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/Chrono.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Compiler.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MathExtras.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"
59 // We only use this for a debug check.
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, but only in
92 class IRBuilderPrefixedInserter : public IRBuilderDefaultInserter {
94 const Twine getNameWithPrefix(const Twine &Name) const {
95 return Name.isTriviallyEmpty() ? Name : Prefix + Name;
99 void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
102 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
103 BasicBlock::iterator InsertPt) const {
104 IRBuilderDefaultInserter::InsertHelper(I, getNameWithPrefix(Name), BB,
109 /// \brief Provide a typedef for IRBuilder that drops names in release builds.
110 using IRBuilderTy = llvm::IRBuilder<ConstantFolder, IRBuilderPrefixedInserter>;
114 /// \brief A used slice of an alloca.
116 /// This structure represents a slice of an alloca used by some instruction. It
117 /// stores both the begin and end offsets of this use, a pointer to the use
118 /// itself, and a flag indicating whether we can classify the use as splittable
119 /// or not when forming partitions of the alloca.
121 /// \brief The beginning offset of the range.
122 uint64_t BeginOffset;
124 /// \brief The ending offset, not included in the range.
127 /// \brief Storage for both the use of this slice and whether it can be
129 PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
132 Slice() : BeginOffset(), EndOffset() {}
133 Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
134 : BeginOffset(BeginOffset), EndOffset(EndOffset),
135 UseAndIsSplittable(U, IsSplittable) {}
137 uint64_t beginOffset() const { return BeginOffset; }
138 uint64_t endOffset() const { return EndOffset; }
140 bool isSplittable() const { return UseAndIsSplittable.getInt(); }
141 void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
143 Use *getUse() const { return UseAndIsSplittable.getPointer(); }
145 bool isDead() const { return getUse() == nullptr; }
146 void kill() { UseAndIsSplittable.setPointer(nullptr); }
148 /// \brief Support for ordering ranges.
150 /// This provides an ordering over ranges such that start offsets are
151 /// always increasing, and within equal start offsets, the end offsets are
152 /// decreasing. Thus the spanning range comes first in a cluster with the
153 /// same start position.
154 bool operator<(const Slice &RHS) const {
155 if (beginOffset() < RHS.beginOffset())
157 if (beginOffset() > RHS.beginOffset())
159 if (isSplittable() != RHS.isSplittable())
160 return !isSplittable();
161 if (endOffset() > RHS.endOffset())
166 /// \brief Support comparison with a single offset to allow binary searches.
167 friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Slice &LHS,
168 uint64_t RHSOffset) {
169 return LHS.beginOffset() < RHSOffset;
171 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
173 return LHSOffset < RHS.beginOffset();
176 bool operator==(const Slice &RHS) const {
177 return isSplittable() == RHS.isSplittable() &&
178 beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
180 bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
182 } // end anonymous namespace
185 template <typename T> struct isPodLike;
186 template <> struct isPodLike<Slice> { static const bool value = true; };
189 /// \brief Representation of the alloca slices.
191 /// This class represents the slices of an alloca which are formed by its
192 /// various uses. If a pointer escapes, we can't fully build a representation
193 /// for the slices used and we reflect that in this structure. The uses are
194 /// stored, sorted by increasing beginning offset and with unsplittable slices
195 /// starting at a particular offset before splittable slices.
196 class llvm::sroa::AllocaSlices {
198 /// \brief Construct the slices of a particular alloca.
199 AllocaSlices(const DataLayout &DL, AllocaInst &AI);
201 /// \brief Test whether a pointer to the allocation escapes our analysis.
203 /// If this is true, the slices are never fully built and should be
205 bool isEscaped() const { return PointerEscapingInstr; }
207 /// \brief Support for iterating over the slices.
209 typedef SmallVectorImpl<Slice>::iterator iterator;
210 typedef iterator_range<iterator> range;
211 iterator begin() { return Slices.begin(); }
212 iterator end() { return Slices.end(); }
214 typedef SmallVectorImpl<Slice>::const_iterator const_iterator;
215 typedef iterator_range<const_iterator> const_range;
216 const_iterator begin() const { return Slices.begin(); }
217 const_iterator end() const { return Slices.end(); }
220 /// \brief Erase a range of slices.
221 void erase(iterator Start, iterator Stop) { Slices.erase(Start, Stop); }
223 /// \brief Insert new slices for this alloca.
225 /// This moves the slices into the alloca's slices collection, and re-sorts
226 /// everything so that the usual ordering properties of the alloca's slices
228 void insert(ArrayRef<Slice> NewSlices) {
229 int OldSize = Slices.size();
230 Slices.append(NewSlices.begin(), NewSlices.end());
231 auto SliceI = Slices.begin() + OldSize;
232 std::sort(SliceI, Slices.end());
233 std::inplace_merge(Slices.begin(), SliceI, Slices.end());
236 // Forward declare the iterator and range accessor for walking the
238 class partition_iterator;
239 iterator_range<partition_iterator> partitions();
241 /// \brief Access the dead users for this alloca.
242 ArrayRef<Instruction *> getDeadUsers() const { return DeadUsers; }
244 /// \brief Access the dead operands referring to this alloca.
246 /// These are operands which have cannot actually be used to refer to the
247 /// alloca as they are outside its range and the user doesn't correct for
248 /// that. These mostly consist of PHI node inputs and the like which we just
249 /// need to replace with undef.
250 ArrayRef<Use *> getDeadOperands() const { return DeadOperands; }
252 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
253 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
254 void printSlice(raw_ostream &OS, const_iterator I,
255 StringRef Indent = " ") const;
256 void printUse(raw_ostream &OS, const_iterator I,
257 StringRef Indent = " ") const;
258 void print(raw_ostream &OS) const;
259 void dump(const_iterator I) const;
264 template <typename DerivedT, typename RetT = void> class BuilderBase;
266 friend class AllocaSlices::SliceBuilder;
268 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
269 /// \brief Handle to alloca instruction to simplify method interfaces.
273 /// \brief The instruction responsible for this alloca not having a known set
276 /// When an instruction (potentially) escapes the pointer to the alloca, we
277 /// store a pointer to that here and abort trying to form slices of the
278 /// alloca. This will be null if the alloca slices are analyzed successfully.
279 Instruction *PointerEscapingInstr;
281 /// \brief The slices of the alloca.
283 /// We store a vector of the slices formed by uses of the alloca here. This
284 /// vector is sorted by increasing begin offset, and then the unsplittable
285 /// slices before the splittable ones. See the Slice inner class for more
287 SmallVector<Slice, 8> Slices;
289 /// \brief Instructions which will become dead if we rewrite the alloca.
291 /// Note that these are not separated by slice. This is because we expect an
292 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
293 /// all these instructions can simply be removed and replaced with undef as
294 /// they come from outside of the allocated space.
295 SmallVector<Instruction *, 8> DeadUsers;
297 /// \brief Operands which will become dead if we rewrite the alloca.
299 /// These are operands that in their particular use can be replaced with
300 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
301 /// to PHI nodes and the like. They aren't entirely dead (there might be
302 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
303 /// want to swap this particular input for undef to simplify the use lists of
305 SmallVector<Use *, 8> DeadOperands;
308 /// \brief A partition of the slices.
310 /// An ephemeral representation for a range of slices which can be viewed as
311 /// a partition of the alloca. This range represents a span of the alloca's
312 /// memory which cannot be split, and provides access to all of the slices
313 /// overlapping some part of the partition.
315 /// Objects of this type are produced by traversing the alloca's slices, but
316 /// are only ephemeral and not persistent.
317 class llvm::sroa::Partition {
319 friend class AllocaSlices;
320 friend class AllocaSlices::partition_iterator;
322 typedef AllocaSlices::iterator iterator;
324 /// \brief The beginning and ending offsets of the alloca for this
326 uint64_t BeginOffset, EndOffset;
328 /// \brief The start end end iterators of this partition.
331 /// \brief A collection of split slice tails overlapping the partition.
332 SmallVector<Slice *, 4> SplitTails;
334 /// \brief Raw constructor builds an empty partition starting and ending at
335 /// the given iterator.
336 Partition(iterator SI) : SI(SI), SJ(SI) {}
339 /// \brief The start offset of this partition.
341 /// All of the contained slices start at or after this offset.
342 uint64_t beginOffset() const { return BeginOffset; }
344 /// \brief The end offset of this partition.
346 /// All of the contained slices end at or before this offset.
347 uint64_t endOffset() const { return EndOffset; }
349 /// \brief The size of the partition.
351 /// Note that this can never be zero.
352 uint64_t size() const {
353 assert(BeginOffset < EndOffset && "Partitions must span some bytes!");
354 return EndOffset - BeginOffset;
357 /// \brief Test whether this partition contains no slices, and merely spans
358 /// a region occupied by split slices.
359 bool empty() const { return SI == SJ; }
361 /// \name Iterate slices that start within the partition.
362 /// These may be splittable or unsplittable. They have a begin offset >= the
363 /// partition begin offset.
365 // FIXME: We should probably define a "concat_iterator" helper and use that
366 // to stitch together pointee_iterators over the split tails and the
367 // contiguous iterators of the partition. That would give a much nicer
368 // interface here. We could then additionally expose filtered iterators for
369 // split, unsplit, and unsplittable splices based on the usage patterns.
370 iterator begin() const { return SI; }
371 iterator end() const { return SJ; }
374 /// \brief Get the sequence of split slice tails.
376 /// These tails are of slices which start before this partition but are
377 /// split and overlap into the partition. We accumulate these while forming
379 ArrayRef<Slice *> splitSliceTails() const { return SplitTails; }
382 /// \brief An iterator over partitions of the alloca's slices.
384 /// This iterator implements the core algorithm for partitioning the alloca's
385 /// slices. It is a forward iterator as we don't support backtracking for
386 /// efficiency reasons, and re-use a single storage area to maintain the
387 /// current set of split slices.
389 /// It is templated on the slice iterator type to use so that it can operate
390 /// with either const or non-const slice iterators.
391 class AllocaSlices::partition_iterator
392 : public iterator_facade_base<partition_iterator, std::forward_iterator_tag,
394 friend class AllocaSlices;
396 /// \brief Most of the state for walking the partitions is held in a class
397 /// with a nice interface for examining them.
400 /// \brief We need to keep the end of the slices to know when to stop.
401 AllocaSlices::iterator SE;
403 /// \brief We also need to keep track of the maximum split end offset seen.
404 /// FIXME: Do we really?
405 uint64_t MaxSplitSliceEndOffset;
407 /// \brief Sets the partition to be empty at given iterator, and sets the
409 partition_iterator(AllocaSlices::iterator SI, AllocaSlices::iterator SE)
410 : P(SI), SE(SE), MaxSplitSliceEndOffset(0) {
411 // If not already at the end, advance our state to form the initial
417 /// \brief Advance the iterator to the next partition.
419 /// Requires that the iterator not be at the end of the slices.
421 assert((P.SI != SE || !P.SplitTails.empty()) &&
422 "Cannot advance past the end of the slices!");
424 // Clear out any split uses which have ended.
425 if (!P.SplitTails.empty()) {
426 if (P.EndOffset >= MaxSplitSliceEndOffset) {
427 // If we've finished all splits, this is easy.
428 P.SplitTails.clear();
429 MaxSplitSliceEndOffset = 0;
431 // Remove the uses which have ended in the prior partition. This
432 // cannot change the max split slice end because we just checked that
433 // the prior partition ended prior to that max.
435 remove_if(P.SplitTails,
436 [&](Slice *S) { return S->endOffset() <= P.EndOffset; }),
438 assert(any_of(P.SplitTails,
440 return S->endOffset() == MaxSplitSliceEndOffset;
442 "Could not find the current max split slice offset!");
443 assert(all_of(P.SplitTails,
445 return S->endOffset() <= MaxSplitSliceEndOffset;
447 "Max split slice end offset is not actually the max!");
451 // If P.SI is already at the end, then we've cleared the split tail and
452 // now have an end iterator.
454 assert(P.SplitTails.empty() && "Failed to clear the split slices!");
458 // If we had a non-empty partition previously, set up the state for
459 // subsequent partitions.
461 // Accumulate all the splittable slices which started in the old
462 // partition into the split list.
464 if (S.isSplittable() && S.endOffset() > P.EndOffset) {
465 P.SplitTails.push_back(&S);
466 MaxSplitSliceEndOffset =
467 std::max(S.endOffset(), MaxSplitSliceEndOffset);
470 // Start from the end of the previous partition.
473 // If P.SI is now at the end, we at most have a tail of split slices.
475 P.BeginOffset = P.EndOffset;
476 P.EndOffset = MaxSplitSliceEndOffset;
480 // If the we have split slices and the next slice is after a gap and is
481 // not splittable immediately form an empty partition for the split
482 // slices up until the next slice begins.
483 if (!P.SplitTails.empty() && P.SI->beginOffset() != P.EndOffset &&
484 !P.SI->isSplittable()) {
485 P.BeginOffset = P.EndOffset;
486 P.EndOffset = P.SI->beginOffset();
491 // OK, we need to consume new slices. Set the end offset based on the
492 // current slice, and step SJ past it. The beginning offset of the
493 // partition is the beginning offset of the next slice unless we have
494 // pre-existing split slices that are continuing, in which case we begin
495 // at the prior end offset.
496 P.BeginOffset = P.SplitTails.empty() ? P.SI->beginOffset() : P.EndOffset;
497 P.EndOffset = P.SI->endOffset();
500 // There are two strategies to form a partition based on whether the
501 // partition starts with an unsplittable slice or a splittable slice.
502 if (!P.SI->isSplittable()) {
503 // When we're forming an unsplittable region, it must always start at
504 // the first slice and will extend through its end.
505 assert(P.BeginOffset == P.SI->beginOffset());
507 // Form a partition including all of the overlapping slices with this
508 // unsplittable slice.
509 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
510 if (!P.SJ->isSplittable())
511 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
515 // We have a partition across a set of overlapping unsplittable
520 // If we're starting with a splittable slice, then we need to form
521 // a synthetic partition spanning it and any other overlapping splittable
523 assert(P.SI->isSplittable() && "Forming a splittable partition!");
525 // Collect all of the overlapping splittable slices.
526 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset &&
527 P.SJ->isSplittable()) {
528 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
532 // Back upiP.EndOffset if we ended the span early when encountering an
533 // unsplittable slice. This synthesizes the early end offset of
534 // a partition spanning only splittable slices.
535 if (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
536 assert(!P.SJ->isSplittable());
537 P.EndOffset = P.SJ->beginOffset();
542 bool operator==(const partition_iterator &RHS) const {
543 assert(SE == RHS.SE &&
544 "End iterators don't match between compared partition iterators!");
546 // The observed positions of partitions is marked by the P.SI iterator and
547 // the emptiness of the split slices. The latter is only relevant when
548 // P.SI == SE, as the end iterator will additionally have an empty split
549 // slices list, but the prior may have the same P.SI and a tail of split
551 if (P.SI == RHS.P.SI && P.SplitTails.empty() == RHS.P.SplitTails.empty()) {
552 assert(P.SJ == RHS.P.SJ &&
553 "Same set of slices formed two different sized partitions!");
554 assert(P.SplitTails.size() == RHS.P.SplitTails.size() &&
555 "Same slice position with differently sized non-empty split "
562 partition_iterator &operator++() {
567 Partition &operator*() { return P; }
570 /// \brief A forward range over the partitions of the alloca's slices.
572 /// This accesses an iterator range over the partitions of the alloca's
573 /// slices. It computes these partitions on the fly based on the overlapping
574 /// offsets of the slices and the ability to split them. It will visit "empty"
575 /// partitions to cover regions of the alloca only accessed via split
577 iterator_range<AllocaSlices::partition_iterator> AllocaSlices::partitions() {
578 return make_range(partition_iterator(begin(), end()),
579 partition_iterator(end(), end()));
582 static Value *foldSelectInst(SelectInst &SI) {
583 // If the condition being selected on is a constant or the same value is
584 // being selected between, fold the select. Yes this does (rarely) happen
586 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
587 return SI.getOperand(1 + CI->isZero());
588 if (SI.getOperand(1) == SI.getOperand(2))
589 return SI.getOperand(1);
594 /// \brief A helper that folds a PHI node or a select.
595 static Value *foldPHINodeOrSelectInst(Instruction &I) {
596 if (PHINode *PN = dyn_cast<PHINode>(&I)) {
597 // If PN merges together the same value, return that value.
598 return PN->hasConstantValue();
600 return foldSelectInst(cast<SelectInst>(I));
603 /// \brief Builder for the alloca slices.
605 /// This class builds a set of alloca slices by recursively visiting the uses
606 /// of an alloca and making a slice for each load and store at each offset.
607 class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
608 friend class PtrUseVisitor<SliceBuilder>;
609 friend class InstVisitor<SliceBuilder>;
610 typedef PtrUseVisitor<SliceBuilder> Base;
612 const uint64_t AllocSize;
615 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
616 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
618 /// \brief Set to de-duplicate dead instructions found in the use walk.
619 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
622 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &AS)
623 : PtrUseVisitor<SliceBuilder>(DL),
624 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), AS(AS) {}
627 void markAsDead(Instruction &I) {
628 if (VisitedDeadInsts.insert(&I).second)
629 AS.DeadUsers.push_back(&I);
632 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
633 bool IsSplittable = false) {
634 // Completely skip uses which have a zero size or start either before or
635 // past the end of the allocation.
636 if (Size == 0 || Offset.uge(AllocSize)) {
637 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
638 << " which has zero size or starts outside of the "
639 << AllocSize << " byte alloca:\n"
640 << " alloca: " << AS.AI << "\n"
641 << " use: " << I << "\n");
642 return markAsDead(I);
645 uint64_t BeginOffset = Offset.getZExtValue();
646 uint64_t EndOffset = BeginOffset + Size;
648 // Clamp the end offset to the end of the allocation. Note that this is
649 // formulated to handle even the case where "BeginOffset + Size" overflows.
650 // This may appear superficially to be something we could ignore entirely,
651 // but that is not so! There may be widened loads or PHI-node uses where
652 // some instructions are dead but not others. We can't completely ignore
653 // them, and so have to record at least the information here.
654 assert(AllocSize >= BeginOffset); // Established above.
655 if (Size > AllocSize - BeginOffset) {
656 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
657 << " to remain within the " << AllocSize << " byte alloca:\n"
658 << " alloca: " << AS.AI << "\n"
659 << " use: " << I << "\n");
660 EndOffset = AllocSize;
663 AS.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
666 void visitBitCastInst(BitCastInst &BC) {
668 return markAsDead(BC);
670 return Base::visitBitCastInst(BC);
673 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
674 if (GEPI.use_empty())
675 return markAsDead(GEPI);
677 if (SROAStrictInbounds && GEPI.isInBounds()) {
678 // FIXME: This is a manually un-factored variant of the basic code inside
679 // of GEPs with checking of the inbounds invariant specified in the
680 // langref in a very strict sense. If we ever want to enable
681 // SROAStrictInbounds, this code should be factored cleanly into
682 // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
683 // by writing out the code here where we have the underlying allocation
684 // size readily available.
685 APInt GEPOffset = Offset;
686 const DataLayout &DL = GEPI.getModule()->getDataLayout();
687 for (gep_type_iterator GTI = gep_type_begin(GEPI),
688 GTE = gep_type_end(GEPI);
690 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
694 // Handle a struct index, which adds its field offset to the pointer.
695 if (StructType *STy = GTI.getStructTypeOrNull()) {
696 unsigned ElementIdx = OpC->getZExtValue();
697 const StructLayout *SL = DL.getStructLayout(STy);
699 APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
701 // For array or vector indices, scale the index by the size of the
703 APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
704 GEPOffset += Index * APInt(Offset.getBitWidth(),
705 DL.getTypeAllocSize(GTI.getIndexedType()));
708 // If this index has computed an intermediate pointer which is not
709 // inbounds, then the result of the GEP is a poison value and we can
710 // delete it and all uses.
711 if (GEPOffset.ugt(AllocSize))
712 return markAsDead(GEPI);
716 return Base::visitGetElementPtrInst(GEPI);
719 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
720 uint64_t Size, bool IsVolatile) {
721 // We allow splitting of non-volatile loads and stores where the type is an
722 // integer type. These may be used to implement 'memcpy' or other "transfer
723 // of bits" patterns.
724 bool IsSplittable = Ty->isIntegerTy() && !IsVolatile;
726 insertUse(I, Offset, Size, IsSplittable);
729 void visitLoadInst(LoadInst &LI) {
730 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
731 "All simple FCA loads should have been pre-split");
734 return PI.setAborted(&LI);
736 const DataLayout &DL = LI.getModule()->getDataLayout();
737 uint64_t Size = DL.getTypeStoreSize(LI.getType());
738 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
741 void visitStoreInst(StoreInst &SI) {
742 Value *ValOp = SI.getValueOperand();
744 return PI.setEscapedAndAborted(&SI);
746 return PI.setAborted(&SI);
748 const DataLayout &DL = SI.getModule()->getDataLayout();
749 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
751 // If this memory access can be shown to *statically* extend outside the
752 // bounds of of the allocation, it's behavior is undefined, so simply
753 // ignore it. Note that this is more strict than the generic clamping
754 // behavior of insertUse. We also try to handle cases which might run the
756 // FIXME: We should instead consider the pointer to have escaped if this
757 // function is being instrumented for addressing bugs or race conditions.
758 if (Size > AllocSize || Offset.ugt(AllocSize - Size)) {
759 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
760 << " which extends past the end of the " << AllocSize
762 << " alloca: " << AS.AI << "\n"
763 << " use: " << SI << "\n");
764 return markAsDead(SI);
767 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
768 "All simple FCA stores should have been pre-split");
769 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
772 void visitMemSetInst(MemSetInst &II) {
773 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
774 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
775 if ((Length && Length->getValue() == 0) ||
776 (IsOffsetKnown && Offset.uge(AllocSize)))
777 // Zero-length mem transfer intrinsics can be ignored entirely.
778 return markAsDead(II);
781 return PI.setAborted(&II);
783 insertUse(II, Offset, Length ? Length->getLimitedValue()
784 : AllocSize - Offset.getLimitedValue(),
788 void visitMemTransferInst(MemTransferInst &II) {
789 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
790 if (Length && Length->getValue() == 0)
791 // Zero-length mem transfer intrinsics can be ignored entirely.
792 return markAsDead(II);
794 // Because we can visit these intrinsics twice, also check to see if the
795 // first time marked this instruction as dead. If so, skip it.
796 if (VisitedDeadInsts.count(&II))
800 return PI.setAborted(&II);
802 // This side of the transfer is completely out-of-bounds, and so we can
803 // nuke the entire transfer. However, we also need to nuke the other side
804 // if already added to our partitions.
805 // FIXME: Yet another place we really should bypass this when
806 // instrumenting for ASan.
807 if (Offset.uge(AllocSize)) {
808 SmallDenseMap<Instruction *, unsigned>::iterator MTPI =
809 MemTransferSliceMap.find(&II);
810 if (MTPI != MemTransferSliceMap.end())
811 AS.Slices[MTPI->second].kill();
812 return markAsDead(II);
815 uint64_t RawOffset = Offset.getLimitedValue();
816 uint64_t Size = Length ? Length->getLimitedValue() : AllocSize - RawOffset;
818 // Check for the special case where the same exact value is used for both
820 if (*U == II.getRawDest() && *U == II.getRawSource()) {
821 // For non-volatile transfers this is a no-op.
822 if (!II.isVolatile())
823 return markAsDead(II);
825 return insertUse(II, Offset, Size, /*IsSplittable=*/false);
828 // If we have seen both source and destination for a mem transfer, then
829 // they both point to the same alloca.
831 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
832 std::tie(MTPI, Inserted) =
833 MemTransferSliceMap.insert(std::make_pair(&II, AS.Slices.size()));
834 unsigned PrevIdx = MTPI->second;
836 Slice &PrevP = AS.Slices[PrevIdx];
838 // Check if the begin offsets match and this is a non-volatile transfer.
839 // In that case, we can completely elide the transfer.
840 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
842 return markAsDead(II);
845 // Otherwise we have an offset transfer within the same alloca. We can't
847 PrevP.makeUnsplittable();
850 // Insert the use now that we've fixed up the splittable nature.
851 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
853 // Check that we ended up with a valid index in the map.
854 assert(AS.Slices[PrevIdx].getUse()->getUser() == &II &&
855 "Map index doesn't point back to a slice with this user.");
858 // Disable SRoA for any intrinsics except for lifetime invariants.
859 // FIXME: What about debug intrinsics? This matches old behavior, but
860 // doesn't make sense.
861 void visitIntrinsicInst(IntrinsicInst &II) {
863 return PI.setAborted(&II);
865 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
866 II.getIntrinsicID() == Intrinsic::lifetime_end) {
867 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
868 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
869 Length->getLimitedValue());
870 insertUse(II, Offset, Size, true);
874 Base::visitIntrinsicInst(II);
877 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
878 // We consider any PHI or select that results in a direct load or store of
879 // the same offset to be a viable use for slicing purposes. These uses
880 // are considered unsplittable and the size is the maximum loaded or stored
882 SmallPtrSet<Instruction *, 4> Visited;
883 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
884 Visited.insert(Root);
885 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
886 const DataLayout &DL = Root->getModule()->getDataLayout();
887 // If there are no loads or stores, the access is dead. We mark that as
888 // a size zero access.
891 Instruction *I, *UsedI;
892 std::tie(UsedI, I) = Uses.pop_back_val();
894 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
895 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
898 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
899 Value *Op = SI->getOperand(0);
902 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
906 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
907 if (!GEP->hasAllZeroIndices())
909 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
910 !isa<SelectInst>(I)) {
914 for (User *U : I->users())
915 if (Visited.insert(cast<Instruction>(U)).second)
916 Uses.push_back(std::make_pair(I, cast<Instruction>(U)));
917 } while (!Uses.empty());
922 void visitPHINodeOrSelectInst(Instruction &I) {
923 assert(isa<PHINode>(I) || isa<SelectInst>(I));
925 return markAsDead(I);
927 // TODO: We could use SimplifyInstruction here to fold PHINodes and
928 // SelectInsts. However, doing so requires to change the current
929 // dead-operand-tracking mechanism. For instance, suppose neither loading
930 // from %U nor %other traps. Then "load (select undef, %U, %other)" does not
931 // trap either. However, if we simply replace %U with undef using the
932 // current dead-operand-tracking mechanism, "load (select undef, undef,
933 // %other)" may trap because the select may return the first operand
935 if (Value *Result = foldPHINodeOrSelectInst(I)) {
937 // If the result of the constant fold will be the pointer, recurse
938 // through the PHI/select as if we had RAUW'ed it.
941 // Otherwise the operand to the PHI/select is dead, and we can replace
943 AS.DeadOperands.push_back(U);
949 return PI.setAborted(&I);
951 // See if we already have computed info on this node.
952 uint64_t &Size = PHIOrSelectSizes[&I];
954 // This is a new PHI/Select, check for an unsafe use of it.
955 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&I, Size))
956 return PI.setAborted(UnsafeI);
959 // For PHI and select operands outside the alloca, we can't nuke the entire
960 // phi or select -- the other side might still be relevant, so we special
961 // case them here and use a separate structure to track the operands
962 // themselves which should be replaced with undef.
963 // FIXME: This should instead be escaped in the event we're instrumenting
964 // for address sanitization.
965 if (Offset.uge(AllocSize)) {
966 AS.DeadOperands.push_back(U);
970 insertUse(I, Offset, Size);
973 void visitPHINode(PHINode &PN) { visitPHINodeOrSelectInst(PN); }
975 void visitSelectInst(SelectInst &SI) { visitPHINodeOrSelectInst(SI); }
977 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
978 void visitInstruction(Instruction &I) { PI.setAborted(&I); }
981 AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
983 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
986 PointerEscapingInstr(nullptr) {
987 SliceBuilder PB(DL, AI, *this);
988 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
989 if (PtrI.isEscaped() || PtrI.isAborted()) {
990 // FIXME: We should sink the escape vs. abort info into the caller nicely,
991 // possibly by just storing the PtrInfo in the AllocaSlices.
992 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
993 : PtrI.getAbortingInst();
994 assert(PointerEscapingInstr && "Did not track a bad instruction");
998 Slices.erase(remove_if(Slices, [](const Slice &S) { return S.isDead(); }),
1002 if (SROARandomShuffleSlices) {
1003 std::mt19937 MT(static_cast<unsigned>(
1004 std::chrono::system_clock::now().time_since_epoch().count()));
1005 std::shuffle(Slices.begin(), Slices.end(), MT);
1009 // Sort the uses. This arranges for the offsets to be in ascending order,
1010 // and the sizes to be in descending order.
1011 std::sort(Slices.begin(), Slices.end());
1014 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1016 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
1017 StringRef Indent) const {
1018 printSlice(OS, I, Indent);
1020 printUse(OS, I, Indent);
1023 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
1024 StringRef Indent) const {
1025 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
1026 << " slice #" << (I - begin())
1027 << (I->isSplittable() ? " (splittable)" : "");
1030 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
1031 StringRef Indent) const {
1032 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
1035 void AllocaSlices::print(raw_ostream &OS) const {
1036 if (PointerEscapingInstr) {
1037 OS << "Can't analyze slices for alloca: " << AI << "\n"
1038 << " A pointer to this alloca escaped by:\n"
1039 << " " << *PointerEscapingInstr << "\n";
1043 OS << "Slices of alloca: " << AI << "\n";
1044 for (const_iterator I = begin(), E = end(); I != E; ++I)
1048 LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
1051 LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
1053 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1055 /// Walk the range of a partitioning looking for a common type to cover this
1056 /// sequence of slices.
1057 static Type *findCommonType(AllocaSlices::const_iterator B,
1058 AllocaSlices::const_iterator E,
1059 uint64_t EndOffset) {
1061 bool TyIsCommon = true;
1062 IntegerType *ITy = nullptr;
1064 // Note that we need to look at *every* alloca slice's Use to ensure we
1065 // always get consistent results regardless of the order of slices.
1066 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1067 Use *U = I->getUse();
1068 if (isa<IntrinsicInst>(*U->getUser()))
1070 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1073 Type *UserTy = nullptr;
1074 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1075 UserTy = LI->getType();
1076 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1077 UserTy = SI->getValueOperand()->getType();
1080 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1081 // If the type is larger than the partition, skip it. We only encounter
1082 // this for split integer operations where we want to use the type of the
1083 // entity causing the split. Also skip if the type is not a byte width
1085 if (UserITy->getBitWidth() % 8 != 0 ||
1086 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1089 // Track the largest bitwidth integer type used in this way in case there
1090 // is no common type.
1091 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1095 // To avoid depending on the order of slices, Ty and TyIsCommon must not
1096 // depend on types skipped above.
1097 if (!UserTy || (Ty && Ty != UserTy))
1098 TyIsCommon = false; // Give up on anything but an iN type.
1103 return TyIsCommon ? Ty : ITy;
1106 /// PHI instructions that use an alloca and are subsequently loaded can be
1107 /// rewritten to load both input pointers in the pred blocks and then PHI the
1108 /// results, allowing the load of the alloca to be promoted.
1110 /// %P2 = phi [i32* %Alloca, i32* %Other]
1111 /// %V = load i32* %P2
1113 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1115 /// %V2 = load i32* %Other
1117 /// %V = phi [i32 %V1, i32 %V2]
1119 /// We can do this to a select if its only uses are loads and if the operands
1120 /// to the select can be loaded unconditionally.
1122 /// FIXME: This should be hoisted into a generic utility, likely in
1123 /// Transforms/Util/Local.h
1124 static bool isSafePHIToSpeculate(PHINode &PN) {
1125 // For now, we can only do this promotion if the load is in the same block
1126 // as the PHI, and if there are no stores between the phi and load.
1127 // TODO: Allow recursive phi users.
1128 // TODO: Allow stores.
1129 BasicBlock *BB = PN.getParent();
1130 unsigned MaxAlign = 0;
1131 bool HaveLoad = false;
1132 for (User *U : PN.users()) {
1133 LoadInst *LI = dyn_cast<LoadInst>(U);
1134 if (!LI || !LI->isSimple())
1137 // For now we only allow loads in the same block as the PHI. This is
1138 // a common case that happens when instcombine merges two loads through
1140 if (LI->getParent() != BB)
1143 // Ensure that there are no instructions between the PHI and the load that
1145 for (BasicBlock::iterator BBI(PN); &*BBI != LI; ++BBI)
1146 if (BBI->mayWriteToMemory())
1149 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1156 const DataLayout &DL = PN.getModule()->getDataLayout();
1158 // We can only transform this if it is safe to push the loads into the
1159 // predecessor blocks. The only thing to watch out for is that we can't put
1160 // a possibly trapping load in the predecessor if it is a critical edge.
1161 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1162 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1163 Value *InVal = PN.getIncomingValue(Idx);
1165 // If the value is produced by the terminator of the predecessor (an
1166 // invoke) or it has side-effects, there is no valid place to put a load
1167 // in the predecessor.
1168 if (TI == InVal || TI->mayHaveSideEffects())
1171 // If the predecessor has a single successor, then the edge isn't
1173 if (TI->getNumSuccessors() == 1)
1176 // If this pointer is always safe to load, or if we can prove that there
1177 // is already a load in the block, then we can move the load to the pred
1179 if (isSafeToLoadUnconditionally(InVal, MaxAlign, DL, TI))
1188 static void speculatePHINodeLoads(PHINode &PN) {
1189 DEBUG(dbgs() << " original: " << PN << "\n");
1191 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1192 IRBuilderTy PHIBuilder(&PN);
1193 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1194 PN.getName() + ".sroa.speculated");
1196 // Get the AA tags and alignment to use from one of the loads. It doesn't
1197 // matter which one we get and if any differ.
1198 LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1201 SomeLoad->getAAMetadata(AATags);
1202 unsigned Align = SomeLoad->getAlignment();
1204 // Rewrite all loads of the PN to use the new PHI.
1205 while (!PN.use_empty()) {
1206 LoadInst *LI = cast<LoadInst>(PN.user_back());
1207 LI->replaceAllUsesWith(NewPN);
1208 LI->eraseFromParent();
1211 // Inject loads into all of the pred blocks.
1212 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1213 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1214 TerminatorInst *TI = Pred->getTerminator();
1215 Value *InVal = PN.getIncomingValue(Idx);
1216 IRBuilderTy PredBuilder(TI);
1218 LoadInst *Load = PredBuilder.CreateLoad(
1219 InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1220 ++NumLoadsSpeculated;
1221 Load->setAlignment(Align);
1223 Load->setAAMetadata(AATags);
1224 NewPN->addIncoming(Load, Pred);
1227 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1228 PN.eraseFromParent();
1231 /// Select instructions that use an alloca and are subsequently loaded can be
1232 /// rewritten to load both input pointers and then select between the result,
1233 /// allowing the load of the alloca to be promoted.
1235 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1236 /// %V = load i32* %P2
1238 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1239 /// %V2 = load i32* %Other
1240 /// %V = select i1 %cond, i32 %V1, i32 %V2
1242 /// We can do this to a select if its only uses are loads and if the operand
1243 /// to the select can be loaded unconditionally.
1244 static bool isSafeSelectToSpeculate(SelectInst &SI) {
1245 Value *TValue = SI.getTrueValue();
1246 Value *FValue = SI.getFalseValue();
1247 const DataLayout &DL = SI.getModule()->getDataLayout();
1249 for (User *U : SI.users()) {
1250 LoadInst *LI = dyn_cast<LoadInst>(U);
1251 if (!LI || !LI->isSimple())
1254 // Both operands to the select need to be dereferencable, either
1255 // absolutely (e.g. allocas) or at this point because we can see other
1257 if (!isSafeToLoadUnconditionally(TValue, LI->getAlignment(), DL, LI))
1259 if (!isSafeToLoadUnconditionally(FValue, LI->getAlignment(), DL, LI))
1266 static void speculateSelectInstLoads(SelectInst &SI) {
1267 DEBUG(dbgs() << " original: " << SI << "\n");
1269 IRBuilderTy IRB(&SI);
1270 Value *TV = SI.getTrueValue();
1271 Value *FV = SI.getFalseValue();
1272 // Replace the loads of the select with a select of two loads.
1273 while (!SI.use_empty()) {
1274 LoadInst *LI = cast<LoadInst>(SI.user_back());
1275 assert(LI->isSimple() && "We only speculate simple loads");
1277 IRB.SetInsertPoint(LI);
1279 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1281 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1282 NumLoadsSpeculated += 2;
1284 // Transfer alignment and AA info if present.
1285 TL->setAlignment(LI->getAlignment());
1286 FL->setAlignment(LI->getAlignment());
1289 LI->getAAMetadata(Tags);
1291 TL->setAAMetadata(Tags);
1292 FL->setAAMetadata(Tags);
1295 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1296 LI->getName() + ".sroa.speculated");
1298 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1299 LI->replaceAllUsesWith(V);
1300 LI->eraseFromParent();
1302 SI.eraseFromParent();
1305 /// \brief Build a GEP out of a base pointer and indices.
1307 /// This will return the BasePtr if that is valid, or build a new GEP
1308 /// instruction using the IRBuilder if GEP-ing is needed.
1309 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1310 SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1311 if (Indices.empty())
1314 // A single zero index is a no-op, so check for this and avoid building a GEP
1316 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1319 return IRB.CreateInBoundsGEP(nullptr, BasePtr, Indices,
1320 NamePrefix + "sroa_idx");
1323 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1324 /// TargetTy without changing the offset of the pointer.
1326 /// This routine assumes we've already established a properly offset GEP with
1327 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1328 /// zero-indices down through type layers until we find one the same as
1329 /// TargetTy. If we can't find one with the same type, we at least try to use
1330 /// one with the same size. If none of that works, we just produce the GEP as
1331 /// indicated by Indices to have the correct offset.
1332 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1333 Value *BasePtr, Type *Ty, Type *TargetTy,
1334 SmallVectorImpl<Value *> &Indices,
1337 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1339 // Pointer size to use for the indices.
1340 unsigned PtrSize = DL.getPointerTypeSizeInBits(BasePtr->getType());
1342 // See if we can descend into a struct and locate a field with the correct
1344 unsigned NumLayers = 0;
1345 Type *ElementTy = Ty;
1347 if (ElementTy->isPointerTy())
1350 if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1351 ElementTy = ArrayTy->getElementType();
1352 Indices.push_back(IRB.getIntN(PtrSize, 0));
1353 } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1354 ElementTy = VectorTy->getElementType();
1355 Indices.push_back(IRB.getInt32(0));
1356 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1357 if (STy->element_begin() == STy->element_end())
1358 break; // Nothing left to descend into.
1359 ElementTy = *STy->element_begin();
1360 Indices.push_back(IRB.getInt32(0));
1365 } while (ElementTy != TargetTy);
1366 if (ElementTy != TargetTy)
1367 Indices.erase(Indices.end() - NumLayers, Indices.end());
1369 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1372 /// \brief Recursively compute indices for a natural GEP.
1374 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1375 /// element types adding appropriate indices for the GEP.
1376 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1377 Value *Ptr, Type *Ty, APInt &Offset,
1379 SmallVectorImpl<Value *> &Indices,
1382 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices,
1385 // We can't recurse through pointer types.
1386 if (Ty->isPointerTy())
1389 // We try to analyze GEPs over vectors here, but note that these GEPs are
1390 // extremely poorly defined currently. The long-term goal is to remove GEPing
1391 // over a vector from the IR completely.
1392 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1393 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1394 if (ElementSizeInBits % 8 != 0) {
1395 // GEPs over non-multiple of 8 size vector elements are invalid.
1398 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1399 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1400 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1402 Offset -= NumSkippedElements * ElementSize;
1403 Indices.push_back(IRB.getInt(NumSkippedElements));
1404 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1405 Offset, TargetTy, Indices, NamePrefix);
1408 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1409 Type *ElementTy = ArrTy->getElementType();
1410 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1411 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1412 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1415 Offset -= NumSkippedElements * ElementSize;
1416 Indices.push_back(IRB.getInt(NumSkippedElements));
1417 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1418 Indices, NamePrefix);
1421 StructType *STy = dyn_cast<StructType>(Ty);
1425 const StructLayout *SL = DL.getStructLayout(STy);
1426 uint64_t StructOffset = Offset.getZExtValue();
1427 if (StructOffset >= SL->getSizeInBytes())
1429 unsigned Index = SL->getElementContainingOffset(StructOffset);
1430 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1431 Type *ElementTy = STy->getElementType(Index);
1432 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1433 return nullptr; // The offset points into alignment padding.
1435 Indices.push_back(IRB.getInt32(Index));
1436 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1437 Indices, NamePrefix);
1440 /// \brief Get a natural GEP from a base pointer to a particular offset and
1441 /// resulting in a particular type.
1443 /// The goal is to produce a "natural" looking GEP that works with the existing
1444 /// composite types to arrive at the appropriate offset and element type for
1445 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1446 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1447 /// Indices, and setting Ty to the result subtype.
1449 /// If no natural GEP can be constructed, this function returns null.
1450 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1451 Value *Ptr, APInt Offset, Type *TargetTy,
1452 SmallVectorImpl<Value *> &Indices,
1454 PointerType *Ty = cast<PointerType>(Ptr->getType());
1456 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1458 if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1461 Type *ElementTy = Ty->getElementType();
1462 if (!ElementTy->isSized())
1463 return nullptr; // We can't GEP through an unsized element.
1464 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1465 if (ElementSize == 0)
1466 return nullptr; // Zero-length arrays can't help us build a natural GEP.
1467 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1469 Offset -= NumSkippedElements * ElementSize;
1470 Indices.push_back(IRB.getInt(NumSkippedElements));
1471 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1472 Indices, NamePrefix);
1475 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1476 /// resulting pointer has PointerTy.
1478 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1479 /// and produces the pointer type desired. Where it cannot, it will try to use
1480 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1481 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1482 /// bitcast to the type.
1484 /// The strategy for finding the more natural GEPs is to peel off layers of the
1485 /// pointer, walking back through bit casts and GEPs, searching for a base
1486 /// pointer from which we can compute a natural GEP with the desired
1487 /// properties. The algorithm tries to fold as many constant indices into
1488 /// a single GEP as possible, thus making each GEP more independent of the
1489 /// surrounding code.
1490 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1491 APInt Offset, Type *PointerTy, Twine NamePrefix) {
1492 // Even though we don't look through PHI nodes, we could be called on an
1493 // instruction in an unreachable block, which may be on a cycle.
1494 SmallPtrSet<Value *, 4> Visited;
1495 Visited.insert(Ptr);
1496 SmallVector<Value *, 4> Indices;
1498 // We may end up computing an offset pointer that has the wrong type. If we
1499 // never are able to compute one directly that has the correct type, we'll
1500 // fall back to it, so keep it and the base it was computed from around here.
1501 Value *OffsetPtr = nullptr;
1502 Value *OffsetBasePtr;
1504 // Remember any i8 pointer we come across to re-use if we need to do a raw
1506 Value *Int8Ptr = nullptr;
1507 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1509 Type *TargetTy = PointerTy->getPointerElementType();
1512 // First fold any existing GEPs into the offset.
1513 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1514 APInt GEPOffset(Offset.getBitWidth(), 0);
1515 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1517 Offset += GEPOffset;
1518 Ptr = GEP->getPointerOperand();
1519 if (!Visited.insert(Ptr).second)
1523 // See if we can perform a natural GEP here.
1525 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1526 Indices, NamePrefix)) {
1527 // If we have a new natural pointer at the offset, clear out any old
1528 // offset pointer we computed. Unless it is the base pointer or
1529 // a non-instruction, we built a GEP we don't need. Zap it.
1530 if (OffsetPtr && OffsetPtr != OffsetBasePtr)
1531 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr)) {
1532 assert(I->use_empty() && "Built a GEP with uses some how!");
1533 I->eraseFromParent();
1536 OffsetBasePtr = Ptr;
1537 // If we also found a pointer of the right type, we're done.
1538 if (P->getType() == PointerTy)
1542 // Stash this pointer if we've found an i8*.
1543 if (Ptr->getType()->isIntegerTy(8)) {
1545 Int8PtrOffset = Offset;
1548 // Peel off a layer of the pointer and update the offset appropriately.
1549 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1550 Ptr = cast<Operator>(Ptr)->getOperand(0);
1551 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1552 if (GA->isInterposable())
1554 Ptr = GA->getAliasee();
1558 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1559 } while (Visited.insert(Ptr).second);
1563 Int8Ptr = IRB.CreateBitCast(
1564 Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1565 NamePrefix + "sroa_raw_cast");
1566 Int8PtrOffset = Offset;
1569 OffsetPtr = Int8PtrOffset == 0
1571 : IRB.CreateInBoundsGEP(IRB.getInt8Ty(), Int8Ptr,
1572 IRB.getInt(Int8PtrOffset),
1573 NamePrefix + "sroa_raw_idx");
1577 // On the off chance we were targeting i8*, guard the bitcast here.
1578 if (Ptr->getType() != PointerTy)
1579 Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
1584 /// \brief Compute the adjusted alignment for a load or store from an offset.
1585 static unsigned getAdjustedAlignment(Instruction *I, uint64_t Offset,
1586 const DataLayout &DL) {
1589 if (auto *LI = dyn_cast<LoadInst>(I)) {
1590 Alignment = LI->getAlignment();
1592 } else if (auto *SI = dyn_cast<StoreInst>(I)) {
1593 Alignment = SI->getAlignment();
1594 Ty = SI->getValueOperand()->getType();
1596 llvm_unreachable("Only loads and stores are allowed!");
1600 Alignment = DL.getABITypeAlignment(Ty);
1602 return MinAlign(Alignment, Offset);
1605 /// \brief Test whether we can convert a value from the old to the new type.
1607 /// This predicate should be used to guard calls to convertValue in order to
1608 /// ensure that we only try to convert viable values. The strategy is that we
1609 /// will peel off single element struct and array wrappings to get to an
1610 /// underlying value, and convert that value.
1611 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1615 // For integer types, we can't handle any bit-width differences. This would
1616 // break both vector conversions with extension and introduce endianness
1617 // issues when in conjunction with loads and stores.
1618 if (isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) {
1619 assert(cast<IntegerType>(OldTy)->getBitWidth() !=
1620 cast<IntegerType>(NewTy)->getBitWidth() &&
1621 "We can't have the same bitwidth for different int types");
1625 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1627 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1630 // We can convert pointers to integers and vice-versa. Same for vectors
1631 // of pointers and integers.
1632 OldTy = OldTy->getScalarType();
1633 NewTy = NewTy->getScalarType();
1634 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1635 if (NewTy->isPointerTy() && OldTy->isPointerTy()) {
1636 return cast<PointerType>(NewTy)->getPointerAddressSpace() ==
1637 cast<PointerType>(OldTy)->getPointerAddressSpace();
1639 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1647 /// \brief Generic routine to convert an SSA value to a value of a different
1650 /// This will try various different casting techniques, such as bitcasts,
1651 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1652 /// two types for viability with this routine.
1653 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1655 Type *OldTy = V->getType();
1656 assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1661 assert(!(isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) &&
1662 "Integer types must be the exact same to convert.");
1664 // See if we need inttoptr for this type pair. A cast involving both scalars
1665 // and vectors requires and additional bitcast.
1666 if (OldTy->getScalarType()->isIntegerTy() &&
1667 NewTy->getScalarType()->isPointerTy()) {
1668 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1669 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1670 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1673 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1674 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1675 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1678 return IRB.CreateIntToPtr(V, NewTy);
1681 // See if we need ptrtoint for this type pair. A cast involving both scalars
1682 // and vectors requires and additional bitcast.
1683 if (OldTy->getScalarType()->isPointerTy() &&
1684 NewTy->getScalarType()->isIntegerTy()) {
1685 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1686 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1687 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1690 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1691 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1692 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1695 return IRB.CreatePtrToInt(V, NewTy);
1698 return IRB.CreateBitCast(V, NewTy);
1701 /// \brief Test whether the given slice use can be promoted to a vector.
1703 /// This function is called to test each entry in a partition which is slated
1704 /// for a single slice.
1705 static bool isVectorPromotionViableForSlice(Partition &P, const Slice &S,
1707 uint64_t ElementSize,
1708 const DataLayout &DL) {
1709 // First validate the slice offsets.
1710 uint64_t BeginOffset =
1711 std::max(S.beginOffset(), P.beginOffset()) - P.beginOffset();
1712 uint64_t BeginIndex = BeginOffset / ElementSize;
1713 if (BeginIndex * ElementSize != BeginOffset ||
1714 BeginIndex >= Ty->getNumElements())
1716 uint64_t EndOffset =
1717 std::min(S.endOffset(), P.endOffset()) - P.beginOffset();
1718 uint64_t EndIndex = EndOffset / ElementSize;
1719 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1722 assert(EndIndex > BeginIndex && "Empty vector!");
1723 uint64_t NumElements = EndIndex - BeginIndex;
1724 Type *SliceTy = (NumElements == 1)
1725 ? Ty->getElementType()
1726 : VectorType::get(Ty->getElementType(), NumElements);
1729 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1731 Use *U = S.getUse();
1733 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1734 if (MI->isVolatile())
1736 if (!S.isSplittable())
1737 return false; // Skip any unsplittable intrinsics.
1738 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1739 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1740 II->getIntrinsicID() != Intrinsic::lifetime_end)
1742 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1743 // Disable vector promotion when there are loads or stores of an FCA.
1745 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1746 if (LI->isVolatile())
1748 Type *LTy = LI->getType();
1749 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1750 assert(LTy->isIntegerTy());
1753 if (!canConvertValue(DL, SliceTy, LTy))
1755 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1756 if (SI->isVolatile())
1758 Type *STy = SI->getValueOperand()->getType();
1759 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1760 assert(STy->isIntegerTy());
1763 if (!canConvertValue(DL, STy, SliceTy))
1772 /// \brief Test whether the given alloca partitioning and range of slices can be
1773 /// promoted to a vector.
1775 /// This is a quick test to check whether we can rewrite a particular alloca
1776 /// partition (and its newly formed alloca) into a vector alloca with only
1777 /// whole-vector loads and stores such that it could be promoted to a vector
1778 /// SSA value. We only can ensure this for a limited set of operations, and we
1779 /// don't want to do the rewrites unless we are confident that the result will
1780 /// be promotable, so we have an early test here.
1781 static VectorType *isVectorPromotionViable(Partition &P, const DataLayout &DL) {
1782 // Collect the candidate types for vector-based promotion. Also track whether
1783 // we have different element types.
1784 SmallVector<VectorType *, 4> CandidateTys;
1785 Type *CommonEltTy = nullptr;
1786 bool HaveCommonEltTy = true;
1787 auto CheckCandidateType = [&](Type *Ty) {
1788 if (auto *VTy = dyn_cast<VectorType>(Ty)) {
1789 CandidateTys.push_back(VTy);
1791 CommonEltTy = VTy->getElementType();
1792 else if (CommonEltTy != VTy->getElementType())
1793 HaveCommonEltTy = false;
1796 // Consider any loads or stores that are the exact size of the slice.
1797 for (const Slice &S : P)
1798 if (S.beginOffset() == P.beginOffset() &&
1799 S.endOffset() == P.endOffset()) {
1800 if (auto *LI = dyn_cast<LoadInst>(S.getUse()->getUser()))
1801 CheckCandidateType(LI->getType());
1802 else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser()))
1803 CheckCandidateType(SI->getValueOperand()->getType());
1806 // If we didn't find a vector type, nothing to do here.
1807 if (CandidateTys.empty())
1810 // Remove non-integer vector types if we had multiple common element types.
1811 // FIXME: It'd be nice to replace them with integer vector types, but we can't
1812 // do that until all the backends are known to produce good code for all
1813 // integer vector types.
1814 if (!HaveCommonEltTy) {
1815 CandidateTys.erase(remove_if(CandidateTys,
1816 [](VectorType *VTy) {
1817 return !VTy->getElementType()->isIntegerTy();
1819 CandidateTys.end());
1821 // If there were no integer vector types, give up.
1822 if (CandidateTys.empty())
1825 // Rank the remaining candidate vector types. This is easy because we know
1826 // they're all integer vectors. We sort by ascending number of elements.
1827 auto RankVectorTypes = [&DL](VectorType *RHSTy, VectorType *LHSTy) {
1829 assert(DL.getTypeSizeInBits(RHSTy) == DL.getTypeSizeInBits(LHSTy) &&
1830 "Cannot have vector types of different sizes!");
1831 assert(RHSTy->getElementType()->isIntegerTy() &&
1832 "All non-integer types eliminated!");
1833 assert(LHSTy->getElementType()->isIntegerTy() &&
1834 "All non-integer types eliminated!");
1835 return RHSTy->getNumElements() < LHSTy->getNumElements();
1837 std::sort(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes);
1839 std::unique(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes),
1840 CandidateTys.end());
1842 // The only way to have the same element type in every vector type is to
1843 // have the same vector type. Check that and remove all but one.
1845 for (VectorType *VTy : CandidateTys) {
1846 assert(VTy->getElementType() == CommonEltTy &&
1847 "Unaccounted for element type!");
1848 assert(VTy == CandidateTys[0] &&
1849 "Different vector types with the same element type!");
1852 CandidateTys.resize(1);
1855 // Try each vector type, and return the one which works.
1856 auto CheckVectorTypeForPromotion = [&](VectorType *VTy) {
1857 uint64_t ElementSize = DL.getTypeSizeInBits(VTy->getElementType());
1859 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1860 // that aren't byte sized.
1861 if (ElementSize % 8)
1863 assert((DL.getTypeSizeInBits(VTy) % 8) == 0 &&
1864 "vector size not a multiple of element size?");
1867 for (const Slice &S : P)
1868 if (!isVectorPromotionViableForSlice(P, S, VTy, ElementSize, DL))
1871 for (const Slice *S : P.splitSliceTails())
1872 if (!isVectorPromotionViableForSlice(P, *S, VTy, ElementSize, DL))
1877 for (VectorType *VTy : CandidateTys)
1878 if (CheckVectorTypeForPromotion(VTy))
1884 /// \brief Test whether a slice of an alloca is valid for integer widening.
1886 /// This implements the necessary checking for the \c isIntegerWideningViable
1887 /// test below on a single slice of the alloca.
1888 static bool isIntegerWideningViableForSlice(const Slice &S,
1889 uint64_t AllocBeginOffset,
1891 const DataLayout &DL,
1892 bool &WholeAllocaOp) {
1893 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1895 uint64_t RelBegin = S.beginOffset() - AllocBeginOffset;
1896 uint64_t RelEnd = S.endOffset() - AllocBeginOffset;
1898 // We can't reasonably handle cases where the load or store extends past
1899 // the end of the alloca's type and into its padding.
1903 Use *U = S.getUse();
1905 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1906 if (LI->isVolatile())
1908 // We can't handle loads that extend past the allocated memory.
1909 if (DL.getTypeStoreSize(LI->getType()) > Size)
1911 // Note that we don't count vector loads or stores as whole-alloca
1912 // operations which enable integer widening because we would prefer to use
1913 // vector widening instead.
1914 if (!isa<VectorType>(LI->getType()) && RelBegin == 0 && RelEnd == Size)
1915 WholeAllocaOp = true;
1916 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1917 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1919 } else if (RelBegin != 0 || RelEnd != Size ||
1920 !canConvertValue(DL, AllocaTy, LI->getType())) {
1921 // Non-integer loads need to be convertible from the alloca type so that
1922 // they are promotable.
1925 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1926 Type *ValueTy = SI->getValueOperand()->getType();
1927 if (SI->isVolatile())
1929 // We can't handle stores that extend past the allocated memory.
1930 if (DL.getTypeStoreSize(ValueTy) > Size)
1932 // Note that we don't count vector loads or stores as whole-alloca
1933 // operations which enable integer widening because we would prefer to use
1934 // vector widening instead.
1935 if (!isa<VectorType>(ValueTy) && RelBegin == 0 && RelEnd == Size)
1936 WholeAllocaOp = true;
1937 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
1938 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1940 } else if (RelBegin != 0 || RelEnd != Size ||
1941 !canConvertValue(DL, ValueTy, AllocaTy)) {
1942 // Non-integer stores need to be convertible to the alloca type so that
1943 // they are promotable.
1946 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1947 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
1949 if (!S.isSplittable())
1950 return false; // Skip any unsplittable intrinsics.
1951 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1952 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1953 II->getIntrinsicID() != Intrinsic::lifetime_end)
1962 /// \brief Test whether the given alloca partition's integer operations can be
1963 /// widened to promotable ones.
1965 /// This is a quick test to check whether we can rewrite the integer loads and
1966 /// stores to a particular alloca into wider loads and stores and be able to
1967 /// promote the resulting alloca.
1968 static bool isIntegerWideningViable(Partition &P, Type *AllocaTy,
1969 const DataLayout &DL) {
1970 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
1971 // Don't create integer types larger than the maximum bitwidth.
1972 if (SizeInBits > IntegerType::MAX_INT_BITS)
1975 // Don't try to handle allocas with bit-padding.
1976 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
1979 // We need to ensure that an integer type with the appropriate bitwidth can
1980 // be converted to the alloca type, whatever that is. We don't want to force
1981 // the alloca itself to have an integer type if there is a more suitable one.
1982 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
1983 if (!canConvertValue(DL, AllocaTy, IntTy) ||
1984 !canConvertValue(DL, IntTy, AllocaTy))
1987 // While examining uses, we ensure that the alloca has a covering load or
1988 // store. We don't want to widen the integer operations only to fail to
1989 // promote due to some other unsplittable entry (which we may make splittable
1990 // later). However, if there are only splittable uses, go ahead and assume
1991 // that we cover the alloca.
1992 // FIXME: We shouldn't consider split slices that happen to start in the
1993 // partition here...
1994 bool WholeAllocaOp =
1995 P.begin() != P.end() ? false : DL.isLegalInteger(SizeInBits);
1997 for (const Slice &S : P)
1998 if (!isIntegerWideningViableForSlice(S, P.beginOffset(), AllocaTy, DL,
2002 for (const Slice *S : P.splitSliceTails())
2003 if (!isIntegerWideningViableForSlice(*S, P.beginOffset(), AllocaTy, DL,
2007 return WholeAllocaOp;
2010 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
2011 IntegerType *Ty, uint64_t Offset,
2012 const Twine &Name) {
2013 DEBUG(dbgs() << " start: " << *V << "\n");
2014 IntegerType *IntTy = cast<IntegerType>(V->getType());
2015 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2016 "Element extends past full value");
2017 uint64_t ShAmt = 8 * Offset;
2018 if (DL.isBigEndian())
2019 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2021 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2022 DEBUG(dbgs() << " shifted: " << *V << "\n");
2024 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2025 "Cannot extract to a larger integer!");
2027 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2028 DEBUG(dbgs() << " trunced: " << *V << "\n");
2033 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
2034 Value *V, uint64_t Offset, const Twine &Name) {
2035 IntegerType *IntTy = cast<IntegerType>(Old->getType());
2036 IntegerType *Ty = cast<IntegerType>(V->getType());
2037 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2038 "Cannot insert a larger integer!");
2039 DEBUG(dbgs() << " start: " << *V << "\n");
2041 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2042 DEBUG(dbgs() << " extended: " << *V << "\n");
2044 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2045 "Element store outside of alloca store");
2046 uint64_t ShAmt = 8 * Offset;
2047 if (DL.isBigEndian())
2048 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2050 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2051 DEBUG(dbgs() << " shifted: " << *V << "\n");
2054 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2055 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2056 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2057 DEBUG(dbgs() << " masked: " << *Old << "\n");
2058 V = IRB.CreateOr(Old, V, Name + ".insert");
2059 DEBUG(dbgs() << " inserted: " << *V << "\n");
2064 static Value *extractVector(IRBuilderTy &IRB, Value *V, unsigned BeginIndex,
2065 unsigned EndIndex, const Twine &Name) {
2066 VectorType *VecTy = cast<VectorType>(V->getType());
2067 unsigned NumElements = EndIndex - BeginIndex;
2068 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2070 if (NumElements == VecTy->getNumElements())
2073 if (NumElements == 1) {
2074 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2076 DEBUG(dbgs() << " extract: " << *V << "\n");
2080 SmallVector<Constant *, 8> Mask;
2081 Mask.reserve(NumElements);
2082 for (unsigned i = BeginIndex; i != EndIndex; ++i)
2083 Mask.push_back(IRB.getInt32(i));
2084 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2085 ConstantVector::get(Mask), Name + ".extract");
2086 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2090 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
2091 unsigned BeginIndex, const Twine &Name) {
2092 VectorType *VecTy = cast<VectorType>(Old->getType());
2093 assert(VecTy && "Can only insert a vector into a vector");
2095 VectorType *Ty = dyn_cast<VectorType>(V->getType());
2097 // Single element to insert.
2098 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2100 DEBUG(dbgs() << " insert: " << *V << "\n");
2104 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
2105 "Too many elements!");
2106 if (Ty->getNumElements() == VecTy->getNumElements()) {
2107 assert(V->getType() == VecTy && "Vector type mismatch");
2110 unsigned EndIndex = BeginIndex + Ty->getNumElements();
2112 // When inserting a smaller vector into the larger to store, we first
2113 // use a shuffle vector to widen it with undef elements, and then
2114 // a second shuffle vector to select between the loaded vector and the
2116 SmallVector<Constant *, 8> Mask;
2117 Mask.reserve(VecTy->getNumElements());
2118 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2119 if (i >= BeginIndex && i < EndIndex)
2120 Mask.push_back(IRB.getInt32(i - BeginIndex));
2122 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2123 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2124 ConstantVector::get(Mask), Name + ".expand");
2125 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2128 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2129 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
2131 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
2133 DEBUG(dbgs() << " blend: " << *V << "\n");
2137 /// \brief Visitor to rewrite instructions using p particular slice of an alloca
2138 /// to use a new alloca.
2140 /// Also implements the rewriting to vector-based accesses when the partition
2141 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2143 class llvm::sroa::AllocaSliceRewriter
2144 : public InstVisitor<AllocaSliceRewriter, bool> {
2145 // Befriend the base class so it can delegate to private visit methods.
2146 friend class llvm::InstVisitor<AllocaSliceRewriter, bool>;
2147 typedef llvm::InstVisitor<AllocaSliceRewriter, bool> Base;
2149 const DataLayout &DL;
2152 AllocaInst &OldAI, &NewAI;
2153 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2156 // This is a convenience and flag variable that will be null unless the new
2157 // alloca's integer operations should be widened to this integer type due to
2158 // passing isIntegerWideningViable above. If it is non-null, the desired
2159 // integer type will be stored here for easy access during rewriting.
2162 // If we are rewriting an alloca partition which can be written as pure
2163 // vector operations, we stash extra information here. When VecTy is
2164 // non-null, we have some strict guarantees about the rewritten alloca:
2165 // - The new alloca is exactly the size of the vector type here.
2166 // - The accesses all either map to the entire vector or to a single
2168 // - The set of accessing instructions is only one of those handled above
2169 // in isVectorPromotionViable. Generally these are the same access kinds
2170 // which are promotable via mem2reg.
2173 uint64_t ElementSize;
2175 // The original offset of the slice currently being rewritten relative to
2176 // the original alloca.
2177 uint64_t BeginOffset, EndOffset;
2178 // The new offsets of the slice currently being rewritten relative to the
2180 uint64_t NewBeginOffset, NewEndOffset;
2186 Instruction *OldPtr;
2188 // Track post-rewrite users which are PHI nodes and Selects.
2189 SmallPtrSetImpl<PHINode *> &PHIUsers;
2190 SmallPtrSetImpl<SelectInst *> &SelectUsers;
2192 // Utility IR builder, whose name prefix is setup for each visited use, and
2193 // the insertion point is set to point to the user.
2197 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &AS, SROA &Pass,
2198 AllocaInst &OldAI, AllocaInst &NewAI,
2199 uint64_t NewAllocaBeginOffset,
2200 uint64_t NewAllocaEndOffset, bool IsIntegerPromotable,
2201 VectorType *PromotableVecTy,
2202 SmallPtrSetImpl<PHINode *> &PHIUsers,
2203 SmallPtrSetImpl<SelectInst *> &SelectUsers)
2204 : DL(DL), AS(AS), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2205 NewAllocaBeginOffset(NewAllocaBeginOffset),
2206 NewAllocaEndOffset(NewAllocaEndOffset),
2207 NewAllocaTy(NewAI.getAllocatedType()),
2208 IntTy(IsIntegerPromotable
2211 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2213 VecTy(PromotableVecTy),
2214 ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2215 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
2216 BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
2217 OldPtr(), PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2218 IRB(NewAI.getContext(), ConstantFolder()) {
2220 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
2221 "Only multiple-of-8 sized vector elements are viable");
2224 assert((!IntTy && !VecTy) || (IntTy && !VecTy) || (!IntTy && VecTy));
2227 bool visit(AllocaSlices::const_iterator I) {
2228 bool CanSROA = true;
2229 BeginOffset = I->beginOffset();
2230 EndOffset = I->endOffset();
2231 IsSplittable = I->isSplittable();
2233 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2234 DEBUG(dbgs() << " rewriting " << (IsSplit ? "split " : ""));
2235 DEBUG(AS.printSlice(dbgs(), I, ""));
2236 DEBUG(dbgs() << "\n");
2238 // Compute the intersecting offset range.
2239 assert(BeginOffset < NewAllocaEndOffset);
2240 assert(EndOffset > NewAllocaBeginOffset);
2241 NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2242 NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2244 SliceSize = NewEndOffset - NewBeginOffset;
2246 OldUse = I->getUse();
2247 OldPtr = cast<Instruction>(OldUse->get());
2249 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2250 IRB.SetInsertPoint(OldUserI);
2251 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2252 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2254 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2261 // Make sure the other visit overloads are visible.
2264 // Every instruction which can end up as a user must have a rewrite rule.
2265 bool visitInstruction(Instruction &I) {
2266 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2267 llvm_unreachable("No rewrite rule for this instruction!");
2270 Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2271 // Note that the offset computation can use BeginOffset or NewBeginOffset
2272 // interchangeably for unsplit slices.
2273 assert(IsSplit || BeginOffset == NewBeginOffset);
2274 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2277 StringRef OldName = OldPtr->getName();
2278 // Skip through the last '.sroa.' component of the name.
2279 size_t LastSROAPrefix = OldName.rfind(".sroa.");
2280 if (LastSROAPrefix != StringRef::npos) {
2281 OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2282 // Look for an SROA slice index.
2283 size_t IndexEnd = OldName.find_first_not_of("0123456789");
2284 if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2285 // Strip the index and look for the offset.
2286 OldName = OldName.substr(IndexEnd + 1);
2287 size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2288 if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2289 // Strip the offset.
2290 OldName = OldName.substr(OffsetEnd + 1);
2293 // Strip any SROA suffixes as well.
2294 OldName = OldName.substr(0, OldName.find(".sroa_"));
2297 return getAdjustedPtr(IRB, DL, &NewAI,
2298 APInt(DL.getPointerSizeInBits(), Offset), PointerTy,
2300 Twine(OldName) + "."
2307 /// \brief Compute suitable alignment to access this slice of the *new*
2310 /// You can optionally pass a type to this routine and if that type's ABI
2311 /// alignment is itself suitable, this will return zero.
2312 unsigned getSliceAlign(Type *Ty = nullptr) {
2313 unsigned NewAIAlign = NewAI.getAlignment();
2315 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2317 MinAlign(NewAIAlign, NewBeginOffset - NewAllocaBeginOffset);
2318 return (Ty && Align == DL.getABITypeAlignment(Ty)) ? 0 : Align;
2321 unsigned getIndex(uint64_t Offset) {
2322 assert(VecTy && "Can only call getIndex when rewriting a vector");
2323 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2324 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2325 uint32_t Index = RelOffset / ElementSize;
2326 assert(Index * ElementSize == RelOffset);
2330 void deleteIfTriviallyDead(Value *V) {
2331 Instruction *I = cast<Instruction>(V);
2332 if (isInstructionTriviallyDead(I))
2333 Pass.DeadInsts.insert(I);
2336 Value *rewriteVectorizedLoadInst() {
2337 unsigned BeginIndex = getIndex(NewBeginOffset);
2338 unsigned EndIndex = getIndex(NewEndOffset);
2339 assert(EndIndex > BeginIndex && "Empty vector!");
2341 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2342 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2345 Value *rewriteIntegerLoad(LoadInst &LI) {
2346 assert(IntTy && "We cannot insert an integer to the alloca");
2347 assert(!LI.isVolatile());
2348 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2349 V = convertValue(DL, IRB, V, IntTy);
2350 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2351 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2352 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset) {
2353 IntegerType *ExtractTy = Type::getIntNTy(LI.getContext(), SliceSize * 8);
2354 V = extractInteger(DL, IRB, V, ExtractTy, Offset, "extract");
2356 // It is possible that the extracted type is not the load type. This
2357 // happens if there is a load past the end of the alloca, and as
2358 // a consequence the slice is narrower but still a candidate for integer
2359 // lowering. To handle this case, we just zero extend the extracted
2361 assert(cast<IntegerType>(LI.getType())->getBitWidth() >= SliceSize * 8 &&
2362 "Can only handle an extract for an overly wide load");
2363 if (cast<IntegerType>(LI.getType())->getBitWidth() > SliceSize * 8)
2364 V = IRB.CreateZExt(V, LI.getType());
2368 bool visitLoadInst(LoadInst &LI) {
2369 DEBUG(dbgs() << " original: " << LI << "\n");
2370 Value *OldOp = LI.getOperand(0);
2371 assert(OldOp == OldPtr);
2373 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2375 const bool IsLoadPastEnd = DL.getTypeStoreSize(TargetTy) > SliceSize;
2376 bool IsPtrAdjusted = false;
2379 V = rewriteVectorizedLoadInst();
2380 } else if (IntTy && LI.getType()->isIntegerTy()) {
2381 V = rewriteIntegerLoad(LI);
2382 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2383 NewEndOffset == NewAllocaEndOffset &&
2384 (canConvertValue(DL, NewAllocaTy, TargetTy) ||
2385 (IsLoadPastEnd && NewAllocaTy->isIntegerTy() &&
2386 TargetTy->isIntegerTy()))) {
2387 LoadInst *NewLI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2388 LI.isVolatile(), LI.getName());
2389 if (LI.isVolatile())
2390 NewLI->setAtomic(LI.getOrdering(), LI.getSynchScope());
2393 // If this is an integer load past the end of the slice (which means the
2394 // bytes outside the slice are undef or this load is dead) just forcibly
2395 // fix the integer size with correct handling of endianness.
2396 if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2397 if (auto *TITy = dyn_cast<IntegerType>(TargetTy))
2398 if (AITy->getBitWidth() < TITy->getBitWidth()) {
2399 V = IRB.CreateZExt(V, TITy, "load.ext");
2400 if (DL.isBigEndian())
2401 V = IRB.CreateShl(V, TITy->getBitWidth() - AITy->getBitWidth(),
2405 Type *LTy = TargetTy->getPointerTo();
2406 LoadInst *NewLI = IRB.CreateAlignedLoad(getNewAllocaSlicePtr(IRB, LTy),
2407 getSliceAlign(TargetTy),
2408 LI.isVolatile(), LI.getName());
2409 if (LI.isVolatile())
2410 NewLI->setAtomic(LI.getOrdering(), LI.getSynchScope());
2413 IsPtrAdjusted = true;
2415 V = convertValue(DL, IRB, V, TargetTy);
2418 assert(!LI.isVolatile());
2419 assert(LI.getType()->isIntegerTy() &&
2420 "Only integer type loads and stores are split");
2421 assert(SliceSize < DL.getTypeStoreSize(LI.getType()) &&
2422 "Split load isn't smaller than original load");
2423 assert(LI.getType()->getIntegerBitWidth() ==
2424 DL.getTypeStoreSizeInBits(LI.getType()) &&
2425 "Non-byte-multiple bit width");
2426 // Move the insertion point just past the load so that we can refer to it.
2427 IRB.SetInsertPoint(&*std::next(BasicBlock::iterator(&LI)));
2428 // Create a placeholder value with the same type as LI to use as the
2429 // basis for the new value. This allows us to replace the uses of LI with
2430 // the computed value, and then replace the placeholder with LI, leaving
2431 // LI only used for this computation.
2432 Value *Placeholder =
2433 new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2434 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset - BeginOffset,
2436 LI.replaceAllUsesWith(V);
2437 Placeholder->replaceAllUsesWith(&LI);
2440 LI.replaceAllUsesWith(V);
2443 Pass.DeadInsts.insert(&LI);
2444 deleteIfTriviallyDead(OldOp);
2445 DEBUG(dbgs() << " to: " << *V << "\n");
2446 return !LI.isVolatile() && !IsPtrAdjusted;
2449 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp) {
2450 if (V->getType() != VecTy) {
2451 unsigned BeginIndex = getIndex(NewBeginOffset);
2452 unsigned EndIndex = getIndex(NewEndOffset);
2453 assert(EndIndex > BeginIndex && "Empty vector!");
2454 unsigned NumElements = EndIndex - BeginIndex;
2455 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2456 Type *SliceTy = (NumElements == 1)
2458 : VectorType::get(ElementTy, NumElements);
2459 if (V->getType() != SliceTy)
2460 V = convertValue(DL, IRB, V, SliceTy);
2462 // Mix in the existing elements.
2463 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2464 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2466 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2467 Pass.DeadInsts.insert(&SI);
2470 DEBUG(dbgs() << " to: " << *Store << "\n");
2474 bool rewriteIntegerStore(Value *V, StoreInst &SI) {
2475 assert(IntTy && "We cannot extract an integer from the alloca");
2476 assert(!SI.isVolatile());
2477 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2479 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2480 Old = convertValue(DL, IRB, Old, IntTy);
2481 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2482 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2483 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset, "insert");
2485 V = convertValue(DL, IRB, V, NewAllocaTy);
2486 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2487 Store->copyMetadata(SI, LLVMContext::MD_mem_parallel_loop_access);
2488 Pass.DeadInsts.insert(&SI);
2489 DEBUG(dbgs() << " to: " << *Store << "\n");
2493 bool visitStoreInst(StoreInst &SI) {
2494 DEBUG(dbgs() << " original: " << SI << "\n");
2495 Value *OldOp = SI.getOperand(1);
2496 assert(OldOp == OldPtr);
2498 Value *V = SI.getValueOperand();
2500 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2501 // alloca that should be re-examined after promoting this alloca.
2502 if (V->getType()->isPointerTy())
2503 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2504 Pass.PostPromotionWorklist.insert(AI);
2506 if (SliceSize < DL.getTypeStoreSize(V->getType())) {
2507 assert(!SI.isVolatile());
2508 assert(V->getType()->isIntegerTy() &&
2509 "Only integer type loads and stores are split");
2510 assert(V->getType()->getIntegerBitWidth() ==
2511 DL.getTypeStoreSizeInBits(V->getType()) &&
2512 "Non-byte-multiple bit width");
2513 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2514 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset - BeginOffset,
2519 return rewriteVectorizedStoreInst(V, SI, OldOp);
2520 if (IntTy && V->getType()->isIntegerTy())
2521 return rewriteIntegerStore(V, SI);
2523 const bool IsStorePastEnd = DL.getTypeStoreSize(V->getType()) > SliceSize;
2525 if (NewBeginOffset == NewAllocaBeginOffset &&
2526 NewEndOffset == NewAllocaEndOffset &&
2527 (canConvertValue(DL, V->getType(), NewAllocaTy) ||
2528 (IsStorePastEnd && NewAllocaTy->isIntegerTy() &&
2529 V->getType()->isIntegerTy()))) {
2530 // If this is an integer store past the end of slice (and thus the bytes
2531 // past that point are irrelevant or this is unreachable), truncate the
2532 // value prior to storing.
2533 if (auto *VITy = dyn_cast<IntegerType>(V->getType()))
2534 if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2535 if (VITy->getBitWidth() > AITy->getBitWidth()) {
2536 if (DL.isBigEndian())
2537 V = IRB.CreateLShr(V, VITy->getBitWidth() - AITy->getBitWidth(),
2539 V = IRB.CreateTrunc(V, AITy, "load.trunc");
2542 V = convertValue(DL, IRB, V, NewAllocaTy);
2543 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2546 Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo());
2547 NewSI = IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(V->getType()),
2550 NewSI->copyMetadata(SI, LLVMContext::MD_mem_parallel_loop_access);
2551 if (SI.isVolatile())
2552 NewSI->setAtomic(SI.getOrdering(), SI.getSynchScope());
2553 Pass.DeadInsts.insert(&SI);
2554 deleteIfTriviallyDead(OldOp);
2556 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2557 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2560 /// \brief Compute an integer value from splatting an i8 across the given
2561 /// number of bytes.
2563 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2564 /// call this routine.
2565 /// FIXME: Heed the advice above.
2567 /// \param V The i8 value to splat.
2568 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2569 Value *getIntegerSplat(Value *V, unsigned Size) {
2570 assert(Size > 0 && "Expected a positive number of bytes.");
2571 IntegerType *VTy = cast<IntegerType>(V->getType());
2572 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2576 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size * 8);
2578 IRB.CreateZExt(V, SplatIntTy, "zext"),
2579 ConstantExpr::getUDiv(
2580 Constant::getAllOnesValue(SplatIntTy),
2581 ConstantExpr::getZExt(Constant::getAllOnesValue(V->getType()),
2587 /// \brief Compute a vector splat for a given element value.
2588 Value *getVectorSplat(Value *V, unsigned NumElements) {
2589 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2590 DEBUG(dbgs() << " splat: " << *V << "\n");
2594 bool visitMemSetInst(MemSetInst &II) {
2595 DEBUG(dbgs() << " original: " << II << "\n");
2596 assert(II.getRawDest() == OldPtr);
2598 // If the memset has a variable size, it cannot be split, just adjust the
2599 // pointer to the new alloca.
2600 if (!isa<Constant>(II.getLength())) {
2602 assert(NewBeginOffset == BeginOffset);
2603 II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2604 Type *CstTy = II.getAlignmentCst()->getType();
2605 II.setAlignment(ConstantInt::get(CstTy, getSliceAlign()));
2607 deleteIfTriviallyDead(OldPtr);
2611 // Record this instruction for deletion.
2612 Pass.DeadInsts.insert(&II);
2614 Type *AllocaTy = NewAI.getAllocatedType();
2615 Type *ScalarTy = AllocaTy->getScalarType();
2617 // If this doesn't map cleanly onto the alloca type, and that type isn't
2618 // a single value type, just emit a memset.
2619 if (!VecTy && !IntTy &&
2620 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2621 SliceSize != DL.getTypeStoreSize(AllocaTy) ||
2622 !AllocaTy->isSingleValueType() ||
2623 !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2624 DL.getTypeSizeInBits(ScalarTy) % 8 != 0)) {
2625 Type *SizeTy = II.getLength()->getType();
2626 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2627 CallInst *New = IRB.CreateMemSet(
2628 getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2629 getSliceAlign(), II.isVolatile());
2631 DEBUG(dbgs() << " to: " << *New << "\n");
2635 // If we can represent this as a simple value, we have to build the actual
2636 // value to store, which requires expanding the byte present in memset to
2637 // a sensible representation for the alloca type. This is essentially
2638 // splatting the byte to a sufficiently wide integer, splatting it across
2639 // any desired vector width, and bitcasting to the final type.
2643 // If this is a memset of a vectorized alloca, insert it.
2644 assert(ElementTy == ScalarTy);
2646 unsigned BeginIndex = getIndex(NewBeginOffset);
2647 unsigned EndIndex = getIndex(NewEndOffset);
2648 assert(EndIndex > BeginIndex && "Empty vector!");
2649 unsigned NumElements = EndIndex - BeginIndex;
2650 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2653 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2654 Splat = convertValue(DL, IRB, Splat, ElementTy);
2655 if (NumElements > 1)
2656 Splat = getVectorSplat(Splat, NumElements);
2659 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2660 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2662 // If this is a memset on an alloca where we can widen stores, insert the
2664 assert(!II.isVolatile());
2666 uint64_t Size = NewEndOffset - NewBeginOffset;
2667 V = getIntegerSplat(II.getValue(), Size);
2669 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2670 EndOffset != NewAllocaBeginOffset)) {
2672 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2673 Old = convertValue(DL, IRB, Old, IntTy);
2674 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2675 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2677 assert(V->getType() == IntTy &&
2678 "Wrong type for an alloca wide integer!");
2680 V = convertValue(DL, IRB, V, AllocaTy);
2682 // Established these invariants above.
2683 assert(NewBeginOffset == NewAllocaBeginOffset);
2684 assert(NewEndOffset == NewAllocaEndOffset);
2686 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2687 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2688 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2690 V = convertValue(DL, IRB, V, AllocaTy);
2693 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2696 DEBUG(dbgs() << " to: " << *New << "\n");
2697 return !II.isVolatile();
2700 bool visitMemTransferInst(MemTransferInst &II) {
2701 // Rewriting of memory transfer instructions can be a bit tricky. We break
2702 // them into two categories: split intrinsics and unsplit intrinsics.
2704 DEBUG(dbgs() << " original: " << II << "\n");
2706 bool IsDest = &II.getRawDestUse() == OldUse;
2707 assert((IsDest && II.getRawDest() == OldPtr) ||
2708 (!IsDest && II.getRawSource() == OldPtr));
2710 unsigned SliceAlign = getSliceAlign();
2712 // For unsplit intrinsics, we simply modify the source and destination
2713 // pointers in place. This isn't just an optimization, it is a matter of
2714 // correctness. With unsplit intrinsics we may be dealing with transfers
2715 // within a single alloca before SROA ran, or with transfers that have
2716 // a variable length. We may also be dealing with memmove instead of
2717 // memcpy, and so simply updating the pointers is the necessary for us to
2718 // update both source and dest of a single call.
2719 if (!IsSplittable) {
2720 Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2722 II.setDest(AdjustedPtr);
2724 II.setSource(AdjustedPtr);
2726 if (II.getAlignment() > SliceAlign) {
2727 Type *CstTy = II.getAlignmentCst()->getType();
2729 ConstantInt::get(CstTy, MinAlign(II.getAlignment(), SliceAlign)));
2732 DEBUG(dbgs() << " to: " << II << "\n");
2733 deleteIfTriviallyDead(OldPtr);
2736 // For split transfer intrinsics we have an incredibly useful assurance:
2737 // the source and destination do not reside within the same alloca, and at
2738 // least one of them does not escape. This means that we can replace
2739 // memmove with memcpy, and we don't need to worry about all manner of
2740 // downsides to splitting and transforming the operations.
2742 // If this doesn't map cleanly onto the alloca type, and that type isn't
2743 // a single value type, just emit a memcpy.
2746 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2747 SliceSize != DL.getTypeStoreSize(NewAI.getAllocatedType()) ||
2748 !NewAI.getAllocatedType()->isSingleValueType());
2750 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2751 // size hasn't been shrunk based on analysis of the viable range, this is
2753 if (EmitMemCpy && &OldAI == &NewAI) {
2754 // Ensure the start lines up.
2755 assert(NewBeginOffset == BeginOffset);
2757 // Rewrite the size as needed.
2758 if (NewEndOffset != EndOffset)
2759 II.setLength(ConstantInt::get(II.getLength()->getType(),
2760 NewEndOffset - NewBeginOffset));
2763 // Record this instruction for deletion.
2764 Pass.DeadInsts.insert(&II);
2766 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2767 // alloca that should be re-examined after rewriting this instruction.
2768 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2769 if (AllocaInst *AI =
2770 dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2771 assert(AI != &OldAI && AI != &NewAI &&
2772 "Splittable transfers cannot reach the same alloca on both ends.");
2773 Pass.Worklist.insert(AI);
2776 Type *OtherPtrTy = OtherPtr->getType();
2777 unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2779 // Compute the relative offset for the other pointer within the transfer.
2780 unsigned IntPtrWidth = DL.getPointerSizeInBits(OtherAS);
2781 APInt OtherOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2782 unsigned OtherAlign = MinAlign(II.getAlignment() ? II.getAlignment() : 1,
2783 OtherOffset.zextOrTrunc(64).getZExtValue());
2786 // Compute the other pointer, folding as much as possible to produce
2787 // a single, simple GEP in most cases.
2788 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2789 OtherPtr->getName() + ".");
2791 Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2792 Type *SizeTy = II.getLength()->getType();
2793 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2795 CallInst *New = IRB.CreateMemCpy(
2796 IsDest ? OurPtr : OtherPtr, IsDest ? OtherPtr : OurPtr, Size,
2797 MinAlign(SliceAlign, OtherAlign), II.isVolatile());
2799 DEBUG(dbgs() << " to: " << *New << "\n");
2803 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2804 NewEndOffset == NewAllocaEndOffset;
2805 uint64_t Size = NewEndOffset - NewBeginOffset;
2806 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2807 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2808 unsigned NumElements = EndIndex - BeginIndex;
2809 IntegerType *SubIntTy =
2810 IntTy ? Type::getIntNTy(IntTy->getContext(), Size * 8) : nullptr;
2812 // Reset the other pointer type to match the register type we're going to
2813 // use, but using the address space of the original other pointer.
2814 if (VecTy && !IsWholeAlloca) {
2815 if (NumElements == 1)
2816 OtherPtrTy = VecTy->getElementType();
2818 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2820 OtherPtrTy = OtherPtrTy->getPointerTo(OtherAS);
2821 } else if (IntTy && !IsWholeAlloca) {
2822 OtherPtrTy = SubIntTy->getPointerTo(OtherAS);
2824 OtherPtrTy = NewAllocaTy->getPointerTo(OtherAS);
2827 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2828 OtherPtr->getName() + ".");
2829 unsigned SrcAlign = OtherAlign;
2830 Value *DstPtr = &NewAI;
2831 unsigned DstAlign = SliceAlign;
2833 std::swap(SrcPtr, DstPtr);
2834 std::swap(SrcAlign, DstAlign);
2838 if (VecTy && !IsWholeAlloca && !IsDest) {
2839 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2840 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2841 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2842 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2843 Src = convertValue(DL, IRB, Src, IntTy);
2844 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2845 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2848 IRB.CreateAlignedLoad(SrcPtr, SrcAlign, II.isVolatile(), "copyload");
2851 if (VecTy && !IsWholeAlloca && IsDest) {
2853 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2854 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2855 } else if (IntTy && !IsWholeAlloca && IsDest) {
2857 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2858 Old = convertValue(DL, IRB, Old, IntTy);
2859 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2860 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
2861 Src = convertValue(DL, IRB, Src, NewAllocaTy);
2864 StoreInst *Store = cast<StoreInst>(
2865 IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
2867 DEBUG(dbgs() << " to: " << *Store << "\n");
2868 return !II.isVolatile();
2871 bool visitIntrinsicInst(IntrinsicInst &II) {
2872 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2873 II.getIntrinsicID() == Intrinsic::lifetime_end);
2874 DEBUG(dbgs() << " original: " << II << "\n");
2875 assert(II.getArgOperand(1) == OldPtr);
2877 // Record this instruction for deletion.
2878 Pass.DeadInsts.insert(&II);
2880 // Lifetime intrinsics are only promotable if they cover the whole alloca.
2881 // Therefore, we drop lifetime intrinsics which don't cover the whole
2883 // (In theory, intrinsics which partially cover an alloca could be
2884 // promoted, but PromoteMemToReg doesn't handle that case.)
2885 // FIXME: Check whether the alloca is promotable before dropping the
2886 // lifetime intrinsics?
2887 if (NewBeginOffset != NewAllocaBeginOffset ||
2888 NewEndOffset != NewAllocaEndOffset)
2892 ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2893 NewEndOffset - NewBeginOffset);
2894 Value *Ptr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2896 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2897 New = IRB.CreateLifetimeStart(Ptr, Size);
2899 New = IRB.CreateLifetimeEnd(Ptr, Size);
2902 DEBUG(dbgs() << " to: " << *New << "\n");
2907 bool visitPHINode(PHINode &PN) {
2908 DEBUG(dbgs() << " original: " << PN << "\n");
2909 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
2910 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
2912 // We would like to compute a new pointer in only one place, but have it be
2913 // as local as possible to the PHI. To do that, we re-use the location of
2914 // the old pointer, which necessarily must be in the right position to
2915 // dominate the PHI.
2916 IRBuilderTy PtrBuilder(IRB);
2917 if (isa<PHINode>(OldPtr))
2918 PtrBuilder.SetInsertPoint(&*OldPtr->getParent()->getFirstInsertionPt());
2920 PtrBuilder.SetInsertPoint(OldPtr);
2921 PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
2923 Value *NewPtr = getNewAllocaSlicePtr(PtrBuilder, OldPtr->getType());
2924 // Replace the operands which were using the old pointer.
2925 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
2927 DEBUG(dbgs() << " to: " << PN << "\n");
2928 deleteIfTriviallyDead(OldPtr);
2930 // PHIs can't be promoted on their own, but often can be speculated. We
2931 // check the speculation outside of the rewriter so that we see the
2932 // fully-rewritten alloca.
2933 PHIUsers.insert(&PN);
2937 bool visitSelectInst(SelectInst &SI) {
2938 DEBUG(dbgs() << " original: " << SI << "\n");
2939 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
2940 "Pointer isn't an operand!");
2941 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
2942 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
2944 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2945 // Replace the operands which were using the old pointer.
2946 if (SI.getOperand(1) == OldPtr)
2947 SI.setOperand(1, NewPtr);
2948 if (SI.getOperand(2) == OldPtr)
2949 SI.setOperand(2, NewPtr);
2951 DEBUG(dbgs() << " to: " << SI << "\n");
2952 deleteIfTriviallyDead(OldPtr);
2954 // Selects can't be promoted on their own, but often can be speculated. We
2955 // check the speculation outside of the rewriter so that we see the
2956 // fully-rewritten alloca.
2957 SelectUsers.insert(&SI);
2963 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2965 /// This pass aggressively rewrites all aggregate loads and stores on
2966 /// a particular pointer (or any pointer derived from it which we can identify)
2967 /// with scalar loads and stores.
2968 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2969 // Befriend the base class so it can delegate to private visit methods.
2970 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2972 /// Queue of pointer uses to analyze and potentially rewrite.
2973 SmallVector<Use *, 8> Queue;
2975 /// Set to prevent us from cycling with phi nodes and loops.
2976 SmallPtrSet<User *, 8> Visited;
2978 /// The current pointer use being rewritten. This is used to dig up the used
2979 /// value (as opposed to the user).
2983 /// Rewrite loads and stores through a pointer and all pointers derived from
2985 bool rewrite(Instruction &I) {
2986 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2988 bool Changed = false;
2989 while (!Queue.empty()) {
2990 U = Queue.pop_back_val();
2991 Changed |= visit(cast<Instruction>(U->getUser()));
2997 /// Enqueue all the users of the given instruction for further processing.
2998 /// This uses a set to de-duplicate users.
2999 void enqueueUsers(Instruction &I) {
3000 for (Use &U : I.uses())
3001 if (Visited.insert(U.getUser()).second)
3002 Queue.push_back(&U);
3005 // Conservative default is to not rewrite anything.
3006 bool visitInstruction(Instruction &I) { return false; }
3008 /// \brief Generic recursive split emission class.
3009 template <typename Derived> class OpSplitter {
3011 /// The builder used to form new instructions.
3013 /// The indices which to be used with insert- or extractvalue to select the
3014 /// appropriate value within the aggregate.
3015 SmallVector<unsigned, 4> Indices;
3016 /// The indices to a GEP instruction which will move Ptr to the correct slot
3017 /// within the aggregate.
3018 SmallVector<Value *, 4> GEPIndices;
3019 /// The base pointer of the original op, used as a base for GEPing the
3020 /// split operations.
3023 /// Initialize the splitter with an insertion point, Ptr and start with a
3024 /// single zero GEP index.
3025 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
3026 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
3029 /// \brief Generic recursive split emission routine.
3031 /// This method recursively splits an aggregate op (load or store) into
3032 /// scalar or vector ops. It splits recursively until it hits a single value
3033 /// and emits that single value operation via the template argument.
3035 /// The logic of this routine relies on GEPs and insertvalue and
3036 /// extractvalue all operating with the same fundamental index list, merely
3037 /// formatted differently (GEPs need actual values).
3039 /// \param Ty The type being split recursively into smaller ops.
3040 /// \param Agg The aggregate value being built up or stored, depending on
3041 /// whether this is splitting a load or a store respectively.
3042 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3043 if (Ty->isSingleValueType())
3044 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
3046 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3047 unsigned OldSize = Indices.size();
3049 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3051 assert(Indices.size() == OldSize && "Did not return to the old size");
3052 Indices.push_back(Idx);
3053 GEPIndices.push_back(IRB.getInt32(Idx));
3054 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3055 GEPIndices.pop_back();
3061 if (StructType *STy = dyn_cast<StructType>(Ty)) {
3062 unsigned OldSize = Indices.size();
3064 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3066 assert(Indices.size() == OldSize && "Did not return to the old size");
3067 Indices.push_back(Idx);
3068 GEPIndices.push_back(IRB.getInt32(Idx));
3069 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3070 GEPIndices.pop_back();
3076 llvm_unreachable("Only arrays and structs are aggregate loadable types");
3080 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3081 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3082 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
3084 /// Emit a leaf load of a single value. This is called at the leaves of the
3085 /// recursive emission to actually load values.
3086 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3087 assert(Ty->isSingleValueType());
3088 // Load the single value and insert it using the indices.
3090 IRB.CreateInBoundsGEP(nullptr, Ptr, GEPIndices, Name + ".gep");
3091 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
3092 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3093 DEBUG(dbgs() << " to: " << *Load << "\n");
3097 bool visitLoadInst(LoadInst &LI) {
3098 assert(LI.getPointerOperand() == *U);
3099 if (!LI.isSimple() || LI.getType()->isSingleValueType())
3102 // We have an aggregate being loaded, split it apart.
3103 DEBUG(dbgs() << " original: " << LI << "\n");
3104 LoadOpSplitter Splitter(&LI, *U);
3105 Value *V = UndefValue::get(LI.getType());
3106 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3107 LI.replaceAllUsesWith(V);
3108 LI.eraseFromParent();
3112 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3113 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3114 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
3116 /// Emit a leaf store of a single value. This is called at the leaves of the
3117 /// recursive emission to actually produce stores.
3118 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3119 assert(Ty->isSingleValueType());
3120 // Extract the single value and store it using the indices.
3122 // The gep and extractvalue values are factored out of the CreateStore
3123 // call to make the output independent of the argument evaluation order.
3124 Value *ExtractValue =
3125 IRB.CreateExtractValue(Agg, Indices, Name + ".extract");
3126 Value *InBoundsGEP =
3127 IRB.CreateInBoundsGEP(nullptr, Ptr, GEPIndices, Name + ".gep");
3128 Value *Store = IRB.CreateStore(ExtractValue, InBoundsGEP);
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 Type *ElementTy = SeqTy->getElementType();
3224 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3225 uint64_t NumSkippedElements = Offset / ElementSize;
3226 if (NumSkippedElements >= SeqTy->getNumElements())
3228 Offset -= NumSkippedElements * ElementSize;
3230 // First check if we need to recurse.
3231 if (Offset > 0 || Size < ElementSize) {
3232 // Bail if the partition ends in a different array element.
3233 if ((Offset + Size) > ElementSize)
3235 // Recurse through the element type trying to peel off offset bytes.
3236 return getTypePartition(DL, ElementTy, Offset, Size);
3238 assert(Offset == 0);
3240 if (Size == ElementSize)
3241 return stripAggregateTypeWrapping(DL, ElementTy);
3242 assert(Size > ElementSize);
3243 uint64_t NumElements = Size / ElementSize;
3244 if (NumElements * ElementSize != Size)
3246 return ArrayType::get(ElementTy, NumElements);
3249 StructType *STy = dyn_cast<StructType>(Ty);
3253 const StructLayout *SL = DL.getStructLayout(STy);
3254 if (Offset >= SL->getSizeInBytes())
3256 uint64_t EndOffset = Offset + Size;
3257 if (EndOffset > SL->getSizeInBytes())
3260 unsigned Index = SL->getElementContainingOffset(Offset);
3261 Offset -= SL->getElementOffset(Index);
3263 Type *ElementTy = STy->getElementType(Index);
3264 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3265 if (Offset >= ElementSize)
3266 return nullptr; // The offset points into alignment padding.
3268 // See if any partition must be contained by the element.
3269 if (Offset > 0 || Size < ElementSize) {
3270 if ((Offset + Size) > ElementSize)
3272 return getTypePartition(DL, ElementTy, Offset, Size);
3274 assert(Offset == 0);
3276 if (Size == ElementSize)
3277 return stripAggregateTypeWrapping(DL, ElementTy);
3279 StructType::element_iterator EI = STy->element_begin() + Index,
3280 EE = STy->element_end();
3281 if (EndOffset < SL->getSizeInBytes()) {
3282 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3283 if (Index == EndIndex)
3284 return nullptr; // Within a single element and its padding.
3286 // Don't try to form "natural" types if the elements don't line up with the
3288 // FIXME: We could potentially recurse down through the last element in the
3289 // sub-struct to find a natural end point.
3290 if (SL->getElementOffset(EndIndex) != EndOffset)
3293 assert(Index < EndIndex);
3294 EE = STy->element_begin() + EndIndex;
3297 // Try to build up a sub-structure.
3299 StructType::get(STy->getContext(), makeArrayRef(EI, EE), STy->isPacked());
3300 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3301 if (Size != SubSL->getSizeInBytes())
3302 return nullptr; // The sub-struct doesn't have quite the size needed.
3307 /// \brief Pre-split loads and stores to simplify rewriting.
3309 /// We want to break up the splittable load+store pairs as much as
3310 /// possible. This is important to do as a preprocessing step, as once we
3311 /// start rewriting the accesses to partitions of the alloca we lose the
3312 /// necessary information to correctly split apart paired loads and stores
3313 /// which both point into this alloca. The case to consider is something like
3316 /// %a = alloca [12 x i8]
3317 /// %gep1 = getelementptr [12 x i8]* %a, i32 0, i32 0
3318 /// %gep2 = getelementptr [12 x i8]* %a, i32 0, i32 4
3319 /// %gep3 = getelementptr [12 x i8]* %a, i32 0, i32 8
3320 /// %iptr1 = bitcast i8* %gep1 to i64*
3321 /// %iptr2 = bitcast i8* %gep2 to i64*
3322 /// %fptr1 = bitcast i8* %gep1 to float*
3323 /// %fptr2 = bitcast i8* %gep2 to float*
3324 /// %fptr3 = bitcast i8* %gep3 to float*
3325 /// store float 0.0, float* %fptr1
3326 /// store float 1.0, float* %fptr2
3327 /// %v = load i64* %iptr1
3328 /// store i64 %v, i64* %iptr2
3329 /// %f1 = load float* %fptr2
3330 /// %f2 = load float* %fptr3
3332 /// Here we want to form 3 partitions of the alloca, each 4 bytes large, and
3333 /// promote everything so we recover the 2 SSA values that should have been
3334 /// there all along.
3336 /// \returns true if any changes are made.
3337 bool SROA::presplitLoadsAndStores(AllocaInst &AI, AllocaSlices &AS) {
3338 DEBUG(dbgs() << "Pre-splitting loads and stores\n");
3340 // Track the loads and stores which are candidates for pre-splitting here, in
3341 // the order they first appear during the partition scan. These give stable
3342 // iteration order and a basis for tracking which loads and stores we
3344 SmallVector<LoadInst *, 4> Loads;
3345 SmallVector<StoreInst *, 4> Stores;
3347 // We need to accumulate the splits required of each load or store where we
3348 // can find them via a direct lookup. This is important to cross-check loads
3349 // and stores against each other. We also track the slice so that we can kill
3350 // all the slices that end up split.
3351 struct SplitOffsets {
3353 std::vector<uint64_t> Splits;
3355 SmallDenseMap<Instruction *, SplitOffsets, 8> SplitOffsetsMap;
3357 // Track loads out of this alloca which cannot, for any reason, be pre-split.
3358 // This is important as we also cannot pre-split stores of those loads!
3359 // FIXME: This is all pretty gross. It means that we can be more aggressive
3360 // in pre-splitting when the load feeding the store happens to come from
3361 // a separate alloca. Put another way, the effectiveness of SROA would be
3362 // decreased by a frontend which just concatenated all of its local allocas
3363 // into one big flat alloca. But defeating such patterns is exactly the job
3364 // SROA is tasked with! Sadly, to not have this discrepancy we would have
3365 // change store pre-splitting to actually force pre-splitting of the load
3366 // that feeds it *and all stores*. That makes pre-splitting much harder, but
3367 // maybe it would make it more principled?
3368 SmallPtrSet<LoadInst *, 8> UnsplittableLoads;
3370 DEBUG(dbgs() << " Searching for candidate loads and stores\n");
3371 for (auto &P : AS.partitions()) {
3372 for (Slice &S : P) {
3373 Instruction *I = cast<Instruction>(S.getUse()->getUser());
3374 if (!S.isSplittable() || S.endOffset() <= P.endOffset()) {
3375 // If this is a load we have to track that it can't participate in any
3376 // pre-splitting. If this is a store of a load we have to track that
3377 // that load also can't participate in any pre-splitting.
3378 if (auto *LI = dyn_cast<LoadInst>(I))
3379 UnsplittableLoads.insert(LI);
3380 else if (auto *SI = dyn_cast<StoreInst>(I))
3381 if (auto *LI = dyn_cast<LoadInst>(SI->getValueOperand()))
3382 UnsplittableLoads.insert(LI);
3385 assert(P.endOffset() > S.beginOffset() &&
3386 "Empty or backwards partition!");
3388 // Determine if this is a pre-splittable slice.
3389 if (auto *LI = dyn_cast<LoadInst>(I)) {
3390 assert(!LI->isVolatile() && "Cannot split volatile loads!");
3392 // The load must be used exclusively to store into other pointers for
3393 // us to be able to arbitrarily pre-split it. The stores must also be
3394 // simple to avoid changing semantics.
3395 auto IsLoadSimplyStored = [](LoadInst *LI) {
3396 for (User *LU : LI->users()) {
3397 auto *SI = dyn_cast<StoreInst>(LU);
3398 if (!SI || !SI->isSimple())
3403 if (!IsLoadSimplyStored(LI)) {
3404 UnsplittableLoads.insert(LI);
3408 Loads.push_back(LI);
3409 } else if (auto *SI = dyn_cast<StoreInst>(I)) {
3410 if (S.getUse() != &SI->getOperandUse(SI->getPointerOperandIndex()))
3411 // Skip stores *of* pointers. FIXME: This shouldn't even be possible!
3413 auto *StoredLoad = dyn_cast<LoadInst>(SI->getValueOperand());
3414 if (!StoredLoad || !StoredLoad->isSimple())
3416 assert(!SI->isVolatile() && "Cannot split volatile stores!");
3418 Stores.push_back(SI);
3420 // Other uses cannot be pre-split.
3424 // Record the initial split.
3425 DEBUG(dbgs() << " Candidate: " << *I << "\n");
3426 auto &Offsets = SplitOffsetsMap[I];
3427 assert(Offsets.Splits.empty() &&
3428 "Should not have splits the first time we see an instruction!");
3430 Offsets.Splits.push_back(P.endOffset() - S.beginOffset());
3433 // Now scan the already split slices, and add a split for any of them which
3434 // we're going to pre-split.
3435 for (Slice *S : P.splitSliceTails()) {
3436 auto SplitOffsetsMapI =
3437 SplitOffsetsMap.find(cast<Instruction>(S->getUse()->getUser()));
3438 if (SplitOffsetsMapI == SplitOffsetsMap.end())
3440 auto &Offsets = SplitOffsetsMapI->second;
3442 assert(Offsets.S == S && "Found a mismatched slice!");
3443 assert(!Offsets.Splits.empty() &&
3444 "Cannot have an empty set of splits on the second partition!");
3445 assert(Offsets.Splits.back() ==
3446 P.beginOffset() - Offsets.S->beginOffset() &&
3447 "Previous split does not end where this one begins!");
3449 // Record each split. The last partition's end isn't needed as the size
3450 // of the slice dictates that.
3451 if (S->endOffset() > P.endOffset())
3452 Offsets.Splits.push_back(P.endOffset() - Offsets.S->beginOffset());
3456 // We may have split loads where some of their stores are split stores. For
3457 // such loads and stores, we can only pre-split them if their splits exactly
3458 // match relative to their starting offset. We have to verify this prior to
3462 [&UnsplittableLoads, &SplitOffsetsMap](StoreInst *SI) {
3463 // Lookup the load we are storing in our map of split
3465 auto *LI = cast<LoadInst>(SI->getValueOperand());
3466 // If it was completely unsplittable, then we're done,
3467 // and this store can't be pre-split.
3468 if (UnsplittableLoads.count(LI))
3471 auto LoadOffsetsI = SplitOffsetsMap.find(LI);
3472 if (LoadOffsetsI == SplitOffsetsMap.end())
3473 return false; // Unrelated loads are definitely safe.
3474 auto &LoadOffsets = LoadOffsetsI->second;
3476 // Now lookup the store's offsets.
3477 auto &StoreOffsets = SplitOffsetsMap[SI];
3479 // If the relative offsets of each split in the load and
3480 // store match exactly, then we can split them and we
3481 // don't need to remove them here.
3482 if (LoadOffsets.Splits == StoreOffsets.Splits)
3485 DEBUG(dbgs() << " Mismatched splits for load and store:\n"
3486 << " " << *LI << "\n"
3487 << " " << *SI << "\n");
3489 // We've found a store and load that we need to split
3490 // with mismatched relative splits. Just give up on them
3491 // and remove both instructions from our list of
3493 UnsplittableLoads.insert(LI);
3497 // Now we have to go *back* through all the stores, because a later store may
3498 // have caused an earlier store's load to become unsplittable and if it is
3499 // unsplittable for the later store, then we can't rely on it being split in
3500 // the earlier store either.
3501 Stores.erase(remove_if(Stores,
3502 [&UnsplittableLoads](StoreInst *SI) {
3503 auto *LI = cast<LoadInst>(SI->getValueOperand());
3504 return UnsplittableLoads.count(LI);
3507 // Once we've established all the loads that can't be split for some reason,
3508 // filter any that made it into our list out.
3509 Loads.erase(remove_if(Loads,
3510 [&UnsplittableLoads](LoadInst *LI) {
3511 return UnsplittableLoads.count(LI);
3515 // If no loads or stores are left, there is no pre-splitting to be done for
3517 if (Loads.empty() && Stores.empty())
3520 // From here on, we can't fail and will be building new accesses, so rig up
3522 IRBuilderTy IRB(&AI);
3524 // Collect the new slices which we will merge into the alloca slices.
3525 SmallVector<Slice, 4> NewSlices;
3527 // Track any allocas we end up splitting loads and stores for so we iterate
3529 SmallPtrSet<AllocaInst *, 4> ResplitPromotableAllocas;
3531 // At this point, we have collected all of the loads and stores we can
3532 // pre-split, and the specific splits needed for them. We actually do the
3533 // splitting in a specific order in order to handle when one of the loads in
3534 // the value operand to one of the stores.
3536 // First, we rewrite all of the split loads, and just accumulate each split
3537 // load in a parallel structure. We also build the slices for them and append
3538 // them to the alloca slices.
3539 SmallDenseMap<LoadInst *, std::vector<LoadInst *>, 1> SplitLoadsMap;
3540 std::vector<LoadInst *> SplitLoads;
3541 const DataLayout &DL = AI.getModule()->getDataLayout();
3542 for (LoadInst *LI : Loads) {
3545 IntegerType *Ty = cast<IntegerType>(LI->getType());
3546 uint64_t LoadSize = Ty->getBitWidth() / 8;
3547 assert(LoadSize > 0 && "Cannot have a zero-sized integer load!");
3549 auto &Offsets = SplitOffsetsMap[LI];
3550 assert(LoadSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3551 "Slice size should always match load size exactly!");
3552 uint64_t BaseOffset = Offsets.S->beginOffset();
3553 assert(BaseOffset + LoadSize > BaseOffset &&
3554 "Cannot represent alloca access size using 64-bit integers!");
3556 Instruction *BasePtr = cast<Instruction>(LI->getPointerOperand());
3557 IRB.SetInsertPoint(LI);
3559 DEBUG(dbgs() << " Splitting load: " << *LI << "\n");
3561 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3562 int Idx = 0, Size = Offsets.Splits.size();
3564 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3565 auto *PartPtrTy = PartTy->getPointerTo(LI->getPointerAddressSpace());
3566 LoadInst *PLoad = IRB.CreateAlignedLoad(
3567 getAdjustedPtr(IRB, DL, BasePtr,
3568 APInt(DL.getPointerSizeInBits(), PartOffset),
3569 PartPtrTy, BasePtr->getName() + "."),
3570 getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3572 PLoad->copyMetadata(*LI, LLVMContext::MD_mem_parallel_loop_access);
3574 // Append this load onto the list of split loads so we can find it later
3575 // to rewrite the stores.
3576 SplitLoads.push_back(PLoad);
3578 // Now build a new slice for the alloca.
3579 NewSlices.push_back(
3580 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3581 &PLoad->getOperandUse(PLoad->getPointerOperandIndex()),
3582 /*IsSplittable*/ false));
3583 DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3584 << ", " << NewSlices.back().endOffset() << "): " << *PLoad
3587 // See if we've handled all the splits.
3591 // Setup the next partition.
3592 PartOffset = Offsets.Splits[Idx];
3594 PartSize = (Idx < Size ? Offsets.Splits[Idx] : LoadSize) - PartOffset;
3597 // Now that we have the split loads, do the slow walk over all uses of the
3598 // load and rewrite them as split stores, or save the split loads to use
3599 // below if the store is going to be split there anyways.
3600 bool DeferredStores = false;
3601 for (User *LU : LI->users()) {
3602 StoreInst *SI = cast<StoreInst>(LU);
3603 if (!Stores.empty() && SplitOffsetsMap.count(SI)) {
3604 DeferredStores = true;
3605 DEBUG(dbgs() << " Deferred splitting of store: " << *SI << "\n");
3609 Value *StoreBasePtr = SI->getPointerOperand();
3610 IRB.SetInsertPoint(SI);
3612 DEBUG(dbgs() << " Splitting store of load: " << *SI << "\n");
3614 for (int Idx = 0, Size = SplitLoads.size(); Idx < Size; ++Idx) {
3615 LoadInst *PLoad = SplitLoads[Idx];
3616 uint64_t PartOffset = Idx == 0 ? 0 : Offsets.Splits[Idx - 1];
3618 PLoad->getType()->getPointerTo(SI->getPointerAddressSpace());
3620 StoreInst *PStore = IRB.CreateAlignedStore(
3621 PLoad, getAdjustedPtr(IRB, DL, StoreBasePtr,
3622 APInt(DL.getPointerSizeInBits(), PartOffset),
3623 PartPtrTy, StoreBasePtr->getName() + "."),
3624 getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3625 PStore->copyMetadata(*LI, LLVMContext::MD_mem_parallel_loop_access);
3626 DEBUG(dbgs() << " +" << PartOffset << ":" << *PStore << "\n");
3629 // We want to immediately iterate on any allocas impacted by splitting
3630 // this store, and we have to track any promotable alloca (indicated by
3631 // a direct store) as needing to be resplit because it is no longer
3633 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(StoreBasePtr)) {
3634 ResplitPromotableAllocas.insert(OtherAI);
3635 Worklist.insert(OtherAI);
3636 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3637 StoreBasePtr->stripInBoundsOffsets())) {
3638 Worklist.insert(OtherAI);
3641 // Mark the original store as dead.
3642 DeadInsts.insert(SI);
3645 // Save the split loads if there are deferred stores among the users.
3647 SplitLoadsMap.insert(std::make_pair(LI, std::move(SplitLoads)));
3649 // Mark the original load as dead and kill the original slice.
3650 DeadInsts.insert(LI);
3654 // Second, we rewrite all of the split stores. At this point, we know that
3655 // all loads from this alloca have been split already. For stores of such
3656 // loads, we can simply look up the pre-existing split loads. For stores of
3657 // other loads, we split those loads first and then write split stores of
3659 for (StoreInst *SI : Stores) {
3660 auto *LI = cast<LoadInst>(SI->getValueOperand());
3661 IntegerType *Ty = cast<IntegerType>(LI->getType());
3662 uint64_t StoreSize = Ty->getBitWidth() / 8;
3663 assert(StoreSize > 0 && "Cannot have a zero-sized integer store!");
3665 auto &Offsets = SplitOffsetsMap[SI];
3666 assert(StoreSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3667 "Slice size should always match load size exactly!");
3668 uint64_t BaseOffset = Offsets.S->beginOffset();
3669 assert(BaseOffset + StoreSize > BaseOffset &&
3670 "Cannot represent alloca access size using 64-bit integers!");
3672 Value *LoadBasePtr = LI->getPointerOperand();
3673 Instruction *StoreBasePtr = cast<Instruction>(SI->getPointerOperand());
3675 DEBUG(dbgs() << " Splitting store: " << *SI << "\n");
3677 // Check whether we have an already split load.
3678 auto SplitLoadsMapI = SplitLoadsMap.find(LI);
3679 std::vector<LoadInst *> *SplitLoads = nullptr;
3680 if (SplitLoadsMapI != SplitLoadsMap.end()) {
3681 SplitLoads = &SplitLoadsMapI->second;
3682 assert(SplitLoads->size() == Offsets.Splits.size() + 1 &&
3683 "Too few split loads for the number of splits in the store!");
3685 DEBUG(dbgs() << " of load: " << *LI << "\n");
3688 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3689 int Idx = 0, Size = Offsets.Splits.size();
3691 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3692 auto *PartPtrTy = PartTy->getPointerTo(SI->getPointerAddressSpace());
3694 // Either lookup a split load or create one.
3697 PLoad = (*SplitLoads)[Idx];
3699 IRB.SetInsertPoint(LI);
3700 PLoad = IRB.CreateAlignedLoad(
3701 getAdjustedPtr(IRB, DL, LoadBasePtr,
3702 APInt(DL.getPointerSizeInBits(), PartOffset),
3703 PartPtrTy, LoadBasePtr->getName() + "."),
3704 getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3708 // And store this partition.
3709 IRB.SetInsertPoint(SI);
3710 StoreInst *PStore = IRB.CreateAlignedStore(
3711 PLoad, getAdjustedPtr(IRB, DL, StoreBasePtr,
3712 APInt(DL.getPointerSizeInBits(), PartOffset),
3713 PartPtrTy, StoreBasePtr->getName() + "."),
3714 getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3716 // Now build a new slice for the alloca.
3717 NewSlices.push_back(
3718 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3719 &PStore->getOperandUse(PStore->getPointerOperandIndex()),
3720 /*IsSplittable*/ false));
3721 DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3722 << ", " << NewSlices.back().endOffset() << "): " << *PStore
3725 DEBUG(dbgs() << " of split load: " << *PLoad << "\n");
3728 // See if we've finished all the splits.
3732 // Setup the next partition.
3733 PartOffset = Offsets.Splits[Idx];
3735 PartSize = (Idx < Size ? Offsets.Splits[Idx] : StoreSize) - PartOffset;
3738 // We want to immediately iterate on any allocas impacted by splitting
3739 // this load, which is only relevant if it isn't a load of this alloca and
3740 // thus we didn't already split the loads above. We also have to keep track
3741 // of any promotable allocas we split loads on as they can no longer be
3744 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(LoadBasePtr)) {
3745 assert(OtherAI != &AI && "We can't re-split our own alloca!");
3746 ResplitPromotableAllocas.insert(OtherAI);
3747 Worklist.insert(OtherAI);
3748 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3749 LoadBasePtr->stripInBoundsOffsets())) {
3750 assert(OtherAI != &AI && "We can't re-split our own alloca!");
3751 Worklist.insert(OtherAI);
3755 // Mark the original store as dead now that we've split it up and kill its
3756 // slice. Note that we leave the original load in place unless this store
3757 // was its only use. It may in turn be split up if it is an alloca load
3758 // for some other alloca, but it may be a normal load. This may introduce
3759 // redundant loads, but where those can be merged the rest of the optimizer
3760 // should handle the merging, and this uncovers SSA splits which is more
3761 // important. In practice, the original loads will almost always be fully
3762 // split and removed eventually, and the splits will be merged by any
3763 // trivial CSE, including instcombine.
3764 if (LI->hasOneUse()) {
3765 assert(*LI->user_begin() == SI && "Single use isn't this store!");
3766 DeadInsts.insert(LI);
3768 DeadInsts.insert(SI);
3772 // Remove the killed slices that have ben pre-split.
3773 AS.erase(remove_if(AS, [](const Slice &S) { return S.isDead(); }), AS.end());
3775 // Insert our new slices. This will sort and merge them into the sorted
3777 AS.insert(NewSlices);
3779 DEBUG(dbgs() << " Pre-split slices:\n");
3781 for (auto I = AS.begin(), E = AS.end(); I != E; ++I)
3782 DEBUG(AS.print(dbgs(), I, " "));
3785 // Finally, don't try to promote any allocas that new require re-splitting.
3786 // They have already been added to the worklist above.
3787 PromotableAllocas.erase(
3790 [&](AllocaInst *AI) { return ResplitPromotableAllocas.count(AI); }),
3791 PromotableAllocas.end());
3796 /// \brief Rewrite an alloca partition's users.
3798 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3799 /// to rewrite uses of an alloca partition to be conducive for SSA value
3800 /// promotion. If the partition needs a new, more refined alloca, this will
3801 /// build that new alloca, preserving as much type information as possible, and
3802 /// rewrite the uses of the old alloca to point at the new one and have the
3803 /// appropriate new offsets. It also evaluates how successful the rewrite was
3804 /// at enabling promotion and if it was successful queues the alloca to be
3806 AllocaInst *SROA::rewritePartition(AllocaInst &AI, AllocaSlices &AS,
3808 // Try to compute a friendly type for this partition of the alloca. This
3809 // won't always succeed, in which case we fall back to a legal integer type
3810 // or an i8 array of an appropriate size.
3811 Type *SliceTy = nullptr;
3812 const DataLayout &DL = AI.getModule()->getDataLayout();
3813 if (Type *CommonUseTy = findCommonType(P.begin(), P.end(), P.endOffset()))
3814 if (DL.getTypeAllocSize(CommonUseTy) >= P.size())
3815 SliceTy = CommonUseTy;
3817 if (Type *TypePartitionTy = getTypePartition(DL, AI.getAllocatedType(),
3818 P.beginOffset(), P.size()))
3819 SliceTy = TypePartitionTy;
3820 if ((!SliceTy || (SliceTy->isArrayTy() &&
3821 SliceTy->getArrayElementType()->isIntegerTy())) &&
3822 DL.isLegalInteger(P.size() * 8))
3823 SliceTy = Type::getIntNTy(*C, P.size() * 8);
3825 SliceTy = ArrayType::get(Type::getInt8Ty(*C), P.size());
3826 assert(DL.getTypeAllocSize(SliceTy) >= P.size());
3828 bool IsIntegerPromotable = isIntegerWideningViable(P, SliceTy, DL);
3831 IsIntegerPromotable ? nullptr : isVectorPromotionViable(P, DL);
3835 // Check for the case where we're going to rewrite to a new alloca of the
3836 // exact same type as the original, and with the same access offsets. In that
3837 // case, re-use the existing alloca, but still run through the rewriter to
3838 // perform phi and select speculation.
3840 if (SliceTy == AI.getAllocatedType()) {
3841 assert(P.beginOffset() == 0 &&
3842 "Non-zero begin offset but same alloca type");
3844 // FIXME: We should be able to bail at this point with "nothing changed".
3845 // FIXME: We might want to defer PHI speculation until after here.
3846 // FIXME: return nullptr;
3848 unsigned Alignment = AI.getAlignment();
3850 // The minimum alignment which users can rely on when the explicit
3851 // alignment is omitted or zero is that required by the ABI for this
3853 Alignment = DL.getABITypeAlignment(AI.getAllocatedType());
3855 Alignment = MinAlign(Alignment, P.beginOffset());
3856 // If we will get at least this much alignment from the type alone, leave
3857 // the alloca's alignment unconstrained.
3858 if (Alignment <= DL.getABITypeAlignment(SliceTy))
3860 NewAI = new AllocaInst(
3861 SliceTy, nullptr, Alignment,
3862 AI.getName() + ".sroa." + Twine(P.begin() - AS.begin()), &AI);
3866 DEBUG(dbgs() << "Rewriting alloca partition "
3867 << "[" << P.beginOffset() << "," << P.endOffset()
3868 << ") to: " << *NewAI << "\n");
3870 // Track the high watermark on the worklist as it is only relevant for
3871 // promoted allocas. We will reset it to this point if the alloca is not in
3872 // fact scheduled for promotion.
3873 unsigned PPWOldSize = PostPromotionWorklist.size();
3874 unsigned NumUses = 0;
3875 SmallPtrSet<PHINode *, 8> PHIUsers;
3876 SmallPtrSet<SelectInst *, 8> SelectUsers;
3878 AllocaSliceRewriter Rewriter(DL, AS, *this, AI, *NewAI, P.beginOffset(),
3879 P.endOffset(), IsIntegerPromotable, VecTy,
3880 PHIUsers, SelectUsers);
3881 bool Promotable = true;
3882 for (Slice *S : P.splitSliceTails()) {
3883 Promotable &= Rewriter.visit(S);
3886 for (Slice &S : P) {
3887 Promotable &= Rewriter.visit(&S);
3891 NumAllocaPartitionUses += NumUses;
3892 MaxUsesPerAllocaPartition =
3893 std::max<unsigned>(NumUses, MaxUsesPerAllocaPartition);
3895 // Now that we've processed all the slices in the new partition, check if any
3896 // PHIs or Selects would block promotion.
3897 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
3900 if (!isSafePHIToSpeculate(**I)) {
3903 SelectUsers.clear();
3906 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
3907 E = SelectUsers.end();
3909 if (!isSafeSelectToSpeculate(**I)) {
3912 SelectUsers.clear();
3917 if (PHIUsers.empty() && SelectUsers.empty()) {
3918 // Promote the alloca.
3919 PromotableAllocas.push_back(NewAI);
3921 // If we have either PHIs or Selects to speculate, add them to those
3922 // worklists and re-queue the new alloca so that we promote in on the
3924 for (PHINode *PHIUser : PHIUsers)
3925 SpeculatablePHIs.insert(PHIUser);
3926 for (SelectInst *SelectUser : SelectUsers)
3927 SpeculatableSelects.insert(SelectUser);
3928 Worklist.insert(NewAI);
3931 // Drop any post-promotion work items if promotion didn't happen.
3932 while (PostPromotionWorklist.size() > PPWOldSize)
3933 PostPromotionWorklist.pop_back();
3935 // We couldn't promote and we didn't create a new partition, nothing
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.
3943 Worklist.insert(NewAI);
3949 /// \brief Walks the slices of an alloca and form partitions based on them,
3950 /// rewriting each of their uses.
3951 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &AS) {
3952 if (AS.begin() == AS.end())
3955 unsigned NumPartitions = 0;
3956 bool Changed = false;
3957 const DataLayout &DL = AI.getModule()->getDataLayout();
3959 // First try to pre-split loads and stores.
3960 Changed |= presplitLoadsAndStores(AI, AS);
3962 // Now that we have identified any pre-splitting opportunities, mark any
3963 // splittable (non-whole-alloca) loads and stores as unsplittable. If we fail
3964 // to split these during pre-splitting, we want to force them to be
3965 // rewritten into a partition.
3966 bool IsSorted = true;
3967 for (Slice &S : AS) {
3968 if (!S.isSplittable())
3970 // FIXME: We currently leave whole-alloca splittable loads and stores. This
3971 // used to be the only splittable loads and stores and we need to be
3972 // confident that the above handling of splittable loads and stores is
3973 // completely sufficient before we forcibly disable the remaining handling.
3974 if (S.beginOffset() == 0 &&
3975 S.endOffset() >= DL.getTypeAllocSize(AI.getAllocatedType()))
3977 if (isa<LoadInst>(S.getUse()->getUser()) ||
3978 isa<StoreInst>(S.getUse()->getUser())) {
3979 S.makeUnsplittable();
3984 std::sort(AS.begin(), AS.end());
3986 /// Describes the allocas introduced by rewritePartition in order to migrate
3992 Fragment(AllocaInst *AI, uint64_t O, uint64_t S)
3993 : Alloca(AI), Offset(O), Size(S) {}
3995 SmallVector<Fragment, 4> Fragments;
3997 // Rewrite each partition.
3998 for (auto &P : AS.partitions()) {
3999 if (AllocaInst *NewAI = rewritePartition(AI, AS, P)) {
4002 uint64_t SizeOfByte = 8;
4003 uint64_t AllocaSize = DL.getTypeSizeInBits(NewAI->getAllocatedType());
4004 // Don't include any padding.
4005 uint64_t Size = std::min(AllocaSize, P.size() * SizeOfByte);
4006 Fragments.push_back(Fragment(NewAI, P.beginOffset() * SizeOfByte, Size));
4012 NumAllocaPartitions += NumPartitions;
4013 MaxPartitionsPerAlloca =
4014 std::max<unsigned>(NumPartitions, MaxPartitionsPerAlloca);
4016 // Migrate debug information from the old alloca to the new alloca(s)
4017 // and the individual partitions.
4018 if (DbgDeclareInst *DbgDecl = FindAllocaDbgDeclare(&AI)) {
4019 auto *Var = DbgDecl->getVariable();
4020 auto *Expr = DbgDecl->getExpression();
4021 DIBuilder DIB(*AI.getModule(), /*AllowUnresolved*/ false);
4022 uint64_t AllocaSize = DL.getTypeSizeInBits(AI.getAllocatedType());
4023 for (auto Fragment : Fragments) {
4024 // Create a fragment expression describing the new partition or reuse AI's
4025 // expression if there is only one partition.
4026 auto *FragmentExpr = Expr;
4027 if (Fragment.Size < AllocaSize || Expr->isFragment()) {
4028 // If this alloca is already a scalar replacement of a larger aggregate,
4029 // Fragment.Offset describes the offset inside the scalar.
4030 auto ExprFragment = Expr->getFragmentInfo();
4031 uint64_t Offset = ExprFragment ? ExprFragment->OffsetInBits : 0;
4032 uint64_t Start = Offset + Fragment.Offset;
4033 uint64_t Size = Fragment.Size;
4036 ExprFragment->OffsetInBits + ExprFragment->SizeInBits;
4037 if (Start >= AbsEnd)
4038 // No need to describe a SROAed padding.
4040 Size = std::min(Size, AbsEnd - Start);
4042 FragmentExpr = DIB.createFragmentExpression(Start, Size);
4045 // Remove any existing dbg.declare intrinsic describing the same alloca.
4046 if (DbgDeclareInst *OldDDI = FindAllocaDbgDeclare(Fragment.Alloca))
4047 OldDDI->eraseFromParent();
4049 DIB.insertDeclare(Fragment.Alloca, Var, FragmentExpr,
4050 DbgDecl->getDebugLoc(), &AI);
4056 /// \brief Clobber a use with undef, deleting the used value if it becomes dead.
4057 void SROA::clobberUse(Use &U) {
4059 // Replace the use with an undef value.
4060 U = UndefValue::get(OldV->getType());
4062 // Check for this making an instruction dead. We have to garbage collect
4063 // all the dead instructions to ensure the uses of any alloca end up being
4065 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
4066 if (isInstructionTriviallyDead(OldI)) {
4067 DeadInsts.insert(OldI);
4071 /// \brief Analyze an alloca for SROA.
4073 /// This analyzes the alloca to ensure we can reason about it, builds
4074 /// the slices of the alloca, and then hands it off to be split and
4075 /// rewritten as needed.
4076 bool SROA::runOnAlloca(AllocaInst &AI) {
4077 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
4078 ++NumAllocasAnalyzed;
4080 // Special case dead allocas, as they're trivial.
4081 if (AI.use_empty()) {
4082 AI.eraseFromParent();
4085 const DataLayout &DL = AI.getModule()->getDataLayout();
4087 // Skip alloca forms that this analysis can't handle.
4088 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
4089 DL.getTypeAllocSize(AI.getAllocatedType()) == 0)
4092 bool Changed = false;
4094 // First, split any FCA loads and stores touching this alloca to promote
4095 // better splitting and promotion opportunities.
4096 AggLoadStoreRewriter AggRewriter;
4097 Changed |= AggRewriter.rewrite(AI);
4099 // Build the slices using a recursive instruction-visiting builder.
4100 AllocaSlices AS(DL, AI);
4101 DEBUG(AS.print(dbgs()));
4105 // Delete all the dead users of this alloca before splitting and rewriting it.
4106 for (Instruction *DeadUser : AS.getDeadUsers()) {
4107 // Free up everything used by this instruction.
4108 for (Use &DeadOp : DeadUser->operands())
4111 // Now replace the uses of this instruction.
4112 DeadUser->replaceAllUsesWith(UndefValue::get(DeadUser->getType()));
4114 // And mark it for deletion.
4115 DeadInsts.insert(DeadUser);
4118 for (Use *DeadOp : AS.getDeadOperands()) {
4119 clobberUse(*DeadOp);
4123 // No slices to split. Leave the dead alloca for a later pass to clean up.
4124 if (AS.begin() == AS.end())
4127 Changed |= splitAlloca(AI, AS);
4129 DEBUG(dbgs() << " Speculating PHIs\n");
4130 while (!SpeculatablePHIs.empty())
4131 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
4133 DEBUG(dbgs() << " Speculating Selects\n");
4134 while (!SpeculatableSelects.empty())
4135 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
4140 /// \brief Delete the dead instructions accumulated in this run.
4142 /// Recursively deletes the dead instructions we've accumulated. This is done
4143 /// at the very end to maximize locality of the recursive delete and to
4144 /// minimize the problems of invalidated instruction pointers as such pointers
4145 /// are used heavily in the intermediate stages of the algorithm.
4147 /// We also record the alloca instructions deleted here so that they aren't
4148 /// subsequently handed to mem2reg to promote.
4149 void SROA::deleteDeadInstructions(
4150 SmallPtrSetImpl<AllocaInst *> &DeletedAllocas) {
4151 while (!DeadInsts.empty()) {
4152 Instruction *I = DeadInsts.pop_back_val();
4153 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
4155 I->replaceAllUsesWith(UndefValue::get(I->getType()));
4157 for (Use &Operand : I->operands())
4158 if (Instruction *U = dyn_cast<Instruction>(Operand)) {
4159 // Zero out the operand and see if it becomes trivially dead.
4161 if (isInstructionTriviallyDead(U))
4162 DeadInsts.insert(U);
4165 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
4166 DeletedAllocas.insert(AI);
4167 if (DbgDeclareInst *DbgDecl = FindAllocaDbgDeclare(AI))
4168 DbgDecl->eraseFromParent();
4172 I->eraseFromParent();
4176 /// \brief Promote the allocas, using the best available technique.
4178 /// This attempts to promote whatever allocas have been identified as viable in
4179 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
4180 /// This function returns whether any promotion occurred.
4181 bool SROA::promoteAllocas(Function &F) {
4182 if (PromotableAllocas.empty())
4185 NumPromoted += PromotableAllocas.size();
4187 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
4188 PromoteMemToReg(PromotableAllocas, *DT, nullptr, AC);
4189 PromotableAllocas.clear();
4193 PreservedAnalyses SROA::runImpl(Function &F, DominatorTree &RunDT,
4194 AssumptionCache &RunAC) {
4195 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
4196 C = &F.getContext();
4200 BasicBlock &EntryBB = F.getEntryBlock();
4201 for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
4203 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
4204 Worklist.insert(AI);
4207 bool Changed = false;
4208 // A set of deleted alloca instruction pointers which should be removed from
4209 // the list of promotable allocas.
4210 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
4213 while (!Worklist.empty()) {
4214 Changed |= runOnAlloca(*Worklist.pop_back_val());
4215 deleteDeadInstructions(DeletedAllocas);
4217 // Remove the deleted allocas from various lists so that we don't try to
4218 // continue processing them.
4219 if (!DeletedAllocas.empty()) {
4220 auto IsInSet = [&](AllocaInst *AI) { return DeletedAllocas.count(AI); };
4221 Worklist.remove_if(IsInSet);
4222 PostPromotionWorklist.remove_if(IsInSet);
4223 PromotableAllocas.erase(remove_if(PromotableAllocas, IsInSet),
4224 PromotableAllocas.end());
4225 DeletedAllocas.clear();
4229 Changed |= promoteAllocas(F);
4231 Worklist = PostPromotionWorklist;
4232 PostPromotionWorklist.clear();
4233 } while (!Worklist.empty());
4236 return PreservedAnalyses::all();
4238 PreservedAnalyses PA;
4239 PA.preserveSet<CFGAnalyses>();
4240 PA.preserve<GlobalsAA>();
4244 PreservedAnalyses SROA::run(Function &F, FunctionAnalysisManager &AM) {
4245 return runImpl(F, AM.getResult<DominatorTreeAnalysis>(F),
4246 AM.getResult<AssumptionAnalysis>(F));
4249 /// A legacy pass for the legacy pass manager that wraps the \c SROA pass.
4251 /// This is in the llvm namespace purely to allow it to be a friend of the \c
4253 class llvm::sroa::SROALegacyPass : public FunctionPass {
4254 /// The SROA implementation.
4258 SROALegacyPass() : FunctionPass(ID) {
4259 initializeSROALegacyPassPass(*PassRegistry::getPassRegistry());
4261 bool runOnFunction(Function &F) override {
4262 if (skipFunction(F))
4265 auto PA = Impl.runImpl(
4266 F, getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
4267 getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F));
4268 return !PA.areAllPreserved();
4270 void getAnalysisUsage(AnalysisUsage &AU) const override {
4271 AU.addRequired<AssumptionCacheTracker>();
4272 AU.addRequired<DominatorTreeWrapperPass>();
4273 AU.addPreserved<GlobalsAAWrapperPass>();
4274 AU.setPreservesCFG();
4277 StringRef getPassName() const override { return "SROA"; }
4281 char SROALegacyPass::ID = 0;
4283 FunctionPass *llvm::createSROAPass() { return new SROALegacyPass(); }
4285 INITIALIZE_PASS_BEGIN(SROALegacyPass, "sroa",
4286 "Scalar Replacement Of Aggregates", false, false)
4287 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
4288 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
4289 INITIALIZE_PASS_END(SROALegacyPass, "sroa", "Scalar Replacement Of Aggregates",