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 #include "llvm/Transforms/Scalar/SROA.h"
27 #include "llvm/ADT/STLExtras.h"
28 #include "llvm/ADT/SmallVector.h"
29 #include "llvm/ADT/Statistic.h"
30 #include "llvm/Analysis/AssumptionCache.h"
31 #include "llvm/Analysis/GlobalsModRef.h"
32 #include "llvm/Analysis/Loads.h"
33 #include "llvm/Analysis/PtrUseVisitor.h"
34 #include "llvm/Analysis/ValueTracking.h"
35 #include "llvm/IR/Constants.h"
36 #include "llvm/IR/DIBuilder.h"
37 #include "llvm/IR/DataLayout.h"
38 #include "llvm/IR/DebugInfo.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/IRBuilder.h"
41 #include "llvm/IR/InstVisitor.h"
42 #include "llvm/IR/Instructions.h"
43 #include "llvm/IR/IntrinsicInst.h"
44 #include "llvm/IR/LLVMContext.h"
45 #include "llvm/IR/Operator.h"
46 #include "llvm/Pass.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Compiler.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/ErrorHandling.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/TimeValue.h"
53 #include "llvm/Support/raw_ostream.h"
54 #include "llvm/Transforms/Scalar.h"
55 #include "llvm/Transforms/Utils/Local.h"
56 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
58 #if __cplusplus >= 201103L && !defined(NDEBUG)
59 // We only use this for a debug check in C++11
64 using namespace llvm::sroa;
66 #define DEBUG_TYPE "sroa"
68 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
69 STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed");
70 STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions per alloca");
71 STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten");
72 STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition");
73 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
74 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
75 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
76 STATISTIC(NumDeleted, "Number of instructions deleted");
77 STATISTIC(NumVectorized, "Number of vectorized aggregates");
79 /// Hidden option to enable randomly shuffling the slices to help uncover
80 /// instability in their order.
81 static cl::opt<bool> SROARandomShuffleSlices("sroa-random-shuffle-slices",
82 cl::init(false), cl::Hidden);
84 /// Hidden option to experiment with completely strict handling of inbounds
86 static cl::opt<bool> SROAStrictInbounds("sroa-strict-inbounds", cl::init(false),
90 /// \brief A custom IRBuilder inserter which prefixes all names if they are
92 template <bool preserveNames = true>
93 class IRBuilderPrefixedInserter
94 : public IRBuilderDefaultInserter<preserveNames> {
98 void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
101 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
102 BasicBlock::iterator InsertPt) const {
103 IRBuilderDefaultInserter<preserveNames>::InsertHelper(
104 I, Name.isTriviallyEmpty() ? Name : Prefix + Name, BB, InsertPt);
108 // Specialization for not preserving the name is trivial.
110 class IRBuilderPrefixedInserter<false>
111 : public IRBuilderDefaultInserter<false> {
113 void SetNamePrefix(const Twine &P) {}
116 /// \brief Provide a typedef for IRBuilder that drops names in release builds.
118 typedef llvm::IRBuilder<true, ConstantFolder, IRBuilderPrefixedInserter<true>>
121 typedef llvm::IRBuilder<false, ConstantFolder, IRBuilderPrefixedInserter<false>>
127 /// \brief A used slice of an alloca.
129 /// This structure represents a slice of an alloca used by some instruction. It
130 /// stores both the begin and end offsets of this use, a pointer to the use
131 /// itself, and a flag indicating whether we can classify the use as splittable
132 /// or not when forming partitions of the alloca.
134 /// \brief The beginning offset of the range.
135 uint64_t BeginOffset;
137 /// \brief The ending offset, not included in the range.
140 /// \brief Storage for both the use of this slice and whether it can be
142 PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
145 Slice() : BeginOffset(), EndOffset() {}
146 Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
147 : BeginOffset(BeginOffset), EndOffset(EndOffset),
148 UseAndIsSplittable(U, IsSplittable) {}
150 uint64_t beginOffset() const { return BeginOffset; }
151 uint64_t endOffset() const { return EndOffset; }
153 bool isSplittable() const { return UseAndIsSplittable.getInt(); }
154 void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
156 Use *getUse() const { return UseAndIsSplittable.getPointer(); }
158 bool isDead() const { return getUse() == nullptr; }
159 void kill() { UseAndIsSplittable.setPointer(nullptr); }
161 /// \brief Support for ordering ranges.
163 /// This provides an ordering over ranges such that start offsets are
164 /// always increasing, and within equal start offsets, the end offsets are
165 /// decreasing. Thus the spanning range comes first in a cluster with the
166 /// same start position.
167 bool operator<(const Slice &RHS) const {
168 if (beginOffset() < RHS.beginOffset())
170 if (beginOffset() > RHS.beginOffset())
172 if (isSplittable() != RHS.isSplittable())
173 return !isSplittable();
174 if (endOffset() > RHS.endOffset())
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> { static const bool value = true; };
202 /// \brief Representation of the alloca slices.
204 /// This class represents the slices of an alloca which are formed by its
205 /// various uses. If a pointer escapes, we can't fully build a representation
206 /// for the slices used and we reflect that in this structure. The uses are
207 /// stored, sorted by increasing beginning offset and with unsplittable slices
208 /// starting at a particular offset before splittable slices.
209 class llvm::sroa::AllocaSlices {
211 /// \brief Construct the slices of a particular alloca.
212 AllocaSlices(const DataLayout &DL, AllocaInst &AI);
214 /// \brief Test whether a pointer to the allocation escapes our analysis.
216 /// If this is true, the slices are never fully built and should be
218 bool isEscaped() const { return PointerEscapingInstr; }
220 /// \brief Support for iterating over the slices.
222 typedef SmallVectorImpl<Slice>::iterator iterator;
223 typedef iterator_range<iterator> range;
224 iterator begin() { return Slices.begin(); }
225 iterator end() { return Slices.end(); }
227 typedef SmallVectorImpl<Slice>::const_iterator const_iterator;
228 typedef iterator_range<const_iterator> const_range;
229 const_iterator begin() const { return Slices.begin(); }
230 const_iterator end() const { return Slices.end(); }
233 /// \brief Erase a range of slices.
234 void erase(iterator Start, iterator Stop) { Slices.erase(Start, Stop); }
236 /// \brief Insert new slices for this alloca.
238 /// This moves the slices into the alloca's slices collection, and re-sorts
239 /// everything so that the usual ordering properties of the alloca's slices
241 void insert(ArrayRef<Slice> NewSlices) {
242 int OldSize = Slices.size();
243 Slices.append(NewSlices.begin(), NewSlices.end());
244 auto SliceI = Slices.begin() + OldSize;
245 std::sort(SliceI, Slices.end());
246 std::inplace_merge(Slices.begin(), SliceI, Slices.end());
249 // Forward declare the iterator and range accessor for walking the
251 class partition_iterator;
252 iterator_range<partition_iterator> partitions();
254 /// \brief Access the dead users for this alloca.
255 ArrayRef<Instruction *> getDeadUsers() const { return DeadUsers; }
257 /// \brief Access the dead operands referring to this alloca.
259 /// These are operands which have cannot actually be used to refer to the
260 /// alloca as they are outside its range and the user doesn't correct for
261 /// that. These mostly consist of PHI node inputs and the like which we just
262 /// need to replace with undef.
263 ArrayRef<Use *> getDeadOperands() const { return DeadOperands; }
265 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
266 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
267 void printSlice(raw_ostream &OS, const_iterator I,
268 StringRef Indent = " ") const;
269 void printUse(raw_ostream &OS, const_iterator I,
270 StringRef Indent = " ") const;
271 void print(raw_ostream &OS) const;
272 void dump(const_iterator I) const;
277 template <typename DerivedT, typename RetT = void> class BuilderBase;
279 friend class AllocaSlices::SliceBuilder;
281 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
282 /// \brief Handle to alloca instruction to simplify method interfaces.
286 /// \brief The instruction responsible for this alloca not having a known set
289 /// When an instruction (potentially) escapes the pointer to the alloca, we
290 /// store a pointer to that here and abort trying to form slices of the
291 /// alloca. This will be null if the alloca slices are analyzed successfully.
292 Instruction *PointerEscapingInstr;
294 /// \brief The slices of the alloca.
296 /// We store a vector of the slices formed by uses of the alloca here. This
297 /// vector is sorted by increasing begin offset, and then the unsplittable
298 /// slices before the splittable ones. See the Slice inner class for more
300 SmallVector<Slice, 8> Slices;
302 /// \brief Instructions which will become dead if we rewrite the alloca.
304 /// Note that these are not separated by slice. This is because we expect an
305 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
306 /// all these instructions can simply be removed and replaced with undef as
307 /// they come from outside of the allocated space.
308 SmallVector<Instruction *, 8> DeadUsers;
310 /// \brief Operands which will become dead if we rewrite the alloca.
312 /// These are operands that in their particular use can be replaced with
313 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
314 /// to PHI nodes and the like. They aren't entirely dead (there might be
315 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
316 /// want to swap this particular input for undef to simplify the use lists of
318 SmallVector<Use *, 8> DeadOperands;
321 /// \brief A partition of the slices.
323 /// An ephemeral representation for a range of slices which can be viewed as
324 /// a partition of the alloca. This range represents a span of the alloca's
325 /// memory which cannot be split, and provides access to all of the slices
326 /// overlapping some part of the partition.
328 /// Objects of this type are produced by traversing the alloca's slices, but
329 /// are only ephemeral and not persistent.
330 class llvm::sroa::Partition {
332 friend class AllocaSlices;
333 friend class AllocaSlices::partition_iterator;
335 typedef AllocaSlices::iterator iterator;
337 /// \brief The beginning and ending offsets of the alloca for this
339 uint64_t BeginOffset, EndOffset;
341 /// \brief The start end end iterators of this partition.
344 /// \brief A collection of split slice tails overlapping the partition.
345 SmallVector<Slice *, 4> SplitTails;
347 /// \brief Raw constructor builds an empty partition starting and ending at
348 /// the given iterator.
349 Partition(iterator SI) : SI(SI), SJ(SI) {}
352 /// \brief The start offset of this partition.
354 /// All of the contained slices start at or after this offset.
355 uint64_t beginOffset() const { return BeginOffset; }
357 /// \brief The end offset of this partition.
359 /// All of the contained slices end at or before this offset.
360 uint64_t endOffset() const { return EndOffset; }
362 /// \brief The size of the partition.
364 /// Note that this can never be zero.
365 uint64_t size() const {
366 assert(BeginOffset < EndOffset && "Partitions must span some bytes!");
367 return EndOffset - BeginOffset;
370 /// \brief Test whether this partition contains no slices, and merely spans
371 /// a region occupied by split slices.
372 bool empty() const { return SI == SJ; }
374 /// \name Iterate slices that start within the partition.
375 /// These may be splittable or unsplittable. They have a begin offset >= the
376 /// partition begin offset.
378 // FIXME: We should probably define a "concat_iterator" helper and use that
379 // to stitch together pointee_iterators over the split tails and the
380 // contiguous iterators of the partition. That would give a much nicer
381 // interface here. We could then additionally expose filtered iterators for
382 // split, unsplit, and unsplittable splices based on the usage patterns.
383 iterator begin() const { return SI; }
384 iterator end() const { return SJ; }
387 /// \brief Get the sequence of split slice tails.
389 /// These tails are of slices which start before this partition but are
390 /// split and overlap into the partition. We accumulate these while forming
392 ArrayRef<Slice *> splitSliceTails() const { return SplitTails; }
395 /// \brief An iterator over partitions of the alloca's slices.
397 /// This iterator implements the core algorithm for partitioning the alloca's
398 /// slices. It is a forward iterator as we don't support backtracking for
399 /// efficiency reasons, and re-use a single storage area to maintain the
400 /// current set of split slices.
402 /// It is templated on the slice iterator type to use so that it can operate
403 /// with either const or non-const slice iterators.
404 class AllocaSlices::partition_iterator
405 : public iterator_facade_base<partition_iterator, std::forward_iterator_tag,
407 friend class AllocaSlices;
409 /// \brief Most of the state for walking the partitions is held in a class
410 /// with a nice interface for examining them.
413 /// \brief We need to keep the end of the slices to know when to stop.
414 AllocaSlices::iterator SE;
416 /// \brief We also need to keep track of the maximum split end offset seen.
417 /// FIXME: Do we really?
418 uint64_t MaxSplitSliceEndOffset;
420 /// \brief Sets the partition to be empty at given iterator, and sets the
422 partition_iterator(AllocaSlices::iterator SI, AllocaSlices::iterator SE)
423 : P(SI), SE(SE), MaxSplitSliceEndOffset(0) {
424 // If not already at the end, advance our state to form the initial
430 /// \brief Advance the iterator to the next partition.
432 /// Requires that the iterator not be at the end of the slices.
434 assert((P.SI != SE || !P.SplitTails.empty()) &&
435 "Cannot advance past the end of the slices!");
437 // Clear out any split uses which have ended.
438 if (!P.SplitTails.empty()) {
439 if (P.EndOffset >= MaxSplitSliceEndOffset) {
440 // If we've finished all splits, this is easy.
441 P.SplitTails.clear();
442 MaxSplitSliceEndOffset = 0;
444 // Remove the uses which have ended in the prior partition. This
445 // cannot change the max split slice end because we just checked that
446 // the prior partition ended prior to that max.
449 P.SplitTails.begin(), P.SplitTails.end(),
450 [&](Slice *S) { return S->endOffset() <= P.EndOffset; }),
452 assert(std::any_of(P.SplitTails.begin(), P.SplitTails.end(),
454 return S->endOffset() == MaxSplitSliceEndOffset;
456 "Could not find the current max split slice offset!");
457 assert(std::all_of(P.SplitTails.begin(), P.SplitTails.end(),
459 return S->endOffset() <= MaxSplitSliceEndOffset;
461 "Max split slice end offset is not actually the max!");
465 // If P.SI is already at the end, then we've cleared the split tail and
466 // now have an end iterator.
468 assert(P.SplitTails.empty() && "Failed to clear the split slices!");
472 // If we had a non-empty partition previously, set up the state for
473 // subsequent partitions.
475 // Accumulate all the splittable slices which started in the old
476 // partition into the split list.
478 if (S.isSplittable() && S.endOffset() > P.EndOffset) {
479 P.SplitTails.push_back(&S);
480 MaxSplitSliceEndOffset =
481 std::max(S.endOffset(), MaxSplitSliceEndOffset);
484 // Start from the end of the previous partition.
487 // If P.SI is now at the end, we at most have a tail of split slices.
489 P.BeginOffset = P.EndOffset;
490 P.EndOffset = MaxSplitSliceEndOffset;
494 // If the we have split slices and the next slice is after a gap and is
495 // not splittable immediately form an empty partition for the split
496 // slices up until the next slice begins.
497 if (!P.SplitTails.empty() && P.SI->beginOffset() != P.EndOffset &&
498 !P.SI->isSplittable()) {
499 P.BeginOffset = P.EndOffset;
500 P.EndOffset = P.SI->beginOffset();
505 // OK, we need to consume new slices. Set the end offset based on the
506 // current slice, and step SJ past it. The beginning offset of the
507 // partition is the beginning offset of the next slice unless we have
508 // pre-existing split slices that are continuing, in which case we begin
509 // at the prior end offset.
510 P.BeginOffset = P.SplitTails.empty() ? P.SI->beginOffset() : P.EndOffset;
511 P.EndOffset = P.SI->endOffset();
514 // There are two strategies to form a partition based on whether the
515 // partition starts with an unsplittable slice or a splittable slice.
516 if (!P.SI->isSplittable()) {
517 // When we're forming an unsplittable region, it must always start at
518 // the first slice and will extend through its end.
519 assert(P.BeginOffset == P.SI->beginOffset());
521 // Form a partition including all of the overlapping slices with this
522 // unsplittable slice.
523 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
524 if (!P.SJ->isSplittable())
525 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
529 // We have a partition across a set of overlapping unsplittable
534 // If we're starting with a splittable slice, then we need to form
535 // a synthetic partition spanning it and any other overlapping splittable
537 assert(P.SI->isSplittable() && "Forming a splittable partition!");
539 // Collect all of the overlapping splittable slices.
540 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset &&
541 P.SJ->isSplittable()) {
542 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
546 // Back upiP.EndOffset if we ended the span early when encountering an
547 // unsplittable slice. This synthesizes the early end offset of
548 // a partition spanning only splittable slices.
549 if (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
550 assert(!P.SJ->isSplittable());
551 P.EndOffset = P.SJ->beginOffset();
556 bool operator==(const partition_iterator &RHS) const {
557 assert(SE == RHS.SE &&
558 "End iterators don't match between compared partition iterators!");
560 // The observed positions of partitions is marked by the P.SI iterator and
561 // the emptiness of the split slices. The latter is only relevant when
562 // P.SI == SE, as the end iterator will additionally have an empty split
563 // slices list, but the prior may have the same P.SI and a tail of split
565 if (P.SI == RHS.P.SI && P.SplitTails.empty() == RHS.P.SplitTails.empty()) {
566 assert(P.SJ == RHS.P.SJ &&
567 "Same set of slices formed two different sized partitions!");
568 assert(P.SplitTails.size() == RHS.P.SplitTails.size() &&
569 "Same slice position with differently sized non-empty split "
576 partition_iterator &operator++() {
581 Partition &operator*() { return P; }
584 /// \brief A forward range over the partitions of the alloca's slices.
586 /// This accesses an iterator range over the partitions of the alloca's
587 /// slices. It computes these partitions on the fly based on the overlapping
588 /// offsets of the slices and the ability to split them. It will visit "empty"
589 /// partitions to cover regions of the alloca only accessed via split
591 iterator_range<AllocaSlices::partition_iterator> AllocaSlices::partitions() {
592 return make_range(partition_iterator(begin(), end()),
593 partition_iterator(end(), end()));
596 static Value *foldSelectInst(SelectInst &SI) {
597 // If the condition being selected on is a constant or the same value is
598 // being selected between, fold the select. Yes this does (rarely) happen
600 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
601 return SI.getOperand(1 + CI->isZero());
602 if (SI.getOperand(1) == SI.getOperand(2))
603 return SI.getOperand(1);
608 /// \brief A helper that folds a PHI node or a select.
609 static Value *foldPHINodeOrSelectInst(Instruction &I) {
610 if (PHINode *PN = dyn_cast<PHINode>(&I)) {
611 // If PN merges together the same value, return that value.
612 return PN->hasConstantValue();
614 return foldSelectInst(cast<SelectInst>(I));
617 /// \brief Builder for the alloca slices.
619 /// This class builds a set of alloca slices by recursively visiting the uses
620 /// of an alloca and making a slice for each load and store at each offset.
621 class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
622 friend class PtrUseVisitor<SliceBuilder>;
623 friend class InstVisitor<SliceBuilder>;
624 typedef PtrUseVisitor<SliceBuilder> Base;
626 const uint64_t AllocSize;
629 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
630 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
632 /// \brief Set to de-duplicate dead instructions found in the use walk.
633 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
636 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &AS)
637 : PtrUseVisitor<SliceBuilder>(DL),
638 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), AS(AS) {}
641 void markAsDead(Instruction &I) {
642 if (VisitedDeadInsts.insert(&I).second)
643 AS.DeadUsers.push_back(&I);
646 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
647 bool IsSplittable = false) {
648 // Completely skip uses which have a zero size or start either before or
649 // past the end of the allocation.
650 if (Size == 0 || Offset.uge(AllocSize)) {
651 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
652 << " which has zero size or starts outside of the "
653 << AllocSize << " byte alloca:\n"
654 << " alloca: " << AS.AI << "\n"
655 << " use: " << I << "\n");
656 return markAsDead(I);
659 uint64_t BeginOffset = Offset.getZExtValue();
660 uint64_t EndOffset = BeginOffset + Size;
662 // Clamp the end offset to the end of the allocation. Note that this is
663 // formulated to handle even the case where "BeginOffset + Size" overflows.
664 // This may appear superficially to be something we could ignore entirely,
665 // but that is not so! There may be widened loads or PHI-node uses where
666 // some instructions are dead but not others. We can't completely ignore
667 // them, and so have to record at least the information here.
668 assert(AllocSize >= BeginOffset); // Established above.
669 if (Size > AllocSize - BeginOffset) {
670 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
671 << " to remain within the " << AllocSize << " byte alloca:\n"
672 << " alloca: " << AS.AI << "\n"
673 << " use: " << I << "\n");
674 EndOffset = AllocSize;
677 AS.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
680 void visitBitCastInst(BitCastInst &BC) {
682 return markAsDead(BC);
684 return Base::visitBitCastInst(BC);
687 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
688 if (GEPI.use_empty())
689 return markAsDead(GEPI);
691 if (SROAStrictInbounds && GEPI.isInBounds()) {
692 // FIXME: This is a manually un-factored variant of the basic code inside
693 // of GEPs with checking of the inbounds invariant specified in the
694 // langref in a very strict sense. If we ever want to enable
695 // SROAStrictInbounds, this code should be factored cleanly into
696 // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
697 // by writing out the code here where we have tho underlying allocation
698 // size readily available.
699 APInt GEPOffset = Offset;
700 const DataLayout &DL = GEPI.getModule()->getDataLayout();
701 for (gep_type_iterator GTI = gep_type_begin(GEPI),
702 GTE = gep_type_end(GEPI);
704 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
708 // Handle a struct index, which adds its field offset to the pointer.
709 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
710 unsigned ElementIdx = OpC->getZExtValue();
711 const StructLayout *SL = DL.getStructLayout(STy);
713 APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
715 // For array or vector indices, scale the index by the size of the
717 APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
718 GEPOffset += Index * APInt(Offset.getBitWidth(),
719 DL.getTypeAllocSize(GTI.getIndexedType()));
722 // If this index has computed an intermediate pointer which is not
723 // inbounds, then the result of the GEP is a poison value and we can
724 // delete it and all uses.
725 if (GEPOffset.ugt(AllocSize))
726 return markAsDead(GEPI);
730 return Base::visitGetElementPtrInst(GEPI);
733 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
734 uint64_t Size, bool IsVolatile) {
735 // We allow splitting of non-volatile loads and stores where the type is an
736 // integer type. These may be used to implement 'memcpy' or other "transfer
737 // of bits" patterns.
738 bool IsSplittable = Ty->isIntegerTy() && !IsVolatile;
740 insertUse(I, Offset, Size, IsSplittable);
743 void visitLoadInst(LoadInst &LI) {
744 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
745 "All simple FCA loads should have been pre-split");
748 return PI.setAborted(&LI);
750 const DataLayout &DL = LI.getModule()->getDataLayout();
751 uint64_t Size = DL.getTypeStoreSize(LI.getType());
752 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
755 void visitStoreInst(StoreInst &SI) {
756 Value *ValOp = SI.getValueOperand();
758 return PI.setEscapedAndAborted(&SI);
760 return PI.setAborted(&SI);
762 const DataLayout &DL = SI.getModule()->getDataLayout();
763 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
765 // If this memory access can be shown to *statically* extend outside the
766 // bounds of of the allocation, it's behavior is undefined, so simply
767 // ignore it. Note that this is more strict than the generic clamping
768 // behavior of insertUse. We also try to handle cases which might run the
770 // FIXME: We should instead consider the pointer to have escaped if this
771 // function is being instrumented for addressing bugs or race conditions.
772 if (Size > AllocSize || Offset.ugt(AllocSize - Size)) {
773 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
774 << " which extends past the end of the " << AllocSize
776 << " alloca: " << AS.AI << "\n"
777 << " use: " << SI << "\n");
778 return markAsDead(SI);
781 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
782 "All simple FCA stores should have been pre-split");
783 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
786 void visitMemSetInst(MemSetInst &II) {
787 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
788 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
789 if ((Length && Length->getValue() == 0) ||
790 (IsOffsetKnown && Offset.uge(AllocSize)))
791 // Zero-length mem transfer intrinsics can be ignored entirely.
792 return markAsDead(II);
795 return PI.setAborted(&II);
797 insertUse(II, Offset, Length ? Length->getLimitedValue()
798 : AllocSize - Offset.getLimitedValue(),
802 void visitMemTransferInst(MemTransferInst &II) {
803 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
804 if (Length && Length->getValue() == 0)
805 // Zero-length mem transfer intrinsics can be ignored entirely.
806 return markAsDead(II);
808 // Because we can visit these intrinsics twice, also check to see if the
809 // first time marked this instruction as dead. If so, skip it.
810 if (VisitedDeadInsts.count(&II))
814 return PI.setAborted(&II);
816 // This side of the transfer is completely out-of-bounds, and so we can
817 // nuke the entire transfer. However, we also need to nuke the other side
818 // if already added to our partitions.
819 // FIXME: Yet another place we really should bypass this when
820 // instrumenting for ASan.
821 if (Offset.uge(AllocSize)) {
822 SmallDenseMap<Instruction *, unsigned>::iterator MTPI =
823 MemTransferSliceMap.find(&II);
824 if (MTPI != MemTransferSliceMap.end())
825 AS.Slices[MTPI->second].kill();
826 return markAsDead(II);
829 uint64_t RawOffset = Offset.getLimitedValue();
830 uint64_t Size = Length ? Length->getLimitedValue() : AllocSize - RawOffset;
832 // Check for the special case where the same exact value is used for both
834 if (*U == II.getRawDest() && *U == II.getRawSource()) {
835 // For non-volatile transfers this is a no-op.
836 if (!II.isVolatile())
837 return markAsDead(II);
839 return insertUse(II, Offset, Size, /*IsSplittable=*/false);
842 // If we have seen both source and destination for a mem transfer, then
843 // they both point to the same alloca.
845 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
846 std::tie(MTPI, Inserted) =
847 MemTransferSliceMap.insert(std::make_pair(&II, AS.Slices.size()));
848 unsigned PrevIdx = MTPI->second;
850 Slice &PrevP = AS.Slices[PrevIdx];
852 // Check if the begin offsets match and this is a non-volatile transfer.
853 // In that case, we can completely elide the transfer.
854 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
856 return markAsDead(II);
859 // Otherwise we have an offset transfer within the same alloca. We can't
861 PrevP.makeUnsplittable();
864 // Insert the use now that we've fixed up the splittable nature.
865 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
867 // Check that we ended up with a valid index in the map.
868 assert(AS.Slices[PrevIdx].getUse()->getUser() == &II &&
869 "Map index doesn't point back to a slice with this user.");
872 // Disable SRoA for any intrinsics except for lifetime invariants.
873 // FIXME: What about debug intrinsics? This matches old behavior, but
874 // doesn't make sense.
875 void visitIntrinsicInst(IntrinsicInst &II) {
877 return PI.setAborted(&II);
879 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
880 II.getIntrinsicID() == Intrinsic::lifetime_end) {
881 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
882 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
883 Length->getLimitedValue());
884 insertUse(II, Offset, Size, true);
888 Base::visitIntrinsicInst(II);
891 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
892 // We consider any PHI or select that results in a direct load or store of
893 // the same offset to be a viable use for slicing purposes. These uses
894 // are considered unsplittable and the size is the maximum loaded or stored
896 SmallPtrSet<Instruction *, 4> Visited;
897 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
898 Visited.insert(Root);
899 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
900 const DataLayout &DL = Root->getModule()->getDataLayout();
901 // If there are no loads or stores, the access is dead. We mark that as
902 // a size zero access.
905 Instruction *I, *UsedI;
906 std::tie(UsedI, I) = Uses.pop_back_val();
908 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
909 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
912 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
913 Value *Op = SI->getOperand(0);
916 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
920 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
921 if (!GEP->hasAllZeroIndices())
923 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
924 !isa<SelectInst>(I)) {
928 for (User *U : I->users())
929 if (Visited.insert(cast<Instruction>(U)).second)
930 Uses.push_back(std::make_pair(I, cast<Instruction>(U)));
931 } while (!Uses.empty());
936 void visitPHINodeOrSelectInst(Instruction &I) {
937 assert(isa<PHINode>(I) || isa<SelectInst>(I));
939 return markAsDead(I);
941 // TODO: We could use SimplifyInstruction here to fold PHINodes and
942 // SelectInsts. However, doing so requires to change the current
943 // dead-operand-tracking mechanism. For instance, suppose neither loading
944 // from %U nor %other traps. Then "load (select undef, %U, %other)" does not
945 // trap either. However, if we simply replace %U with undef using the
946 // current dead-operand-tracking mechanism, "load (select undef, undef,
947 // %other)" may trap because the select may return the first operand
949 if (Value *Result = foldPHINodeOrSelectInst(I)) {
951 // If the result of the constant fold will be the pointer, recurse
952 // through the PHI/select as if we had RAUW'ed it.
955 // Otherwise the operand to the PHI/select is dead, and we can replace
957 AS.DeadOperands.push_back(U);
963 return PI.setAborted(&I);
965 // See if we already have computed info on this node.
966 uint64_t &Size = PHIOrSelectSizes[&I];
968 // This is a new PHI/Select, check for an unsafe use of it.
969 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&I, Size))
970 return PI.setAborted(UnsafeI);
973 // For PHI and select operands outside the alloca, we can't nuke the entire
974 // phi or select -- the other side might still be relevant, so we special
975 // case them here and use a separate structure to track the operands
976 // themselves which should be replaced with undef.
977 // FIXME: This should instead be escaped in the event we're instrumenting
978 // for address sanitization.
979 if (Offset.uge(AllocSize)) {
980 AS.DeadOperands.push_back(U);
984 insertUse(I, Offset, Size);
987 void visitPHINode(PHINode &PN) { visitPHINodeOrSelectInst(PN); }
989 void visitSelectInst(SelectInst &SI) { visitPHINodeOrSelectInst(SI); }
991 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
992 void visitInstruction(Instruction &I) { PI.setAborted(&I); }
995 AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
997 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1000 PointerEscapingInstr(nullptr) {
1001 SliceBuilder PB(DL, AI, *this);
1002 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
1003 if (PtrI.isEscaped() || PtrI.isAborted()) {
1004 // FIXME: We should sink the escape vs. abort info into the caller nicely,
1005 // possibly by just storing the PtrInfo in the AllocaSlices.
1006 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
1007 : PtrI.getAbortingInst();
1008 assert(PointerEscapingInstr && "Did not track a bad instruction");
1012 Slices.erase(std::remove_if(Slices.begin(), Slices.end(),
1013 [](const Slice &S) {
1018 #if __cplusplus >= 201103L && !defined(NDEBUG)
1019 if (SROARandomShuffleSlices) {
1020 std::mt19937 MT(static_cast<unsigned>(sys::TimeValue::now().msec()));
1021 std::shuffle(Slices.begin(), Slices.end(), MT);
1025 // Sort the uses. This arranges for the offsets to be in ascending order,
1026 // and the sizes to be in descending order.
1027 std::sort(Slices.begin(), Slices.end());
1030 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1032 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
1033 StringRef Indent) const {
1034 printSlice(OS, I, Indent);
1036 printUse(OS, I, Indent);
1039 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
1040 StringRef Indent) const {
1041 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
1042 << " slice #" << (I - begin())
1043 << (I->isSplittable() ? " (splittable)" : "");
1046 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
1047 StringRef Indent) const {
1048 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
1051 void AllocaSlices::print(raw_ostream &OS) const {
1052 if (PointerEscapingInstr) {
1053 OS << "Can't analyze slices for alloca: " << AI << "\n"
1054 << " A pointer to this alloca escaped by:\n"
1055 << " " << *PointerEscapingInstr << "\n";
1059 OS << "Slices of alloca: " << AI << "\n";
1060 for (const_iterator I = begin(), E = end(); I != E; ++I)
1064 LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
1067 LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
1069 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1071 /// Walk the range of a partitioning looking for a common type to cover this
1072 /// sequence of slices.
1073 static Type *findCommonType(AllocaSlices::const_iterator B,
1074 AllocaSlices::const_iterator E,
1075 uint64_t EndOffset) {
1077 bool TyIsCommon = true;
1078 IntegerType *ITy = nullptr;
1080 // Note that we need to look at *every* alloca slice's Use to ensure we
1081 // always get consistent results regardless of the order of slices.
1082 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1083 Use *U = I->getUse();
1084 if (isa<IntrinsicInst>(*U->getUser()))
1086 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1089 Type *UserTy = nullptr;
1090 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1091 UserTy = LI->getType();
1092 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1093 UserTy = SI->getValueOperand()->getType();
1096 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1097 // If the type is larger than the partition, skip it. We only encounter
1098 // this for split integer operations where we want to use the type of the
1099 // entity causing the split. Also skip if the type is not a byte width
1101 if (UserITy->getBitWidth() % 8 != 0 ||
1102 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1105 // Track the largest bitwidth integer type used in this way in case there
1106 // is no common type.
1107 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1111 // To avoid depending on the order of slices, Ty and TyIsCommon must not
1112 // depend on types skipped above.
1113 if (!UserTy || (Ty && Ty != UserTy))
1114 TyIsCommon = false; // Give up on anything but an iN type.
1119 return TyIsCommon ? Ty : ITy;
1122 /// PHI instructions that use an alloca and are subsequently loaded can be
1123 /// rewritten to load both input pointers in the pred blocks and then PHI the
1124 /// results, allowing the load of the alloca to be promoted.
1126 /// %P2 = phi [i32* %Alloca, i32* %Other]
1127 /// %V = load i32* %P2
1129 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1131 /// %V2 = load i32* %Other
1133 /// %V = phi [i32 %V1, i32 %V2]
1135 /// We can do this to a select if its only uses are loads and if the operands
1136 /// to the select can be loaded unconditionally.
1138 /// FIXME: This should be hoisted into a generic utility, likely in
1139 /// Transforms/Util/Local.h
1140 static bool isSafePHIToSpeculate(PHINode &PN) {
1141 // For now, we can only do this promotion if the load is in the same block
1142 // as the PHI, and if there are no stores between the phi and load.
1143 // TODO: Allow recursive phi users.
1144 // TODO: Allow stores.
1145 BasicBlock *BB = PN.getParent();
1146 unsigned MaxAlign = 0;
1147 bool HaveLoad = false;
1148 for (User *U : PN.users()) {
1149 LoadInst *LI = dyn_cast<LoadInst>(U);
1150 if (!LI || !LI->isSimple())
1153 // For now we only allow loads in the same block as the PHI. This is
1154 // a common case that happens when instcombine merges two loads through
1156 if (LI->getParent() != BB)
1159 // Ensure that there are no instructions between the PHI and the load that
1161 for (BasicBlock::iterator BBI(PN); &*BBI != LI; ++BBI)
1162 if (BBI->mayWriteToMemory())
1165 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1172 const DataLayout &DL = PN.getModule()->getDataLayout();
1174 // We can only transform this if it is safe to push the loads into the
1175 // predecessor blocks. The only thing to watch out for is that we can't put
1176 // a possibly trapping load in the predecessor if it is a critical edge.
1177 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1178 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1179 Value *InVal = PN.getIncomingValue(Idx);
1181 // If the value is produced by the terminator of the predecessor (an
1182 // invoke) or it has side-effects, there is no valid place to put a load
1183 // in the predecessor.
1184 if (TI == InVal || TI->mayHaveSideEffects())
1187 // If the predecessor has a single successor, then the edge isn't
1189 if (TI->getNumSuccessors() == 1)
1192 // If this pointer is always safe to load, or if we can prove that there
1193 // is already a load in the block, then we can move the load to the pred
1195 if (isDereferenceablePointer(InVal, DL) ||
1196 isSafeToLoadUnconditionally(InVal, TI, MaxAlign))
1205 static void speculatePHINodeLoads(PHINode &PN) {
1206 DEBUG(dbgs() << " original: " << PN << "\n");
1208 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1209 IRBuilderTy PHIBuilder(&PN);
1210 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1211 PN.getName() + ".sroa.speculated");
1213 // Get the AA tags and alignment to use from one of the loads. It doesn't
1214 // matter which one we get and if any differ.
1215 LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1218 SomeLoad->getAAMetadata(AATags);
1219 unsigned Align = SomeLoad->getAlignment();
1221 // Rewrite all loads of the PN to use the new PHI.
1222 while (!PN.use_empty()) {
1223 LoadInst *LI = cast<LoadInst>(PN.user_back());
1224 LI->replaceAllUsesWith(NewPN);
1225 LI->eraseFromParent();
1228 // Inject loads into all of the pred blocks.
1229 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1230 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1231 TerminatorInst *TI = Pred->getTerminator();
1232 Value *InVal = PN.getIncomingValue(Idx);
1233 IRBuilderTy PredBuilder(TI);
1235 LoadInst *Load = PredBuilder.CreateLoad(
1236 InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1237 ++NumLoadsSpeculated;
1238 Load->setAlignment(Align);
1240 Load->setAAMetadata(AATags);
1241 NewPN->addIncoming(Load, Pred);
1244 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1245 PN.eraseFromParent();
1248 /// Select instructions that use an alloca and are subsequently loaded can be
1249 /// rewritten to load both input pointers and then select between the result,
1250 /// allowing the load of the alloca to be promoted.
1252 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1253 /// %V = load i32* %P2
1255 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1256 /// %V2 = load i32* %Other
1257 /// %V = select i1 %cond, i32 %V1, i32 %V2
1259 /// We can do this to a select if its only uses are loads and if the operand
1260 /// to the select can be loaded unconditionally.
1261 static bool isSafeSelectToSpeculate(SelectInst &SI) {
1262 Value *TValue = SI.getTrueValue();
1263 Value *FValue = SI.getFalseValue();
1264 const DataLayout &DL = SI.getModule()->getDataLayout();
1265 bool TDerefable = isDereferenceablePointer(TValue, DL);
1266 bool FDerefable = isDereferenceablePointer(FValue, DL);
1268 for (User *U : SI.users()) {
1269 LoadInst *LI = dyn_cast<LoadInst>(U);
1270 if (!LI || !LI->isSimple())
1273 // Both operands to the select need to be dereferencable, either
1274 // absolutely (e.g. allocas) or at this point because we can see other
1277 !isSafeToLoadUnconditionally(TValue, LI, LI->getAlignment()))
1280 !isSafeToLoadUnconditionally(FValue, LI, LI->getAlignment()))
1287 static void speculateSelectInstLoads(SelectInst &SI) {
1288 DEBUG(dbgs() << " original: " << SI << "\n");
1290 IRBuilderTy IRB(&SI);
1291 Value *TV = SI.getTrueValue();
1292 Value *FV = SI.getFalseValue();
1293 // Replace the loads of the select with a select of two loads.
1294 while (!SI.use_empty()) {
1295 LoadInst *LI = cast<LoadInst>(SI.user_back());
1296 assert(LI->isSimple() && "We only speculate simple loads");
1298 IRB.SetInsertPoint(LI);
1300 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1302 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1303 NumLoadsSpeculated += 2;
1305 // Transfer alignment and AA info if present.
1306 TL->setAlignment(LI->getAlignment());
1307 FL->setAlignment(LI->getAlignment());
1310 LI->getAAMetadata(Tags);
1312 TL->setAAMetadata(Tags);
1313 FL->setAAMetadata(Tags);
1316 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1317 LI->getName() + ".sroa.speculated");
1319 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1320 LI->replaceAllUsesWith(V);
1321 LI->eraseFromParent();
1323 SI.eraseFromParent();
1326 /// \brief Build a GEP out of a base pointer and indices.
1328 /// This will return the BasePtr if that is valid, or build a new GEP
1329 /// instruction using the IRBuilder if GEP-ing is needed.
1330 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1331 SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1332 if (Indices.empty())
1335 // A single zero index is a no-op, so check for this and avoid building a GEP
1337 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1340 return IRB.CreateInBoundsGEP(nullptr, BasePtr, Indices,
1341 NamePrefix + "sroa_idx");
1344 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1345 /// TargetTy without changing the offset of the pointer.
1347 /// This routine assumes we've already established a properly offset GEP with
1348 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1349 /// zero-indices down through type layers until we find one the same as
1350 /// TargetTy. If we can't find one with the same type, we at least try to use
1351 /// one with the same size. If none of that works, we just produce the GEP as
1352 /// indicated by Indices to have the correct offset.
1353 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1354 Value *BasePtr, Type *Ty, Type *TargetTy,
1355 SmallVectorImpl<Value *> &Indices,
1358 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1360 // Pointer size to use for the indices.
1361 unsigned PtrSize = DL.getPointerTypeSizeInBits(BasePtr->getType());
1363 // See if we can descend into a struct and locate a field with the correct
1365 unsigned NumLayers = 0;
1366 Type *ElementTy = Ty;
1368 if (ElementTy->isPointerTy())
1371 if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1372 ElementTy = ArrayTy->getElementType();
1373 Indices.push_back(IRB.getIntN(PtrSize, 0));
1374 } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1375 ElementTy = VectorTy->getElementType();
1376 Indices.push_back(IRB.getInt32(0));
1377 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1378 if (STy->element_begin() == STy->element_end())
1379 break; // Nothing left to descend into.
1380 ElementTy = *STy->element_begin();
1381 Indices.push_back(IRB.getInt32(0));
1386 } while (ElementTy != TargetTy);
1387 if (ElementTy != TargetTy)
1388 Indices.erase(Indices.end() - NumLayers, Indices.end());
1390 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1393 /// \brief Recursively compute indices for a natural GEP.
1395 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1396 /// element types adding appropriate indices for the GEP.
1397 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1398 Value *Ptr, Type *Ty, APInt &Offset,
1400 SmallVectorImpl<Value *> &Indices,
1403 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices,
1406 // We can't recurse through pointer types.
1407 if (Ty->isPointerTy())
1410 // We try to analyze GEPs over vectors here, but note that these GEPs are
1411 // extremely poorly defined currently. The long-term goal is to remove GEPing
1412 // over a vector from the IR completely.
1413 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1414 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1415 if (ElementSizeInBits % 8 != 0) {
1416 // GEPs over non-multiple of 8 size vector elements are invalid.
1419 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1420 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1421 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1423 Offset -= NumSkippedElements * ElementSize;
1424 Indices.push_back(IRB.getInt(NumSkippedElements));
1425 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1426 Offset, TargetTy, Indices, NamePrefix);
1429 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1430 Type *ElementTy = ArrTy->getElementType();
1431 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1432 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1433 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1436 Offset -= NumSkippedElements * ElementSize;
1437 Indices.push_back(IRB.getInt(NumSkippedElements));
1438 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1439 Indices, NamePrefix);
1442 StructType *STy = dyn_cast<StructType>(Ty);
1446 const StructLayout *SL = DL.getStructLayout(STy);
1447 uint64_t StructOffset = Offset.getZExtValue();
1448 if (StructOffset >= SL->getSizeInBytes())
1450 unsigned Index = SL->getElementContainingOffset(StructOffset);
1451 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1452 Type *ElementTy = STy->getElementType(Index);
1453 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1454 return nullptr; // The offset points into alignment padding.
1456 Indices.push_back(IRB.getInt32(Index));
1457 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1458 Indices, NamePrefix);
1461 /// \brief Get a natural GEP from a base pointer to a particular offset and
1462 /// resulting in a particular type.
1464 /// The goal is to produce a "natural" looking GEP that works with the existing
1465 /// composite types to arrive at the appropriate offset and element type for
1466 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1467 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1468 /// Indices, and setting Ty to the result subtype.
1470 /// If no natural GEP can be constructed, this function returns null.
1471 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1472 Value *Ptr, APInt Offset, Type *TargetTy,
1473 SmallVectorImpl<Value *> &Indices,
1475 PointerType *Ty = cast<PointerType>(Ptr->getType());
1477 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1479 if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1482 Type *ElementTy = Ty->getElementType();
1483 if (!ElementTy->isSized())
1484 return nullptr; // We can't GEP through an unsized element.
1485 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1486 if (ElementSize == 0)
1487 return nullptr; // Zero-length arrays can't help us build a natural GEP.
1488 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1490 Offset -= NumSkippedElements * ElementSize;
1491 Indices.push_back(IRB.getInt(NumSkippedElements));
1492 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1493 Indices, NamePrefix);
1496 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1497 /// resulting pointer has PointerTy.
1499 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1500 /// and produces the pointer type desired. Where it cannot, it will try to use
1501 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1502 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1503 /// bitcast to the type.
1505 /// The strategy for finding the more natural GEPs is to peel off layers of the
1506 /// pointer, walking back through bit casts and GEPs, searching for a base
1507 /// pointer from which we can compute a natural GEP with the desired
1508 /// properties. The algorithm tries to fold as many constant indices into
1509 /// a single GEP as possible, thus making each GEP more independent of the
1510 /// surrounding code.
1511 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1512 APInt Offset, Type *PointerTy, Twine NamePrefix) {
1513 // Even though we don't look through PHI nodes, we could be called on an
1514 // instruction in an unreachable block, which may be on a cycle.
1515 SmallPtrSet<Value *, 4> Visited;
1516 Visited.insert(Ptr);
1517 SmallVector<Value *, 4> Indices;
1519 // We may end up computing an offset pointer that has the wrong type. If we
1520 // never are able to compute one directly that has the correct type, we'll
1521 // fall back to it, so keep it and the base it was computed from around here.
1522 Value *OffsetPtr = nullptr;
1523 Value *OffsetBasePtr;
1525 // Remember any i8 pointer we come across to re-use if we need to do a raw
1527 Value *Int8Ptr = nullptr;
1528 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1530 Type *TargetTy = PointerTy->getPointerElementType();
1533 // First fold any existing GEPs into the offset.
1534 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1535 APInt GEPOffset(Offset.getBitWidth(), 0);
1536 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1538 Offset += GEPOffset;
1539 Ptr = GEP->getPointerOperand();
1540 if (!Visited.insert(Ptr).second)
1544 // See if we can perform a natural GEP here.
1546 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1547 Indices, NamePrefix)) {
1548 // If we have a new natural pointer at the offset, clear out any old
1549 // offset pointer we computed. Unless it is the base pointer or
1550 // a non-instruction, we built a GEP we don't need. Zap it.
1551 if (OffsetPtr && OffsetPtr != OffsetBasePtr)
1552 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr)) {
1553 assert(I->use_empty() && "Built a GEP with uses some how!");
1554 I->eraseFromParent();
1557 OffsetBasePtr = Ptr;
1558 // If we also found a pointer of the right type, we're done.
1559 if (P->getType() == PointerTy)
1563 // Stash this pointer if we've found an i8*.
1564 if (Ptr->getType()->isIntegerTy(8)) {
1566 Int8PtrOffset = Offset;
1569 // Peel off a layer of the pointer and update the offset appropriately.
1570 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1571 Ptr = cast<Operator>(Ptr)->getOperand(0);
1572 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1573 if (GA->mayBeOverridden())
1575 Ptr = GA->getAliasee();
1579 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1580 } while (Visited.insert(Ptr).second);
1584 Int8Ptr = IRB.CreateBitCast(
1585 Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1586 NamePrefix + "sroa_raw_cast");
1587 Int8PtrOffset = Offset;
1590 OffsetPtr = Int8PtrOffset == 0
1592 : IRB.CreateInBoundsGEP(IRB.getInt8Ty(), Int8Ptr,
1593 IRB.getInt(Int8PtrOffset),
1594 NamePrefix + "sroa_raw_idx");
1598 // On the off chance we were targeting i8*, guard the bitcast here.
1599 if (Ptr->getType() != PointerTy)
1600 Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
1605 /// \brief Compute the adjusted alignment for a load or store from an offset.
1606 static unsigned getAdjustedAlignment(Instruction *I, uint64_t Offset,
1607 const DataLayout &DL) {
1610 if (auto *LI = dyn_cast<LoadInst>(I)) {
1611 Alignment = LI->getAlignment();
1613 } else if (auto *SI = dyn_cast<StoreInst>(I)) {
1614 Alignment = SI->getAlignment();
1615 Ty = SI->getValueOperand()->getType();
1617 llvm_unreachable("Only loads and stores are allowed!");
1621 Alignment = DL.getABITypeAlignment(Ty);
1623 return MinAlign(Alignment, Offset);
1626 /// \brief Test whether we can convert a value from the old to the new type.
1628 /// This predicate should be used to guard calls to convertValue in order to
1629 /// ensure that we only try to convert viable values. The strategy is that we
1630 /// will peel off single element struct and array wrappings to get to an
1631 /// underlying value, and convert that value.
1632 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1636 // For integer types, we can't handle any bit-width differences. This would
1637 // break both vector conversions with extension and introduce endianness
1638 // issues when in conjunction with loads and stores.
1639 if (isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) {
1640 assert(cast<IntegerType>(OldTy)->getBitWidth() !=
1641 cast<IntegerType>(NewTy)->getBitWidth() &&
1642 "We can't have the same bitwidth for different int types");
1646 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1648 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1651 // We can convert pointers to integers and vice-versa. Same for vectors
1652 // of pointers and integers.
1653 OldTy = OldTy->getScalarType();
1654 NewTy = NewTy->getScalarType();
1655 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1656 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1658 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1666 /// \brief Generic routine to convert an SSA value to a value of a different
1669 /// This will try various different casting techniques, such as bitcasts,
1670 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1671 /// two types for viability with this routine.
1672 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1674 Type *OldTy = V->getType();
1675 assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1680 assert(!(isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) &&
1681 "Integer types must be the exact same to convert.");
1683 // See if we need inttoptr for this type pair. A cast involving both scalars
1684 // and vectors requires and additional bitcast.
1685 if (OldTy->getScalarType()->isIntegerTy() &&
1686 NewTy->getScalarType()->isPointerTy()) {
1687 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1688 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1689 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1692 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1693 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1694 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1697 return IRB.CreateIntToPtr(V, NewTy);
1700 // See if we need ptrtoint for this type pair. A cast involving both scalars
1701 // and vectors requires and additional bitcast.
1702 if (OldTy->getScalarType()->isPointerTy() &&
1703 NewTy->getScalarType()->isIntegerTy()) {
1704 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1705 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1706 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1709 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1710 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1711 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1714 return IRB.CreatePtrToInt(V, NewTy);
1717 return IRB.CreateBitCast(V, NewTy);
1720 /// \brief Test whether the given slice use can be promoted to a vector.
1722 /// This function is called to test each entry in a partition which is slated
1723 /// for a single slice.
1724 static bool isVectorPromotionViableForSlice(Partition &P, const Slice &S,
1726 uint64_t ElementSize,
1727 const DataLayout &DL) {
1728 // First validate the slice offsets.
1729 uint64_t BeginOffset =
1730 std::max(S.beginOffset(), P.beginOffset()) - P.beginOffset();
1731 uint64_t BeginIndex = BeginOffset / ElementSize;
1732 if (BeginIndex * ElementSize != BeginOffset ||
1733 BeginIndex >= Ty->getNumElements())
1735 uint64_t EndOffset =
1736 std::min(S.endOffset(), P.endOffset()) - P.beginOffset();
1737 uint64_t EndIndex = EndOffset / ElementSize;
1738 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1741 assert(EndIndex > BeginIndex && "Empty vector!");
1742 uint64_t NumElements = EndIndex - BeginIndex;
1743 Type *SliceTy = (NumElements == 1)
1744 ? Ty->getElementType()
1745 : VectorType::get(Ty->getElementType(), NumElements);
1748 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1750 Use *U = S.getUse();
1752 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1753 if (MI->isVolatile())
1755 if (!S.isSplittable())
1756 return false; // Skip any unsplittable intrinsics.
1757 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1758 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1759 II->getIntrinsicID() != Intrinsic::lifetime_end)
1761 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1762 // Disable vector promotion when there are loads or stores of an FCA.
1764 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1765 if (LI->isVolatile())
1767 Type *LTy = LI->getType();
1768 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1769 assert(LTy->isIntegerTy());
1772 if (!canConvertValue(DL, SliceTy, LTy))
1774 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1775 if (SI->isVolatile())
1777 Type *STy = SI->getValueOperand()->getType();
1778 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1779 assert(STy->isIntegerTy());
1782 if (!canConvertValue(DL, STy, SliceTy))
1791 /// \brief Test whether the given alloca partitioning and range of slices can be
1792 /// promoted to a vector.
1794 /// This is a quick test to check whether we can rewrite a particular alloca
1795 /// partition (and its newly formed alloca) into a vector alloca with only
1796 /// whole-vector loads and stores such that it could be promoted to a vector
1797 /// SSA value. We only can ensure this for a limited set of operations, and we
1798 /// don't want to do the rewrites unless we are confident that the result will
1799 /// be promotable, so we have an early test here.
1800 static VectorType *isVectorPromotionViable(Partition &P, const DataLayout &DL) {
1801 // Collect the candidate types for vector-based promotion. Also track whether
1802 // we have different element types.
1803 SmallVector<VectorType *, 4> CandidateTys;
1804 Type *CommonEltTy = nullptr;
1805 bool HaveCommonEltTy = true;
1806 auto CheckCandidateType = [&](Type *Ty) {
1807 if (auto *VTy = dyn_cast<VectorType>(Ty)) {
1808 CandidateTys.push_back(VTy);
1810 CommonEltTy = VTy->getElementType();
1811 else if (CommonEltTy != VTy->getElementType())
1812 HaveCommonEltTy = false;
1815 // Consider any loads or stores that are the exact size of the slice.
1816 for (const Slice &S : P)
1817 if (S.beginOffset() == P.beginOffset() &&
1818 S.endOffset() == P.endOffset()) {
1819 if (auto *LI = dyn_cast<LoadInst>(S.getUse()->getUser()))
1820 CheckCandidateType(LI->getType());
1821 else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser()))
1822 CheckCandidateType(SI->getValueOperand()->getType());
1825 // If we didn't find a vector type, nothing to do here.
1826 if (CandidateTys.empty())
1829 // Remove non-integer vector types if we had multiple common element types.
1830 // FIXME: It'd be nice to replace them with integer vector types, but we can't
1831 // do that until all the backends are known to produce good code for all
1832 // integer vector types.
1833 if (!HaveCommonEltTy) {
1834 CandidateTys.erase(std::remove_if(CandidateTys.begin(), CandidateTys.end(),
1835 [](VectorType *VTy) {
1836 return !VTy->getElementType()->isIntegerTy();
1838 CandidateTys.end());
1840 // If there were no integer vector types, give up.
1841 if (CandidateTys.empty())
1844 // Rank the remaining candidate vector types. This is easy because we know
1845 // they're all integer vectors. We sort by ascending number of elements.
1846 auto RankVectorTypes = [&DL](VectorType *RHSTy, VectorType *LHSTy) {
1847 assert(DL.getTypeSizeInBits(RHSTy) == DL.getTypeSizeInBits(LHSTy) &&
1848 "Cannot have vector types of different sizes!");
1849 assert(RHSTy->getElementType()->isIntegerTy() &&
1850 "All non-integer types eliminated!");
1851 assert(LHSTy->getElementType()->isIntegerTy() &&
1852 "All non-integer types eliminated!");
1853 return RHSTy->getNumElements() < LHSTy->getNumElements();
1855 std::sort(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes);
1857 std::unique(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes),
1858 CandidateTys.end());
1860 // The only way to have the same element type in every vector type is to
1861 // have the same vector type. Check that and remove all but one.
1863 for (VectorType *VTy : CandidateTys) {
1864 assert(VTy->getElementType() == CommonEltTy &&
1865 "Unaccounted for element type!");
1866 assert(VTy == CandidateTys[0] &&
1867 "Different vector types with the same element type!");
1870 CandidateTys.resize(1);
1873 // Try each vector type, and return the one which works.
1874 auto CheckVectorTypeForPromotion = [&](VectorType *VTy) {
1875 uint64_t ElementSize = DL.getTypeSizeInBits(VTy->getElementType());
1877 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1878 // that aren't byte sized.
1879 if (ElementSize % 8)
1881 assert((DL.getTypeSizeInBits(VTy) % 8) == 0 &&
1882 "vector size not a multiple of element size?");
1885 for (const Slice &S : P)
1886 if (!isVectorPromotionViableForSlice(P, S, VTy, ElementSize, DL))
1889 for (const Slice *S : P.splitSliceTails())
1890 if (!isVectorPromotionViableForSlice(P, *S, VTy, ElementSize, DL))
1895 for (VectorType *VTy : CandidateTys)
1896 if (CheckVectorTypeForPromotion(VTy))
1902 /// \brief Test whether a slice of an alloca is valid for integer widening.
1904 /// This implements the necessary checking for the \c isIntegerWideningViable
1905 /// test below on a single slice of the alloca.
1906 static bool isIntegerWideningViableForSlice(const Slice &S,
1907 uint64_t AllocBeginOffset,
1909 const DataLayout &DL,
1910 bool &WholeAllocaOp) {
1911 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1913 uint64_t RelBegin = S.beginOffset() - AllocBeginOffset;
1914 uint64_t RelEnd = S.endOffset() - AllocBeginOffset;
1916 // We can't reasonably handle cases where the load or store extends past
1917 // the end of the alloca's type and into its padding.
1921 Use *U = S.getUse();
1923 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1924 if (LI->isVolatile())
1926 // We can't handle loads that extend past the allocated memory.
1927 if (DL.getTypeStoreSize(LI->getType()) > Size)
1929 // Note that we don't count vector loads or stores as whole-alloca
1930 // operations which enable integer widening because we would prefer to use
1931 // vector widening instead.
1932 if (!isa<VectorType>(LI->getType()) && RelBegin == 0 && RelEnd == Size)
1933 WholeAllocaOp = true;
1934 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1935 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1937 } else if (RelBegin != 0 || RelEnd != Size ||
1938 !canConvertValue(DL, AllocaTy, LI->getType())) {
1939 // Non-integer loads need to be convertible from the alloca type so that
1940 // they are promotable.
1943 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1944 Type *ValueTy = SI->getValueOperand()->getType();
1945 if (SI->isVolatile())
1947 // We can't handle stores that extend past the allocated memory.
1948 if (DL.getTypeStoreSize(ValueTy) > Size)
1950 // Note that we don't count vector loads or stores as whole-alloca
1951 // operations which enable integer widening because we would prefer to use
1952 // vector widening instead.
1953 if (!isa<VectorType>(ValueTy) && RelBegin == 0 && RelEnd == Size)
1954 WholeAllocaOp = true;
1955 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
1956 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1958 } else if (RelBegin != 0 || RelEnd != Size ||
1959 !canConvertValue(DL, ValueTy, AllocaTy)) {
1960 // Non-integer stores need to be convertible to the alloca type so that
1961 // they are promotable.
1964 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1965 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
1967 if (!S.isSplittable())
1968 return false; // Skip any unsplittable intrinsics.
1969 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1970 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1971 II->getIntrinsicID() != Intrinsic::lifetime_end)
1980 /// \brief Test whether the given alloca partition's integer operations can be
1981 /// widened to promotable ones.
1983 /// This is a quick test to check whether we can rewrite the integer loads and
1984 /// stores to a particular alloca into wider loads and stores and be able to
1985 /// promote the resulting alloca.
1986 static bool isIntegerWideningViable(Partition &P, Type *AllocaTy,
1987 const DataLayout &DL) {
1988 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
1989 // Don't create integer types larger than the maximum bitwidth.
1990 if (SizeInBits > IntegerType::MAX_INT_BITS)
1993 // Don't try to handle allocas with bit-padding.
1994 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
1997 // We need to ensure that an integer type with the appropriate bitwidth can
1998 // be converted to the alloca type, whatever that is. We don't want to force
1999 // the alloca itself to have an integer type if there is a more suitable one.
2000 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
2001 if (!canConvertValue(DL, AllocaTy, IntTy) ||
2002 !canConvertValue(DL, IntTy, AllocaTy))
2005 // While examining uses, we ensure that the alloca has a covering load or
2006 // store. We don't want to widen the integer operations only to fail to
2007 // promote due to some other unsplittable entry (which we may make splittable
2008 // later). However, if there are only splittable uses, go ahead and assume
2009 // that we cover the alloca.
2010 // FIXME: We shouldn't consider split slices that happen to start in the
2011 // partition here...
2012 bool WholeAllocaOp =
2013 P.begin() != P.end() ? false : DL.isLegalInteger(SizeInBits);
2015 for (const Slice &S : P)
2016 if (!isIntegerWideningViableForSlice(S, P.beginOffset(), AllocaTy, DL,
2020 for (const Slice *S : P.splitSliceTails())
2021 if (!isIntegerWideningViableForSlice(*S, P.beginOffset(), AllocaTy, DL,
2025 return WholeAllocaOp;
2028 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
2029 IntegerType *Ty, uint64_t Offset,
2030 const Twine &Name) {
2031 DEBUG(dbgs() << " start: " << *V << "\n");
2032 IntegerType *IntTy = cast<IntegerType>(V->getType());
2033 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2034 "Element extends past full value");
2035 uint64_t ShAmt = 8 * Offset;
2036 if (DL.isBigEndian())
2037 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2039 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2040 DEBUG(dbgs() << " shifted: " << *V << "\n");
2042 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2043 "Cannot extract to a larger integer!");
2045 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2046 DEBUG(dbgs() << " trunced: " << *V << "\n");
2051 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
2052 Value *V, uint64_t Offset, const Twine &Name) {
2053 IntegerType *IntTy = cast<IntegerType>(Old->getType());
2054 IntegerType *Ty = cast<IntegerType>(V->getType());
2055 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2056 "Cannot insert a larger integer!");
2057 DEBUG(dbgs() << " start: " << *V << "\n");
2059 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2060 DEBUG(dbgs() << " extended: " << *V << "\n");
2062 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2063 "Element store outside of alloca store");
2064 uint64_t ShAmt = 8 * Offset;
2065 if (DL.isBigEndian())
2066 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2068 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2069 DEBUG(dbgs() << " shifted: " << *V << "\n");
2072 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2073 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2074 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2075 DEBUG(dbgs() << " masked: " << *Old << "\n");
2076 V = IRB.CreateOr(Old, V, Name + ".insert");
2077 DEBUG(dbgs() << " inserted: " << *V << "\n");
2082 static Value *extractVector(IRBuilderTy &IRB, Value *V, unsigned BeginIndex,
2083 unsigned EndIndex, const Twine &Name) {
2084 VectorType *VecTy = cast<VectorType>(V->getType());
2085 unsigned NumElements = EndIndex - BeginIndex;
2086 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2088 if (NumElements == VecTy->getNumElements())
2091 if (NumElements == 1) {
2092 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2094 DEBUG(dbgs() << " extract: " << *V << "\n");
2098 SmallVector<Constant *, 8> Mask;
2099 Mask.reserve(NumElements);
2100 for (unsigned i = BeginIndex; i != EndIndex; ++i)
2101 Mask.push_back(IRB.getInt32(i));
2102 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2103 ConstantVector::get(Mask), Name + ".extract");
2104 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2108 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
2109 unsigned BeginIndex, const Twine &Name) {
2110 VectorType *VecTy = cast<VectorType>(Old->getType());
2111 assert(VecTy && "Can only insert a vector into a vector");
2113 VectorType *Ty = dyn_cast<VectorType>(V->getType());
2115 // Single element to insert.
2116 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2118 DEBUG(dbgs() << " insert: " << *V << "\n");
2122 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
2123 "Too many elements!");
2124 if (Ty->getNumElements() == VecTy->getNumElements()) {
2125 assert(V->getType() == VecTy && "Vector type mismatch");
2128 unsigned EndIndex = BeginIndex + Ty->getNumElements();
2130 // When inserting a smaller vector into the larger to store, we first
2131 // use a shuffle vector to widen it with undef elements, and then
2132 // a second shuffle vector to select between the loaded vector and the
2134 SmallVector<Constant *, 8> Mask;
2135 Mask.reserve(VecTy->getNumElements());
2136 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2137 if (i >= BeginIndex && i < EndIndex)
2138 Mask.push_back(IRB.getInt32(i - BeginIndex));
2140 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2141 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2142 ConstantVector::get(Mask), Name + ".expand");
2143 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2146 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2147 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
2149 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
2151 DEBUG(dbgs() << " blend: " << *V << "\n");
2155 /// \brief Visitor to rewrite instructions using p particular slice of an alloca
2156 /// to use a new alloca.
2158 /// Also implements the rewriting to vector-based accesses when the partition
2159 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2161 class llvm::sroa::AllocaSliceRewriter
2162 : public InstVisitor<AllocaSliceRewriter, bool> {
2163 // Befriend the base class so it can delegate to private visit methods.
2164 friend class llvm::InstVisitor<AllocaSliceRewriter, bool>;
2165 typedef llvm::InstVisitor<AllocaSliceRewriter, bool> Base;
2167 const DataLayout &DL;
2170 AllocaInst &OldAI, &NewAI;
2171 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2174 // This is a convenience and flag variable that will be null unless the new
2175 // alloca's integer operations should be widened to this integer type due to
2176 // passing isIntegerWideningViable above. If it is non-null, the desired
2177 // integer type will be stored here for easy access during rewriting.
2180 // If we are rewriting an alloca partition which can be written as pure
2181 // vector operations, we stash extra information here. When VecTy is
2182 // non-null, we have some strict guarantees about the rewritten alloca:
2183 // - The new alloca is exactly the size of the vector type here.
2184 // - The accesses all either map to the entire vector or to a single
2186 // - The set of accessing instructions is only one of those handled above
2187 // in isVectorPromotionViable. Generally these are the same access kinds
2188 // which are promotable via mem2reg.
2191 uint64_t ElementSize;
2193 // The original offset of the slice currently being rewritten relative to
2194 // the original alloca.
2195 uint64_t BeginOffset, EndOffset;
2196 // The new offsets of the slice currently being rewritten relative to the
2198 uint64_t NewBeginOffset, NewEndOffset;
2204 Instruction *OldPtr;
2206 // Track post-rewrite users which are PHI nodes and Selects.
2207 SmallPtrSetImpl<PHINode *> &PHIUsers;
2208 SmallPtrSetImpl<SelectInst *> &SelectUsers;
2210 // Utility IR builder, whose name prefix is setup for each visited use, and
2211 // the insertion point is set to point to the user.
2215 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &AS, SROA &Pass,
2216 AllocaInst &OldAI, AllocaInst &NewAI,
2217 uint64_t NewAllocaBeginOffset,
2218 uint64_t NewAllocaEndOffset, bool IsIntegerPromotable,
2219 VectorType *PromotableVecTy,
2220 SmallPtrSetImpl<PHINode *> &PHIUsers,
2221 SmallPtrSetImpl<SelectInst *> &SelectUsers)
2222 : DL(DL), AS(AS), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2223 NewAllocaBeginOffset(NewAllocaBeginOffset),
2224 NewAllocaEndOffset(NewAllocaEndOffset),
2225 NewAllocaTy(NewAI.getAllocatedType()),
2226 IntTy(IsIntegerPromotable
2229 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2231 VecTy(PromotableVecTy),
2232 ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2233 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
2234 BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
2235 OldPtr(), PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2236 IRB(NewAI.getContext(), ConstantFolder()) {
2238 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
2239 "Only multiple-of-8 sized vector elements are viable");
2242 assert((!IntTy && !VecTy) || (IntTy && !VecTy) || (!IntTy && VecTy));
2245 bool visit(AllocaSlices::const_iterator I) {
2246 bool CanSROA = true;
2247 BeginOffset = I->beginOffset();
2248 EndOffset = I->endOffset();
2249 IsSplittable = I->isSplittable();
2251 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2252 DEBUG(dbgs() << " rewriting " << (IsSplit ? "split " : ""));
2253 DEBUG(AS.printSlice(dbgs(), I, ""));
2254 DEBUG(dbgs() << "\n");
2256 // Compute the intersecting offset range.
2257 assert(BeginOffset < NewAllocaEndOffset);
2258 assert(EndOffset > NewAllocaBeginOffset);
2259 NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2260 NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2262 SliceSize = NewEndOffset - NewBeginOffset;
2264 OldUse = I->getUse();
2265 OldPtr = cast<Instruction>(OldUse->get());
2267 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2268 IRB.SetInsertPoint(OldUserI);
2269 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2270 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2272 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2279 // Make sure the other visit overloads are visible.
2282 // Every instruction which can end up as a user must have a rewrite rule.
2283 bool visitInstruction(Instruction &I) {
2284 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2285 llvm_unreachable("No rewrite rule for this instruction!");
2288 Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2289 // Note that the offset computation can use BeginOffset or NewBeginOffset
2290 // interchangeably for unsplit slices.
2291 assert(IsSplit || BeginOffset == NewBeginOffset);
2292 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2295 StringRef OldName = OldPtr->getName();
2296 // Skip through the last '.sroa.' component of the name.
2297 size_t LastSROAPrefix = OldName.rfind(".sroa.");
2298 if (LastSROAPrefix != StringRef::npos) {
2299 OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2300 // Look for an SROA slice index.
2301 size_t IndexEnd = OldName.find_first_not_of("0123456789");
2302 if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2303 // Strip the index and look for the offset.
2304 OldName = OldName.substr(IndexEnd + 1);
2305 size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2306 if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2307 // Strip the offset.
2308 OldName = OldName.substr(OffsetEnd + 1);
2311 // Strip any SROA suffixes as well.
2312 OldName = OldName.substr(0, OldName.find(".sroa_"));
2315 return getAdjustedPtr(IRB, DL, &NewAI,
2316 APInt(DL.getPointerSizeInBits(), Offset), PointerTy,
2318 Twine(OldName) + "."
2325 /// \brief Compute suitable alignment to access this slice of the *new*
2328 /// You can optionally pass a type to this routine and if that type's ABI
2329 /// alignment is itself suitable, this will return zero.
2330 unsigned getSliceAlign(Type *Ty = nullptr) {
2331 unsigned NewAIAlign = NewAI.getAlignment();
2333 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2335 MinAlign(NewAIAlign, NewBeginOffset - NewAllocaBeginOffset);
2336 return (Ty && Align == DL.getABITypeAlignment(Ty)) ? 0 : Align;
2339 unsigned getIndex(uint64_t Offset) {
2340 assert(VecTy && "Can only call getIndex when rewriting a vector");
2341 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2342 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2343 uint32_t Index = RelOffset / ElementSize;
2344 assert(Index * ElementSize == RelOffset);
2348 void deleteIfTriviallyDead(Value *V) {
2349 Instruction *I = cast<Instruction>(V);
2350 if (isInstructionTriviallyDead(I))
2351 Pass.DeadInsts.insert(I);
2354 Value *rewriteVectorizedLoadInst() {
2355 unsigned BeginIndex = getIndex(NewBeginOffset);
2356 unsigned EndIndex = getIndex(NewEndOffset);
2357 assert(EndIndex > BeginIndex && "Empty vector!");
2359 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2360 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2363 Value *rewriteIntegerLoad(LoadInst &LI) {
2364 assert(IntTy && "We cannot insert an integer to the alloca");
2365 assert(!LI.isVolatile());
2366 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2367 V = convertValue(DL, IRB, V, IntTy);
2368 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2369 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2370 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset) {
2371 IntegerType *ExtractTy = Type::getIntNTy(LI.getContext(), SliceSize * 8);
2372 V = extractInteger(DL, IRB, V, ExtractTy, Offset, "extract");
2374 // It is possible that the extracted type is not the load type. This
2375 // happens if there is a load past the end of the alloca, and as
2376 // a consequence the slice is narrower but still a candidate for integer
2377 // lowering. To handle this case, we just zero extend the extracted
2379 assert(cast<IntegerType>(LI.getType())->getBitWidth() >= SliceSize * 8 &&
2380 "Can only handle an extract for an overly wide load");
2381 if (cast<IntegerType>(LI.getType())->getBitWidth() > SliceSize * 8)
2382 V = IRB.CreateZExt(V, LI.getType());
2386 bool visitLoadInst(LoadInst &LI) {
2387 DEBUG(dbgs() << " original: " << LI << "\n");
2388 Value *OldOp = LI.getOperand(0);
2389 assert(OldOp == OldPtr);
2391 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2393 const bool IsLoadPastEnd = DL.getTypeStoreSize(TargetTy) > SliceSize;
2394 bool IsPtrAdjusted = false;
2397 V = rewriteVectorizedLoadInst();
2398 } else if (IntTy && LI.getType()->isIntegerTy()) {
2399 V = rewriteIntegerLoad(LI);
2400 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2401 NewEndOffset == NewAllocaEndOffset &&
2402 (canConvertValue(DL, NewAllocaTy, TargetTy) ||
2403 (IsLoadPastEnd && NewAllocaTy->isIntegerTy() &&
2404 TargetTy->isIntegerTy()))) {
2405 LoadInst *NewLI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2406 LI.isVolatile(), LI.getName());
2407 if (LI.isVolatile())
2408 NewLI->setAtomic(LI.getOrdering(), LI.getSynchScope());
2411 // If this is an integer load past the end of the slice (which means the
2412 // bytes outside the slice are undef or this load is dead) just forcibly
2413 // fix the integer size with correct handling of endianness.
2414 if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2415 if (auto *TITy = dyn_cast<IntegerType>(TargetTy))
2416 if (AITy->getBitWidth() < TITy->getBitWidth()) {
2417 V = IRB.CreateZExt(V, TITy, "load.ext");
2418 if (DL.isBigEndian())
2419 V = IRB.CreateShl(V, TITy->getBitWidth() - AITy->getBitWidth(),
2423 Type *LTy = TargetTy->getPointerTo();
2424 LoadInst *NewLI = IRB.CreateAlignedLoad(getNewAllocaSlicePtr(IRB, LTy),
2425 getSliceAlign(TargetTy),
2426 LI.isVolatile(), LI.getName());
2427 if (LI.isVolatile())
2428 NewLI->setAtomic(LI.getOrdering(), LI.getSynchScope());
2431 IsPtrAdjusted = true;
2433 V = convertValue(DL, IRB, V, TargetTy);
2436 assert(!LI.isVolatile());
2437 assert(LI.getType()->isIntegerTy() &&
2438 "Only integer type loads and stores are split");
2439 assert(SliceSize < DL.getTypeStoreSize(LI.getType()) &&
2440 "Split load isn't smaller than original load");
2441 assert(LI.getType()->getIntegerBitWidth() ==
2442 DL.getTypeStoreSizeInBits(LI.getType()) &&
2443 "Non-byte-multiple bit width");
2444 // Move the insertion point just past the load so that we can refer to it.
2445 IRB.SetInsertPoint(&*std::next(BasicBlock::iterator(&LI)));
2446 // Create a placeholder value with the same type as LI to use as the
2447 // basis for the new value. This allows us to replace the uses of LI with
2448 // the computed value, and then replace the placeholder with LI, leaving
2449 // LI only used for this computation.
2450 Value *Placeholder =
2451 new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2452 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset - BeginOffset,
2454 LI.replaceAllUsesWith(V);
2455 Placeholder->replaceAllUsesWith(&LI);
2458 LI.replaceAllUsesWith(V);
2461 Pass.DeadInsts.insert(&LI);
2462 deleteIfTriviallyDead(OldOp);
2463 DEBUG(dbgs() << " to: " << *V << "\n");
2464 return !LI.isVolatile() && !IsPtrAdjusted;
2467 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp) {
2468 if (V->getType() != VecTy) {
2469 unsigned BeginIndex = getIndex(NewBeginOffset);
2470 unsigned EndIndex = getIndex(NewEndOffset);
2471 assert(EndIndex > BeginIndex && "Empty vector!");
2472 unsigned NumElements = EndIndex - BeginIndex;
2473 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2474 Type *SliceTy = (NumElements == 1)
2476 : VectorType::get(ElementTy, NumElements);
2477 if (V->getType() != SliceTy)
2478 V = convertValue(DL, IRB, V, SliceTy);
2480 // Mix in the existing elements.
2481 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2482 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2484 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2485 Pass.DeadInsts.insert(&SI);
2488 DEBUG(dbgs() << " to: " << *Store << "\n");
2492 bool rewriteIntegerStore(Value *V, StoreInst &SI) {
2493 assert(IntTy && "We cannot extract an integer from the alloca");
2494 assert(!SI.isVolatile());
2495 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2497 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2498 Old = convertValue(DL, IRB, Old, IntTy);
2499 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2500 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2501 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset, "insert");
2503 V = convertValue(DL, IRB, V, NewAllocaTy);
2504 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2505 Pass.DeadInsts.insert(&SI);
2507 DEBUG(dbgs() << " to: " << *Store << "\n");
2511 bool visitStoreInst(StoreInst &SI) {
2512 DEBUG(dbgs() << " original: " << SI << "\n");
2513 Value *OldOp = SI.getOperand(1);
2514 assert(OldOp == OldPtr);
2516 Value *V = SI.getValueOperand();
2518 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2519 // alloca that should be re-examined after promoting this alloca.
2520 if (V->getType()->isPointerTy())
2521 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2522 Pass.PostPromotionWorklist.insert(AI);
2524 if (SliceSize < DL.getTypeStoreSize(V->getType())) {
2525 assert(!SI.isVolatile());
2526 assert(V->getType()->isIntegerTy() &&
2527 "Only integer type loads and stores are split");
2528 assert(V->getType()->getIntegerBitWidth() ==
2529 DL.getTypeStoreSizeInBits(V->getType()) &&
2530 "Non-byte-multiple bit width");
2531 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2532 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset - BeginOffset,
2537 return rewriteVectorizedStoreInst(V, SI, OldOp);
2538 if (IntTy && V->getType()->isIntegerTy())
2539 return rewriteIntegerStore(V, SI);
2541 const bool IsStorePastEnd = DL.getTypeStoreSize(V->getType()) > SliceSize;
2543 if (NewBeginOffset == NewAllocaBeginOffset &&
2544 NewEndOffset == NewAllocaEndOffset &&
2545 (canConvertValue(DL, V->getType(), NewAllocaTy) ||
2546 (IsStorePastEnd && NewAllocaTy->isIntegerTy() &&
2547 V->getType()->isIntegerTy()))) {
2548 // If this is an integer store past the end of slice (and thus the bytes
2549 // past that point are irrelevant or this is unreachable), truncate the
2550 // value prior to storing.
2551 if (auto *VITy = dyn_cast<IntegerType>(V->getType()))
2552 if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2553 if (VITy->getBitWidth() > AITy->getBitWidth()) {
2554 if (DL.isBigEndian())
2555 V = IRB.CreateLShr(V, VITy->getBitWidth() - AITy->getBitWidth(),
2557 V = IRB.CreateTrunc(V, AITy, "load.trunc");
2560 V = convertValue(DL, IRB, V, NewAllocaTy);
2561 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2564 Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo());
2565 NewSI = IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(V->getType()),
2568 if (SI.isVolatile())
2569 NewSI->setAtomic(SI.getOrdering(), SI.getSynchScope());
2570 Pass.DeadInsts.insert(&SI);
2571 deleteIfTriviallyDead(OldOp);
2573 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2574 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2577 /// \brief Compute an integer value from splatting an i8 across the given
2578 /// number of bytes.
2580 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2581 /// call this routine.
2582 /// FIXME: Heed the advice above.
2584 /// \param V The i8 value to splat.
2585 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2586 Value *getIntegerSplat(Value *V, unsigned Size) {
2587 assert(Size > 0 && "Expected a positive number of bytes.");
2588 IntegerType *VTy = cast<IntegerType>(V->getType());
2589 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2593 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size * 8);
2595 IRB.CreateZExt(V, SplatIntTy, "zext"),
2596 ConstantExpr::getUDiv(
2597 Constant::getAllOnesValue(SplatIntTy),
2598 ConstantExpr::getZExt(Constant::getAllOnesValue(V->getType()),
2604 /// \brief Compute a vector splat for a given element value.
2605 Value *getVectorSplat(Value *V, unsigned NumElements) {
2606 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2607 DEBUG(dbgs() << " splat: " << *V << "\n");
2611 bool visitMemSetInst(MemSetInst &II) {
2612 DEBUG(dbgs() << " original: " << II << "\n");
2613 assert(II.getRawDest() == OldPtr);
2615 // If the memset has a variable size, it cannot be split, just adjust the
2616 // pointer to the new alloca.
2617 if (!isa<Constant>(II.getLength())) {
2619 assert(NewBeginOffset == BeginOffset);
2620 II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2621 Type *CstTy = II.getAlignmentCst()->getType();
2622 II.setAlignment(ConstantInt::get(CstTy, getSliceAlign()));
2624 deleteIfTriviallyDead(OldPtr);
2628 // Record this instruction for deletion.
2629 Pass.DeadInsts.insert(&II);
2631 Type *AllocaTy = NewAI.getAllocatedType();
2632 Type *ScalarTy = AllocaTy->getScalarType();
2634 // If this doesn't map cleanly onto the alloca type, and that type isn't
2635 // a single value type, just emit a memset.
2636 if (!VecTy && !IntTy &&
2637 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2638 SliceSize != DL.getTypeStoreSize(AllocaTy) ||
2639 !AllocaTy->isSingleValueType() ||
2640 !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2641 DL.getTypeSizeInBits(ScalarTy) % 8 != 0)) {
2642 Type *SizeTy = II.getLength()->getType();
2643 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2644 CallInst *New = IRB.CreateMemSet(
2645 getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2646 getSliceAlign(), II.isVolatile());
2648 DEBUG(dbgs() << " to: " << *New << "\n");
2652 // If we can represent this as a simple value, we have to build the actual
2653 // value to store, which requires expanding the byte present in memset to
2654 // a sensible representation for the alloca type. This is essentially
2655 // splatting the byte to a sufficiently wide integer, splatting it across
2656 // any desired vector width, and bitcasting to the final type.
2660 // If this is a memset of a vectorized alloca, insert it.
2661 assert(ElementTy == ScalarTy);
2663 unsigned BeginIndex = getIndex(NewBeginOffset);
2664 unsigned EndIndex = getIndex(NewEndOffset);
2665 assert(EndIndex > BeginIndex && "Empty vector!");
2666 unsigned NumElements = EndIndex - BeginIndex;
2667 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2670 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2671 Splat = convertValue(DL, IRB, Splat, ElementTy);
2672 if (NumElements > 1)
2673 Splat = getVectorSplat(Splat, NumElements);
2676 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2677 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2679 // If this is a memset on an alloca where we can widen stores, insert the
2681 assert(!II.isVolatile());
2683 uint64_t Size = NewEndOffset - NewBeginOffset;
2684 V = getIntegerSplat(II.getValue(), Size);
2686 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2687 EndOffset != NewAllocaBeginOffset)) {
2689 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2690 Old = convertValue(DL, IRB, Old, IntTy);
2691 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2692 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2694 assert(V->getType() == IntTy &&
2695 "Wrong type for an alloca wide integer!");
2697 V = convertValue(DL, IRB, V, AllocaTy);
2699 // Established these invariants above.
2700 assert(NewBeginOffset == NewAllocaBeginOffset);
2701 assert(NewEndOffset == NewAllocaEndOffset);
2703 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2704 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2705 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2707 V = convertValue(DL, IRB, V, AllocaTy);
2710 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2713 DEBUG(dbgs() << " to: " << *New << "\n");
2714 return !II.isVolatile();
2717 bool visitMemTransferInst(MemTransferInst &II) {
2718 // Rewriting of memory transfer instructions can be a bit tricky. We break
2719 // them into two categories: split intrinsics and unsplit intrinsics.
2721 DEBUG(dbgs() << " original: " << II << "\n");
2723 bool IsDest = &II.getRawDestUse() == OldUse;
2724 assert((IsDest && II.getRawDest() == OldPtr) ||
2725 (!IsDest && II.getRawSource() == OldPtr));
2727 unsigned SliceAlign = getSliceAlign();
2729 // For unsplit intrinsics, we simply modify the source and destination
2730 // pointers in place. This isn't just an optimization, it is a matter of
2731 // correctness. With unsplit intrinsics we may be dealing with transfers
2732 // within a single alloca before SROA ran, or with transfers that have
2733 // a variable length. We may also be dealing with memmove instead of
2734 // memcpy, and so simply updating the pointers is the necessary for us to
2735 // update both source and dest of a single call.
2736 if (!IsSplittable) {
2737 Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2739 II.setDest(AdjustedPtr);
2741 II.setSource(AdjustedPtr);
2743 if (II.getAlignment() > SliceAlign) {
2744 Type *CstTy = II.getAlignmentCst()->getType();
2746 ConstantInt::get(CstTy, MinAlign(II.getAlignment(), SliceAlign)));
2749 DEBUG(dbgs() << " to: " << II << "\n");
2750 deleteIfTriviallyDead(OldPtr);
2753 // For split transfer intrinsics we have an incredibly useful assurance:
2754 // the source and destination do not reside within the same alloca, and at
2755 // least one of them does not escape. This means that we can replace
2756 // memmove with memcpy, and we don't need to worry about all manner of
2757 // downsides to splitting and transforming the operations.
2759 // If this doesn't map cleanly onto the alloca type, and that type isn't
2760 // a single value type, just emit a memcpy.
2763 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2764 SliceSize != DL.getTypeStoreSize(NewAI.getAllocatedType()) ||
2765 !NewAI.getAllocatedType()->isSingleValueType());
2767 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2768 // size hasn't been shrunk based on analysis of the viable range, this is
2770 if (EmitMemCpy && &OldAI == &NewAI) {
2771 // Ensure the start lines up.
2772 assert(NewBeginOffset == BeginOffset);
2774 // Rewrite the size as needed.
2775 if (NewEndOffset != EndOffset)
2776 II.setLength(ConstantInt::get(II.getLength()->getType(),
2777 NewEndOffset - NewBeginOffset));
2780 // Record this instruction for deletion.
2781 Pass.DeadInsts.insert(&II);
2783 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2784 // alloca that should be re-examined after rewriting this instruction.
2785 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2786 if (AllocaInst *AI =
2787 dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2788 assert(AI != &OldAI && AI != &NewAI &&
2789 "Splittable transfers cannot reach the same alloca on both ends.");
2790 Pass.Worklist.insert(AI);
2793 Type *OtherPtrTy = OtherPtr->getType();
2794 unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2796 // Compute the relative offset for the other pointer within the transfer.
2797 unsigned IntPtrWidth = DL.getPointerSizeInBits(OtherAS);
2798 APInt OtherOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2799 unsigned OtherAlign = MinAlign(II.getAlignment() ? II.getAlignment() : 1,
2800 OtherOffset.zextOrTrunc(64).getZExtValue());
2803 // Compute the other pointer, folding as much as possible to produce
2804 // a single, simple GEP in most cases.
2805 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2806 OtherPtr->getName() + ".");
2808 Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2809 Type *SizeTy = II.getLength()->getType();
2810 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2812 CallInst *New = IRB.CreateMemCpy(
2813 IsDest ? OurPtr : OtherPtr, IsDest ? OtherPtr : OurPtr, Size,
2814 MinAlign(SliceAlign, OtherAlign), II.isVolatile());
2816 DEBUG(dbgs() << " to: " << *New << "\n");
2820 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2821 NewEndOffset == NewAllocaEndOffset;
2822 uint64_t Size = NewEndOffset - NewBeginOffset;
2823 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2824 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2825 unsigned NumElements = EndIndex - BeginIndex;
2826 IntegerType *SubIntTy =
2827 IntTy ? Type::getIntNTy(IntTy->getContext(), Size * 8) : nullptr;
2829 // Reset the other pointer type to match the register type we're going to
2830 // use, but using the address space of the original other pointer.
2831 if (VecTy && !IsWholeAlloca) {
2832 if (NumElements == 1)
2833 OtherPtrTy = VecTy->getElementType();
2835 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2837 OtherPtrTy = OtherPtrTy->getPointerTo(OtherAS);
2838 } else if (IntTy && !IsWholeAlloca) {
2839 OtherPtrTy = SubIntTy->getPointerTo(OtherAS);
2841 OtherPtrTy = NewAllocaTy->getPointerTo(OtherAS);
2844 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2845 OtherPtr->getName() + ".");
2846 unsigned SrcAlign = OtherAlign;
2847 Value *DstPtr = &NewAI;
2848 unsigned DstAlign = SliceAlign;
2850 std::swap(SrcPtr, DstPtr);
2851 std::swap(SrcAlign, DstAlign);
2855 if (VecTy && !IsWholeAlloca && !IsDest) {
2856 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2857 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2858 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2859 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2860 Src = convertValue(DL, IRB, Src, IntTy);
2861 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2862 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2865 IRB.CreateAlignedLoad(SrcPtr, SrcAlign, II.isVolatile(), "copyload");
2868 if (VecTy && !IsWholeAlloca && IsDest) {
2870 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2871 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2872 } else if (IntTy && !IsWholeAlloca && IsDest) {
2874 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2875 Old = convertValue(DL, IRB, Old, IntTy);
2876 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2877 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
2878 Src = convertValue(DL, IRB, Src, NewAllocaTy);
2881 StoreInst *Store = cast<StoreInst>(
2882 IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
2884 DEBUG(dbgs() << " to: " << *Store << "\n");
2885 return !II.isVolatile();
2888 bool visitIntrinsicInst(IntrinsicInst &II) {
2889 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2890 II.getIntrinsicID() == Intrinsic::lifetime_end);
2891 DEBUG(dbgs() << " original: " << II << "\n");
2892 assert(II.getArgOperand(1) == OldPtr);
2894 // Record this instruction for deletion.
2895 Pass.DeadInsts.insert(&II);
2898 ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2899 NewEndOffset - NewBeginOffset);
2900 Value *Ptr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2902 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2903 New = IRB.CreateLifetimeStart(Ptr, Size);
2905 New = IRB.CreateLifetimeEnd(Ptr, Size);
2908 DEBUG(dbgs() << " to: " << *New << "\n");
2912 bool visitPHINode(PHINode &PN) {
2913 DEBUG(dbgs() << " original: " << PN << "\n");
2914 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
2915 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
2917 // We would like to compute a new pointer in only one place, but have it be
2918 // as local as possible to the PHI. To do that, we re-use the location of
2919 // the old pointer, which necessarily must be in the right position to
2920 // dominate the PHI.
2921 IRBuilderTy PtrBuilder(IRB);
2922 if (isa<PHINode>(OldPtr))
2923 PtrBuilder.SetInsertPoint(&*OldPtr->getParent()->getFirstInsertionPt());
2925 PtrBuilder.SetInsertPoint(OldPtr);
2926 PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
2928 Value *NewPtr = getNewAllocaSlicePtr(PtrBuilder, OldPtr->getType());
2929 // Replace the operands which were using the old pointer.
2930 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
2932 DEBUG(dbgs() << " to: " << PN << "\n");
2933 deleteIfTriviallyDead(OldPtr);
2935 // PHIs can't be promoted on their own, but often can be speculated. We
2936 // check the speculation outside of the rewriter so that we see the
2937 // fully-rewritten alloca.
2938 PHIUsers.insert(&PN);
2942 bool visitSelectInst(SelectInst &SI) {
2943 DEBUG(dbgs() << " original: " << SI << "\n");
2944 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
2945 "Pointer isn't an operand!");
2946 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
2947 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
2949 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2950 // Replace the operands which were using the old pointer.
2951 if (SI.getOperand(1) == OldPtr)
2952 SI.setOperand(1, NewPtr);
2953 if (SI.getOperand(2) == OldPtr)
2954 SI.setOperand(2, NewPtr);
2956 DEBUG(dbgs() << " to: " << SI << "\n");
2957 deleteIfTriviallyDead(OldPtr);
2959 // Selects can't be promoted on their own, but often can be speculated. We
2960 // check the speculation outside of the rewriter so that we see the
2961 // fully-rewritten alloca.
2962 SelectUsers.insert(&SI);
2968 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2970 /// This pass aggressively rewrites all aggregate loads and stores on
2971 /// a particular pointer (or any pointer derived from it which we can identify)
2972 /// with scalar loads and stores.
2973 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2974 // Befriend the base class so it can delegate to private visit methods.
2975 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2977 /// Queue of pointer uses to analyze and potentially rewrite.
2978 SmallVector<Use *, 8> Queue;
2980 /// Set to prevent us from cycling with phi nodes and loops.
2981 SmallPtrSet<User *, 8> Visited;
2983 /// The current pointer use being rewritten. This is used to dig up the used
2984 /// value (as opposed to the user).
2988 /// Rewrite loads and stores through a pointer and all pointers derived from
2990 bool rewrite(Instruction &I) {
2991 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2993 bool Changed = false;
2994 while (!Queue.empty()) {
2995 U = Queue.pop_back_val();
2996 Changed |= visit(cast<Instruction>(U->getUser()));
3002 /// Enqueue all the users of the given instruction for further processing.
3003 /// This uses a set to de-duplicate users.
3004 void enqueueUsers(Instruction &I) {
3005 for (Use &U : I.uses())
3006 if (Visited.insert(U.getUser()).second)
3007 Queue.push_back(&U);
3010 // Conservative default is to not rewrite anything.
3011 bool visitInstruction(Instruction &I) { return false; }
3013 /// \brief Generic recursive split emission class.
3014 template <typename Derived> class OpSplitter {
3016 /// The builder used to form new instructions.
3018 /// The indices which to be used with insert- or extractvalue to select the
3019 /// appropriate value within the aggregate.
3020 SmallVector<unsigned, 4> Indices;
3021 /// The indices to a GEP instruction which will move Ptr to the correct slot
3022 /// within the aggregate.
3023 SmallVector<Value *, 4> GEPIndices;
3024 /// The base pointer of the original op, used as a base for GEPing the
3025 /// split operations.
3028 /// Initialize the splitter with an insertion point, Ptr and start with a
3029 /// single zero GEP index.
3030 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
3031 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
3034 /// \brief Generic recursive split emission routine.
3036 /// This method recursively splits an aggregate op (load or store) into
3037 /// scalar or vector ops. It splits recursively until it hits a single value
3038 /// and emits that single value operation via the template argument.
3040 /// The logic of this routine relies on GEPs and insertvalue and
3041 /// extractvalue all operating with the same fundamental index list, merely
3042 /// formatted differently (GEPs need actual values).
3044 /// \param Ty The type being split recursively into smaller ops.
3045 /// \param Agg The aggregate value being built up or stored, depending on
3046 /// whether this is splitting a load or a store respectively.
3047 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3048 if (Ty->isSingleValueType())
3049 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
3051 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3052 unsigned OldSize = Indices.size();
3054 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3056 assert(Indices.size() == OldSize && "Did not return to the old size");
3057 Indices.push_back(Idx);
3058 GEPIndices.push_back(IRB.getInt32(Idx));
3059 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3060 GEPIndices.pop_back();
3066 if (StructType *STy = dyn_cast<StructType>(Ty)) {
3067 unsigned OldSize = Indices.size();
3069 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3071 assert(Indices.size() == OldSize && "Did not return to the old size");
3072 Indices.push_back(Idx);
3073 GEPIndices.push_back(IRB.getInt32(Idx));
3074 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3075 GEPIndices.pop_back();
3081 llvm_unreachable("Only arrays and structs are aggregate loadable types");
3085 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3086 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3087 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
3089 /// Emit a leaf load of a single value. This is called at the leaves of the
3090 /// recursive emission to actually load values.
3091 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3092 assert(Ty->isSingleValueType());
3093 // Load the single value and insert it using the indices.
3095 IRB.CreateInBoundsGEP(nullptr, Ptr, GEPIndices, Name + ".gep");
3096 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
3097 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3098 DEBUG(dbgs() << " to: " << *Load << "\n");
3102 bool visitLoadInst(LoadInst &LI) {
3103 assert(LI.getPointerOperand() == *U);
3104 if (!LI.isSimple() || LI.getType()->isSingleValueType())
3107 // We have an aggregate being loaded, split it apart.
3108 DEBUG(dbgs() << " original: " << LI << "\n");
3109 LoadOpSplitter Splitter(&LI, *U);
3110 Value *V = UndefValue::get(LI.getType());
3111 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3112 LI.replaceAllUsesWith(V);
3113 LI.eraseFromParent();
3117 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3118 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3119 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
3121 /// Emit a leaf store of a single value. This is called at the leaves of the
3122 /// recursive emission to actually produce stores.
3123 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3124 assert(Ty->isSingleValueType());
3125 // Extract the single value and store it using the indices.
3126 Value *Store = IRB.CreateStore(
3127 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
3128 IRB.CreateInBoundsGEP(nullptr, Ptr, GEPIndices, Name + ".gep"));
3130 DEBUG(dbgs() << " to: " << *Store << "\n");
3134 bool visitStoreInst(StoreInst &SI) {
3135 if (!SI.isSimple() || SI.getPointerOperand() != *U)
3137 Value *V = SI.getValueOperand();
3138 if (V->getType()->isSingleValueType())
3141 // We have an aggregate being stored, split it apart.
3142 DEBUG(dbgs() << " original: " << SI << "\n");
3143 StoreOpSplitter Splitter(&SI, *U);
3144 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3145 SI.eraseFromParent();
3149 bool visitBitCastInst(BitCastInst &BC) {
3154 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3159 bool visitPHINode(PHINode &PN) {
3164 bool visitSelectInst(SelectInst &SI) {
3171 /// \brief Strip aggregate type wrapping.
3173 /// This removes no-op aggregate types wrapping an underlying type. It will
3174 /// strip as many layers of types as it can without changing either the type
3175 /// size or the allocated size.
3176 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
3177 if (Ty->isSingleValueType())
3180 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3181 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
3184 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3185 InnerTy = ArrTy->getElementType();
3186 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3187 const StructLayout *SL = DL.getStructLayout(STy);
3188 unsigned Index = SL->getElementContainingOffset(0);
3189 InnerTy = STy->getElementType(Index);
3194 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3195 TypeSize > DL.getTypeSizeInBits(InnerTy))
3198 return stripAggregateTypeWrapping(DL, InnerTy);
3201 /// \brief Try to find a partition of the aggregate type passed in for a given
3202 /// offset and size.
3204 /// This recurses through the aggregate type and tries to compute a subtype
3205 /// based on the offset and size. When the offset and size span a sub-section
3206 /// of an array, it will even compute a new array type for that sub-section,
3207 /// and the same for structs.
3209 /// Note that this routine is very strict and tries to find a partition of the
3210 /// type which produces the *exact* right offset and size. It is not forgiving
3211 /// when the size or offset cause either end of type-based partition to be off.
3212 /// Also, this is a best-effort routine. It is reasonable to give up and not
3213 /// return a type if necessary.
3214 static Type *getTypePartition(const DataLayout &DL, Type *Ty, uint64_t Offset,
3216 if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
3217 return stripAggregateTypeWrapping(DL, Ty);
3218 if (Offset > DL.getTypeAllocSize(Ty) ||
3219 (DL.getTypeAllocSize(Ty) - Offset) < Size)
3222 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3223 // We can't partition pointers...
3224 if (SeqTy->isPointerTy())
3227 Type *ElementTy = SeqTy->getElementType();
3228 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3229 uint64_t NumSkippedElements = Offset / ElementSize;
3230 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) {
3231 if (NumSkippedElements >= ArrTy->getNumElements())
3233 } else if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) {
3234 if (NumSkippedElements >= VecTy->getNumElements())
3237 Offset -= NumSkippedElements * ElementSize;
3239 // First check if we need to recurse.
3240 if (Offset > 0 || Size < ElementSize) {
3241 // Bail if the partition ends in a different array element.
3242 if ((Offset + Size) > ElementSize)
3244 // Recurse through the element type trying to peel off offset bytes.
3245 return getTypePartition(DL, ElementTy, Offset, Size);
3247 assert(Offset == 0);
3249 if (Size == ElementSize)
3250 return stripAggregateTypeWrapping(DL, ElementTy);
3251 assert(Size > ElementSize);
3252 uint64_t NumElements = Size / ElementSize;
3253 if (NumElements * ElementSize != Size)
3255 return ArrayType::get(ElementTy, NumElements);
3258 StructType *STy = dyn_cast<StructType>(Ty);
3262 const StructLayout *SL = DL.getStructLayout(STy);
3263 if (Offset >= SL->getSizeInBytes())
3265 uint64_t EndOffset = Offset + Size;
3266 if (EndOffset > SL->getSizeInBytes())
3269 unsigned Index = SL->getElementContainingOffset(Offset);
3270 Offset -= SL->getElementOffset(Index);
3272 Type *ElementTy = STy->getElementType(Index);
3273 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3274 if (Offset >= ElementSize)
3275 return nullptr; // The offset points into alignment padding.
3277 // See if any partition must be contained by the element.
3278 if (Offset > 0 || Size < ElementSize) {
3279 if ((Offset + Size) > ElementSize)
3281 return getTypePartition(DL, ElementTy, Offset, Size);
3283 assert(Offset == 0);
3285 if (Size == ElementSize)
3286 return stripAggregateTypeWrapping(DL, ElementTy);
3288 StructType::element_iterator EI = STy->element_begin() + Index,
3289 EE = STy->element_end();
3290 if (EndOffset < SL->getSizeInBytes()) {
3291 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3292 if (Index == EndIndex)
3293 return nullptr; // Within a single element and its padding.
3295 // Don't try to form "natural" types if the elements don't line up with the
3297 // FIXME: We could potentially recurse down through the last element in the
3298 // sub-struct to find a natural end point.
3299 if (SL->getElementOffset(EndIndex) != EndOffset)
3302 assert(Index < EndIndex);
3303 EE = STy->element_begin() + EndIndex;
3306 // Try to build up a sub-structure.
3308 StructType::get(STy->getContext(), makeArrayRef(EI, EE), STy->isPacked());
3309 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3310 if (Size != SubSL->getSizeInBytes())
3311 return nullptr; // The sub-struct doesn't have quite the size needed.
3316 /// \brief Pre-split loads and stores to simplify rewriting.
3318 /// We want to break up the splittable load+store pairs as much as
3319 /// possible. This is important to do as a preprocessing step, as once we
3320 /// start rewriting the accesses to partitions of the alloca we lose the
3321 /// necessary information to correctly split apart paired loads and stores
3322 /// which both point into this alloca. The case to consider is something like
3325 /// %a = alloca [12 x i8]
3326 /// %gep1 = getelementptr [12 x i8]* %a, i32 0, i32 0
3327 /// %gep2 = getelementptr [12 x i8]* %a, i32 0, i32 4
3328 /// %gep3 = getelementptr [12 x i8]* %a, i32 0, i32 8
3329 /// %iptr1 = bitcast i8* %gep1 to i64*
3330 /// %iptr2 = bitcast i8* %gep2 to i64*
3331 /// %fptr1 = bitcast i8* %gep1 to float*
3332 /// %fptr2 = bitcast i8* %gep2 to float*
3333 /// %fptr3 = bitcast i8* %gep3 to float*
3334 /// store float 0.0, float* %fptr1
3335 /// store float 1.0, float* %fptr2
3336 /// %v = load i64* %iptr1
3337 /// store i64 %v, i64* %iptr2
3338 /// %f1 = load float* %fptr2
3339 /// %f2 = load float* %fptr3
3341 /// Here we want to form 3 partitions of the alloca, each 4 bytes large, and
3342 /// promote everything so we recover the 2 SSA values that should have been
3343 /// there all along.
3345 /// \returns true if any changes are made.
3346 bool SROA::presplitLoadsAndStores(AllocaInst &AI, AllocaSlices &AS) {
3347 DEBUG(dbgs() << "Pre-splitting loads and stores\n");
3349 // Track the loads and stores which are candidates for pre-splitting here, in
3350 // the order they first appear during the partition scan. These give stable
3351 // iteration order and a basis for tracking which loads and stores we
3353 SmallVector<LoadInst *, 4> Loads;
3354 SmallVector<StoreInst *, 4> Stores;
3356 // We need to accumulate the splits required of each load or store where we
3357 // can find them via a direct lookup. This is important to cross-check loads
3358 // and stores against each other. We also track the slice so that we can kill
3359 // all the slices that end up split.
3360 struct SplitOffsets {
3362 std::vector<uint64_t> Splits;
3364 SmallDenseMap<Instruction *, SplitOffsets, 8> SplitOffsetsMap;
3366 // Track loads out of this alloca which cannot, for any reason, be pre-split.
3367 // This is important as we also cannot pre-split stores of those loads!
3368 // FIXME: This is all pretty gross. It means that we can be more aggressive
3369 // in pre-splitting when the load feeding the store happens to come from
3370 // a separate alloca. Put another way, the effectiveness of SROA would be
3371 // decreased by a frontend which just concatenated all of its local allocas
3372 // into one big flat alloca. But defeating such patterns is exactly the job
3373 // SROA is tasked with! Sadly, to not have this discrepancy we would have
3374 // change store pre-splitting to actually force pre-splitting of the load
3375 // that feeds it *and all stores*. That makes pre-splitting much harder, but
3376 // maybe it would make it more principled?
3377 SmallPtrSet<LoadInst *, 8> UnsplittableLoads;
3379 DEBUG(dbgs() << " Searching for candidate loads and stores\n");
3380 for (auto &P : AS.partitions()) {
3381 for (Slice &S : P) {
3382 Instruction *I = cast<Instruction>(S.getUse()->getUser());
3383 if (!S.isSplittable() ||S.endOffset() <= P.endOffset()) {
3384 // If this was a load we have to track that it can't participate in any
3386 if (auto *LI = dyn_cast<LoadInst>(I))
3387 UnsplittableLoads.insert(LI);
3390 assert(P.endOffset() > S.beginOffset() &&
3391 "Empty or backwards partition!");
3393 // Determine if this is a pre-splittable slice.
3394 if (auto *LI = dyn_cast<LoadInst>(I)) {
3395 assert(!LI->isVolatile() && "Cannot split volatile loads!");
3397 // The load must be used exclusively to store into other pointers for
3398 // us to be able to arbitrarily pre-split it. The stores must also be
3399 // simple to avoid changing semantics.
3400 auto IsLoadSimplyStored = [](LoadInst *LI) {
3401 for (User *LU : LI->users()) {
3402 auto *SI = dyn_cast<StoreInst>(LU);
3403 if (!SI || !SI->isSimple())
3408 if (!IsLoadSimplyStored(LI)) {
3409 UnsplittableLoads.insert(LI);
3413 Loads.push_back(LI);
3414 } else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser())) {
3416 S.getUse() != &SI->getOperandUse(SI->getPointerOperandIndex()))
3418 auto *StoredLoad = dyn_cast<LoadInst>(SI->getValueOperand());
3419 if (!StoredLoad || !StoredLoad->isSimple())
3421 assert(!SI->isVolatile() && "Cannot split volatile stores!");
3423 Stores.push_back(SI);
3425 // Other uses cannot be pre-split.
3429 // Record the initial split.
3430 DEBUG(dbgs() << " Candidate: " << *I << "\n");
3431 auto &Offsets = SplitOffsetsMap[I];
3432 assert(Offsets.Splits.empty() &&
3433 "Should not have splits the first time we see an instruction!");
3435 Offsets.Splits.push_back(P.endOffset() - S.beginOffset());
3438 // Now scan the already split slices, and add a split for any of them which
3439 // we're going to pre-split.
3440 for (Slice *S : P.splitSliceTails()) {
3441 auto SplitOffsetsMapI =
3442 SplitOffsetsMap.find(cast<Instruction>(S->getUse()->getUser()));
3443 if (SplitOffsetsMapI == SplitOffsetsMap.end())
3445 auto &Offsets = SplitOffsetsMapI->second;
3447 assert(Offsets.S == S && "Found a mismatched slice!");
3448 assert(!Offsets.Splits.empty() &&
3449 "Cannot have an empty set of splits on the second partition!");
3450 assert(Offsets.Splits.back() ==
3451 P.beginOffset() - Offsets.S->beginOffset() &&
3452 "Previous split does not end where this one begins!");
3454 // Record each split. The last partition's end isn't needed as the size
3455 // of the slice dictates that.
3456 if (S->endOffset() > P.endOffset())
3457 Offsets.Splits.push_back(P.endOffset() - Offsets.S->beginOffset());
3461 // We may have split loads where some of their stores are split stores. For
3462 // such loads and stores, we can only pre-split them if their splits exactly
3463 // match relative to their starting offset. We have to verify this prior to
3466 std::remove_if(Stores.begin(), Stores.end(),
3467 [&UnsplittableLoads, &SplitOffsetsMap](StoreInst *SI) {
3468 // Lookup the load we are storing in our map of split
3470 auto *LI = cast<LoadInst>(SI->getValueOperand());
3471 // If it was completely unsplittable, then we're done,
3472 // and this store can't be pre-split.
3473 if (UnsplittableLoads.count(LI))
3476 auto LoadOffsetsI = SplitOffsetsMap.find(LI);
3477 if (LoadOffsetsI == SplitOffsetsMap.end())
3478 return false; // Unrelated loads are definitely safe.
3479 auto &LoadOffsets = LoadOffsetsI->second;
3481 // Now lookup the store's offsets.
3482 auto &StoreOffsets = SplitOffsetsMap[SI];
3484 // If the relative offsets of each split in the load and
3485 // store match exactly, then we can split them and we
3486 // don't need to remove them here.
3487 if (LoadOffsets.Splits == StoreOffsets.Splits)
3491 << " Mismatched splits for load and store:\n"
3492 << " " << *LI << "\n"
3493 << " " << *SI << "\n");
3495 // We've found a store and load that we need to split
3496 // with mismatched relative splits. Just give up on them
3497 // and remove both instructions from our list of
3499 UnsplittableLoads.insert(LI);
3503 // Now we have to go *back* through all the stores, because a later store may
3504 // have caused an earlier store's load to become unsplittable and if it is
3505 // unsplittable for the later store, then we can't rely on it being split in
3506 // the earlier store either.
3507 Stores.erase(std::remove_if(Stores.begin(), Stores.end(),
3508 [&UnsplittableLoads](StoreInst *SI) {
3510 cast<LoadInst>(SI->getValueOperand());
3511 return UnsplittableLoads.count(LI);
3514 // Once we've established all the loads that can't be split for some reason,
3515 // filter any that made it into our list out.
3516 Loads.erase(std::remove_if(Loads.begin(), Loads.end(),
3517 [&UnsplittableLoads](LoadInst *LI) {
3518 return UnsplittableLoads.count(LI);
3523 // If no loads or stores are left, there is no pre-splitting to be done for
3525 if (Loads.empty() && Stores.empty())
3528 // From here on, we can't fail and will be building new accesses, so rig up
3530 IRBuilderTy IRB(&AI);
3532 // Collect the new slices which we will merge into the alloca slices.
3533 SmallVector<Slice, 4> NewSlices;
3535 // Track any allocas we end up splitting loads and stores for so we iterate
3537 SmallPtrSet<AllocaInst *, 4> ResplitPromotableAllocas;
3539 // At this point, we have collected all of the loads and stores we can
3540 // pre-split, and the specific splits needed for them. We actually do the
3541 // splitting in a specific order in order to handle when one of the loads in
3542 // the value operand to one of the stores.
3544 // First, we rewrite all of the split loads, and just accumulate each split
3545 // load in a parallel structure. We also build the slices for them and append
3546 // them to the alloca slices.
3547 SmallDenseMap<LoadInst *, std::vector<LoadInst *>, 1> SplitLoadsMap;
3548 std::vector<LoadInst *> SplitLoads;
3549 const DataLayout &DL = AI.getModule()->getDataLayout();
3550 for (LoadInst *LI : Loads) {
3553 IntegerType *Ty = cast<IntegerType>(LI->getType());
3554 uint64_t LoadSize = Ty->getBitWidth() / 8;
3555 assert(LoadSize > 0 && "Cannot have a zero-sized integer load!");
3557 auto &Offsets = SplitOffsetsMap[LI];
3558 assert(LoadSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3559 "Slice size should always match load size exactly!");
3560 uint64_t BaseOffset = Offsets.S->beginOffset();
3561 assert(BaseOffset + LoadSize > BaseOffset &&
3562 "Cannot represent alloca access size using 64-bit integers!");
3564 Instruction *BasePtr = cast<Instruction>(LI->getPointerOperand());
3565 IRB.SetInsertPoint(LI);
3567 DEBUG(dbgs() << " Splitting load: " << *LI << "\n");
3569 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3570 int Idx = 0, Size = Offsets.Splits.size();
3572 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3573 auto *PartPtrTy = PartTy->getPointerTo(LI->getPointerAddressSpace());
3574 LoadInst *PLoad = IRB.CreateAlignedLoad(
3575 getAdjustedPtr(IRB, DL, BasePtr,
3576 APInt(DL.getPointerSizeInBits(), PartOffset),
3577 PartPtrTy, BasePtr->getName() + "."),
3578 getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3581 // Append this load onto the list of split loads so we can find it later
3582 // to rewrite the stores.
3583 SplitLoads.push_back(PLoad);
3585 // Now build a new slice for the alloca.
3586 NewSlices.push_back(
3587 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3588 &PLoad->getOperandUse(PLoad->getPointerOperandIndex()),
3589 /*IsSplittable*/ false));
3590 DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3591 << ", " << NewSlices.back().endOffset() << "): " << *PLoad
3594 // See if we've handled all the splits.
3598 // Setup the next partition.
3599 PartOffset = Offsets.Splits[Idx];
3601 PartSize = (Idx < Size ? Offsets.Splits[Idx] : LoadSize) - PartOffset;
3604 // Now that we have the split loads, do the slow walk over all uses of the
3605 // load and rewrite them as split stores, or save the split loads to use
3606 // below if the store is going to be split there anyways.
3607 bool DeferredStores = false;
3608 for (User *LU : LI->users()) {
3609 StoreInst *SI = cast<StoreInst>(LU);
3610 if (!Stores.empty() && SplitOffsetsMap.count(SI)) {
3611 DeferredStores = true;
3612 DEBUG(dbgs() << " Deferred splitting of store: " << *SI << "\n");
3616 Value *StoreBasePtr = SI->getPointerOperand();
3617 IRB.SetInsertPoint(SI);
3619 DEBUG(dbgs() << " Splitting store of load: " << *SI << "\n");
3621 for (int Idx = 0, Size = SplitLoads.size(); Idx < Size; ++Idx) {
3622 LoadInst *PLoad = SplitLoads[Idx];
3623 uint64_t PartOffset = Idx == 0 ? 0 : Offsets.Splits[Idx - 1];
3625 PLoad->getType()->getPointerTo(SI->getPointerAddressSpace());
3627 StoreInst *PStore = IRB.CreateAlignedStore(
3628 PLoad, getAdjustedPtr(IRB, DL, StoreBasePtr,
3629 APInt(DL.getPointerSizeInBits(), PartOffset),
3630 PartPtrTy, StoreBasePtr->getName() + "."),
3631 getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3633 DEBUG(dbgs() << " +" << PartOffset << ":" << *PStore << "\n");
3636 // We want to immediately iterate on any allocas impacted by splitting
3637 // this store, and we have to track any promotable alloca (indicated by
3638 // a direct store) as needing to be resplit because it is no longer
3640 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(StoreBasePtr)) {
3641 ResplitPromotableAllocas.insert(OtherAI);
3642 Worklist.insert(OtherAI);
3643 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3644 StoreBasePtr->stripInBoundsOffsets())) {
3645 Worklist.insert(OtherAI);
3648 // Mark the original store as dead.
3649 DeadInsts.insert(SI);
3652 // Save the split loads if there are deferred stores among the users.
3654 SplitLoadsMap.insert(std::make_pair(LI, std::move(SplitLoads)));
3656 // Mark the original load as dead and kill the original slice.
3657 DeadInsts.insert(LI);
3661 // Second, we rewrite all of the split stores. At this point, we know that
3662 // all loads from this alloca have been split already. For stores of such
3663 // loads, we can simply look up the pre-existing split loads. For stores of
3664 // other loads, we split those loads first and then write split stores of
3666 for (StoreInst *SI : Stores) {
3667 auto *LI = cast<LoadInst>(SI->getValueOperand());
3668 IntegerType *Ty = cast<IntegerType>(LI->getType());
3669 uint64_t StoreSize = Ty->getBitWidth() / 8;
3670 assert(StoreSize > 0 && "Cannot have a zero-sized integer store!");
3672 auto &Offsets = SplitOffsetsMap[SI];
3673 assert(StoreSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3674 "Slice size should always match load size exactly!");
3675 uint64_t BaseOffset = Offsets.S->beginOffset();
3676 assert(BaseOffset + StoreSize > BaseOffset &&
3677 "Cannot represent alloca access size using 64-bit integers!");
3679 Value *LoadBasePtr = LI->getPointerOperand();
3680 Instruction *StoreBasePtr = cast<Instruction>(SI->getPointerOperand());
3682 DEBUG(dbgs() << " Splitting store: " << *SI << "\n");
3684 // Check whether we have an already split load.
3685 auto SplitLoadsMapI = SplitLoadsMap.find(LI);
3686 std::vector<LoadInst *> *SplitLoads = nullptr;
3687 if (SplitLoadsMapI != SplitLoadsMap.end()) {
3688 SplitLoads = &SplitLoadsMapI->second;
3689 assert(SplitLoads->size() == Offsets.Splits.size() + 1 &&
3690 "Too few split loads for the number of splits in the store!");
3692 DEBUG(dbgs() << " of load: " << *LI << "\n");
3695 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3696 int Idx = 0, Size = Offsets.Splits.size();
3698 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3699 auto *PartPtrTy = PartTy->getPointerTo(SI->getPointerAddressSpace());
3701 // Either lookup a split load or create one.
3704 PLoad = (*SplitLoads)[Idx];
3706 IRB.SetInsertPoint(LI);
3707 PLoad = IRB.CreateAlignedLoad(
3708 getAdjustedPtr(IRB, DL, LoadBasePtr,
3709 APInt(DL.getPointerSizeInBits(), PartOffset),
3710 PartPtrTy, LoadBasePtr->getName() + "."),
3711 getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3715 // And store this partition.
3716 IRB.SetInsertPoint(SI);
3717 StoreInst *PStore = IRB.CreateAlignedStore(
3718 PLoad, getAdjustedPtr(IRB, DL, StoreBasePtr,
3719 APInt(DL.getPointerSizeInBits(), PartOffset),
3720 PartPtrTy, StoreBasePtr->getName() + "."),
3721 getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3723 // Now build a new slice for the alloca.
3724 NewSlices.push_back(
3725 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3726 &PStore->getOperandUse(PStore->getPointerOperandIndex()),
3727 /*IsSplittable*/ false));
3728 DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3729 << ", " << NewSlices.back().endOffset() << "): " << *PStore
3732 DEBUG(dbgs() << " of split load: " << *PLoad << "\n");
3735 // See if we've finished all the splits.
3739 // Setup the next partition.
3740 PartOffset = Offsets.Splits[Idx];
3742 PartSize = (Idx < Size ? Offsets.Splits[Idx] : StoreSize) - PartOffset;
3745 // We want to immediately iterate on any allocas impacted by splitting
3746 // this load, which is only relevant if it isn't a load of this alloca and
3747 // thus we didn't already split the loads above. We also have to keep track
3748 // of any promotable allocas we split loads on as they can no longer be
3751 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(LoadBasePtr)) {
3752 assert(OtherAI != &AI && "We can't re-split our own alloca!");
3753 ResplitPromotableAllocas.insert(OtherAI);
3754 Worklist.insert(OtherAI);
3755 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3756 LoadBasePtr->stripInBoundsOffsets())) {
3757 assert(OtherAI != &AI && "We can't re-split our own alloca!");
3758 Worklist.insert(OtherAI);
3762 // Mark the original store as dead now that we've split it up and kill its
3763 // slice. Note that we leave the original load in place unless this store
3764 // was its only use. It may in turn be split up if it is an alloca load
3765 // for some other alloca, but it may be a normal load. This may introduce
3766 // redundant loads, but where those can be merged the rest of the optimizer
3767 // should handle the merging, and this uncovers SSA splits which is more
3768 // important. In practice, the original loads will almost always be fully
3769 // split and removed eventually, and the splits will be merged by any
3770 // trivial CSE, including instcombine.
3771 if (LI->hasOneUse()) {
3772 assert(*LI->user_begin() == SI && "Single use isn't this store!");
3773 DeadInsts.insert(LI);
3775 DeadInsts.insert(SI);
3779 // Remove the killed slices that have ben pre-split.
3780 AS.erase(std::remove_if(AS.begin(), AS.end(), [](const Slice &S) {
3784 // Insert our new slices. This will sort and merge them into the sorted
3786 AS.insert(NewSlices);
3788 DEBUG(dbgs() << " Pre-split slices:\n");
3790 for (auto I = AS.begin(), E = AS.end(); I != E; ++I)
3791 DEBUG(AS.print(dbgs(), I, " "));
3794 // Finally, don't try to promote any allocas that new require re-splitting.
3795 // They have already been added to the worklist above.
3796 PromotableAllocas.erase(
3798 PromotableAllocas.begin(), PromotableAllocas.end(),
3799 [&](AllocaInst *AI) { return ResplitPromotableAllocas.count(AI); }),
3800 PromotableAllocas.end());
3805 /// \brief Rewrite an alloca partition's users.
3807 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3808 /// to rewrite uses of an alloca partition to be conducive for SSA value
3809 /// promotion. If the partition needs a new, more refined alloca, this will
3810 /// build that new alloca, preserving as much type information as possible, and
3811 /// rewrite the uses of the old alloca to point at the new one and have the
3812 /// appropriate new offsets. It also evaluates how successful the rewrite was
3813 /// at enabling promotion and if it was successful queues the alloca to be
3815 AllocaInst *SROA::rewritePartition(AllocaInst &AI, AllocaSlices &AS,
3817 // Try to compute a friendly type for this partition of the alloca. This
3818 // won't always succeed, in which case we fall back to a legal integer type
3819 // or an i8 array of an appropriate size.
3820 Type *SliceTy = nullptr;
3821 const DataLayout &DL = AI.getModule()->getDataLayout();
3822 if (Type *CommonUseTy = findCommonType(P.begin(), P.end(), P.endOffset()))
3823 if (DL.getTypeAllocSize(CommonUseTy) >= P.size())
3824 SliceTy = CommonUseTy;
3826 if (Type *TypePartitionTy = getTypePartition(DL, AI.getAllocatedType(),
3827 P.beginOffset(), P.size()))
3828 SliceTy = TypePartitionTy;
3829 if ((!SliceTy || (SliceTy->isArrayTy() &&
3830 SliceTy->getArrayElementType()->isIntegerTy())) &&
3831 DL.isLegalInteger(P.size() * 8))
3832 SliceTy = Type::getIntNTy(*C, P.size() * 8);
3834 SliceTy = ArrayType::get(Type::getInt8Ty(*C), P.size());
3835 assert(DL.getTypeAllocSize(SliceTy) >= P.size());
3837 bool IsIntegerPromotable = isIntegerWideningViable(P, SliceTy, DL);
3840 IsIntegerPromotable ? nullptr : isVectorPromotionViable(P, DL);
3844 // Check for the case where we're going to rewrite to a new alloca of the
3845 // exact same type as the original, and with the same access offsets. In that
3846 // case, re-use the existing alloca, but still run through the rewriter to
3847 // perform phi and select speculation.
3849 if (SliceTy == AI.getAllocatedType()) {
3850 assert(P.beginOffset() == 0 &&
3851 "Non-zero begin offset but same alloca type");
3853 // FIXME: We should be able to bail at this point with "nothing changed".
3854 // FIXME: We might want to defer PHI speculation until after here.
3855 // FIXME: return nullptr;
3857 unsigned Alignment = AI.getAlignment();
3859 // The minimum alignment which users can rely on when the explicit
3860 // alignment is omitted or zero is that required by the ABI for this
3862 Alignment = DL.getABITypeAlignment(AI.getAllocatedType());
3864 Alignment = MinAlign(Alignment, P.beginOffset());
3865 // If we will get at least this much alignment from the type alone, leave
3866 // the alloca's alignment unconstrained.
3867 if (Alignment <= DL.getABITypeAlignment(SliceTy))
3869 NewAI = new AllocaInst(
3870 SliceTy, nullptr, Alignment,
3871 AI.getName() + ".sroa." + Twine(P.begin() - AS.begin()), &AI);
3875 DEBUG(dbgs() << "Rewriting alloca partition "
3876 << "[" << P.beginOffset() << "," << P.endOffset()
3877 << ") to: " << *NewAI << "\n");
3879 // Track the high watermark on the worklist as it is only relevant for
3880 // promoted allocas. We will reset it to this point if the alloca is not in
3881 // fact scheduled for promotion.
3882 unsigned PPWOldSize = PostPromotionWorklist.size();
3883 unsigned NumUses = 0;
3884 SmallPtrSet<PHINode *, 8> PHIUsers;
3885 SmallPtrSet<SelectInst *, 8> SelectUsers;
3887 AllocaSliceRewriter Rewriter(DL, AS, *this, AI, *NewAI, P.beginOffset(),
3888 P.endOffset(), IsIntegerPromotable, VecTy,
3889 PHIUsers, SelectUsers);
3890 bool Promotable = true;
3891 for (Slice *S : P.splitSliceTails()) {
3892 Promotable &= Rewriter.visit(S);
3895 for (Slice &S : P) {
3896 Promotable &= Rewriter.visit(&S);
3900 NumAllocaPartitionUses += NumUses;
3901 MaxUsesPerAllocaPartition =
3902 std::max<unsigned>(NumUses, MaxUsesPerAllocaPartition);
3904 // Now that we've processed all the slices in the new partition, check if any
3905 // PHIs or Selects would block promotion.
3906 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
3909 if (!isSafePHIToSpeculate(**I)) {
3912 SelectUsers.clear();
3915 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
3916 E = SelectUsers.end();
3918 if (!isSafeSelectToSpeculate(**I)) {
3921 SelectUsers.clear();
3926 if (PHIUsers.empty() && SelectUsers.empty()) {
3927 // Promote the alloca.
3928 PromotableAllocas.push_back(NewAI);
3930 // If we have either PHIs or Selects to speculate, add them to those
3931 // worklists and re-queue the new alloca so that we promote in on the
3933 for (PHINode *PHIUser : PHIUsers)
3934 SpeculatablePHIs.insert(PHIUser);
3935 for (SelectInst *SelectUser : SelectUsers)
3936 SpeculatableSelects.insert(SelectUser);
3937 Worklist.insert(NewAI);
3940 // If we can't promote the alloca, iterate on it to check for new
3941 // refinements exposed by splitting the current alloca. Don't iterate on an
3942 // alloca which didn't actually change and didn't get promoted.
3944 Worklist.insert(NewAI);
3946 // Drop any post-promotion work items if promotion didn't happen.
3947 while (PostPromotionWorklist.size() > PPWOldSize)
3948 PostPromotionWorklist.pop_back();
3954 /// \brief Walks the slices of an alloca and form partitions based on them,
3955 /// rewriting each of their uses.
3956 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &AS) {
3957 if (AS.begin() == AS.end())
3960 unsigned NumPartitions = 0;
3961 bool Changed = false;
3962 const DataLayout &DL = AI.getModule()->getDataLayout();
3964 // First try to pre-split loads and stores.
3965 Changed |= presplitLoadsAndStores(AI, AS);
3967 // Now that we have identified any pre-splitting opportunities, mark any
3968 // splittable (non-whole-alloca) loads and stores as unsplittable. If we fail
3969 // to split these during pre-splitting, we want to force them to be
3970 // rewritten into a partition.
3971 bool IsSorted = true;
3972 for (Slice &S : AS) {
3973 if (!S.isSplittable())
3975 // FIXME: We currently leave whole-alloca splittable loads and stores. This
3976 // used to be the only splittable loads and stores and we need to be
3977 // confident that the above handling of splittable loads and stores is
3978 // completely sufficient before we forcibly disable the remaining handling.
3979 if (S.beginOffset() == 0 &&
3980 S.endOffset() >= DL.getTypeAllocSize(AI.getAllocatedType()))
3982 if (isa<LoadInst>(S.getUse()->getUser()) ||
3983 isa<StoreInst>(S.getUse()->getUser())) {
3984 S.makeUnsplittable();
3989 std::sort(AS.begin(), AS.end());
3991 /// \brief Describes the allocas introduced by rewritePartition
3992 /// in order to migrate the debug info.
3997 Piece(AllocaInst *AI, uint64_t O, uint64_t S)
3998 : Alloca(AI), Offset(O), Size(S) {}
4000 SmallVector<Piece, 4> Pieces;
4002 // Rewrite each partition.
4003 for (auto &P : AS.partitions()) {
4004 if (AllocaInst *NewAI = rewritePartition(AI, AS, P)) {
4007 uint64_t SizeOfByte = 8;
4008 uint64_t AllocaSize = DL.getTypeSizeInBits(NewAI->getAllocatedType());
4009 // Don't include any padding.
4010 uint64_t Size = std::min(AllocaSize, P.size() * SizeOfByte);
4011 Pieces.push_back(Piece(NewAI, P.beginOffset() * SizeOfByte, Size));
4017 NumAllocaPartitions += NumPartitions;
4018 MaxPartitionsPerAlloca =
4019 std::max<unsigned>(NumPartitions, MaxPartitionsPerAlloca);
4021 // Migrate debug information from the old alloca to the new alloca(s)
4022 // and the individual partitions.
4023 if (DbgDeclareInst *DbgDecl = FindAllocaDbgDeclare(&AI)) {
4024 auto *Var = DbgDecl->getVariable();
4025 auto *Expr = DbgDecl->getExpression();
4026 DIBuilder DIB(*AI.getParent()->getParent()->getParent(),
4027 /*AllowUnresolved*/ false);
4028 bool IsSplit = Pieces.size() > 1;
4029 for (auto Piece : Pieces) {
4030 // Create a piece expression describing the new partition or reuse AI's
4031 // expression if there is only one partition.
4032 auto *PieceExpr = Expr;
4033 if (IsSplit || Expr->isBitPiece()) {
4034 // If this alloca is already a scalar replacement of a larger aggregate,
4035 // Piece.Offset describes the offset inside the scalar.
4036 uint64_t Offset = Expr->isBitPiece() ? Expr->getBitPieceOffset() : 0;
4037 uint64_t Start = Offset + Piece.Offset;
4038 uint64_t Size = Piece.Size;
4039 if (Expr->isBitPiece()) {
4040 uint64_t AbsEnd = Expr->getBitPieceOffset() + Expr->getBitPieceSize();
4041 if (Start >= AbsEnd)
4042 // No need to describe a SROAed padding.
4044 Size = std::min(Size, AbsEnd - Start);
4046 PieceExpr = DIB.createBitPieceExpression(Start, Size);
4049 // Remove any existing dbg.declare intrinsic describing the same alloca.
4050 if (DbgDeclareInst *OldDDI = FindAllocaDbgDeclare(Piece.Alloca))
4051 OldDDI->eraseFromParent();
4053 DIB.insertDeclare(Piece.Alloca, Var, PieceExpr, DbgDecl->getDebugLoc(),
4060 /// \brief Clobber a use with undef, deleting the used value if it becomes dead.
4061 void SROA::clobberUse(Use &U) {
4063 // Replace the use with an undef value.
4064 U = UndefValue::get(OldV->getType());
4066 // Check for this making an instruction dead. We have to garbage collect
4067 // all the dead instructions to ensure the uses of any alloca end up being
4069 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
4070 if (isInstructionTriviallyDead(OldI)) {
4071 DeadInsts.insert(OldI);
4075 /// \brief Analyze an alloca for SROA.
4077 /// This analyzes the alloca to ensure we can reason about it, builds
4078 /// the slices of the alloca, and then hands it off to be split and
4079 /// rewritten as needed.
4080 bool SROA::runOnAlloca(AllocaInst &AI) {
4081 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
4082 ++NumAllocasAnalyzed;
4084 // Special case dead allocas, as they're trivial.
4085 if (AI.use_empty()) {
4086 AI.eraseFromParent();
4089 const DataLayout &DL = AI.getModule()->getDataLayout();
4091 // Skip alloca forms that this analysis can't handle.
4092 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
4093 DL.getTypeAllocSize(AI.getAllocatedType()) == 0)
4096 bool Changed = false;
4098 // First, split any FCA loads and stores touching this alloca to promote
4099 // better splitting and promotion opportunities.
4100 AggLoadStoreRewriter AggRewriter;
4101 Changed |= AggRewriter.rewrite(AI);
4103 // Build the slices using a recursive instruction-visiting builder.
4104 AllocaSlices AS(DL, AI);
4105 DEBUG(AS.print(dbgs()));
4109 // Delete all the dead users of this alloca before splitting and rewriting it.
4110 for (Instruction *DeadUser : AS.getDeadUsers()) {
4111 // Free up everything used by this instruction.
4112 for (Use &DeadOp : DeadUser->operands())
4115 // Now replace the uses of this instruction.
4116 DeadUser->replaceAllUsesWith(UndefValue::get(DeadUser->getType()));
4118 // And mark it for deletion.
4119 DeadInsts.insert(DeadUser);
4122 for (Use *DeadOp : AS.getDeadOperands()) {
4123 clobberUse(*DeadOp);
4127 // No slices to split. Leave the dead alloca for a later pass to clean up.
4128 if (AS.begin() == AS.end())
4131 Changed |= splitAlloca(AI, AS);
4133 DEBUG(dbgs() << " Speculating PHIs\n");
4134 while (!SpeculatablePHIs.empty())
4135 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
4137 DEBUG(dbgs() << " Speculating Selects\n");
4138 while (!SpeculatableSelects.empty())
4139 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
4144 /// \brief Delete the dead instructions accumulated in this run.
4146 /// Recursively deletes the dead instructions we've accumulated. This is done
4147 /// at the very end to maximize locality of the recursive delete and to
4148 /// minimize the problems of invalidated instruction pointers as such pointers
4149 /// are used heavily in the intermediate stages of the algorithm.
4151 /// We also record the alloca instructions deleted here so that they aren't
4152 /// subsequently handed to mem2reg to promote.
4153 void SROA::deleteDeadInstructions(
4154 SmallPtrSetImpl<AllocaInst *> &DeletedAllocas) {
4155 while (!DeadInsts.empty()) {
4156 Instruction *I = DeadInsts.pop_back_val();
4157 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
4159 I->replaceAllUsesWith(UndefValue::get(I->getType()));
4161 for (Use &Operand : I->operands())
4162 if (Instruction *U = dyn_cast<Instruction>(Operand)) {
4163 // Zero out the operand and see if it becomes trivially dead.
4165 if (isInstructionTriviallyDead(U))
4166 DeadInsts.insert(U);
4169 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
4170 DeletedAllocas.insert(AI);
4171 if (DbgDeclareInst *DbgDecl = FindAllocaDbgDeclare(AI))
4172 DbgDecl->eraseFromParent();
4176 I->eraseFromParent();
4180 /// \brief Promote the allocas, using the best available technique.
4182 /// This attempts to promote whatever allocas have been identified as viable in
4183 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
4184 /// This function returns whether any promotion occurred.
4185 bool SROA::promoteAllocas(Function &F) {
4186 if (PromotableAllocas.empty())
4189 NumPromoted += PromotableAllocas.size();
4191 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
4192 PromoteMemToReg(PromotableAllocas, *DT, nullptr, AC);
4193 PromotableAllocas.clear();
4197 PreservedAnalyses SROA::runImpl(Function &F, DominatorTree &RunDT,
4198 AssumptionCache &RunAC) {
4199 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
4200 C = &F.getContext();
4204 BasicBlock &EntryBB = F.getEntryBlock();
4205 for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
4207 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
4208 Worklist.insert(AI);
4211 bool Changed = false;
4212 // A set of deleted alloca instruction pointers which should be removed from
4213 // the list of promotable allocas.
4214 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
4217 while (!Worklist.empty()) {
4218 Changed |= runOnAlloca(*Worklist.pop_back_val());
4219 deleteDeadInstructions(DeletedAllocas);
4221 // Remove the deleted allocas from various lists so that we don't try to
4222 // continue processing them.
4223 if (!DeletedAllocas.empty()) {
4224 auto IsInSet = [&](AllocaInst *AI) { return DeletedAllocas.count(AI); };
4225 Worklist.remove_if(IsInSet);
4226 PostPromotionWorklist.remove_if(IsInSet);
4227 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
4228 PromotableAllocas.end(),
4230 PromotableAllocas.end());
4231 DeletedAllocas.clear();
4235 Changed |= promoteAllocas(F);
4237 Worklist = PostPromotionWorklist;
4238 PostPromotionWorklist.clear();
4239 } while (!Worklist.empty());
4241 // FIXME: Even when promoting allocas we should preserve some abstract set of
4242 // CFG-specific analyses.
4243 return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all();
4246 PreservedAnalyses SROA::run(Function &F, AnalysisManager<Function> *AM) {
4247 return runImpl(F, AM->getResult<DominatorTreeAnalysis>(F),
4248 AM->getResult<AssumptionAnalysis>(F));
4251 /// A legacy pass for the legacy pass manager that wraps the \c SROA pass.
4253 /// This is in the llvm namespace purely to allow it to be a friend of the \c
4255 class llvm::sroa::SROALegacyPass : public FunctionPass {
4256 /// The SROA implementation.
4260 SROALegacyPass() : FunctionPass(ID) {
4261 initializeSROALegacyPassPass(*PassRegistry::getPassRegistry());
4263 bool runOnFunction(Function &F) override {
4264 if (skipOptnoneFunction(F))
4267 auto PA = Impl.runImpl(
4268 F, getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
4269 getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F));
4270 return !PA.areAllPreserved();
4272 void getAnalysisUsage(AnalysisUsage &AU) const override {
4273 AU.addRequired<AssumptionCacheTracker>();
4274 AU.addRequired<DominatorTreeWrapperPass>();
4275 AU.addPreserved<GlobalsAAWrapperPass>();
4276 AU.setPreservesCFG();
4279 const char *getPassName() const override { return "SROA"; }
4283 char SROALegacyPass::ID = 0;
4285 FunctionPass *llvm::createSROAPass() { return new SROALegacyPass(); }
4287 INITIALIZE_PASS_BEGIN(SROALegacyPass, "sroa",
4288 "Scalar Replacement Of Aggregates", false, false)
4289 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
4290 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
4291 INITIALIZE_PASS_END(SROALegacyPass, "sroa", "Scalar Replacement Of Aggregates",