op_iterator dead_op_end() const { return DeadOperands.end(); }
257 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
258 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
259 void printSlice(raw_ostream &OS, const_iterator I,
260 StringRef Indent = " ") const;
261 void printUse(raw_ostream &OS, const_iterator I,
262 StringRef Indent = " ") const;
263 void print(raw_ostream &OS) const;
264 void dump(const_iterator I) const;
269 template <typename DerivedT, typename RetT = void> class BuilderBase;
271 friend class AllocaSlices::SliceBuilder;
273 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
274 /// \brief Handle to alloca instruction to simplify method interfaces.
278 /// \brief The instruction responsible for this alloca not having a known set
281 /// When an instruction (potentially) escapes the pointer to the alloca, we
282 /// store a pointer to that here and abort trying to form slices of the
283 /// alloca. This will be null if the alloca slices are analyzed successfully.
284 Instruction *PointerEscapingInstr;
286 /// \brief The slices of the alloca.
288 /// We store a vector of the slices formed by uses of the alloca here. This
289 /// vector is sorted by increasing begin offset, and then the unsplittable
290 /// slices before the splittable ones. See the Slice inner class for more
292 SmallVector<Slice, 8> Slices;
294 /// \brief Instructions which will become dead if we rewrite the alloca.
296 /// Note that these are not separated by slice. This is because we expect an
297 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
298 /// all these instructions can simply be removed and replaced with undef as
299 /// they come from outside of the allocated space.
300 SmallVector<Instruction *, 8> DeadUsers;
302 /// \brief Operands which will become dead if we rewrite the alloca.
304 /// These are operands that in their particular use can be replaced with
305 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
306 /// to PHI nodes and the like. They aren't entirely dead (there might be
307 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
308 /// want to swap this particular input for undef to simplify the use lists of
310 SmallVector<Use *, 8> DeadOperands;
314 static Value *foldSelectInst(SelectInst &SI) {
315 // If the condition being selected on is a constant or the same value is
316 // being selected between, fold the select. Yes this does (rarely) happen
318 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
319 return SI.getOperand(1+CI->isZero());
320 if (SI.getOperand(1) == SI.getOperand(2))
321 return SI.getOperand(1);
326 /// \brief Builder for the alloca slices.
328 /// This class builds a set of alloca slices by recursively visiting the uses
329 /// of an alloca and making a slice for each load and store at each offset.
330 class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
331 friend class PtrUseVisitor<SliceBuilder>;
332 friend class InstVisitor<SliceBuilder>;
333 typedef PtrUseVisitor<SliceBuilder> Base;
335 const uint64_t AllocSize;
338 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
339 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
341 /// \brief Set to de-duplicate dead instructions found in the use walk.
342 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
345 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &S)
346 : PtrUseVisitor<SliceBuilder>(DL),
347 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), S(S) {}
350 void markAsDead(Instruction &I) {
351 if (VisitedDeadInsts.insert(&I))
352 S.DeadUsers.push_back(&I);
355 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
356 bool IsSplittable = false) {
357 // Completely skip uses which have a zero size or start either before or
358 // past the end of the allocation.
359 if (Size == 0 || Offset.uge(AllocSize)) {
360 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
361 << " which has zero size or starts outside of the "
362 << AllocSize << " byte alloca:\n"
363 << " alloca: " << S.AI << "\n"
364 << " use: " << I << "\n");
365 return markAsDead(I);
368 uint64_t BeginOffset = Offset.getZExtValue();
369 uint64_t EndOffset = BeginOffset + Size;
371 // Clamp the end offset to the end of the allocation. Note that this is
372 // formulated to handle even the case where "BeginOffset + Size" overflows.
373 // This may appear superficially to be something we could ignore entirely,
374 // but that is not so! There may be widened loads or PHI-node uses where
375 // some instructions are dead but not others. We can't completely ignore
376 // them, and so have to record at least the information here.
377 assert(AllocSize >= BeginOffset); // Established above.
378 if (Size > AllocSize - BeginOffset) {
379 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
380 << " to remain within the " << AllocSize << " byte alloca:\n"
381 << " alloca: " << S.AI << "\n"
382 << " use: " << I << "\n");
383 EndOffset = AllocSize;
386 S.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
389 void visitBitCastInst(BitCastInst &BC) {
391 return markAsDead(BC);
393 return Base::visitBitCastInst(BC);
396 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
397 if (GEPI.use_empty())
398 return markAsDead(GEPI);
400 if (SROAStrictInbounds && GEPI.isInBounds()) {
401 // FIXME: This is a manually un-factored variant of the basic code inside
402 // of GEPs with checking of the inbounds invariant specified in the
403 // langref in a very strict sense. If we ever want to enable
404 // SROAStrictInbounds, this code should be factored cleanly into
405 // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
406 // by writing out the code here where we have tho underlying allocation
407 // size readily available.
408 APInt GEPOffset = Offset;
409 for (gep_type_iterator GTI = gep_type_begin(GEPI),
410 GTE = gep_type_end(GEPI);
412 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
416 // Handle a struct index, which adds its field offset to the pointer.
417 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
418 unsigned ElementIdx = OpC->getZExtValue();
419 const StructLayout *SL = DL.getStructLayout(STy);
421 APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
423 // For array or vector indices, scale the index by the size of the type.
424 APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
425 GEPOffset += Index * APInt(Offset.getBitWidth(),
426 DL.getTypeAllocSize(GTI.getIndexedType()));
429 // If this index has computed an intermediate pointer which is not
430 // inbounds, then the result of the GEP is a poison value and we can
431 // delete it and all uses.
432 if (GEPOffset.ugt(AllocSize))
433 return markAsDead(GEPI);
437 return Base::visitGetElementPtrInst(GEPI);
440 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
441 uint64_t Size, bool IsVolatile) {
442 // We allow splitting of loads and stores where the type is an integer type
443 // and cover the entire alloca. This prevents us from splitting over
445 // FIXME: In the great blue eventually, we should eagerly split all integer
446 // loads and stores, and then have a separate step that merges adjacent
447 // alloca partitions into a single partition suitable for integer widening.
448 // Or we should skip the merge step and rely on GVN and other passes to
449 // merge adjacent loads and stores that survive mem2reg.
451 Ty->isIntegerTy() && !IsVolatile && Offset == 0 && Size >= AllocSize;
453 insertUse(I, Offset, Size, IsSplittable);
456 void visitLoadInst(LoadInst &LI) {
457 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
458 "All simple FCA loads should have been pre-split");
461 return PI.setAborted(&LI);
463 uint64_t Size = DL.getTypeStoreSize(LI.getType());
464 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
467 void visitStoreInst(StoreInst &SI) {
468 Value *ValOp = SI.getValueOperand();
470 return PI.setEscapedAndAborted(&SI);
472 return PI.setAborted(&SI);
474 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
476 // If this memory access can be shown to *statically* extend outside the
477 // bounds of of the allocation, it's behavior is undefined, so simply
478 // ignore it. Note that this is more strict than the generic clamping
479 // behavior of insertUse. We also try to handle cases which might run the
481 // FIXME: We should instead consider the pointer to have escaped if this
482 // function is being instrumented for addressing bugs or race conditions.
483 if (Size > AllocSize || Offset.ugt(AllocSize - Size)) {
484 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
485 << " which extends past the end of the " << AllocSize
487 << " alloca: " << S.AI << "\n"
488 << " use: " << SI << "\n");
489 return markAsDead(SI);
492 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
493 "All simple FCA stores should have been pre-split");
494 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
498 void visitMemSetInst(MemSetInst &II) {
499 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
500 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
501 if ((Length && Length->getValue() == 0) ||
502 (IsOffsetKnown && Offset.uge(AllocSize)))
503 // Zero-length mem transfer intrinsics can be ignored entirely.
504 return markAsDead(II);
507 return PI.setAborted(&II);
509 insertUse(II, Offset,
510 Length ? Length->getLimitedValue()
511 : AllocSize - Offset.getLimitedValue(),
515 void visitMemTransferInst(MemTransferInst &II) {
516 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
517 if (Length && Length->getValue() == 0)
518 // Zero-length mem transfer intrinsics can be ignored entirely.
519 return markAsDead(II);
521 // Because we can visit these intrinsics twice, also check to see if the
522 // first time marked this instruction as dead. If so, skip it.
523 if (VisitedDeadInsts.count(&II))
527 return PI.setAborted(&II);
529 // This side of the transfer is completely out-of-bounds, and so we can
530 // nuke the entire transfer. However, we also need to nuke the other side
531 // if already added to our partitions.
532 // FIXME: Yet another place we really should bypass this when
533 // instrumenting for ASan.
534 if (Offset.uge(AllocSize)) {
535 SmallDenseMap<Instruction *, unsigned>::iterator MTPI = MemTransferSliceMap.find(&II);
536 if (MTPI != MemTransferSliceMap.end())
537 S.Slices[MTPI->second].kill();
538 return markAsDead(II);
541 uint64_t RawOffset = Offset.getLimitedValue();
542 uint64_t Size = Length ? Length->getLimitedValue()
543 : AllocSize - RawOffset;
545 // Check for the special case where the same exact value is used for both
547 if (*U == II.getRawDest() && *U == II.getRawSource()) {
548 // For non-volatile transfers this is a no-op.
549 if (!II.isVolatile())
550 return markAsDead(II);
552 return insertUse(II, Offset, Size, /*IsSplittable=*/false);
555 // If we have seen both source and destination for a mem transfer, then
556 // they both point to the same alloca.
558 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
559 llvm::tie(MTPI, Inserted) =
560 MemTransferSliceMap.insert(std::make_pair(&II, S.Slices.size()));
561 unsigned PrevIdx = MTPI->second;
563 Slice &PrevP = S.Slices[PrevIdx];
565 // Check if the begin offsets match and this is a non-volatile transfer.
566 // In that case, we can completely elide the transfer.
567 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
569 return markAsDead(II);
572 // Otherwise we have an offset transfer within the same alloca. We can't
574 PrevP.makeUnsplittable();
577 // Insert the use now that we've fixed up the splittable nature.
578 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
580 // Check that we ended up with a valid index in the map.
581 assert(S.Slices[PrevIdx].getUse()->getUser() == &II &&
582 "Map index doesn't point back to a slice with this user.");
585 // Disable SRoA for any intrinsics except for lifetime invariants.
586 // FIXME: What about debug intrinsics? This matches old behavior, but
587 // doesn't make sense.
588 void visitIntrinsicInst(IntrinsicInst &II) {
590 return PI.setAborted(&II);
592 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
593 II.getIntrinsicID() == Intrinsic::lifetime_end) {
594 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
595 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
596 Length->getLimitedValue());
597 insertUse(II, Offset, Size, true);
601 Base::visitIntrinsicInst(II);
604 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
605 // We consider any PHI or select that results in a direct load or store of
606 // the same offset to be a viable use for slicing purposes. These uses
607 // are considered unsplittable and the size is the maximum loaded or stored
609 SmallPtrSet<Instruction *, 4> Visited;
610 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
611 Visited.insert(Root);
612 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
613 // If there are no loads or stores, the access is dead. We mark that as
614 // a size zero access.
617 Instruction *I, *UsedI;
618 llvm::tie(UsedI, I) = Uses.pop_back_val();
620 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
621 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
624 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
625 Value *Op = SI->getOperand(0);
628 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
632 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
633 if (!GEP->hasAllZeroIndices())
635 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
636 !isa<SelectInst>(I)) {
640 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
642 if (Visited.insert(cast<Instruction>(*UI)))
643 Uses.push_back(std::make_pair(I, cast<Instruction>(*UI)));
644 } while (!Uses.empty());
649 void visitPHINode(PHINode &PN) {
651 return markAsDead(PN);
653 return PI.setAborted(&PN);
655 // See if we already have computed info on this node.
656 uint64_t &PHISize = PHIOrSelectSizes[&PN];
658 // This is a new PHI node, check for an unsafe use of the PHI node.
659 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&PN, PHISize))
660 return PI.setAborted(UnsafeI);
663 // For PHI and select operands outside the alloca, we can't nuke the entire
664 // phi or select -- the other side might still be relevant, so we special
665 // case them here and use a separate structure to track the operands
666 // themselves which should be replaced with undef.
667 // FIXME: This should instead be escaped in the event we're instrumenting
668 // for address sanitization.
669 if (Offset.uge(AllocSize)) {
670 S.DeadOperands.push_back(U);
674 insertUse(PN, Offset, PHISize);
677 void visitSelectInst(SelectInst &SI) {
679 return markAsDead(SI);
680 if (Value *Result = foldSelectInst(SI)) {
682 // If the result of the constant fold will be the pointer, recurse
683 // through the select as if we had RAUW'ed it.
686 // Otherwise the operand to the select is dead, and we can replace it
688 S.DeadOperands.push_back(U);
693 return PI.setAborted(&SI);
695 // See if we already have computed info on this node.
696 uint64_t &SelectSize = PHIOrSelectSizes[&SI];
698 // This is a new Select, check for an unsafe use of it.
699 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&SI, SelectSize))
700 return PI.setAborted(UnsafeI);
703 // For PHI and select operands outside the alloca, we can't nuke the entire
704 // phi or select -- the other side might still be relevant, so we special
705 // case them here and use a separate structure to track the operands
706 // themselves which should be replaced with undef.
707 // FIXME: This should instead be escaped in the event we're instrumenting
708 // for address sanitization.
709 if (Offset.uge(AllocSize)) {
710 S.DeadOperands.push_back(U);
714 insertUse(SI, Offset, SelectSize);
717 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
718 void visitInstruction(Instruction &I) {
723 AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
725 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
728 PointerEscapingInstr(0) {
729 SliceBuilder PB(DL, AI, *this);
730 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
731 if (PtrI.isEscaped() || PtrI.isAborted()) {
732 // FIXME: We should sink the escape vs. abort info into the caller nicely,
733 // possibly by just storing the PtrInfo in the AllocaSlices.
734 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
735 : PtrI.getAbortingInst();
736 assert(PointerEscapingInstr && "Did not track a bad instruction");
740 Slices.erase(std::remove_if(Slices.begin(), Slices.end(),
741 std::mem_fun_ref(&Slice::isDead)),
744 #if __cplusplus >= 201103L && !defined(NDEBUG)
745 if (SROARandomShuffleSlices) {
746 std::mt19937 MT(static_cast<unsigned>(sys::TimeValue::now().msec()));
747 std::shuffle(Slices.begin(), Slices.end(), MT);
751 // Sort the uses. This arranges for the offsets to be in ascending order,
752 // and the sizes to be in descending order.
753 std::sort(Slices.begin(), Slices.end());
756 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
758 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
759 StringRef Indent) const {
760 printSlice(OS, I, Indent);
761 printUse(OS, I, Indent);
764 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
765 StringRef Indent) const {
766 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
767 << " slice #" << (I - begin())
768 << (I->isSplittable() ? " (splittable)" : "") << "\n";
771 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
772 StringRef Indent) const {
773 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
776 void AllocaSlices::print(raw_ostream &OS) const {
777 if (PointerEscapingInstr) {
778 OS << "Can't analyze slices for alloca: " << AI << "\n"
779 << " A pointer to this alloca escaped by:\n"
780 << " " << *PointerEscapingInstr << "\n";
784 OS << "Slices of alloca: " << AI << "\n";
785 for (const_iterator I = begin(), E = end(); I != E; ++I)
789 LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
792 LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
794 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
797 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
799 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
800 /// the loads and stores of an alloca instruction, as well as updating its
801 /// debug information. This is used when a domtree is unavailable and thus
802 /// mem2reg in its full form can't be used to handle promotion of allocas to
804 class AllocaPromoter : public LoadAndStorePromoter {
808 SmallVector<DbgDeclareInst *, 4> DDIs;
809 SmallVector<DbgValueInst *, 4> DVIs;
812 AllocaPromoter(const SmallVectorImpl<Instruction *> &Insts, SSAUpdater &S,
813 AllocaInst &AI, DIBuilder &DIB)
814 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
816 void run(const SmallVectorImpl<Instruction*> &Insts) {
817 // Retain the debug information attached to the alloca for use when
818 // rewriting loads and stores.
819 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
820 for (Value::use_iterator UI = DebugNode->use_begin(),
821 UE = DebugNode->use_end();
823 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
825 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
829 LoadAndStorePromoter::run(Insts);
831 // While we have the debug information, clear it off of the alloca. The
832 // caller takes care of deleting the alloca.
833 while (!DDIs.empty())
834 DDIs.pop_back_val()->eraseFromParent();
835 while (!DVIs.empty())
836 DVIs.pop_back_val()->eraseFromParent();
839 virtual bool isInstInList(Instruction *I,
840 const SmallVectorImpl<Instruction*> &Insts) const {
842 if (LoadInst *LI = dyn_cast<LoadInst>(I))
843 Ptr = LI->getOperand(0);
845 Ptr = cast<StoreInst>(I)->getPointerOperand();
847 // Only used to detect cycles, which will be rare and quickly found as
848 // we're walking up a chain of defs rather than down through uses.
849 SmallPtrSet<Value *, 4> Visited;
855 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr))
856 Ptr = BCI->getOperand(0);
857 else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr))
858 Ptr = GEPI->getPointerOperand();
862 } while (Visited.insert(Ptr));
867 virtual void updateDebugInfo(Instruction *Inst) const {
868 for (SmallVectorImpl<DbgDeclareInst *>::const_iterator I = DDIs.begin(),
869 E = DDIs.end(); I != E; ++I) {
870 DbgDeclareInst *DDI = *I;
871 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
872 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
873 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
874 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
876 for (SmallVectorImpl<DbgValueInst *>::const_iterator I = DVIs.begin(),
877 E = DVIs.end(); I != E; ++I) {
878 DbgValueInst *DVI = *I;
880 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
881 // If an argument is zero extended then use argument directly. The ZExt
882 // may be zapped by an optimization pass in future.
883 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
884 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
885 else if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
886 Arg = dyn_cast<Argument>(SExt->getOperand(0));
888 Arg = SI->getValueOperand();
889 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
890 Arg = LI->getPointerOperand();
894 Instruction *DbgVal =
895 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
897 DbgVal->setDebugLoc(DVI->getDebugLoc());
901 } // end anon namespace
905 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
907 /// This pass takes allocations which can be completely analyzed (that is, they
908 /// don't escape) and tries to turn them into scalar SSA values. There are
909 /// a few steps to this process.
911 /// 1) It takes allocations of aggregates and analyzes the ways in which they
912 /// are used to try to split them into smaller allocations, ideally of
913 /// a single scalar data type. It will split up memcpy and memset accesses
914 /// as necessary and try to isolate individual scalar accesses.
915 /// 2) It will transform accesses into forms which are suitable for SSA value
916 /// promotion. This can be replacing a memset with a scalar store of an
917 /// integer value, or it can involve speculating operations on a PHI or
918 /// select to be a PHI or select of the results.
919 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
920 /// onto insert and extract operations on a vector value, and convert them to
921 /// this form. By doing so, it will enable promotion of vector aggregates to
922 /// SSA vector values.
923 class SROA : public FunctionPass {
924 const bool RequiresDomTree;
927 const DataLayout *DL;
930 /// \brief Worklist of alloca instructions to simplify.
932 /// Each alloca in the function is added to this. Each new alloca formed gets
933 /// added to it as well to recursively simplify unless that alloca can be
934 /// directly promoted. Finally, each time we rewrite a use of an alloca other
935 /// the one being actively rewritten, we add it back onto the list if not
936 /// already present to ensure it is re-visited.
937 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
939 /// \brief A collection of instructions to delete.
940 /// We try to batch deletions to simplify code and make things a bit more
942 SetVector<Instruction *, SmallVector<Instruction *, 8> > DeadInsts;
944 /// \brief Post-promotion worklist.
946 /// Sometimes we discover an alloca which has a high probability of becoming
947 /// viable for SROA after a round of promotion takes place. In those cases,
948 /// the alloca is enqueued here for re-processing.
950 /// Note that we have to be very careful to clear allocas out of this list in
951 /// the event they are deleted.
952 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist;
954 /// \brief A collection of alloca instructions we can directly promote.
955 std::vector<AllocaInst *> PromotableAllocas;
957 /// \brief A worklist of PHIs to speculate prior to promoting allocas.
959 /// All of these PHIs have been checked for the safety of speculation and by
960 /// being speculated will allow promoting allocas currently in the promotable
962 SetVector<PHINode *, SmallVector<PHINode *, 2> > SpeculatablePHIs;
964 /// \brief A worklist of select instructions to speculate prior to promoting
967 /// All of these select instructions have been checked for the safety of
968 /// speculation and by being speculated will allow promoting allocas
969 /// currently in the promotable queue.
970 SetVector<SelectInst *, SmallVector<SelectInst *, 2> > SpeculatableSelects;
973 SROA(bool RequiresDomTree = true)
974 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
976 initializeSROAPass(*PassRegistry::getPassRegistry());
978 bool runOnFunction(Function &F);
979 void getAnalysisUsage(AnalysisUsage &AU) const;
981 const char *getPassName() const { return "SROA"; }
985 friend class PHIOrSelectSpeculator;
986 friend class AllocaSliceRewriter;
988 bool rewritePartition(AllocaInst &AI, AllocaSlices &S,
989 AllocaSlices::iterator B, AllocaSlices::iterator E,
990 int64_t BeginOffset, int64_t EndOffset,
991 ArrayRef<AllocaSlices::iterator> SplitUses);
992 bool splitAlloca(AllocaInst &AI, AllocaSlices &S);
993 bool runOnAlloca(AllocaInst &AI);
994 void clobberUse(Use &U);
995 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
996 bool promoteAllocas(Function &F);
1002 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
1003 return new SROA(RequiresDomTree);
1006 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
1008 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1009 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
1012 /// Walk the range of a partitioning looking for a common type to cover this
1013 /// sequence of slices.
1014 static Type *findCommonType(AllocaSlices::const_iterator B,
1015 AllocaSlices::const_iterator E,
1016 uint64_t EndOffset) {
1018 bool TyIsCommon = true;
1019 IntegerType *ITy = 0;
1021 // Note that we need to look at *every* alloca slice's Use to ensure we
1022 // always get consistent results regardless of the order of slices.
1023 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1024 Use *U = I->getUse();
1025 if (isa<IntrinsicInst>(*U->getUser()))
1027 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1031 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1032 UserTy = LI->getType();
1033 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1034 UserTy = SI->getValueOperand()->getType();
1037 if (!UserTy || (Ty && Ty != UserTy))
1038 TyIsCommon = false; // Give up on anything but an iN type.
1042 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1043 // If the type is larger than the partition, skip it. We only encounter
1044 // this for split integer operations where we want to use the type of the
1045 // entity causing the split. Also skip if the type is not a byte width
1047 if (UserITy->getBitWidth() % 8 != 0 ||
1048 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1051 // Track the largest bitwidth integer type used in this way in case there
1052 // is no common type.
1053 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1058 return TyIsCommon ? Ty : ITy;
1061 /// PHI instructions that use an alloca and are subsequently loaded can be
1062 /// rewritten to load both input pointers in the pred blocks and then PHI the
1063 /// results, allowing the load of the alloca to be promoted.
1065 /// %P2 = phi [i32* %Alloca, i32* %Other]
1066 /// %V = load i32* %P2
1068 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1070 /// %V2 = load i32* %Other
1072 /// %V = phi [i32 %V1, i32 %V2]
1074 /// We can do this to a select if its only uses are loads and if the operands
1075 /// to the select can be loaded unconditionally.
1077 /// FIXME: This should be hoisted into a generic utility, likely in
1078 /// Transforms/Util/Local.h
1079 static bool isSafePHIToSpeculate(PHINode &PN,
1080 const DataLayout *DL = 0) {
1081 // For now, we can only do this promotion if the load is in the same block
1082 // as the PHI, and if there are no stores between the phi and load.
1083 // TODO: Allow recursive phi users.
1084 // TODO: Allow stores.
1085 BasicBlock *BB = PN.getParent();
1086 unsigned MaxAlign = 0;
1087 bool HaveLoad = false;
1088 for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end(); UI != UE;
1090 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1091 if (LI == 0 || !LI->isSimple())
1094 // For now we only allow loads in the same block as the PHI. This is
1095 // a common case that happens when instcombine merges two loads through
1097 if (LI->getParent() != BB)
1100 // Ensure that there are no instructions between the PHI and the load that
1102 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1103 if (BBI->mayWriteToMemory())
1106 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1113 // We can only transform this if it is safe to push the loads into the
1114 // predecessor blocks. The only thing to watch out for is that we can't put
1115 // a possibly trapping load in the predecessor if it is a critical edge.
1116 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1117 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1118 Value *InVal = PN.getIncomingValue(Idx);
1120 // If the value is produced by the terminator of the predecessor (an
1121 // invoke) or it has side-effects, there is no valid place to put a load
1122 // in the predecessor.
1123 if (TI == InVal || TI->mayHaveSideEffects())
1126 // If the predecessor has a single successor, then the edge isn't
1128 if (TI->getNumSuccessors() == 1)
1131 // If this pointer is always safe to load, or if we can prove that there
1132 // is already a load in the block, then we can move the load to the pred
1134 if (InVal->isDereferenceablePointer() ||
1135 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, DL))
1144 static void speculatePHINodeLoads(PHINode &PN) {
1145 DEBUG(dbgs() << " original: " << PN << "\n");
1147 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1148 IRBuilderTy PHIBuilder(&PN);
1149 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1150 PN.getName() + ".sroa.speculated");
1152 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1153 // matter which one we get and if any differ.
1154 LoadInst *SomeLoad = cast<LoadInst>(*PN.use_begin());
1155 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1156 unsigned Align = SomeLoad->getAlignment();
1158 // Rewrite all loads of the PN to use the new PHI.
1159 while (!PN.use_empty()) {
1160 LoadInst *LI = cast<LoadInst>(*PN.use_begin());
1161 LI->replaceAllUsesWith(NewPN);
1162 LI->eraseFromParent();
1165 // Inject loads into all of the pred blocks.
1166 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1167 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1168 TerminatorInst *TI = Pred->getTerminator();
1169 Value *InVal = PN.getIncomingValue(Idx);
1170 IRBuilderTy PredBuilder(TI);
1172 LoadInst *Load = PredBuilder.CreateLoad(
1173 InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1174 ++NumLoadsSpeculated;
1175 Load->setAlignment(Align);
1177 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1178 NewPN->addIncoming(Load, Pred);
1181 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1182 PN.eraseFromParent();
1185 /// Select instructions that use an alloca and are subsequently loaded can be
1186 /// rewritten to load both input pointers and then select between the result,
1187 /// allowing the load of the alloca to be promoted.
1189 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1190 /// %V = load i32* %P2
1192 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1193 /// %V2 = load i32* %Other
1194 /// %V = select i1 %cond, i32 %V1, i32 %V2
1196 /// We can do this to a select if its only uses are loads and if the operand
1197 /// to the select can be loaded unconditionally.
1198 static bool isSafeSelectToSpeculate(SelectInst &SI, const DataLayout *DL = 0) {
1199 Value *TValue = SI.getTrueValue();
1200 Value *FValue = SI.getFalseValue();
1201 bool TDerefable = TValue->isDereferenceablePointer();
1202 bool FDerefable = FValue->isDereferenceablePointer();
1204 for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end(); UI != UE;
1206 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1207 if (LI == 0 || !LI->isSimple())
1210 // Both operands to the select need to be dereferencable, either
1211 // absolutely (e.g. allocas) or at this point because we can see other
1214 !isSafeToLoadUnconditionally(TValue, LI, LI->getAlignment(), DL))
1217 !isSafeToLoadUnconditionally(FValue, LI, LI->getAlignment(), DL))
1224 static void speculateSelectInstLoads(SelectInst &SI) {
1225 DEBUG(dbgs() << " original: " << SI << "\n");
1227 IRBuilderTy IRB(&SI);
1228 Value *TV = SI.getTrueValue();
1229 Value *FV = SI.getFalseValue();
1230 // Replace the loads of the select with a select of two loads.
1231 while (!SI.use_empty()) {
1232 LoadInst *LI = cast<LoadInst>(*SI.use_begin());
1233 assert(LI->isSimple() && "We only speculate simple loads");
1235 IRB.SetInsertPoint(LI);
1237 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1239 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1240 NumLoadsSpeculated += 2;
1242 // Transfer alignment and TBAA info if present.
1243 TL->setAlignment(LI->getAlignment());
1244 FL->setAlignment(LI->getAlignment());
1245 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1246 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
1247 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
1250 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1251 LI->getName() + ".sroa.speculated");
1253 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1254 LI->replaceAllUsesWith(V);
1255 LI->eraseFromParent();
1257 SI.eraseFromParent();
1260 /// \brief Build a GEP out of a base pointer and indices.
1262 /// This will return the BasePtr if that is valid, or build a new GEP
1263 /// instruction using the IRBuilder if GEP-ing is needed.
1264 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1265 SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1266 if (Indices.empty())
1269 // A single zero index is a no-op, so check for this and avoid building a GEP
1271 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1274 return IRB.CreateInBoundsGEP(BasePtr, Indices, NamePrefix + "sroa_idx");
1277 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1278 /// TargetTy without changing the offset of the pointer.
1280 /// This routine assumes we've already established a properly offset GEP with
1281 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1282 /// zero-indices down through type layers until we find one the same as
1283 /// TargetTy. If we can't find one with the same type, we at least try to use
1284 /// one with the same size. If none of that works, we just produce the GEP as
1285 /// indicated by Indices to have the correct offset.
1286 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1287 Value *BasePtr, Type *Ty, Type *TargetTy,
1288 SmallVectorImpl<Value *> &Indices,
1291 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1293 // See if we can descend into a struct and locate a field with the correct
1295 unsigned NumLayers = 0;
1296 Type *ElementTy = Ty;
1298 if (ElementTy->isPointerTy())
1300 if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
1301 ElementTy = SeqTy->getElementType();
1302 // Note that we use the default address space as this index is over an
1303 // array or a vector, not a pointer.
1304 Indices.push_back(IRB.getInt(APInt(DL.getPointerSizeInBits(0), 0)));
1305 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1306 if (STy->element_begin() == STy->element_end())
1307 break; // Nothing left to descend into.
1308 ElementTy = *STy->element_begin();
1309 Indices.push_back(IRB.getInt32(0));
1314 } while (ElementTy != TargetTy);
1315 if (ElementTy != TargetTy)
1316 Indices.erase(Indices.end() - NumLayers, Indices.end());
1318 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1321 /// \brief Recursively compute indices for a natural GEP.
1323 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1324 /// element types adding appropriate indices for the GEP.
1325 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1326 Value *Ptr, Type *Ty, APInt &Offset,
1328 SmallVectorImpl<Value *> &Indices,
1331 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices, NamePrefix);
1333 // We can't recurse through pointer types.
1334 if (Ty->isPointerTy())
1337 // We try to analyze GEPs over vectors here, but note that these GEPs are
1338 // extremely poorly defined currently. The long-term goal is to remove GEPing
1339 // over a vector from the IR completely.
1340 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1341 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1342 if (ElementSizeInBits % 8)
1343 return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
1344 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1345 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1346 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1348 Offset -= NumSkippedElements * ElementSize;
1349 Indices.push_back(IRB.getInt(NumSkippedElements));
1350 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1351 Offset, TargetTy, Indices, NamePrefix);
1354 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1355 Type *ElementTy = ArrTy->getElementType();
1356 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1357 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1358 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1361 Offset -= NumSkippedElements * ElementSize;
1362 Indices.push_back(IRB.getInt(NumSkippedElements));
1363 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1364 Indices, NamePrefix);
1367 StructType *STy = dyn_cast<StructType>(Ty);
1371 const StructLayout *SL = DL.getStructLayout(STy);
1372 uint64_t StructOffset = Offset.getZExtValue();
1373 if (StructOffset >= SL->getSizeInBytes())
1375 unsigned Index = SL->getElementContainingOffset(StructOffset);
1376 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1377 Type *ElementTy = STy->getElementType(Index);
1378 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1379 return 0; // The offset points into alignment padding.
1381 Indices.push_back(IRB.getInt32(Index));
1382 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1383 Indices, NamePrefix);
1386 /// \brief Get a natural GEP from a base pointer to a particular offset and
1387 /// resulting in a particular type.
1389 /// The goal is to produce a "natural" looking GEP that works with the existing
1390 /// composite types to arrive at the appropriate offset and element type for
1391 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1392 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1393 /// Indices, and setting Ty to the result subtype.
1395 /// If no natural GEP can be constructed, this function returns null.
1396 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1397 Value *Ptr, APInt Offset, Type *TargetTy,
1398 SmallVectorImpl<Value *> &Indices,
1400 PointerType *Ty = cast<PointerType>(Ptr->getType());
1402 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1404 if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
1407 Type *ElementTy = Ty->getElementType();
1408 if (!ElementTy->isSized())
1409 return 0; // We can't GEP through an unsized element.
1410 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1411 if (ElementSize == 0)
1412 return 0; // Zero-length arrays can't help us build a natural GEP.
1413 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1415 Offset -= NumSkippedElements * ElementSize;
1416 Indices.push_back(IRB.getInt(NumSkippedElements));
1417 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1418 Indices, NamePrefix);
1421 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1422 /// resulting pointer has PointerTy.
1424 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1425 /// and produces the pointer type desired. Where it cannot, it will try to use
1426 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1427 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1428 /// bitcast to the type.
1430 /// The strategy for finding the more natural GEPs is to peel off layers of the
1431 /// pointer, walking back through bit casts and GEPs, searching for a base
1432 /// pointer from which we can compute a natural GEP with the desired
1433 /// properties. The algorithm tries to fold as many constant indices into
1434 /// a single GEP as possible, thus making each GEP more independent of the
1435 /// surrounding code.
1436 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1437 APInt Offset, Type *PointerTy,
1439 // Even though we don't look through PHI nodes, we could be called on an
1440 // instruction in an unreachable block, which may be on a cycle.
1441 SmallPtrSet<Value *, 4> Visited;
1442 Visited.insert(Ptr);
1443 SmallVector<Value *, 4> Indices;
1445 // We may end up computing an offset pointer that has the wrong type. If we
1446 // never are able to compute one directly that has the correct type, we'll
1447 // fall back to it, so keep it around here.
1448 Value *OffsetPtr = 0;
1450 // Remember any i8 pointer we come across to re-use if we need to do a raw
1453 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1455 Type *TargetTy = PointerTy->getPointerElementType();
1458 // First fold any existing GEPs into the offset.
1459 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1460 APInt GEPOffset(Offset.getBitWidth(), 0);
1461 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1463 Offset += GEPOffset;
1464 Ptr = GEP->getPointerOperand();
1465 if (!Visited.insert(Ptr))
1469 // See if we can perform a natural GEP here.
1471 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1472 Indices, NamePrefix)) {
1473 if (P->getType() == PointerTy) {
1474 // Zap any offset pointer that we ended up computing in previous rounds.
1475 if (OffsetPtr && OffsetPtr->use_empty())
1476 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1477 I->eraseFromParent();
1485 // Stash this pointer if we've found an i8*.
1486 if (Ptr->getType()->isIntegerTy(8)) {
1488 Int8PtrOffset = Offset;
1491 // Peel off a layer of the pointer and update the offset appropriately.
1492 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1493 Ptr = cast<Operator>(Ptr)->getOperand(0);
1494 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1495 if (GA->mayBeOverridden())
1497 Ptr = GA->getAliasee();
1501 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1502 } while (Visited.insert(Ptr));
1506 Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
1507 NamePrefix + "sroa_raw_cast");
1508 Int8PtrOffset = Offset;
1511 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1512 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1513 NamePrefix + "sroa_raw_idx");
1517 // On the off chance we were targeting i8*, guard the bitcast here.
1518 if (Ptr->getType() != PointerTy)
1519 Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
1524 /// \brief Test whether we can convert a value from the old to the new type.
1526 /// This predicate should be used to guard calls to convertValue in order to
1527 /// ensure that we only try to convert viable values. The strategy is that we
1528 /// will peel off single element struct and array wrappings to get to an
1529 /// underlying value, and convert that value.
1530 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1533 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1534 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1535 if (NewITy->getBitWidth() >= OldITy->getBitWidth())
1537 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1539 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1542 // We can convert pointers to integers and vice-versa. Same for vectors
1543 // of pointers and integers.
1544 OldTy = OldTy->getScalarType();
1545 NewTy = NewTy->getScalarType();
1546 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1547 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1549 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1557 /// \brief Generic routine to convert an SSA value to a value of a different
1560 /// This will try various different casting techniques, such as bitcasts,
1561 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1562 /// two types for viability with this routine.
1563 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1565 Type *OldTy = V->getType();
1566 assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1571 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1572 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1573 if (NewITy->getBitWidth() > OldITy->getBitWidth())
1574 return IRB.CreateZExt(V, NewITy);
1576 // See if we need inttoptr for this type pair. A cast involving both scalars
1577 // and vectors requires and additional bitcast.
1578 if (OldTy->getScalarType()->isIntegerTy() &&
1579 NewTy->getScalarType()->isPointerTy()) {
1580 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1581 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1582 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1585 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1586 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1587 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1590 return IRB.CreateIntToPtr(V, NewTy);
1593 // See if we need ptrtoint for this type pair. A cast involving both scalars
1594 // and vectors requires and additional bitcast.
1595 if (OldTy->getScalarType()->isPointerTy() &&
1596 NewTy->getScalarType()->isIntegerTy()) {
1597 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1598 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1599 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1602 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1603 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1604 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1607 return IRB.CreatePtrToInt(V, NewTy);
1610 return IRB.CreateBitCast(V, NewTy);
1613 /// \brief Test whether the given slice use can be promoted to a vector.
1615 /// This function is called to test each entry in a partioning which is slated
1616 /// for a single slice.
1617 static bool isVectorPromotionViableForSlice(
1618 const DataLayout &DL, AllocaSlices &S, uint64_t SliceBeginOffset,
1619 uint64_t SliceEndOffset, VectorType *Ty, uint64_t ElementSize,
1620 AllocaSlices::const_iterator I) {
1621 // First validate the slice offsets.
1622 uint64_t BeginOffset =
1623 std::max(I->beginOffset(), SliceBeginOffset) - SliceBeginOffset;
1624 uint64_t BeginIndex = BeginOffset / ElementSize;
1625 if (BeginIndex * ElementSize != BeginOffset ||
1626 BeginIndex >= Ty->getNumElements())
1628 uint64_t EndOffset =
1629 std::min(I->endOffset(), SliceEndOffset) - SliceBeginOffset;
1630 uint64_t EndIndex = EndOffset / ElementSize;
1631 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1634 assert(EndIndex > BeginIndex && "Empty vector!");
1635 uint64_t NumElements = EndIndex - BeginIndex;
1637 (NumElements == 1) ? Ty->getElementType()
1638 : VectorType::get(Ty->getElementType(), NumElements);
1641 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1643 Use *U = I->getUse();
1645 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1646 if (MI->isVolatile())
1648 if (!I->isSplittable())
1649 return false; // Skip any unsplittable intrinsics.
1650 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1651 // Disable vector promotion when there are loads or stores of an FCA.
1653 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1654 if (LI->isVolatile())
1656 Type *LTy = LI->getType();
1657 if (SliceBeginOffset > I->beginOffset() ||
1658 SliceEndOffset < I->endOffset()) {
1659 assert(LTy->isIntegerTy());
1662 if (!canConvertValue(DL, SliceTy, LTy))
1664 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1665 if (SI->isVolatile())
1667 Type *STy = SI->getValueOperand()->getType();
1668 if (SliceBeginOffset > I->beginOffset() ||
1669 SliceEndOffset < I->endOffset()) {
1670 assert(STy->isIntegerTy());
1673 if (!canConvertValue(DL, STy, SliceTy))
1682 /// \brief Test whether the given alloca partitioning and range of slices can be
1683 /// promoted to a vector.
1685 /// This is a quick test to check whether we can rewrite a particular alloca
1686 /// partition (and its newly formed alloca) into a vector alloca with only
1687 /// whole-vector loads and stores such that it could be promoted to a vector
1688 /// SSA value. We only can ensure this for a limited set of operations, and we
1689 /// don't want to do the rewrites unless we are confident that the result will
1690 /// be promotable, so we have an early test here.
1692 isVectorPromotionViable(const DataLayout &DL, Type *AllocaTy, AllocaSlices &S,
1693 uint64_t SliceBeginOffset, uint64_t SliceEndOffset,
1694 AllocaSlices::const_iterator I,
1695 AllocaSlices::const_iterator E,
1696 ArrayRef<AllocaSlices::iterator> SplitUses) {
1697 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
1701 uint64_t ElementSize = DL.getTypeSizeInBits(Ty->getScalarType());
1703 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1704 // that aren't byte sized.
1705 if (ElementSize % 8)
1707 assert((DL.getTypeSizeInBits(Ty) % 8) == 0 &&
1708 "vector size not a multiple of element size?");
1712 if (!isVectorPromotionViableForSlice(DL, S, SliceBeginOffset,
1713 SliceEndOffset, Ty, ElementSize, I))
1716 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
1717 SUE = SplitUses.end();
1719 if (!isVectorPromotionViableForSlice(DL, S, SliceBeginOffset,
1720 SliceEndOffset, Ty, ElementSize, *SUI))
1726 /// \brief Test whether a slice of an alloca is valid for integer widening.
1728 /// This implements the necessary checking for the \c isIntegerWideningViable
1729 /// test below on a single slice of the alloca.
1730 static bool isIntegerWideningViableForSlice(const DataLayout &DL,
1732 uint64_t AllocBeginOffset,
1733 uint64_t Size, AllocaSlices &S,
1734 AllocaSlices::const_iterator I,
1735 bool &WholeAllocaOp) {
1736 uint64_t RelBegin = I->beginOffset() - AllocBeginOffset;
1737 uint64_t RelEnd = I->endOffset() - AllocBeginOffset;
1739 // We can't reasonably handle cases where the load or store extends past
1740 // the end of the aloca's type and into its padding.
1744 Use *U = I->getUse();
1746 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1747 if (LI->isVolatile())
1749 if (RelBegin == 0 && RelEnd == Size)
1750 WholeAllocaOp = true;
1751 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1752 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1754 } else if (RelBegin != 0 || RelEnd != Size ||
1755 !canConvertValue(DL, AllocaTy, LI->getType())) {
1756 // Non-integer loads need to be convertible from the alloca type so that
1757 // they are promotable.
1760 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1761 Type *ValueTy = SI->getValueOperand()->getType();
1762 if (SI->isVolatile())
1764 if (RelBegin == 0 && RelEnd == Size)
1765 WholeAllocaOp = true;
1766 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
1767 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1769 } else if (RelBegin != 0 || RelEnd != Size ||
1770 !canConvertValue(DL, ValueTy, AllocaTy)) {
1771 // Non-integer stores need to be convertible to the alloca type so that
1772 // they are promotable.
1775 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1776 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
1778 if (!I->isSplittable())
1779 return false; // Skip any unsplittable intrinsics.
1780 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1781 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1782 II->getIntrinsicID() != Intrinsic::lifetime_end)
1791 /// \brief Test whether the given alloca partition's integer operations can be
1792 /// widened to promotable ones.
1794 /// This is a quick test to check whether we can rewrite the integer loads and
1795 /// stores to a particular alloca into wider loads and stores and be able to
1796 /// promote the resulting alloca.
1798 isIntegerWideningViable(const DataLayout &DL, Type *AllocaTy,
1799 uint64_t AllocBeginOffset, AllocaSlices &S,
1800 AllocaSlices::const_iterator I,
1801 AllocaSlices::const_iterator E,
1802 ArrayRef<AllocaSlices::iterator> SplitUses) {
1803 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
1804 // Don't create integer types larger than the maximum bitwidth.
1805 if (SizeInBits > IntegerType::MAX_INT_BITS)
1808 // Don't try to handle allocas with bit-padding.
1809 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
1812 // We need to ensure that an integer type with the appropriate bitwidth can
1813 // be converted to the alloca type, whatever that is. We don't want to force
1814 // the alloca itself to have an integer type if there is a more suitable one.
1815 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
1816 if (!canConvertValue(DL, AllocaTy, IntTy) ||
1817 !canConvertValue(DL, IntTy, AllocaTy))
1820 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1822 // While examining uses, we ensure that the alloca has a covering load or
1823 // store. We don't want to widen the integer operations only to fail to
1824 // promote due to some other unsplittable entry (which we may make splittable
1825 // later). However, if there are only splittable uses, go ahead and assume
1826 // that we cover the alloca.
1827 bool WholeAllocaOp = (I != E) ? false : DL.isLegalInteger(SizeInBits);
1830 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1831 S, I, WholeAllocaOp))
1834 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
1835 SUE = SplitUses.end();
1837 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1838 S, *SUI, WholeAllocaOp))
1841 return WholeAllocaOp;
1844 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1845 IntegerType *Ty, uint64_t Offset,
1846 const Twine &Name) {
1847 DEBUG(dbgs() << " start: " << *V << "\n");
1848 IntegerType *IntTy = cast<IntegerType>(V->getType());
1849 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1850 "Element extends past full value");
1851 uint64_t ShAmt = 8*Offset;
1852 if (DL.isBigEndian())
1853 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1855 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
1856 DEBUG(dbgs() << " shifted: " << *V << "\n");
1858 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1859 "Cannot extract to a larger integer!");
1861 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
1862 DEBUG(dbgs() << " trunced: " << *V << "\n");
1867 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
1868 Value *V, uint64_t Offset, const Twine &Name) {
1869 IntegerType *IntTy = cast<IntegerType>(Old->getType());
1870 IntegerType *Ty = cast<IntegerType>(V->getType());
1871 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1872 "Cannot insert a larger integer!");
1873 DEBUG(dbgs() << " start: " << *V << "\n");
1875 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
1876 DEBUG(dbgs() << " extended: " << *V << "\n");
1878 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1879 "Element store outside of alloca store");
1880 uint64_t ShAmt = 8*Offset;
1881 if (DL.isBigEndian())
1882 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1884 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
1885 DEBUG(dbgs() << " shifted: " << *V << "\n");
1888 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
1889 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
1890 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
1891 DEBUG(dbgs() << " masked: " << *Old << "\n");
1892 V = IRB.CreateOr(Old, V, Name + ".insert");
1893 DEBUG(dbgs() << " inserted: " << *V << "\n");
1898 static Value *extractVector(IRBuilderTy &IRB, Value *V,
1899 unsigned BeginIndex, unsigned EndIndex,
1900 const Twine &Name) {
1901 VectorType *VecTy = cast<VectorType>(V->getType());
1902 unsigned NumElements = EndIndex - BeginIndex;
1903 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
1905 if (NumElements == VecTy->getNumElements())
1908 if (NumElements == 1) {
1909 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
1911 DEBUG(dbgs() << " extract: " << *V << "\n");
1915 SmallVector<Constant*, 8> Mask;
1916 Mask.reserve(NumElements);
1917 for (unsigned i = BeginIndex; i != EndIndex; ++i)
1918 Mask.push_back(IRB.getInt32(i));
1919 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1920 ConstantVector::get(Mask),
1922 DEBUG(dbgs() << " shuffle: " << *V << "\n");
1926 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
1927 unsigned BeginIndex, const Twine &Name) {
1928 VectorType *VecTy = cast<VectorType>(Old->getType());
1929 assert(VecTy && "Can only insert a vector into a vector");
1931 VectorType *Ty = dyn_cast<VectorType>(V->getType());
1933 // Single element to insert.
1934 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
1936 DEBUG(dbgs() << " insert: " << *V << "\n");
1940 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
1941 "Too many elements!");
1942 if (Ty->getNumElements() == VecTy->getNumElements()) {
1943 assert(V->getType() == VecTy && "Vector type mismatch");
1946 unsigned EndIndex = BeginIndex + Ty->getNumElements();
1948 // When inserting a smaller vector into the larger to store, we first
1949 // use a shuffle vector to widen it with undef elements, and then
1950 // a second shuffle vector to select between the loaded vector and the
1952 SmallVector<Constant*, 8> Mask;
1953 Mask.reserve(VecTy->getNumElements());
1954 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
1955 if (i >= BeginIndex && i < EndIndex)
1956 Mask.push_back(IRB.getInt32(i - BeginIndex));
1958 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
1959 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1960 ConstantVector::get(Mask),
1962 DEBUG(dbgs() << " shuffle: " << *V << "\n");
1965 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
1966 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
1968 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
1970 DEBUG(dbgs() << " blend: " << *V << "\n");
1975 /// \brief Visitor to rewrite instructions using p particular slice of an alloca
1976 /// to use a new alloca.
1978 /// Also implements the rewriting to vector-based accesses when the partition
1979 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
1981 class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
1982 // Befriend the base class so it can delegate to private visit methods.
1983 friend class llvm::InstVisitor<AllocaSliceRewriter, bool>;
1984 typedef llvm::InstVisitor<AllocaSliceRewriter, bool> Base;
1986 const DataLayout &DL;
1989 AllocaInst &OldAI, &NewAI;
1990 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
1993 // If we are rewriting an alloca partition which can be written as pure
1994 // vector operations, we stash extra information here. When VecTy is
1995 // non-null, we have some strict guarantees about the rewritten alloca:
1996 // - The new alloca is exactly the size of the vector type here.
1997 // - The accesses all either map to the entire vector or to a single
1999 // - The set of accessing instructions is only one of those handled above
2000 // in isVectorPromotionViable. Generally these are the same access kinds
2001 // which are promotable via mem2reg.
2004 uint64_t ElementSize;
2006 // This is a convenience and flag variable that will be null unless the new
2007 // alloca's integer operations should be widened to this integer type due to
2008 // passing isIntegerWideningViable above. If it is non-null, the desired
2009 // integer type will be stored here for easy access during rewriting.
2012 // The original offset of the slice currently being rewritten relative to
2013 // the original alloca.
2014 uint64_t BeginOffset, EndOffset;
2015 // The new offsets of the slice currently being rewritten relative to the
2017 uint64_t NewBeginOffset, NewEndOffset;
2023 Instruction *OldPtr;
2025 // Track post-rewrite users which are PHI nodes and Selects.
2026 SmallPtrSetImpl<PHINode *> &PHIUsers;
2027 SmallPtrSetImpl<SelectInst *> &SelectUsers;
2029 // Utility IR builder, whose name prefix is setup for each visited use, and
2030 // the insertion point is set to point to the user.
2034 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &S, SROA &Pass,
2035 AllocaInst &OldAI, AllocaInst &NewAI,
2036 uint64_t NewAllocaBeginOffset,
2037 uint64_t NewAllocaEndOffset, bool IsVectorPromotable,
2038 bool IsIntegerPromotable,
2039 SmallPtrSetImpl<PHINode *> &PHIUsers,
2040 SmallPtrSetImpl<SelectInst *> &SelectUsers)
2041 : DL(DL), S(S), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2042 NewAllocaBeginOffset(NewAllocaBeginOffset),
2043 NewAllocaEndOffset(NewAllocaEndOffset),
2044 NewAllocaTy(NewAI.getAllocatedType()),
2045 VecTy(IsVectorPromotable ? cast<VectorType>(NewAllocaTy) : 0),
2046 ElementTy(VecTy ? VecTy->getElementType() : 0),
2047 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
2048 IntTy(IsIntegerPromotable
2051 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2053 BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
2054 OldPtr(), PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2055 IRB(NewAI.getContext(), ConstantFolder()) {
2057 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
2058 "Only multiple-of-8 sized vector elements are viable");
2061 assert((!IsVectorPromotable && !IsIntegerPromotable) ||
2062 IsVectorPromotable != IsIntegerPromotable);
2065 bool visit(AllocaSlices::const_iterator I) {
2066 bool CanSROA = true;
2067 BeginOffset = I->beginOffset();
2068 EndOffset = I->endOffset();
2069 IsSplittable = I->isSplittable();
2071 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2073 // Compute the intersecting offset range.
2074 assert(BeginOffset < NewAllocaEndOffset);
2075 assert(EndOffset > NewAllocaBeginOffset);
2076 NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2077 NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2079 SliceSize = NewEndOffset - NewBeginOffset;
2081 OldUse = I->getUse();
2082 OldPtr = cast<Instruction>(OldUse->get());
2084 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2085 IRB.SetInsertPoint(OldUserI);
2086 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2087 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2089 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2096 // Make sure the other visit overloads are visible.
2099 // Every instruction which can end up as a user must have a rewrite rule.
2100 bool visitInstruction(Instruction &I) {
2101 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2102 llvm_unreachable("No rewrite rule for this instruction!");
2105 Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2106 // Note that the offset computation can use BeginOffset or NewBeginOffset
2107 // interchangeably for unsplit slices.
2108 assert(IsSplit || BeginOffset == NewBeginOffset);
2109 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2112 StringRef OldName = OldPtr->getName();
2113 // Skip through the last '.sroa.' component of the name.
2114 size_t LastSROAPrefix = OldName.rfind(".sroa.");
2115 if (LastSROAPrefix != StringRef::npos) {
2116 OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2117 // Look for an SROA slice index.
2118 size_t IndexEnd = OldName.find_first_not_of("0123456789");
2119 if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2120 // Strip the index and look for the offset.
2121 OldName = OldName.substr(IndexEnd + 1);
2122 size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2123 if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2124 // Strip the offset.
2125 OldName = OldName.substr(OffsetEnd + 1);
2128 // Strip any SROA suffixes as well.
2129 OldName = OldName.substr(0, OldName.find(".sroa_"));
2132 return getAdjustedPtr(IRB, DL, &NewAI,
2133 APInt(DL.getPointerSizeInBits(), Offset), PointerTy,
2135 Twine(OldName) + "."
2142 /// \brief Compute suitable alignment to access this slice of the *new* alloca.
2144 /// You can optionally pass a type to this routine and if that type's ABI
2145 /// alignment is itself suitable, this will return zero.
2146 unsigned getSliceAlign(Type *Ty = 0) {
2147 unsigned NewAIAlign = NewAI.getAlignment();
2149 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2150 unsigned Align = MinAlign(NewAIAlign, NewBeginOffset - NewAllocaBeginOffset);
2151 return (Ty && Align == DL.getABITypeAlignment(Ty)) ? 0 : Align;
2154 unsigned getIndex(uint64_t Offset) {
2155 assert(VecTy && "Can only call getIndex when rewriting a vector");
2156 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2157 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2158 uint32_t Index = RelOffset / ElementSize;
2159 assert(Index * ElementSize == RelOffset);
2163 void deleteIfTriviallyDead(Value *V) {
2164 Instruction *I = cast<Instruction>(V);
2165 if (isInstructionTriviallyDead(I))
2166 Pass.DeadInsts.insert(I);
2169 Value *rewriteVectorizedLoadInst() {
2170 unsigned BeginIndex = getIndex(NewBeginOffset);
2171 unsigned EndIndex = getIndex(NewEndOffset);
2172 assert(EndIndex > BeginIndex && "Empty vector!");
2174 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2176 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2179 Value *rewriteIntegerLoad(LoadInst &LI) {
2180 assert(IntTy && "We cannot insert an integer to the alloca");
2181 assert(!LI.isVolatile());
2182 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2184 V = convertValue(DL, IRB, V, IntTy);
2185 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2186 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2187 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset)
2188 V = extractInteger(DL, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2193 bool visitLoadInst(LoadInst &LI) {
2194 DEBUG(dbgs() << " original: " << LI << "\n");
2195 Value *OldOp = LI.getOperand(0);
2196 assert(OldOp == OldPtr);
2198 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2200 bool IsPtrAdjusted = false;
2203 V = rewriteVectorizedLoadInst();
2204 } else if (IntTy && LI.getType()->isIntegerTy()) {
2205 V = rewriteIntegerLoad(LI);
2206 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2207 canConvertValue(DL, NewAllocaTy, LI.getType())) {
2208 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2209 LI.isVolatile(), LI.getName());
2211 Type *LTy = TargetTy->getPointerTo();
2212 V = IRB.CreateAlignedLoad(getNewAllocaSlicePtr(IRB, LTy),
2213 getSliceAlign(TargetTy), LI.isVolatile(),
2215 IsPtrAdjusted = true;
2217 V = convertValue(DL, IRB, V, TargetTy);
2220 assert(!LI.isVolatile());
2221 assert(LI.getType()->isIntegerTy() &&
2222 "Only integer type loads and stores are split");
2223 assert(SliceSize < DL.getTypeStoreSize(LI.getType()) &&
2224 "Split load isn't smaller than original load");
2225 assert(LI.getType()->getIntegerBitWidth() ==
2226 DL.getTypeStoreSizeInBits(LI.getType()) &&
2227 "Non-byte-multiple bit width");
2228 // Move the insertion point just past the load so that we can refer to it.
2229 IRB.SetInsertPoint(llvm::next(BasicBlock::iterator(&LI)));
2230 // Create a placeholder value with the same type as LI to use as the
2231 // basis for the new value. This allows us to replace the uses of LI with
2232 // the computed value, and then replace the placeholder with LI, leaving
2233 // LI only used for this computation.
2235 = new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2236 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset,
2238 LI.replaceAllUsesWith(V);
2239 Placeholder->replaceAllUsesWith(&LI);
2242 LI.replaceAllUsesWith(V);
2245 Pass.DeadInsts.insert(&LI);
2246 deleteIfTriviallyDead(OldOp);
2247 DEBUG(dbgs() << " to: " << *V << "\n");
2248 return !LI.isVolatile() && !IsPtrAdjusted;
2251 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp) {
2252 if (V->getType() != VecTy) {
2253 unsigned BeginIndex = getIndex(NewBeginOffset);
2254 unsigned EndIndex = getIndex(NewEndOffset);
2255 assert(EndIndex > BeginIndex && "Empty vector!");
2256 unsigned NumElements = EndIndex - BeginIndex;
2257 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2259 (NumElements == 1) ? ElementTy
2260 : VectorType::get(ElementTy, NumElements);
2261 if (V->getType() != SliceTy)
2262 V = convertValue(DL, IRB, V, SliceTy);
2264 // Mix in the existing elements.
2265 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2267 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2269 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2270 Pass.DeadInsts.insert(&SI);
2273 DEBUG(dbgs() << " to: " << *Store << "\n");
2277 bool rewriteIntegerStore(Value *V, StoreInst &SI) {
2278 assert(IntTy && "We cannot extract an integer from the alloca");
2279 assert(!SI.isVolatile());
2280 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2281 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2283 Old = convertValue(DL, IRB, Old, IntTy);
2284 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2285 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2286 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset,
2289 V = convertValue(DL, IRB, V, NewAllocaTy);
2290 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2291 Pass.DeadInsts.insert(&SI);
2293 DEBUG(dbgs() << " to: " << *Store << "\n");
2297 bool visitStoreInst(StoreInst &SI) {
2298 DEBUG(dbgs() << " original: " << SI << "\n");
2299 Value *OldOp = SI.getOperand(1);
2300 assert(OldOp == OldPtr);
2302 Value *V = SI.getValueOperand();
2304 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2305 // alloca that should be re-examined after promoting this alloca.
2306 if (V->getType()->isPointerTy())
2307 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2308 Pass.PostPromotionWorklist.insert(AI);
2310 if (SliceSize < DL.getTypeStoreSize(V->getType())) {
2311 assert(!SI.isVolatile());
2312 assert(V->getType()->isIntegerTy() &&
2313 "Only integer type loads and stores are split");
2314 assert(V->getType()->getIntegerBitWidth() ==
2315 DL.getTypeStoreSizeInBits(V->getType()) &&
2316 "Non-byte-multiple bit width");
2317 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2318 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset,
2323 return rewriteVectorizedStoreInst(V, SI, OldOp);
2324 if (IntTy && V->getType()->isIntegerTy())
2325 return rewriteIntegerStore(V, SI);
2328 if (NewBeginOffset == NewAllocaBeginOffset &&
2329 NewEndOffset == NewAllocaEndOffset &&
2330 canConvertValue(DL, V->getType(), NewAllocaTy)) {
2331 V = convertValue(DL, IRB, V, NewAllocaTy);
2332 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2335 Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo());
2336 NewSI = IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(V->getType()),
2340 Pass.DeadInsts.insert(&SI);
2341 deleteIfTriviallyDead(OldOp);
2343 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2344 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2347 /// \brief Compute an integer value from splatting an i8 across the given
2348 /// number of bytes.
2350 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2351 /// call this routine.
2352 /// FIXME: Heed the advice above.
2354 /// \param V The i8 value to splat.
2355 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2356 Value *getIntegerSplat(Value *V, unsigned Size) {
2357 assert(Size > 0 && "Expected a positive number of bytes.");
2358 IntegerType *VTy = cast<IntegerType>(V->getType());
2359 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2363 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size*8);
2364 V = IRB.CreateMul(IRB.CreateZExt(V, SplatIntTy, "zext"),
2365 ConstantExpr::getUDiv(
2366 Constant::getAllOnesValue(SplatIntTy),
2367 ConstantExpr::getZExt(
2368 Constant::getAllOnesValue(V->getType()),
2374 /// \brief Compute a vector splat for a given element value.
2375 Value *getVectorSplat(Value *V, unsigned NumElements) {
2376 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2377 DEBUG(dbgs() << " splat: " << *V << "\n");
2381 bool visitMemSetInst(MemSetInst &II) {
2382 DEBUG(dbgs() << " original: " << II << "\n");
2383 assert(II.getRawDest() == OldPtr);
2385 // If the memset has a variable size, it cannot be split, just adjust the
2386 // pointer to the new alloca.
2387 if (!isa<Constant>(II.getLength())) {
2389 assert(NewBeginOffset == BeginOffset);
2390 II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2391 Type *CstTy = II.getAlignmentCst()->getType();
2392 II.setAlignment(ConstantInt::get(CstTy, getSliceAlign()));
2394 deleteIfTriviallyDead(OldPtr);
2398 // Record this instruction for deletion.
2399 Pass.DeadInsts.insert(&II);
2401 Type *AllocaTy = NewAI.getAllocatedType();
2402 Type *ScalarTy = AllocaTy->getScalarType();
2404 // If this doesn't map cleanly onto the alloca type, and that type isn't
2405 // a single value type, just emit a memset.
2406 if (!VecTy && !IntTy &&
2407 (BeginOffset > NewAllocaBeginOffset ||
2408 EndOffset < NewAllocaEndOffset ||
2409 !AllocaTy->isSingleValueType() ||
2410 !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2411 DL.getTypeSizeInBits(ScalarTy)%8 != 0)) {
2412 Type *SizeTy = II.getLength()->getType();
2413 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2414 CallInst *New = IRB.CreateMemSet(
2415 getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2416 getSliceAlign(), II.isVolatile());
2418 DEBUG(dbgs() << " to: " << *New << "\n");
2422 // If we can represent this as a simple value, we have to build the actual
2423 // value to store, which requires expanding the byte present in memset to
2424 // a sensible representation for the alloca type. This is essentially
2425 // splatting the byte to a sufficiently wide integer, splatting it across
2426 // any desired vector width, and bitcasting to the final type.
2430 // If this is a memset of a vectorized alloca, insert it.
2431 assert(ElementTy == ScalarTy);
2433 unsigned BeginIndex = getIndex(NewBeginOffset);
2434 unsigned EndIndex = getIndex(NewEndOffset);
2435 assert(EndIndex > BeginIndex && "Empty vector!");
2436 unsigned NumElements = EndIndex - BeginIndex;
2437 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2440 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2441 Splat = convertValue(DL, IRB, Splat, ElementTy);
2442 if (NumElements > 1)
2443 Splat = getVectorSplat(Splat, NumElements);
2445 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2447 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2449 // If this is a memset on an alloca where we can widen stores, insert the
2451 assert(!II.isVolatile());
2453 uint64_t Size = NewEndOffset - NewBeginOffset;
2454 V = getIntegerSplat(II.getValue(), Size);
2456 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2457 EndOffset != NewAllocaBeginOffset)) {
2458 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2460 Old = convertValue(DL, IRB, Old, IntTy);
2461 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2462 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2464 assert(V->getType() == IntTy &&
2465 "Wrong type for an alloca wide integer!");
2467 V = convertValue(DL, IRB, V, AllocaTy);
2469 // Established these invariants above.
2470 assert(NewBeginOffset == NewAllocaBeginOffset);
2471 assert(NewEndOffset == NewAllocaEndOffset);
2473 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2474 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2475 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2477 V = convertValue(DL, IRB, V, AllocaTy);
2480 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2483 DEBUG(dbgs() << " to: " << *New << "\n");
2484 return !II.isVolatile();
2487 bool visitMemTransferInst(MemTransferInst &II) {
2488 // Rewriting of memory transfer instructions can be a bit tricky. We break
2489 // them into two categories: split intrinsics and unsplit intrinsics.
2491 DEBUG(dbgs() << " original: " << II << "\n");
2493 bool IsDest = &II.getRawDestUse() == OldUse;
2494 assert((IsDest && II.getRawDest() == OldPtr) ||
2495 (!IsDest && II.getRawSource() == OldPtr));
2497 // Compute the relative offset within the transfer.
2498 unsigned IntPtrWidth = DL.getPointerSizeInBits();
2499 APInt RelOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2501 unsigned Align = II.getAlignment();
2503 Align = MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(),
2504 MinAlign(II.getAlignment(), getSliceAlign()));
2506 // For unsplit intrinsics, we simply modify the source and destination
2507 // pointers in place. This isn't just an optimization, it is a matter of
2508 // correctness. With unsplit intrinsics we may be dealing with transfers
2509 // within a single alloca before SROA ran, or with transfers that have
2510 // a variable length. We may also be dealing with memmove instead of
2511 // memcpy, and so simply updating the pointers is the necessary for us to
2512 // update both source and dest of a single call.
2513 if (!IsSplittable) {
2514 Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2516 II.setDest(AdjustedPtr);
2518 II.setSource(AdjustedPtr);
2520 Type *CstTy = II.getAlignmentCst()->getType();
2521 II.setAlignment(ConstantInt::get(CstTy, Align));
2523 DEBUG(dbgs() << " to: " << II << "\n");
2524 deleteIfTriviallyDead(OldPtr);
2527 // For split transfer intrinsics we have an incredibly useful assurance:
2528 // the source and destination do not reside within the same alloca, and at
2529 // least one of them does not escape. This means that we can replace
2530 // memmove with memcpy, and we don't need to worry about all manner of
2531 // downsides to splitting and transforming the operations.
2533 // If this doesn't map cleanly onto the alloca type, and that type isn't
2534 // a single value type, just emit a memcpy.
2536 = !VecTy && !IntTy && (BeginOffset > NewAllocaBeginOffset ||
2537 EndOffset < NewAllocaEndOffset ||
2538 !NewAI.getAllocatedType()->isSingleValueType());
2540 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2541 // size hasn't been shrunk based on analysis of the viable range, this is
2543 if (EmitMemCpy && &OldAI == &NewAI) {
2544 // Ensure the start lines up.
2545 assert(NewBeginOffset == BeginOffset);
2547 // Rewrite the size as needed.
2548 if (NewEndOffset != EndOffset)
2549 II.setLength(ConstantInt::get(II.getLength()->getType(),
2550 NewEndOffset - NewBeginOffset));
2553 // Record this instruction for deletion.
2554 Pass.DeadInsts.insert(&II);
2556 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2557 // alloca that should be re-examined after rewriting this instruction.
2558 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2560 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2561 assert(AI != &OldAI && AI != &NewAI &&
2562 "Splittable transfers cannot reach the same alloca on both ends.");
2563 Pass.Worklist.insert(AI);
2567 Type *OtherPtrTy = OtherPtr->getType();
2569 // Compute the other pointer, folding as much as possible to produce
2570 // a single, simple GEP in most cases.
2571 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, RelOffset, OtherPtrTy,
2572 OtherPtr->getName() + ".");
2574 Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2575 Type *SizeTy = II.getLength()->getType();
2576 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2578 CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
2579 IsDest ? OtherPtr : OurPtr,
2580 Size, Align, II.isVolatile());
2582 DEBUG(dbgs() << " to: " << *New << "\n");
2586 // Note that we clamp the alignment to 1 here as a 0 alignment for a memcpy
2587 // is equivalent to 1, but that isn't true if we end up rewriting this as
2592 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2593 NewEndOffset == NewAllocaEndOffset;
2594 uint64_t Size = NewEndOffset - NewBeginOffset;
2595 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2596 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2597 unsigned NumElements = EndIndex - BeginIndex;
2598 IntegerType *SubIntTy
2599 = IntTy ? Type::getIntNTy(IntTy->getContext(), Size*8) : 0;
2601 Type *OtherPtrTy = NewAI.getType();
2602 if (VecTy && !IsWholeAlloca) {
2603 if (NumElements == 1)
2604 OtherPtrTy = VecTy->getElementType();
2606 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2608 OtherPtrTy = OtherPtrTy->getPointerTo();
2609 } else if (IntTy && !IsWholeAlloca) {
2610 OtherPtrTy = SubIntTy->getPointerTo();
2613 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, RelOffset, OtherPtrTy,
2614 OtherPtr->getName() + ".");
2615 Value *DstPtr = &NewAI;
2617 std::swap(SrcPtr, DstPtr);
2620 if (VecTy && !IsWholeAlloca && !IsDest) {
2621 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2623 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2624 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2625 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2627 Src = convertValue(DL, IRB, Src, IntTy);
2628 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2629 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2631 Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(),
2635 if (VecTy && !IsWholeAlloca && IsDest) {
2636 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2638 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2639 } else if (IntTy && !IsWholeAlloca && IsDest) {
2640 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2642 Old = convertValue(DL, IRB, Old, IntTy);
2643 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2644 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
2645 Src = convertValue(DL, IRB, Src, NewAllocaTy);
2648 StoreInst *Store = cast<StoreInst>(
2649 IRB.CreateAlignedStore(Src, DstPtr, Align, II.isVolatile()));
2651 DEBUG(dbgs() << " to: " << *Store << "\n");
2652 return !II.isVolatile();
2655 bool visitIntrinsicInst(IntrinsicInst &II) {
2656 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2657 II.getIntrinsicID() == Intrinsic::lifetime_end);
2658 DEBUG(dbgs() << " original: " << II << "\n");
2659 assert(II.getArgOperand(1) == OldPtr);
2661 // Record this instruction for deletion.
2662 Pass.DeadInsts.insert(&II);
2665 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2666 NewEndOffset - NewBeginOffset);
2667 Value *Ptr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2669 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2670 New = IRB.CreateLifetimeStart(Ptr, Size);
2672 New = IRB.CreateLifetimeEnd(Ptr, Size);
2675 DEBUG(dbgs() << " to: " << *New << "\n");
2679 bool visitPHINode(PHINode &PN) {
2680 DEBUG(dbgs() << " original: " << PN << "\n");
2681 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
2682 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
2684 // We would like to compute a new pointer in only one place, but have it be
2685 // as local as possible to the PHI. To do that, we re-use the location of
2686 // the old pointer, which necessarily must be in the right position to
2687 // dominate the PHI.
2688 IRBuilderTy PtrBuilder(IRB);
2689 PtrBuilder.SetInsertPoint(OldPtr);
2690 PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
2692 Value *NewPtr = getNewAllocaSlicePtr(PtrBuilder, OldPtr->getType());
2693 // Replace the operands which were using the old pointer.
2694 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
2696 DEBUG(dbgs() << " to: " << PN << "\n");
2697 deleteIfTriviallyDead(OldPtr);
2699 // PHIs can't be promoted on their own, but often can be speculated. We
2700 // check the speculation outside of the rewriter so that we see the
2701 // fully-rewritten alloca.
2702 PHIUsers.insert(&PN);
2706 bool visitSelectInst(SelectInst &SI) {
2707 DEBUG(dbgs() << " original: " << SI << "\n");
2708 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
2709 "Pointer isn't an operand!");
2710 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
2711 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
2713 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2714 // Replace the operands which were using the old pointer.
2715 if (SI.getOperand(1) == OldPtr)
2716 SI.setOperand(1, NewPtr);
2717 if (SI.getOperand(2) == OldPtr)
2718 SI.setOperand(2, NewPtr);
2720 DEBUG(dbgs() << " to: " << SI << "\n");
2721 deleteIfTriviallyDead(OldPtr);
2723 // Selects can't be promoted on their own, but often can be speculated. We
2724 // check the speculation outside of the rewriter so that we see the
2725 // fully-rewritten alloca.
2726 SelectUsers.insert(&SI);
2734 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2736 /// This pass aggressively rewrites all aggregate loads and stores on
2737 /// a particular pointer (or any pointer derived from it which we can identify)
2738 /// with scalar loads and stores.
2739 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2740 // Befriend the base class so it can delegate to private visit methods.
2741 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2743 const DataLayout &DL;
2745 /// Queue of pointer uses to analyze and potentially rewrite.
2746 SmallVector<Use *, 8> Queue;
2748 /// Set to prevent us from cycling with phi nodes and loops.
2749 SmallPtrSet<User *, 8> Visited;
2751 /// The current pointer use being rewritten. This is used to dig up the used
2752 /// value (as opposed to the user).
2756 AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
2758 /// Rewrite loads and stores through a pointer and all pointers derived from
2760 bool rewrite(Instruction &I) {
2761 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2763 bool Changed = false;
2764 while (!Queue.empty()) {
2765 U = Queue.pop_back_val();
2766 Changed |= visit(cast<Instruction>(U->getUser()));
2772 /// Enqueue all the users of the given instruction for further processing.
2773 /// This uses a set to de-duplicate users.
2774 void enqueueUsers(Instruction &I) {
2775 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
2777 if (Visited.insert(*UI))
2778 Queue.push_back(&UI.getUse());
2781 // Conservative default is to not rewrite anything.
2782 bool visitInstruction(Instruction &I) { return false; }
2784 /// \brief Generic recursive split emission class.
2785 template <typename Derived>
2788 /// The builder used to form new instructions.
2790 /// The indices which to be used with insert- or extractvalue to select the
2791 /// appropriate value within the aggregate.
2792 SmallVector<unsigned, 4> Indices;
2793 /// The indices to a GEP instruction which will move Ptr to the correct slot
2794 /// within the aggregate.
2795 SmallVector<Value *, 4> GEPIndices;
2796 /// The base pointer of the original op, used as a base for GEPing the
2797 /// split operations.
2800 /// Initialize the splitter with an insertion point, Ptr and start with a
2801 /// single zero GEP index.
2802 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
2803 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
2806 /// \brief Generic recursive split emission routine.
2808 /// This method recursively splits an aggregate op (load or store) into
2809 /// scalar or vector ops. It splits recursively until it hits a single value
2810 /// and emits that single value operation via the template argument.
2812 /// The logic of this routine relies on GEPs and insertvalue and
2813 /// extractvalue all operating with the same fundamental index list, merely
2814 /// formatted differently (GEPs need actual values).
2816 /// \param Ty The type being split recursively into smaller ops.
2817 /// \param Agg The aggregate value being built up or stored, depending on
2818 /// whether this is splitting a load or a store respectively.
2819 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
2820 if (Ty->isSingleValueType())
2821 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
2823 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2824 unsigned OldSize = Indices.size();
2826 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
2828 assert(Indices.size() == OldSize && "Did not return to the old size");
2829 Indices.push_back(Idx);
2830 GEPIndices.push_back(IRB.getInt32(Idx));
2831 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
2832 GEPIndices.pop_back();
2838 if (StructType *STy = dyn_cast<StructType>(Ty)) {
2839 unsigned OldSize = Indices.size();
2841 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
2843 assert(Indices.size() == OldSize && "Did not return to the old size");
2844 Indices.push_back(Idx);
2845 GEPIndices.push_back(IRB.getInt32(Idx));
2846 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
2847 GEPIndices.pop_back();
2853 llvm_unreachable("Only arrays and structs are aggregate loadable types");
2857 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
2858 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2859 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
2861 /// Emit a leaf load of a single value. This is called at the leaves of the
2862 /// recursive emission to actually load values.
2863 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2864 assert(Ty->isSingleValueType());
2865 // Load the single value and insert it using the indices.
2866 Value *GEP = IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep");
2867 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
2868 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
2869 DEBUG(dbgs() << " to: " << *Load << "\n");
2873 bool visitLoadInst(LoadInst &LI) {
2874 assert(LI.getPointerOperand() == *U);
2875 if (!LI.isSimple() || LI.getType()->isSingleValueType())
2878 // We have an aggregate being loaded, split it apart.
2879 DEBUG(dbgs() << " original: " << LI << "\n");
2880 LoadOpSplitter Splitter(&LI, *U);
2881 Value *V = UndefValue::get(LI.getType());
2882 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
2883 LI.replaceAllUsesWith(V);
2884 LI.eraseFromParent();
2888 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
2889 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2890 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
2892 /// Emit a leaf store of a single value. This is called at the leaves of the
2893 /// recursive emission to actually produce stores.
2894 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2895 assert(Ty->isSingleValueType());
2896 // Extract the single value and store it using the indices.
2897 Value *Store = IRB.CreateStore(
2898 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
2899 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
2901 DEBUG(dbgs() << " to: " << *Store << "\n");
2905 bool visitStoreInst(StoreInst &SI) {
2906 if (!SI.isSimple() || SI.getPointerOperand() != *U)
2908 Value *V = SI.getValueOperand();
2909 if (V->getType()->isSingleValueType())
2912 // We have an aggregate being stored, split it apart.
2913 DEBUG(dbgs() << " original: " << SI << "\n");
2914 StoreOpSplitter Splitter(&SI, *U);
2915 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
2916 SI.eraseFromParent();
2920 bool visitBitCastInst(BitCastInst &BC) {
2925 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
2930 bool visitPHINode(PHINode &PN) {
2935 bool visitSelectInst(SelectInst &SI) {
2942 /// \brief Strip aggregate type wrapping.
2944 /// This removes no-op aggregate types wrapping an underlying type. It will
2945 /// strip as many layers of types as it can without changing either the type
2946 /// size or the allocated size.
2947 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
2948 if (Ty->isSingleValueType())
2951 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
2952 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
2955 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
2956 InnerTy = ArrTy->getElementType();
2957 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
2958 const StructLayout *SL = DL.getStructLayout(STy);
2959 unsigned Index = SL->getElementContainingOffset(0);
2960 InnerTy = STy->getElementType(Index);
2965 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
2966 TypeSize > DL.getTypeSizeInBits(InnerTy))
2969 return stripAggregateTypeWrapping(DL, InnerTy);
2972 /// \brief Try to find a partition of the aggregate type passed in for a given
2973 /// offset and size.
2975 /// This recurses through the aggregate type and tries to compute a subtype
2976 /// based on the offset and size. When the offset and size span a sub-section
2977 /// of an array, it will even compute a new array type for that sub-section,
2978 /// and the same for structs.
2980 /// Note that this routine is very strict and tries to find a partition of the
2981 /// type which produces the *exact* right offset and size. It is not forgiving
2982 /// when the size or offset cause either end of type-based partition to be off.
2983 /// Also, this is a best-effort routine. It is reasonable to give up and not
2984 /// return a type if necessary.
2985 static Type *getTypePartition(const DataLayout &DL, Type *Ty,
2986 uint64_t Offset, uint64_t Size) {
2987 if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
2988 return stripAggregateTypeWrapping(DL, Ty);
2989 if (Offset > DL.getTypeAllocSize(Ty) ||
2990 (DL.getTypeAllocSize(Ty) - Offset) < Size)
2993 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
2994 // We can't partition pointers...
2995 if (SeqTy->isPointerTy())
2998 Type *ElementTy = SeqTy->getElementType();
2999 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3000 uint64_t NumSkippedElements = Offset / ElementSize;
3001 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) {
3002 if (NumSkippedElements >= ArrTy->getNumElements())
3004 } else if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) {
3005 if (NumSkippedElements >= VecTy->getNumElements())
3008 Offset -= NumSkippedElements * ElementSize;
3010 // First check if we need to recurse.
3011 if (Offset > 0 || Size < ElementSize) {
3012 // Bail if the partition ends in a different array element.
3013 if ((Offset + Size) > ElementSize)
3015 // Recurse through the element type trying to peel off offset bytes.
3016 return getTypePartition(DL, ElementTy, Offset, Size);
3018 assert(Offset == 0);
3020 if (Size == ElementSize)
3021 return stripAggregateTypeWrapping(DL, ElementTy);
3022 assert(Size > ElementSize);
3023 uint64_t NumElements = Size / ElementSize;
3024 if (NumElements * ElementSize != Size)
3026 return ArrayType::get(ElementTy, NumElements);
3029 StructType *STy = dyn_cast<StructType>(Ty);
3033 const StructLayout *SL = DL.getStructLayout(STy);
3034 if (Offset >= SL->getSizeInBytes())
3036 uint64_t EndOffset = Offset + Size;
3037 if (EndOffset > SL->getSizeInBytes())
3040 unsigned Index = SL->getElementContainingOffset(Offset);
3041 Offset -= SL->getElementOffset(Index);
3043 Type *ElementTy = STy->getElementType(Index);
3044 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3045 if (Offset >= ElementSize)
3046 return 0; // The offset points into alignment padding.
3048 // See if any partition must be contained by the element.
3049 if (Offset > 0 || Size < ElementSize) {
3050 if ((Offset + Size) > ElementSize)
3052 return getTypePartition(DL, ElementTy, Offset, Size);
3054 assert(Offset == 0);
3056 if (Size == ElementSize)
3057 return stripAggregateTypeWrapping(DL, ElementTy);
3059 StructType::element_iterator EI = STy->element_begin() + Index,
3060 EE = STy->element_end();
3061 if (EndOffset < SL->getSizeInBytes()) {
3062 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3063 if (Index == EndIndex)
3064 return 0; // Within a single element and its padding.
3066 // Don't try to form "natural" types if the elements don't line up with the
3068 // FIXME: We could potentially recurse down through the last element in the
3069 // sub-struct to find a natural end point.
3070 if (SL->getElementOffset(EndIndex) != EndOffset)
3073 assert(Index < EndIndex);
3074 EE = STy->element_begin() + EndIndex;
3077 // Try to build up a sub-structure.
3078 StructType *SubTy = StructType::get(STy->getContext(), makeArrayRef(EI, EE),
3080 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3081 if (Size != SubSL->getSizeInBytes())
3082 return 0; // The sub-struct doesn't have quite the size needed.
3087 /// \brief Rewrite an alloca partition's users.
3089 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3090 /// to rewrite uses of an alloca partition to be conducive for SSA value
3091 /// promotion. If the partition needs a new, more refined alloca, this will
3092 /// build that new alloca, preserving as much type information as possible, and
3093 /// rewrite the uses of the old alloca to point at the new one and have the
3094 /// appropriate new offsets. It also evaluates how successful the rewrite was
3095 /// at enabling promotion and if it was successful queues the alloca to be
3097 bool SROA::rewritePartition(AllocaInst &AI, AllocaSlices &S,
3098 AllocaSlices::iterator B, AllocaSlices::iterator E,
3099 int64_t BeginOffset, int64_t EndOffset,
3100 ArrayRef<AllocaSlices::iterator> SplitUses) {
3101 assert(BeginOffset < EndOffset);
3102 uint64_t SliceSize = EndOffset - BeginOffset;
3104 // Try to compute a friendly type for this partition of the alloca. This
3105 // won't always succeed, in which case we fall back to a legal integer type
3106 // or an i8 array of an appropriate size.
3108 if (Type *CommonUseTy = findCommonType(B, E, EndOffset))
3109 if (DL->getTypeAllocSize(CommonUseTy) >= SliceSize)
3110 SliceTy = CommonUseTy;
3112 if (Type *TypePartitionTy = getTypePartition(*DL, AI.getAllocatedType(),
3113 BeginOffset, SliceSize))
3114 SliceTy = TypePartitionTy;
3115 if ((!SliceTy || (SliceTy->isArrayTy() &&
3116 SliceTy->getArrayElementType()->isIntegerTy())) &&
3117 DL->isLegalInteger(SliceSize * 8))
3118 SliceTy = Type::getIntNTy(*C, SliceSize * 8);
3120 SliceTy = ArrayType::get(Type::getInt8Ty(*C), SliceSize);
3121 assert(DL->getTypeAllocSize(SliceTy) >= SliceSize);
3123 bool IsVectorPromotable = isVectorPromotionViable(
3124 *DL, SliceTy, S, BeginOffset, EndOffset, B, E, SplitUses);
3126 bool IsIntegerPromotable =
3127 !IsVectorPromotable &&
3128 isIntegerWideningViable(*DL, SliceTy, BeginOffset, S, B, E, SplitUses);
3130 // Check for the case where we're going to rewrite to a new alloca of the
3131 // exact same type as the original, and with the same access offsets. In that
3132 // case, re-use the existing alloca, but still run through the rewriter to
3133 // perform phi and select speculation.
3135 if (SliceTy == AI.getAllocatedType()) {
3136 assert(BeginOffset == 0 &&
3137 "Non-zero begin offset but same alloca type");
3139 // FIXME: We should be able to bail at this point with "nothing changed".
3140 // FIXME: We might want to defer PHI speculation until after here.
3142 unsigned Alignment = AI.getAlignment();
3144 // The minimum alignment which users can rely on when the explicit
3145 // alignment is omitted or zero is that required by the ABI for this
3147 Alignment = DL->getABITypeAlignment(AI.getAllocatedType());
3149 Alignment = MinAlign(Alignment, BeginOffset);
3150 // If we will get at least this much alignment from the type alone, leave
3151 // the alloca's alignment unconstrained.
3152 if (Alignment <= DL->getABITypeAlignment(SliceTy))
3154 NewAI = new AllocaInst(SliceTy, 0, Alignment,
3155 AI.getName() + ".sroa." + Twine(B - S.begin()), &AI);
3159 DEBUG(dbgs() << "Rewriting alloca partition "
3160 << "[" << BeginOffset << "," << EndOffset << ") to: " << *NewAI
3163 // Track the high watermark on the worklist as it is only relevant for
3164 // promoted allocas. We will reset it to this point if the alloca is not in
3165 // fact scheduled for promotion.
3166 unsigned PPWOldSize = PostPromotionWorklist.size();
3167 unsigned NumUses = 0;
3168 SmallPtrSet<PHINode *, 8> PHIUsers;
3169 SmallPtrSet<SelectInst *, 8> SelectUsers;
3171 AllocaSliceRewriter Rewriter(*DL, S, *this, AI, *NewAI, BeginOffset,
3172 EndOffset, IsVectorPromotable,
3173 IsIntegerPromotable, PHIUsers, SelectUsers);
3174 bool Promotable = true;
3175 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
3176 SUE = SplitUses.end();
3177 SUI != SUE; ++SUI) {
3178 DEBUG(dbgs() << " rewriting split ");
3179 DEBUG(S.printSlice(dbgs(), *SUI, ""));
3180 Promotable &= Rewriter.visit(*SUI);
3183 for (AllocaSlices::iterator I = B; I != E; ++I) {
3184 DEBUG(dbgs() << " rewriting ");
3185 DEBUG(S.printSlice(dbgs(), I, ""));
3186 Promotable &= Rewriter.visit(I);
3190 NumAllocaPartitionUses += NumUses;
3191 MaxUsesPerAllocaPartition =
3192 std::max<unsigned>(NumUses, MaxUsesPerAllocaPartition);
3194 // Now that we've processed all the slices in the new partition, check if any
3195 // PHIs or Selects would block promotion.
3196 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
3199 if (!isSafePHIToSpeculate(**I, DL)) {
3202 SelectUsers.clear();
3205 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
3206 E = SelectUsers.end();
3208 if (!isSafeSelectToSpeculate(**I, DL)) {
3211 SelectUsers.clear();
3216 if (PHIUsers.empty() && SelectUsers.empty()) {
3217 // Promote the alloca.
3218 PromotableAllocas.push_back(NewAI);
3220 // If we have either PHIs or Selects to speculate, add them to those
3221 // worklists and re-queue the new alloca so that we promote in on the
3223 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
3226 SpeculatablePHIs.insert(*I);
3227 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
3228 E = SelectUsers.end();
3230 SpeculatableSelects.insert(*I);
3231 Worklist.insert(NewAI);
3234 // If we can't promote the alloca, iterate on it to check for new
3235 // refinements exposed by splitting the current alloca. Don't iterate on an
3236 // alloca which didn't actually change and didn't get promoted.
3238 Worklist.insert(NewAI);
3240 // Drop any post-promotion work items if promotion didn't happen.
3241 while (PostPromotionWorklist.size() > PPWOldSize)
3242 PostPromotionWorklist.pop_back();
3249 struct IsSliceEndLessOrEqualTo {
3250 uint64_t UpperBound;
3252 IsSliceEndLessOrEqualTo(uint64_t UpperBound) : UpperBound(UpperBound) {}
3254 bool operator()(const AllocaSlices::iterator &I) {
3255 return I->endOffset() <= UpperBound;
3261 removeFinishedSplitUses(SmallVectorImpl<AllocaSlices::iterator> &SplitUses,
3262 uint64_t &MaxSplitUseEndOffset, uint64_t Offset) {
3263 if (Offset >= MaxSplitUseEndOffset) {
3265 MaxSplitUseEndOffset = 0;
3269 size_t SplitUsesOldSize = SplitUses.size();
3270 SplitUses.erase(std::remove_if(SplitUses.begin(), SplitUses.end(),
3271 IsSliceEndLessOrEqualTo(Offset)),
3273 if (SplitUsesOldSize == SplitUses.size())
3276 // Recompute the max. While this is linear, so is remove_if.
3277 MaxSplitUseEndOffset = 0;
3278 for (SmallVectorImpl<AllocaSlices::iterator>::iterator
3279 SUI = SplitUses.begin(),
3280 SUE = SplitUses.end();
3282 MaxSplitUseEndOffset = std::max((*SUI)->endOffset(), MaxSplitUseEndOffset);
3285 /// \brief Walks the slices of an alloca and form partitions based on them,
3286 /// rewriting each of their uses.
3287 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &S) {
3288 if (S.begin() == S.end())
3291 unsigned NumPartitions = 0;
3292 bool Changed = false;
3293 SmallVector<AllocaSlices::iterator, 4> SplitUses;
3294 uint64_t MaxSplitUseEndOffset = 0;
3296 uint64_t BeginOffset = S.begin()->beginOffset();
3298 for (AllocaSlices::iterator SI = S.begin(), SJ = llvm::next(SI), SE = S.end();
3299 SI != SE; SI = SJ) {
3300 uint64_t MaxEndOffset = SI->endOffset();
3302 if (!SI->isSplittable()) {
3303 // When we're forming an unsplittable region, it must always start at the
3304 // first slice and will extend through its end.
3305 assert(BeginOffset == SI->beginOffset());
3307 // Form a partition including all of the overlapping slices with this
3308 // unsplittable slice.
3309 while (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3310 if (!SJ->isSplittable())
3311 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3315 assert(SI->isSplittable()); // Established above.
3317 // Collect all of the overlapping splittable slices.
3318 while (SJ != SE && SJ->beginOffset() < MaxEndOffset &&
3319 SJ->isSplittable()) {
3320 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3324 // Back up MaxEndOffset and SJ if we ended the span early when
3325 // encountering an unsplittable slice.
3326 if (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3327 assert(!SJ->isSplittable());
3328 MaxEndOffset = SJ->beginOffset();
3332 // Check if we have managed to move the end offset forward yet. If so,
3333 // we'll have to rewrite uses and erase old split uses.
3334 if (BeginOffset < MaxEndOffset) {
3335 // Rewrite a sequence of overlapping slices.
3337 rewritePartition(AI, S, SI, SJ, BeginOffset, MaxEndOffset, SplitUses);
3340 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset, MaxEndOffset);
3343 // Accumulate all the splittable slices from the [SI,SJ) region which
3344 // overlap going forward.
3345 for (AllocaSlices::iterator SK = SI; SK != SJ; ++SK)
3346 if (SK->isSplittable() && SK->endOffset() > MaxEndOffset) {
3347 SplitUses.push_back(SK);
3348 MaxSplitUseEndOffset = std::max(SK->endOffset(), MaxSplitUseEndOffset);
3351 // If we're already at the end and we have no split uses, we're done.
3352 if (SJ == SE && SplitUses.empty())
3355 // If we have no split uses or no gap in offsets, we're ready to move to
3357 if (SplitUses.empty() || (SJ != SE && MaxEndOffset == SJ->beginOffset())) {
3358 BeginOffset = SJ->beginOffset();
3362 // Even if we have split slices, if the next slice is splittable and the
3363 // split slices reach it, we can simply set up the beginning offset of the
3364 // next iteration to bridge between them.
3365 if (SJ != SE && SJ->isSplittable() &&
3366 MaxSplitUseEndOffset > SJ->beginOffset()) {
3367 BeginOffset = MaxEndOffset;
3371 // Otherwise, we have a tail of split slices. Rewrite them with an empty
3373 uint64_t PostSplitEndOffset =
3374 SJ == SE ? MaxSplitUseEndOffset : SJ->beginOffset();
3376 Changed |= rewritePartition(AI, S, SJ, SJ, MaxEndOffset, PostSplitEndOffset,
3381 break; // Skip the rest, we don't need to do any cleanup.
3383 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset,
3384 PostSplitEndOffset);
3386 // Now just reset the begin offset for the next iteration.
3387 BeginOffset = SJ->beginOffset();
3390 NumAllocaPartitions += NumPartitions;
3391 MaxPartitionsPerAlloca =
3392 std::max<unsigned>(NumPartitions, MaxPartitionsPerAlloca);
3397 /// \brief Clobber a use with undef, deleting the used value if it becomes dead.
3398 void SROA::clobberUse(Use &U) {
3400 // Replace the use with an undef value.
3401 U = UndefValue::get(OldV->getType());
3403 // Check for this making an instruction dead. We have to garbage collect
3404 // all the dead instructions to ensure the uses of any alloca end up being
3406 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3407 if (isInstructionTriviallyDead(OldI)) {
3408 DeadInsts.insert(OldI);
3412 /// \brief Analyze an alloca for SROA.
3414 /// This analyzes the alloca to ensure we can reason about it, builds
3415 /// the slices of the alloca, and then hands it off to be split and
3416 /// rewritten as needed.
3417 bool SROA::runOnAlloca(AllocaInst &AI) {
3418 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3419 ++NumAllocasAnalyzed;
3421 // Special case dead allocas, as they're trivial.
3422 if (AI.use_empty()) {
3423 AI.eraseFromParent();
3427 // Skip alloca forms that this analysis can't handle.
3428 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3429 DL->getTypeAllocSize(AI.getAllocatedType()) == 0)
3432 bool Changed = false;
3434 // First, split any FCA loads and stores touching this alloca to promote
3435 // better splitting and promotion opportunities.
3436 AggLoadStoreRewriter AggRewriter(*DL);
3437 Changed |= AggRewriter.rewrite(AI);
3439 // Build the slices using a recursive instruction-visiting builder.
3440 AllocaSlices S(*DL, AI);
3441 DEBUG(S.print(dbgs()));
3445 // Delete all the dead users of this alloca before splitting and rewriting it.
3446 for (AllocaSlices::dead_user_iterator DI = S.dead_user_begin(),
3447 DE = S.dead_user_end();
3449 // Free up everything used by this instruction.
3450 for (User::op_iterator DOI = (*DI)->op_begin(), DOE = (*DI)->op_end();
3454 // Now replace the uses of this instruction.
3455 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3457 // And mark it for deletion.
3458 DeadInsts.insert(*DI);
3461 for (AllocaSlices::dead_op_iterator DO = S.dead_op_begin(),
3462 DE = S.dead_op_end();
3468 // No slices to split. Leave the dead alloca for a later pass to clean up.
3469 if (S.begin() == S.end())
3472 Changed |= splitAlloca(AI, S);
3474 DEBUG(dbgs() << " Speculating PHIs\n");
3475 while (!SpeculatablePHIs.empty())
3476 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
3478 DEBUG(dbgs() << " Speculating Selects\n");
3479 while (!SpeculatableSelects.empty())
3480 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
3485 /// \brief Delete the dead instructions accumulated in this run.
3487 /// Recursively deletes the dead instructions we've accumulated. This is done
3488 /// at the very end to maximize locality of the recursive delete and to
3489 /// minimize the problems of invalidated instruction pointers as such pointers
3490 /// are used heavily in the intermediate stages of the algorithm.
3492 /// We also record the alloca instructions deleted here so that they aren't
3493 /// subsequently handed to mem2reg to promote.
3494 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3495 while (!DeadInsts.empty()) {
3496 Instruction *I = DeadInsts.pop_back_val();
3497 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3499 I->replaceAllUsesWith(UndefValue::get(I->getType()));
3501 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
3502 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
3503 // Zero out the operand and see if it becomes trivially dead.
3505 if (isInstructionTriviallyDead(U))
3506 DeadInsts.insert(U);
3509 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3510 DeletedAllocas.insert(AI);
3513 I->eraseFromParent();
3517 static void enqueueUsersInWorklist(Instruction &I,
3518 SmallVectorImpl<Instruction *> &Worklist,
3519 SmallPtrSet<Instruction *, 8> &Visited) {
3520 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
3522 if (Visited.insert(cast<Instruction>(*UI)))
3523 Worklist.push_back(cast<Instruction>(*UI));
3526 /// \brief Promote the allocas, using the best available technique.
3528 /// This attempts to promote whatever allocas have been identified as viable in
3529 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3530 /// If there is a domtree available, we attempt to promote using the full power
3531 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3532 /// based on the SSAUpdater utilities. This function returns whether any
3533 /// promotion occurred.
3534 bool SROA::promoteAllocas(Function &F) {
3535 if (PromotableAllocas.empty())
3538 NumPromoted += PromotableAllocas.size();
3540 if (DT && !ForceSSAUpdater) {
3541 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3542 PromoteMemToReg(PromotableAllocas, *DT);
3543 PromotableAllocas.clear();
3547 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3549 DIBuilder DIB(*F.getParent());
3550 SmallVector<Instruction *, 64> Insts;
3552 // We need a worklist to walk the uses of each alloca.
3553 SmallVector<Instruction *, 8> Worklist;
3554 SmallPtrSet<Instruction *, 8> Visited;
3555 SmallVector<Instruction *, 32> DeadInsts;
3557 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3558 AllocaInst *AI = PromotableAllocas[Idx];
3563 enqueueUsersInWorklist(*AI, Worklist, Visited);
3565 while (!Worklist.empty()) {
3566 Instruction *I = Worklist.pop_back_val();
3568 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3569 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3570 // leading to them) here. Eventually it should use them to optimize the
3571 // scalar values produced.
3572 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3573 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3574 II->getIntrinsicID() == Intrinsic::lifetime_end);
3575 II->eraseFromParent();
3579 // Push the loads and stores we find onto the list. SROA will already
3580 // have validated that all loads and stores are viable candidates for
3582 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3583 assert(LI->getType() == AI->getAllocatedType());
3584 Insts.push_back(LI);
3587 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3588 assert(SI->getValueOperand()->getType() == AI->getAllocatedType());
3589 Insts.push_back(SI);
3593 // For everything else, we know that only no-op bitcasts and GEPs will
3594 // make it this far, just recurse through them and recall them for later
3596 DeadInsts.push_back(I);
3597 enqueueUsersInWorklist(*I, Worklist, Visited);
3599 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3600 while (!DeadInsts.empty())
3601 DeadInsts.pop_back_val()->eraseFromParent();
3602 AI->eraseFromParent();
3605 PromotableAllocas.clear();
3610 /// \brief A predicate to test whether an alloca belongs to a set.
3611 class IsAllocaInSet {
3612 typedef SmallPtrSet<AllocaInst *, 4> SetType;
3616 typedef AllocaInst *argument_type;
3618 IsAllocaInSet(const SetType &Set) : Set(Set) {}
3619 bool operator()(AllocaInst *AI) const { return Set.count(AI); }
3623 bool SROA::runOnFunction(Function &F) {
3624 if (skipOptnoneFunction(F))
3627 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3628 C = &F.getContext();
3629 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
3631 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3634 DL = &DLP->getDataLayout();
3635 DominatorTreeWrapperPass *DTWP =
3636 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
3637 DT = DTWP ? &DTWP->getDomTree() : 0;
3639 BasicBlock &EntryBB = F.getEntryBlock();
3640 for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end());
3642 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3643 Worklist.insert(AI);
3645 bool Changed = false;
3646 // A set of deleted alloca instruction pointers which should be removed from
3647 // the list of promotable allocas.
3648 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3651 while (!Worklist.empty()) {
3652 Changed |= runOnAlloca(*Worklist.pop_back_val());
3653 deleteDeadInstructions(DeletedAllocas);
3655 // Remove the deleted allocas from various lists so that we don't try to
3656 // continue processing them.
3657 if (!DeletedAllocas.empty()) {
3658 Worklist.remove_if(IsAllocaInSet(DeletedAllocas));
3659 PostPromotionWorklist.remove_if(IsAllocaInSet(DeletedAllocas));
3660 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3661 PromotableAllocas.end(),
3662 IsAllocaInSet(DeletedAllocas)),
3663 PromotableAllocas.end());
3664 DeletedAllocas.clear();
3668 Changed |= promoteAllocas(F);
3670 Worklist = PostPromotionWorklist;
3671 PostPromotionWorklist.clear();
3672 } while (!Worklist.empty());
3677 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3678 if (RequiresDomTree)
3679 AU.addRequired<DominatorTreeWrapperPass>();
3680 AU.setPreservesCFG();