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.h"
27 #include "llvm/ADT/STLExtras.h"
28 #include "llvm/ADT/SetVector.h"
29 #include "llvm/ADT/SmallVector.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/Analysis/AssumptionCache.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/Dominators.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/IRBuilder.h"
43 #include "llvm/IR/InstVisitor.h"
44 #include "llvm/IR/Instructions.h"
45 #include "llvm/IR/IntrinsicInst.h"
46 #include "llvm/IR/LLVMContext.h"
47 #include "llvm/IR/Operator.h"
48 #include "llvm/Pass.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Compiler.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/ErrorHandling.h"
53 #include "llvm/Support/MathExtras.h"
54 #include "llvm/Support/TimeValue.h"
55 #include "llvm/Support/raw_ostream.h"
56 #include "llvm/Transforms/Utils/Local.h"
57 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
58 #include "llvm/Transforms/Utils/SSAUpdater.h"
60 #if __cplusplus >= 201103L && !defined(NDEBUG)
61 // We only use this for a debug check in C++11
67 #define DEBUG_TYPE "sroa"
69 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
70 STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed");
71 STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions per alloca");
72 STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten");
73 STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition");
74 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
75 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
76 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
77 STATISTIC(NumDeleted, "Number of instructions deleted");
78 STATISTIC(NumVectorized, "Number of vectorized aggregates");
80 /// Hidden option to force the pass to not use DomTree and mem2reg, instead
81 /// forming SSA values through the SSAUpdater infrastructure.
82 static cl::opt<bool> ForceSSAUpdater("force-ssa-updater", cl::init(false),
85 /// Hidden option to enable randomly shuffling the slices to help uncover
86 /// instability in their order.
87 static cl::opt<bool> SROARandomShuffleSlices("sroa-random-shuffle-slices",
88 cl::init(false), cl::Hidden);
90 /// Hidden option to experiment with completely strict handling of inbounds
92 static cl::opt<bool> SROAStrictInbounds("sroa-strict-inbounds", cl::init(false),
96 /// \brief A custom IRBuilder inserter which prefixes all names if they are
98 template <bool preserveNames = true>
99 class IRBuilderPrefixedInserter
100 : public IRBuilderDefaultInserter<preserveNames> {
104 void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
107 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
108 BasicBlock::iterator InsertPt) const {
109 IRBuilderDefaultInserter<preserveNames>::InsertHelper(
110 I, Name.isTriviallyEmpty() ? Name : Prefix + Name, BB, InsertPt);
114 // Specialization for not preserving the name is trivial.
116 class IRBuilderPrefixedInserter<false>
117 : public IRBuilderDefaultInserter<false> {
119 void SetNamePrefix(const Twine &P) {}
122 /// \brief Provide a typedef for IRBuilder that drops names in release builds.
124 typedef llvm::IRBuilder<true, ConstantFolder, IRBuilderPrefixedInserter<true>>
127 typedef llvm::IRBuilder<false, ConstantFolder, IRBuilderPrefixedInserter<false>>
133 /// \brief A used slice of an alloca.
135 /// This structure represents a slice of an alloca used by some instruction. It
136 /// stores both the begin and end offsets of this use, a pointer to the use
137 /// itself, and a flag indicating whether we can classify the use as splittable
138 /// or not when forming partitions of the alloca.
140 /// \brief The beginning offset of the range.
141 uint64_t BeginOffset;
143 /// \brief The ending offset, not included in the range.
146 /// \brief Storage for both the use of this slice and whether it can be
148 PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
151 Slice() : BeginOffset(), EndOffset() {}
152 Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
153 : BeginOffset(BeginOffset), EndOffset(EndOffset),
154 UseAndIsSplittable(U, IsSplittable) {}
156 uint64_t beginOffset() const { return BeginOffset; }
157 uint64_t endOffset() const { return EndOffset; }
159 bool isSplittable() const { return UseAndIsSplittable.getInt(); }
160 void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
162 Use *getUse() const { return UseAndIsSplittable.getPointer(); }
164 bool isDead() const { return getUse() == nullptr; }
165 void kill() { UseAndIsSplittable.setPointer(nullptr); }
167 /// \brief Support for ordering ranges.
169 /// This provides an ordering over ranges such that start offsets are
170 /// always increasing, and within equal start offsets, the end offsets are
171 /// decreasing. Thus the spanning range comes first in a cluster with the
172 /// same start position.
173 bool operator<(const Slice &RHS) const {
174 if (beginOffset() < RHS.beginOffset())
176 if (beginOffset() > RHS.beginOffset())
178 if (isSplittable() != RHS.isSplittable())
179 return !isSplittable();
180 if (endOffset() > RHS.endOffset())
185 /// \brief Support comparison with a single offset to allow binary searches.
186 friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Slice &LHS,
187 uint64_t RHSOffset) {
188 return LHS.beginOffset() < RHSOffset;
190 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
192 return LHSOffset < RHS.beginOffset();
195 bool operator==(const Slice &RHS) const {
196 return isSplittable() == RHS.isSplittable() &&
197 beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
199 bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
201 } // end anonymous namespace
204 template <typename T> struct isPodLike;
205 template <> struct isPodLike<Slice> { static const bool value = true; };
209 /// \brief Representation of the alloca slices.
211 /// This class represents the slices of an alloca which are formed by its
212 /// various uses. If a pointer escapes, we can't fully build a representation
213 /// for the slices used and we reflect that in this structure. The uses are
214 /// stored, sorted by increasing beginning offset and with unsplittable slices
215 /// starting at a particular offset before splittable slices.
218 /// \brief Construct the slices of a particular alloca.
219 AllocaSlices(const DataLayout &DL, AllocaInst &AI);
221 /// \brief Test whether a pointer to the allocation escapes our analysis.
223 /// If this is true, the slices are never fully built and should be
225 bool isEscaped() const { return PointerEscapingInstr; }
227 /// \brief Support for iterating over the slices.
229 typedef SmallVectorImpl<Slice>::iterator iterator;
230 typedef iterator_range<iterator> range;
231 iterator begin() { return Slices.begin(); }
232 iterator end() { return Slices.end(); }
234 typedef SmallVectorImpl<Slice>::const_iterator const_iterator;
235 typedef iterator_range<const_iterator> const_range;
236 const_iterator begin() const { return Slices.begin(); }
237 const_iterator end() const { return Slices.end(); }
240 /// \brief Erase a range of slices.
241 void erase(iterator Start, iterator Stop) { Slices.erase(Start, Stop); }
243 /// \brief Insert new slices for this alloca.
245 /// This moves the slices into the alloca's slices collection, and re-sorts
246 /// everything so that the usual ordering properties of the alloca's slices
248 void insert(ArrayRef<Slice> NewSlices) {
249 int OldSize = Slices.size();
250 Slices.append(NewSlices.begin(), NewSlices.end());
251 auto SliceI = Slices.begin() + OldSize;
252 std::sort(SliceI, Slices.end());
253 std::inplace_merge(Slices.begin(), SliceI, Slices.end());
256 // Forward declare an iterator to befriend it.
257 class partition_iterator;
259 /// \brief A partition of the slices.
261 /// An ephemeral representation for a range of slices which can be viewed as
262 /// a partition of the alloca. This range represents a span of the alloca's
263 /// memory which cannot be split, and provides access to all of the slices
264 /// overlapping some part of the partition.
266 /// Objects of this type are produced by traversing the alloca's slices, but
267 /// are only ephemeral and not persistent.
270 friend class AllocaSlices;
271 friend class AllocaSlices::partition_iterator;
273 /// \brief The begining and ending offsets of the alloca for this partition.
274 uint64_t BeginOffset, EndOffset;
276 /// \brief The start end end iterators of this partition.
279 /// \brief A collection of split slice tails overlapping the partition.
280 SmallVector<Slice *, 4> SplitTails;
282 /// \brief Raw constructor builds an empty partition starting and ending at
283 /// the given iterator.
284 Partition(iterator SI) : SI(SI), SJ(SI) {}
287 /// \brief The start offset of this partition.
289 /// All of the contained slices start at or after this offset.
290 uint64_t beginOffset() const { return BeginOffset; }
292 /// \brief The end offset of this partition.
294 /// All of the contained slices end at or before this offset.
295 uint64_t endOffset() const { return EndOffset; }
297 /// \brief The size of the partition.
299 /// Note that this can never be zero.
300 uint64_t size() const {
301 assert(BeginOffset < EndOffset && "Partitions must span some bytes!");
302 return EndOffset - BeginOffset;
305 /// \brief Test whether this partition contains no slices, and merely spans
306 /// a region occupied by split slices.
307 bool empty() const { return SI == SJ; }
309 /// \name Iterate slices that start within the partition.
310 /// These may be splittable or unsplittable. They have a begin offset >= the
311 /// partition begin offset.
313 // FIXME: We should probably define a "concat_iterator" helper and use that
314 // to stitch together pointee_iterators over the split tails and the
315 // contiguous iterators of the partition. That would give a much nicer
316 // interface here. We could then additionally expose filtered iterators for
317 // split, unsplit, and unsplittable splices based on the usage patterns.
318 iterator begin() const { return SI; }
319 iterator end() const { return SJ; }
322 /// \brief Get the sequence of split slice tails.
324 /// These tails are of slices which start before this partition but are
325 /// split and overlap into the partition. We accumulate these while forming
327 ArrayRef<Slice *> splitSliceTails() const { return SplitTails; }
330 /// \brief An iterator over partitions of the alloca's slices.
332 /// This iterator implements the core algorithm for partitioning the alloca's
333 /// slices. It is a forward iterator as we don't support backtracking for
334 /// efficiency reasons, and re-use a single storage area to maintain the
335 /// current set of split slices.
337 /// It is templated on the slice iterator type to use so that it can operate
338 /// with either const or non-const slice iterators.
339 class partition_iterator
340 : public iterator_facade_base<partition_iterator,
341 std::forward_iterator_tag, Partition> {
342 friend class AllocaSlices;
344 /// \brief Most of the state for walking the partitions is held in a class
345 /// with a nice interface for examining them.
348 /// \brief We need to keep the end of the slices to know when to stop.
349 AllocaSlices::iterator SE;
351 /// \brief We also need to keep track of the maximum split end offset seen.
352 /// FIXME: Do we really?
353 uint64_t MaxSplitSliceEndOffset;
355 /// \brief Sets the partition to be empty at given iterator, and sets the
357 partition_iterator(AllocaSlices::iterator SI, AllocaSlices::iterator SE)
358 : P(SI), SE(SE), MaxSplitSliceEndOffset(0) {
359 // If not already at the end, advance our state to form the initial
365 /// \brief Advance the iterator to the next partition.
367 /// Requires that the iterator not be at the end of the slices.
369 assert((P.SI != SE || !P.SplitTails.empty()) &&
370 "Cannot advance past the end of the slices!");
372 // Clear out any split uses which have ended.
373 if (!P.SplitTails.empty()) {
374 if (P.EndOffset >= MaxSplitSliceEndOffset) {
375 // If we've finished all splits, this is easy.
376 P.SplitTails.clear();
377 MaxSplitSliceEndOffset = 0;
379 // Remove the uses which have ended in the prior partition. This
380 // cannot change the max split slice end because we just checked that
381 // the prior partition ended prior to that max.
384 P.SplitTails.begin(), P.SplitTails.end(),
385 [&](Slice *S) { return S->endOffset() <= P.EndOffset; }),
387 assert(std::any_of(P.SplitTails.begin(), P.SplitTails.end(),
389 return S->endOffset() == MaxSplitSliceEndOffset;
391 "Could not find the current max split slice offset!");
392 assert(std::all_of(P.SplitTails.begin(), P.SplitTails.end(),
394 return S->endOffset() <= MaxSplitSliceEndOffset;
396 "Max split slice end offset is not actually the max!");
400 // If P.SI is already at the end, then we've cleared the split tail and
401 // now have an end iterator.
403 assert(P.SplitTails.empty() && "Failed to clear the split slices!");
407 // If we had a non-empty partition previously, set up the state for
408 // subsequent partitions.
410 // Accumulate all the splittable slices which started in the old
411 // partition into the split list.
413 if (S.isSplittable() && S.endOffset() > P.EndOffset) {
414 P.SplitTails.push_back(&S);
415 MaxSplitSliceEndOffset =
416 std::max(S.endOffset(), MaxSplitSliceEndOffset);
419 // Start from the end of the previous partition.
422 // If P.SI is now at the end, we at most have a tail of split slices.
424 P.BeginOffset = P.EndOffset;
425 P.EndOffset = MaxSplitSliceEndOffset;
429 // If the we have split slices and the next slice is after a gap and is
430 // not splittable immediately form an empty partition for the split
431 // slices up until the next slice begins.
432 if (!P.SplitTails.empty() && P.SI->beginOffset() != P.EndOffset &&
433 !P.SI->isSplittable()) {
434 P.BeginOffset = P.EndOffset;
435 P.EndOffset = P.SI->beginOffset();
440 // OK, we need to consume new slices. Set the end offset based on the
441 // current slice, and step SJ past it. The beginning offset of the
442 // parttion is the beginning offset of the next slice unless we have
443 // pre-existing split slices that are continuing, in which case we begin
444 // at the prior end offset.
445 P.BeginOffset = P.SplitTails.empty() ? P.SI->beginOffset() : P.EndOffset;
446 P.EndOffset = P.SI->endOffset();
449 // There are two strategies to form a partition based on whether the
450 // partition starts with an unsplittable slice or a splittable slice.
451 if (!P.SI->isSplittable()) {
452 // When we're forming an unsplittable region, it must always start at
453 // the first slice and will extend through its end.
454 assert(P.BeginOffset == P.SI->beginOffset());
456 // Form a partition including all of the overlapping slices with this
457 // unsplittable slice.
458 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
459 if (!P.SJ->isSplittable())
460 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
464 // We have a partition across a set of overlapping unsplittable
469 // If we're starting with a splittable slice, then we need to form
470 // a synthetic partition spanning it and any other overlapping splittable
472 assert(P.SI->isSplittable() && "Forming a splittable partition!");
474 // Collect all of the overlapping splittable slices.
475 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset &&
476 P.SJ->isSplittable()) {
477 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
481 // Back upiP.EndOffset if we ended the span early when encountering an
482 // unsplittable slice. This synthesizes the early end offset of
483 // a partition spanning only splittable slices.
484 if (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
485 assert(!P.SJ->isSplittable());
486 P.EndOffset = P.SJ->beginOffset();
491 bool operator==(const partition_iterator &RHS) const {
492 assert(SE == RHS.SE &&
493 "End iterators don't match between compared partition iterators!");
495 // The observed positions of partitions is marked by the P.SI iterator and
496 // the emptyness of the split slices. The latter is only relevant when
497 // P.SI == SE, as the end iterator will additionally have an empty split
498 // slices list, but the prior may have the same P.SI and a tail of split
500 if (P.SI == RHS.P.SI &&
501 P.SplitTails.empty() == RHS.P.SplitTails.empty()) {
502 assert(P.SJ == RHS.P.SJ &&
503 "Same set of slices formed two different sized partitions!");
504 assert(P.SplitTails.size() == RHS.P.SplitTails.size() &&
505 "Same slice position with differently sized non-empty split "
512 partition_iterator &operator++() {
517 Partition &operator*() { return P; }
520 /// \brief A forward range over the partitions of the alloca's slices.
522 /// This accesses an iterator range over the partitions of the alloca's
523 /// slices. It computes these partitions on the fly based on the overlapping
524 /// offsets of the slices and the ability to split them. It will visit "empty"
525 /// partitions to cover regions of the alloca only accessed via split
527 iterator_range<partition_iterator> partitions() {
528 return make_range(partition_iterator(begin(), end()),
529 partition_iterator(end(), end()));
532 /// \brief Access the dead users for this alloca.
533 ArrayRef<Instruction *> getDeadUsers() const { return DeadUsers; }
535 /// \brief Access the dead operands referring to this alloca.
537 /// These are operands which have cannot actually be used to refer to the
538 /// alloca as they are outside its range and the user doesn't correct for
539 /// that. These mostly consist of PHI node inputs and the like which we just
540 /// need to replace with undef.
541 ArrayRef<Use *> getDeadOperands() const { return DeadOperands; }
543 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
544 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
545 void printSlice(raw_ostream &OS, const_iterator I,
546 StringRef Indent = " ") const;
547 void printUse(raw_ostream &OS, const_iterator I,
548 StringRef Indent = " ") const;
549 void print(raw_ostream &OS) const;
550 void dump(const_iterator I) const;
555 template <typename DerivedT, typename RetT = void> class BuilderBase;
557 friend class AllocaSlices::SliceBuilder;
559 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
560 /// \brief Handle to alloca instruction to simplify method interfaces.
564 /// \brief The instruction responsible for this alloca not having a known set
567 /// When an instruction (potentially) escapes the pointer to the alloca, we
568 /// store a pointer to that here and abort trying to form slices of the
569 /// alloca. This will be null if the alloca slices are analyzed successfully.
570 Instruction *PointerEscapingInstr;
572 /// \brief The slices of the alloca.
574 /// We store a vector of the slices formed by uses of the alloca here. This
575 /// vector is sorted by increasing begin offset, and then the unsplittable
576 /// slices before the splittable ones. See the Slice inner class for more
578 SmallVector<Slice, 8> Slices;
580 /// \brief Instructions which will become dead if we rewrite the alloca.
582 /// Note that these are not separated by slice. This is because we expect an
583 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
584 /// all these instructions can simply be removed and replaced with undef as
585 /// they come from outside of the allocated space.
586 SmallVector<Instruction *, 8> DeadUsers;
588 /// \brief Operands which will become dead if we rewrite the alloca.
590 /// These are operands that in their particular use can be replaced with
591 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
592 /// to PHI nodes and the like. They aren't entirely dead (there might be
593 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
594 /// want to swap this particular input for undef to simplify the use lists of
596 SmallVector<Use *, 8> DeadOperands;
600 static Value *foldSelectInst(SelectInst &SI) {
601 // If the condition being selected on is a constant or the same value is
602 // being selected between, fold the select. Yes this does (rarely) happen
604 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
605 return SI.getOperand(1 + CI->isZero());
606 if (SI.getOperand(1) == SI.getOperand(2))
607 return SI.getOperand(1);
612 /// \brief A helper that folds a PHI node or a select.
613 static Value *foldPHINodeOrSelectInst(Instruction &I) {
614 if (PHINode *PN = dyn_cast<PHINode>(&I)) {
615 // If PN merges together the same value, return that value.
616 return PN->hasConstantValue();
618 return foldSelectInst(cast<SelectInst>(I));
621 /// \brief Builder for the alloca slices.
623 /// This class builds a set of alloca slices by recursively visiting the uses
624 /// of an alloca and making a slice for each load and store at each offset.
625 class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
626 friend class PtrUseVisitor<SliceBuilder>;
627 friend class InstVisitor<SliceBuilder>;
628 typedef PtrUseVisitor<SliceBuilder> Base;
630 const uint64_t AllocSize;
633 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
634 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
636 /// \brief Set to de-duplicate dead instructions found in the use walk.
637 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
640 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &AS)
641 : PtrUseVisitor<SliceBuilder>(DL),
642 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), AS(AS) {}
645 void markAsDead(Instruction &I) {
646 if (VisitedDeadInsts.insert(&I).second)
647 AS.DeadUsers.push_back(&I);
650 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
651 bool IsSplittable = false) {
652 // Completely skip uses which have a zero size or start either before or
653 // past the end of the allocation.
654 if (Size == 0 || Offset.uge(AllocSize)) {
655 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
656 << " which has zero size or starts outside of the "
657 << AllocSize << " byte alloca:\n"
658 << " alloca: " << AS.AI << "\n"
659 << " use: " << I << "\n");
660 return markAsDead(I);
663 uint64_t BeginOffset = Offset.getZExtValue();
664 uint64_t EndOffset = BeginOffset + Size;
666 // Clamp the end offset to the end of the allocation. Note that this is
667 // formulated to handle even the case where "BeginOffset + Size" overflows.
668 // This may appear superficially to be something we could ignore entirely,
669 // but that is not so! There may be widened loads or PHI-node uses where
670 // some instructions are dead but not others. We can't completely ignore
671 // them, and so have to record at least the information here.
672 assert(AllocSize >= BeginOffset); // Established above.
673 if (Size > AllocSize - BeginOffset) {
674 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
675 << " to remain within the " << AllocSize << " byte alloca:\n"
676 << " alloca: " << AS.AI << "\n"
677 << " use: " << I << "\n");
678 EndOffset = AllocSize;
681 AS.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
684 void visitBitCastInst(BitCastInst &BC) {
686 return markAsDead(BC);
688 return Base::visitBitCastInst(BC);
691 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
692 if (GEPI.use_empty())
693 return markAsDead(GEPI);
695 if (SROAStrictInbounds && GEPI.isInBounds()) {
696 // FIXME: This is a manually un-factored variant of the basic code inside
697 // of GEPs with checking of the inbounds invariant specified in the
698 // langref in a very strict sense. If we ever want to enable
699 // SROAStrictInbounds, this code should be factored cleanly into
700 // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
701 // by writing out the code here where we have tho underlying allocation
702 // size readily available.
703 APInt GEPOffset = Offset;
704 const DataLayout &DL = GEPI.getModule()->getDataLayout();
705 for (gep_type_iterator GTI = gep_type_begin(GEPI),
706 GTE = gep_type_end(GEPI);
708 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
712 // Handle a struct index, which adds its field offset to the pointer.
713 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
714 unsigned ElementIdx = OpC->getZExtValue();
715 const StructLayout *SL = DL.getStructLayout(STy);
717 APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
719 // For array or vector indices, scale the index by the size of the
721 APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
722 GEPOffset += Index * APInt(Offset.getBitWidth(),
723 DL.getTypeAllocSize(GTI.getIndexedType()));
726 // If this index has computed an intermediate pointer which is not
727 // inbounds, then the result of the GEP is a poison value and we can
728 // delete it and all uses.
729 if (GEPOffset.ugt(AllocSize))
730 return markAsDead(GEPI);
734 return Base::visitGetElementPtrInst(GEPI);
737 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
738 uint64_t Size, bool IsVolatile) {
739 // We allow splitting of non-volatile loads and stores where the type is an
740 // integer type. These may be used to implement 'memcpy' or other "transfer
741 // of bits" patterns.
742 bool IsSplittable = Ty->isIntegerTy() && !IsVolatile;
744 insertUse(I, Offset, Size, IsSplittable);
747 void visitLoadInst(LoadInst &LI) {
748 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
749 "All simple FCA loads should have been pre-split");
752 return PI.setAborted(&LI);
754 const DataLayout &DL = LI.getModule()->getDataLayout();
755 uint64_t Size = DL.getTypeStoreSize(LI.getType());
756 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
759 void visitStoreInst(StoreInst &SI) {
760 Value *ValOp = SI.getValueOperand();
762 return PI.setEscapedAndAborted(&SI);
764 return PI.setAborted(&SI);
766 const DataLayout &DL = SI.getModule()->getDataLayout();
767 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
769 // If this memory access can be shown to *statically* extend outside the
770 // bounds of of the allocation, it's behavior is undefined, so simply
771 // ignore it. Note that this is more strict than the generic clamping
772 // behavior of insertUse. We also try to handle cases which might run the
774 // FIXME: We should instead consider the pointer to have escaped if this
775 // function is being instrumented for addressing bugs or race conditions.
776 if (Size > AllocSize || Offset.ugt(AllocSize - Size)) {
777 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
778 << " which extends past the end of the " << AllocSize
780 << " alloca: " << AS.AI << "\n"
781 << " use: " << SI << "\n");
782 return markAsDead(SI);
785 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
786 "All simple FCA stores should have been pre-split");
787 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
790 void visitMemSetInst(MemSetInst &II) {
791 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
792 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
793 if ((Length && Length->getValue() == 0) ||
794 (IsOffsetKnown && Offset.uge(AllocSize)))
795 // Zero-length mem transfer intrinsics can be ignored entirely.
796 return markAsDead(II);
799 return PI.setAborted(&II);
801 insertUse(II, Offset, Length ? Length->getLimitedValue()
802 : AllocSize - Offset.getLimitedValue(),
806 void visitMemTransferInst(MemTransferInst &II) {
807 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
808 if (Length && Length->getValue() == 0)
809 // Zero-length mem transfer intrinsics can be ignored entirely.
810 return markAsDead(II);
812 // Because we can visit these intrinsics twice, also check to see if the
813 // first time marked this instruction as dead. If so, skip it.
814 if (VisitedDeadInsts.count(&II))
818 return PI.setAborted(&II);
820 // This side of the transfer is completely out-of-bounds, and so we can
821 // nuke the entire transfer. However, we also need to nuke the other side
822 // if already added to our partitions.
823 // FIXME: Yet another place we really should bypass this when
824 // instrumenting for ASan.
825 if (Offset.uge(AllocSize)) {
826 SmallDenseMap<Instruction *, unsigned>::iterator MTPI =
827 MemTransferSliceMap.find(&II);
828 if (MTPI != MemTransferSliceMap.end())
829 AS.Slices[MTPI->second].kill();
830 return markAsDead(II);
833 uint64_t RawOffset = Offset.getLimitedValue();
834 uint64_t Size = Length ? Length->getLimitedValue() : AllocSize - RawOffset;
836 // Check for the special case where the same exact value is used for both
838 if (*U == II.getRawDest() && *U == II.getRawSource()) {
839 // For non-volatile transfers this is a no-op.
840 if (!II.isVolatile())
841 return markAsDead(II);
843 return insertUse(II, Offset, Size, /*IsSplittable=*/false);
846 // If we have seen both source and destination for a mem transfer, then
847 // they both point to the same alloca.
849 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
850 std::tie(MTPI, Inserted) =
851 MemTransferSliceMap.insert(std::make_pair(&II, AS.Slices.size()));
852 unsigned PrevIdx = MTPI->second;
854 Slice &PrevP = AS.Slices[PrevIdx];
856 // Check if the begin offsets match and this is a non-volatile transfer.
857 // In that case, we can completely elide the transfer.
858 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
860 return markAsDead(II);
863 // Otherwise we have an offset transfer within the same alloca. We can't
865 PrevP.makeUnsplittable();
868 // Insert the use now that we've fixed up the splittable nature.
869 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
871 // Check that we ended up with a valid index in the map.
872 assert(AS.Slices[PrevIdx].getUse()->getUser() == &II &&
873 "Map index doesn't point back to a slice with this user.");
876 // Disable SRoA for any intrinsics except for lifetime invariants.
877 // FIXME: What about debug intrinsics? This matches old behavior, but
878 // doesn't make sense.
879 void visitIntrinsicInst(IntrinsicInst &II) {
881 return PI.setAborted(&II);
883 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
884 II.getIntrinsicID() == Intrinsic::lifetime_end) {
885 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
886 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
887 Length->getLimitedValue());
888 insertUse(II, Offset, Size, true);
892 Base::visitIntrinsicInst(II);
895 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
896 // We consider any PHI or select that results in a direct load or store of
897 // the same offset to be a viable use for slicing purposes. These uses
898 // are considered unsplittable and the size is the maximum loaded or stored
900 SmallPtrSet<Instruction *, 4> Visited;
901 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
902 Visited.insert(Root);
903 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
904 const DataLayout &DL = Root->getModule()->getDataLayout();
905 // If there are no loads or stores, the access is dead. We mark that as
906 // a size zero access.
909 Instruction *I, *UsedI;
910 std::tie(UsedI, I) = Uses.pop_back_val();
912 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
913 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
916 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
917 Value *Op = SI->getOperand(0);
920 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
924 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
925 if (!GEP->hasAllZeroIndices())
927 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
928 !isa<SelectInst>(I)) {
932 for (User *U : I->users())
933 if (Visited.insert(cast<Instruction>(U)).second)
934 Uses.push_back(std::make_pair(I, cast<Instruction>(U)));
935 } while (!Uses.empty());
940 void visitPHINodeOrSelectInst(Instruction &I) {
941 assert(isa<PHINode>(I) || isa<SelectInst>(I));
943 return markAsDead(I);
945 // TODO: We could use SimplifyInstruction here to fold PHINodes and
946 // SelectInsts. However, doing so requires to change the current
947 // dead-operand-tracking mechanism. For instance, suppose neither loading
948 // from %U nor %other traps. Then "load (select undef, %U, %other)" does not
949 // trap either. However, if we simply replace %U with undef using the
950 // current dead-operand-tracking mechanism, "load (select undef, undef,
951 // %other)" may trap because the select may return the first operand
953 if (Value *Result = foldPHINodeOrSelectInst(I)) {
955 // If the result of the constant fold will be the pointer, recurse
956 // through the PHI/select as if we had RAUW'ed it.
959 // Otherwise the operand to the PHI/select is dead, and we can replace
961 AS.DeadOperands.push_back(U);
967 return PI.setAborted(&I);
969 // See if we already have computed info on this node.
970 uint64_t &Size = PHIOrSelectSizes[&I];
972 // This is a new PHI/Select, check for an unsafe use of it.
973 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&I, Size))
974 return PI.setAborted(UnsafeI);
977 // For PHI and select operands outside the alloca, we can't nuke the entire
978 // phi or select -- the other side might still be relevant, so we special
979 // case them here and use a separate structure to track the operands
980 // themselves which should be replaced with undef.
981 // FIXME: This should instead be escaped in the event we're instrumenting
982 // for address sanitization.
983 if (Offset.uge(AllocSize)) {
984 AS.DeadOperands.push_back(U);
988 insertUse(I, Offset, Size);
991 void visitPHINode(PHINode &PN) { visitPHINodeOrSelectInst(PN); }
993 void visitSelectInst(SelectInst &SI) { visitPHINodeOrSelectInst(SI); }
995 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
996 void visitInstruction(Instruction &I) { PI.setAborted(&I); }
999 AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
1001 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1004 PointerEscapingInstr(nullptr) {
1005 SliceBuilder PB(DL, AI, *this);
1006 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
1007 if (PtrI.isEscaped() || PtrI.isAborted()) {
1008 // FIXME: We should sink the escape vs. abort info into the caller nicely,
1009 // possibly by just storing the PtrInfo in the AllocaSlices.
1010 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
1011 : PtrI.getAbortingInst();
1012 assert(PointerEscapingInstr && "Did not track a bad instruction");
1016 Slices.erase(std::remove_if(Slices.begin(), Slices.end(),
1017 [](const Slice &S) {
1022 #if __cplusplus >= 201103L && !defined(NDEBUG)
1023 if (SROARandomShuffleSlices) {
1024 std::mt19937 MT(static_cast<unsigned>(sys::TimeValue::now().msec()));
1025 std::shuffle(Slices.begin(), Slices.end(), MT);
1029 // Sort the uses. This arranges for the offsets to be in ascending order,
1030 // and the sizes to be in descending order.
1031 std::sort(Slices.begin(), Slices.end());
1034 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1036 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
1037 StringRef Indent) const {
1038 printSlice(OS, I, Indent);
1040 printUse(OS, I, Indent);
1043 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
1044 StringRef Indent) const {
1045 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
1046 << " slice #" << (I - begin())
1047 << (I->isSplittable() ? " (splittable)" : "");
1050 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
1051 StringRef Indent) const {
1052 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
1055 void AllocaSlices::print(raw_ostream &OS) const {
1056 if (PointerEscapingInstr) {
1057 OS << "Can't analyze slices for alloca: " << AI << "\n"
1058 << " A pointer to this alloca escaped by:\n"
1059 << " " << *PointerEscapingInstr << "\n";
1063 OS << "Slices of alloca: " << AI << "\n";
1064 for (const_iterator I = begin(), E = end(); I != E; ++I)
1068 LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
1071 LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
1073 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1076 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
1078 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
1079 /// the loads and stores of an alloca instruction, as well as updating its
1080 /// debug information. This is used when a domtree is unavailable and thus
1081 /// mem2reg in its full form can't be used to handle promotion of allocas to
1083 class AllocaPromoter : public LoadAndStorePromoter {
1087 SmallVector<DbgDeclareInst *, 4> DDIs;
1088 SmallVector<DbgValueInst *, 4> DVIs;
1091 AllocaPromoter(const SmallVectorImpl<Instruction *> &Insts, SSAUpdater &S,
1092 AllocaInst &AI, DIBuilder &DIB)
1093 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
1095 void run(const SmallVectorImpl<Instruction *> &Insts) {
1096 // Retain the debug information attached to the alloca for use when
1097 // rewriting loads and stores.
1098 if (auto *L = LocalAsMetadata::getIfExists(&AI)) {
1099 if (auto *DebugNode = MetadataAsValue::getIfExists(AI.getContext(), L)) {
1100 for (User *U : DebugNode->users())
1101 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
1102 DDIs.push_back(DDI);
1103 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(U))
1104 DVIs.push_back(DVI);
1108 LoadAndStorePromoter::run(Insts);
1110 // While we have the debug information, clear it off of the alloca. The
1111 // caller takes care of deleting the alloca.
1112 while (!DDIs.empty())
1113 DDIs.pop_back_val()->eraseFromParent();
1114 while (!DVIs.empty())
1115 DVIs.pop_back_val()->eraseFromParent();
1119 isInstInList(Instruction *I,
1120 const SmallVectorImpl<Instruction *> &Insts) const override {
1122 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1123 Ptr = LI->getOperand(0);
1125 Ptr = cast<StoreInst>(I)->getPointerOperand();
1127 // Only used to detect cycles, which will be rare and quickly found as
1128 // we're walking up a chain of defs rather than down through uses.
1129 SmallPtrSet<Value *, 4> Visited;
1135 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr))
1136 Ptr = BCI->getOperand(0);
1137 else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr))
1138 Ptr = GEPI->getPointerOperand();
1142 } while (Visited.insert(Ptr).second);
1147 void updateDebugInfo(Instruction *Inst) const override {
1148 for (DbgDeclareInst *DDI : DDIs)
1149 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
1150 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1151 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
1152 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1153 for (DbgValueInst *DVI : DVIs) {
1154 Value *Arg = nullptr;
1155 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1156 // If an argument is zero extended then use argument directly. The ZExt
1157 // may be zapped by an optimization pass in future.
1158 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1159 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
1160 else if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1161 Arg = dyn_cast<Argument>(SExt->getOperand(0));
1163 Arg = SI->getValueOperand();
1164 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
1165 Arg = LI->getPointerOperand();
1169 Instruction *DbgVal =
1170 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
1171 DIExpression(DVI->getExpression()), Inst);
1172 DbgVal->setDebugLoc(DVI->getDebugLoc());
1176 } // end anon namespace
1179 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
1181 /// This pass takes allocations which can be completely analyzed (that is, they
1182 /// don't escape) and tries to turn them into scalar SSA values. There are
1183 /// a few steps to this process.
1185 /// 1) It takes allocations of aggregates and analyzes the ways in which they
1186 /// are used to try to split them into smaller allocations, ideally of
1187 /// a single scalar data type. It will split up memcpy and memset accesses
1188 /// as necessary and try to isolate individual scalar accesses.
1189 /// 2) It will transform accesses into forms which are suitable for SSA value
1190 /// promotion. This can be replacing a memset with a scalar store of an
1191 /// integer value, or it can involve speculating operations on a PHI or
1192 /// select to be a PHI or select of the results.
1193 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
1194 /// onto insert and extract operations on a vector value, and convert them to
1195 /// this form. By doing so, it will enable promotion of vector aggregates to
1196 /// SSA vector values.
1197 class SROA : public FunctionPass {
1198 const bool RequiresDomTree;
1202 AssumptionCache *AC;
1204 /// \brief Worklist of alloca instructions to simplify.
1206 /// Each alloca in the function is added to this. Each new alloca formed gets
1207 /// added to it as well to recursively simplify unless that alloca can be
1208 /// directly promoted. Finally, each time we rewrite a use of an alloca other
1209 /// the one being actively rewritten, we add it back onto the list if not
1210 /// already present to ensure it is re-visited.
1211 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16>> Worklist;
1213 /// \brief A collection of instructions to delete.
1214 /// We try to batch deletions to simplify code and make things a bit more
1216 SetVector<Instruction *, SmallVector<Instruction *, 8>> DeadInsts;
1218 /// \brief Post-promotion worklist.
1220 /// Sometimes we discover an alloca which has a high probability of becoming
1221 /// viable for SROA after a round of promotion takes place. In those cases,
1222 /// the alloca is enqueued here for re-processing.
1224 /// Note that we have to be very careful to clear allocas out of this list in
1225 /// the event they are deleted.
1226 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16>> PostPromotionWorklist;
1228 /// \brief A collection of alloca instructions we can directly promote.
1229 std::vector<AllocaInst *> PromotableAllocas;
1231 /// \brief A worklist of PHIs to speculate prior to promoting allocas.
1233 /// All of these PHIs have been checked for the safety of speculation and by
1234 /// being speculated will allow promoting allocas currently in the promotable
1236 SetVector<PHINode *, SmallVector<PHINode *, 2>> SpeculatablePHIs;
1238 /// \brief A worklist of select instructions to speculate prior to promoting
1241 /// All of these select instructions have been checked for the safety of
1242 /// speculation and by being speculated will allow promoting allocas
1243 /// currently in the promotable queue.
1244 SetVector<SelectInst *, SmallVector<SelectInst *, 2>> SpeculatableSelects;
1247 SROA(bool RequiresDomTree = true)
1248 : FunctionPass(ID), RequiresDomTree(RequiresDomTree), C(nullptr),
1250 initializeSROAPass(*PassRegistry::getPassRegistry());
1252 bool runOnFunction(Function &F) override;
1253 void getAnalysisUsage(AnalysisUsage &AU) const override;
1255 const char *getPassName() const override { return "SROA"; }
1259 friend class PHIOrSelectSpeculator;
1260 friend class AllocaSliceRewriter;
1262 bool presplitLoadsAndStores(AllocaInst &AI, AllocaSlices &AS);
1263 AllocaInst *rewritePartition(AllocaInst &AI, AllocaSlices &AS,
1264 AllocaSlices::Partition &P);
1265 bool splitAlloca(AllocaInst &AI, AllocaSlices &AS);
1266 bool runOnAlloca(AllocaInst &AI);
1267 void clobberUse(Use &U);
1268 void deleteDeadInstructions(SmallPtrSetImpl<AllocaInst *> &DeletedAllocas);
1269 bool promoteAllocas(Function &F);
1275 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
1276 return new SROA(RequiresDomTree);
1279 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates", false,
1281 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1282 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1283 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates", false,
1286 /// Walk the range of a partitioning looking for a common type to cover this
1287 /// sequence of slices.
1288 static Type *findCommonType(AllocaSlices::const_iterator B,
1289 AllocaSlices::const_iterator E,
1290 uint64_t EndOffset) {
1292 bool TyIsCommon = true;
1293 IntegerType *ITy = nullptr;
1295 // Note that we need to look at *every* alloca slice's Use to ensure we
1296 // always get consistent results regardless of the order of slices.
1297 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1298 Use *U = I->getUse();
1299 if (isa<IntrinsicInst>(*U->getUser()))
1301 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1304 Type *UserTy = nullptr;
1305 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1306 UserTy = LI->getType();
1307 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1308 UserTy = SI->getValueOperand()->getType();
1311 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1312 // If the type is larger than the partition, skip it. We only encounter
1313 // this for split integer operations where we want to use the type of the
1314 // entity causing the split. Also skip if the type is not a byte width
1316 if (UserITy->getBitWidth() % 8 != 0 ||
1317 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1320 // Track the largest bitwidth integer type used in this way in case there
1321 // is no common type.
1322 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1326 // To avoid depending on the order of slices, Ty and TyIsCommon must not
1327 // depend on types skipped above.
1328 if (!UserTy || (Ty && Ty != UserTy))
1329 TyIsCommon = false; // Give up on anything but an iN type.
1334 return TyIsCommon ? Ty : ITy;
1337 /// PHI instructions that use an alloca and are subsequently loaded can be
1338 /// rewritten to load both input pointers in the pred blocks and then PHI the
1339 /// results, allowing the load of the alloca to be promoted.
1341 /// %P2 = phi [i32* %Alloca, i32* %Other]
1342 /// %V = load i32* %P2
1344 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1346 /// %V2 = load i32* %Other
1348 /// %V = phi [i32 %V1, i32 %V2]
1350 /// We can do this to a select if its only uses are loads and if the operands
1351 /// to the select can be loaded unconditionally.
1353 /// FIXME: This should be hoisted into a generic utility, likely in
1354 /// Transforms/Util/Local.h
1355 static bool isSafePHIToSpeculate(PHINode &PN) {
1356 // For now, we can only do this promotion if the load is in the same block
1357 // as the PHI, and if there are no stores between the phi and load.
1358 // TODO: Allow recursive phi users.
1359 // TODO: Allow stores.
1360 BasicBlock *BB = PN.getParent();
1361 unsigned MaxAlign = 0;
1362 bool HaveLoad = false;
1363 for (User *U : PN.users()) {
1364 LoadInst *LI = dyn_cast<LoadInst>(U);
1365 if (!LI || !LI->isSimple())
1368 // For now we only allow loads in the same block as the PHI. This is
1369 // a common case that happens when instcombine merges two loads through
1371 if (LI->getParent() != BB)
1374 // Ensure that there are no instructions between the PHI and the load that
1376 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1377 if (BBI->mayWriteToMemory())
1380 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1387 const DataLayout &DL = PN.getModule()->getDataLayout();
1389 // We can only transform this if it is safe to push the loads into the
1390 // predecessor blocks. The only thing to watch out for is that we can't put
1391 // a possibly trapping load in the predecessor if it is a critical edge.
1392 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1393 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1394 Value *InVal = PN.getIncomingValue(Idx);
1396 // If the value is produced by the terminator of the predecessor (an
1397 // invoke) or it has side-effects, there is no valid place to put a load
1398 // in the predecessor.
1399 if (TI == InVal || TI->mayHaveSideEffects())
1402 // If the predecessor has a single successor, then the edge isn't
1404 if (TI->getNumSuccessors() == 1)
1407 // If this pointer is always safe to load, or if we can prove that there
1408 // is already a load in the block, then we can move the load to the pred
1410 if (InVal->isDereferenceablePointer(DL) ||
1411 isSafeToLoadUnconditionally(InVal, TI, MaxAlign))
1420 static void speculatePHINodeLoads(PHINode &PN) {
1421 DEBUG(dbgs() << " original: " << PN << "\n");
1423 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1424 IRBuilderTy PHIBuilder(&PN);
1425 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1426 PN.getName() + ".sroa.speculated");
1428 // Get the AA tags and alignment to use from one of the loads. It doesn't
1429 // matter which one we get and if any differ.
1430 LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1433 SomeLoad->getAAMetadata(AATags);
1434 unsigned Align = SomeLoad->getAlignment();
1436 // Rewrite all loads of the PN to use the new PHI.
1437 while (!PN.use_empty()) {
1438 LoadInst *LI = cast<LoadInst>(PN.user_back());
1439 LI->replaceAllUsesWith(NewPN);
1440 LI->eraseFromParent();
1443 // Inject loads into all of the pred blocks.
1444 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1445 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1446 TerminatorInst *TI = Pred->getTerminator();
1447 Value *InVal = PN.getIncomingValue(Idx);
1448 IRBuilderTy PredBuilder(TI);
1450 LoadInst *Load = PredBuilder.CreateLoad(
1451 InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1452 ++NumLoadsSpeculated;
1453 Load->setAlignment(Align);
1455 Load->setAAMetadata(AATags);
1456 NewPN->addIncoming(Load, Pred);
1459 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1460 PN.eraseFromParent();
1463 /// Select instructions that use an alloca and are subsequently loaded can be
1464 /// rewritten to load both input pointers and then select between the result,
1465 /// allowing the load of the alloca to be promoted.
1467 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1468 /// %V = load i32* %P2
1470 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1471 /// %V2 = load i32* %Other
1472 /// %V = select i1 %cond, i32 %V1, i32 %V2
1474 /// We can do this to a select if its only uses are loads and if the operand
1475 /// to the select can be loaded unconditionally.
1476 static bool isSafeSelectToSpeculate(SelectInst &SI) {
1477 Value *TValue = SI.getTrueValue();
1478 Value *FValue = SI.getFalseValue();
1479 const DataLayout &DL = SI.getModule()->getDataLayout();
1480 bool TDerefable = TValue->isDereferenceablePointer(DL);
1481 bool FDerefable = FValue->isDereferenceablePointer(DL);
1483 for (User *U : SI.users()) {
1484 LoadInst *LI = dyn_cast<LoadInst>(U);
1485 if (!LI || !LI->isSimple())
1488 // Both operands to the select need to be dereferencable, either
1489 // absolutely (e.g. allocas) or at this point because we can see other
1492 !isSafeToLoadUnconditionally(TValue, LI, LI->getAlignment()))
1495 !isSafeToLoadUnconditionally(FValue, LI, LI->getAlignment()))
1502 static void speculateSelectInstLoads(SelectInst &SI) {
1503 DEBUG(dbgs() << " original: " << SI << "\n");
1505 IRBuilderTy IRB(&SI);
1506 Value *TV = SI.getTrueValue();
1507 Value *FV = SI.getFalseValue();
1508 // Replace the loads of the select with a select of two loads.
1509 while (!SI.use_empty()) {
1510 LoadInst *LI = cast<LoadInst>(SI.user_back());
1511 assert(LI->isSimple() && "We only speculate simple loads");
1513 IRB.SetInsertPoint(LI);
1515 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1517 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1518 NumLoadsSpeculated += 2;
1520 // Transfer alignment and AA info if present.
1521 TL->setAlignment(LI->getAlignment());
1522 FL->setAlignment(LI->getAlignment());
1525 LI->getAAMetadata(Tags);
1527 TL->setAAMetadata(Tags);
1528 FL->setAAMetadata(Tags);
1531 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1532 LI->getName() + ".sroa.speculated");
1534 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1535 LI->replaceAllUsesWith(V);
1536 LI->eraseFromParent();
1538 SI.eraseFromParent();
1541 /// \brief Build a GEP out of a base pointer and indices.
1543 /// This will return the BasePtr if that is valid, or build a new GEP
1544 /// instruction using the IRBuilder if GEP-ing is needed.
1545 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1546 SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1547 if (Indices.empty())
1550 // A single zero index is a no-op, so check for this and avoid building a GEP
1552 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1555 return IRB.CreateInBoundsGEP(BasePtr, Indices, NamePrefix + "sroa_idx");
1558 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1559 /// TargetTy without changing the offset of the pointer.
1561 /// This routine assumes we've already established a properly offset GEP with
1562 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1563 /// zero-indices down through type layers until we find one the same as
1564 /// TargetTy. If we can't find one with the same type, we at least try to use
1565 /// one with the same size. If none of that works, we just produce the GEP as
1566 /// indicated by Indices to have the correct offset.
1567 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1568 Value *BasePtr, Type *Ty, Type *TargetTy,
1569 SmallVectorImpl<Value *> &Indices,
1572 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1574 // Pointer size to use for the indices.
1575 unsigned PtrSize = DL.getPointerTypeSizeInBits(BasePtr->getType());
1577 // See if we can descend into a struct and locate a field with the correct
1579 unsigned NumLayers = 0;
1580 Type *ElementTy = Ty;
1582 if (ElementTy->isPointerTy())
1585 if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1586 ElementTy = ArrayTy->getElementType();
1587 Indices.push_back(IRB.getIntN(PtrSize, 0));
1588 } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1589 ElementTy = VectorTy->getElementType();
1590 Indices.push_back(IRB.getInt32(0));
1591 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1592 if (STy->element_begin() == STy->element_end())
1593 break; // Nothing left to descend into.
1594 ElementTy = *STy->element_begin();
1595 Indices.push_back(IRB.getInt32(0));
1600 } while (ElementTy != TargetTy);
1601 if (ElementTy != TargetTy)
1602 Indices.erase(Indices.end() - NumLayers, Indices.end());
1604 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1607 /// \brief Recursively compute indices for a natural GEP.
1609 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1610 /// element types adding appropriate indices for the GEP.
1611 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1612 Value *Ptr, Type *Ty, APInt &Offset,
1614 SmallVectorImpl<Value *> &Indices,
1617 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices,
1620 // We can't recurse through pointer types.
1621 if (Ty->isPointerTy())
1624 // We try to analyze GEPs over vectors here, but note that these GEPs are
1625 // extremely poorly defined currently. The long-term goal is to remove GEPing
1626 // over a vector from the IR completely.
1627 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1628 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1629 if (ElementSizeInBits % 8 != 0) {
1630 // GEPs over non-multiple of 8 size vector elements are invalid.
1633 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1634 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1635 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1637 Offset -= NumSkippedElements * ElementSize;
1638 Indices.push_back(IRB.getInt(NumSkippedElements));
1639 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1640 Offset, TargetTy, Indices, NamePrefix);
1643 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1644 Type *ElementTy = ArrTy->getElementType();
1645 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1646 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1647 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1650 Offset -= NumSkippedElements * ElementSize;
1651 Indices.push_back(IRB.getInt(NumSkippedElements));
1652 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1653 Indices, NamePrefix);
1656 StructType *STy = dyn_cast<StructType>(Ty);
1660 const StructLayout *SL = DL.getStructLayout(STy);
1661 uint64_t StructOffset = Offset.getZExtValue();
1662 if (StructOffset >= SL->getSizeInBytes())
1664 unsigned Index = SL->getElementContainingOffset(StructOffset);
1665 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1666 Type *ElementTy = STy->getElementType(Index);
1667 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1668 return nullptr; // The offset points into alignment padding.
1670 Indices.push_back(IRB.getInt32(Index));
1671 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1672 Indices, NamePrefix);
1675 /// \brief Get a natural GEP from a base pointer to a particular offset and
1676 /// resulting in a particular type.
1678 /// The goal is to produce a "natural" looking GEP that works with the existing
1679 /// composite types to arrive at the appropriate offset and element type for
1680 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1681 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1682 /// Indices, and setting Ty to the result subtype.
1684 /// If no natural GEP can be constructed, this function returns null.
1685 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1686 Value *Ptr, APInt Offset, Type *TargetTy,
1687 SmallVectorImpl<Value *> &Indices,
1689 PointerType *Ty = cast<PointerType>(Ptr->getType());
1691 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1693 if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1696 Type *ElementTy = Ty->getElementType();
1697 if (!ElementTy->isSized())
1698 return nullptr; // We can't GEP through an unsized element.
1699 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1700 if (ElementSize == 0)
1701 return nullptr; // Zero-length arrays can't help us build a natural GEP.
1702 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1704 Offset -= NumSkippedElements * ElementSize;
1705 Indices.push_back(IRB.getInt(NumSkippedElements));
1706 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1707 Indices, NamePrefix);
1710 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1711 /// resulting pointer has PointerTy.
1713 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1714 /// and produces the pointer type desired. Where it cannot, it will try to use
1715 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1716 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1717 /// bitcast to the type.
1719 /// The strategy for finding the more natural GEPs is to peel off layers of the
1720 /// pointer, walking back through bit casts and GEPs, searching for a base
1721 /// pointer from which we can compute a natural GEP with the desired
1722 /// properties. The algorithm tries to fold as many constant indices into
1723 /// a single GEP as possible, thus making each GEP more independent of the
1724 /// surrounding code.
1725 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1726 APInt Offset, Type *PointerTy, Twine NamePrefix) {
1727 // Even though we don't look through PHI nodes, we could be called on an
1728 // instruction in an unreachable block, which may be on a cycle.
1729 SmallPtrSet<Value *, 4> Visited;
1730 Visited.insert(Ptr);
1731 SmallVector<Value *, 4> Indices;
1733 // We may end up computing an offset pointer that has the wrong type. If we
1734 // never are able to compute one directly that has the correct type, we'll
1735 // fall back to it, so keep it and the base it was computed from around here.
1736 Value *OffsetPtr = nullptr;
1737 Value *OffsetBasePtr;
1739 // Remember any i8 pointer we come across to re-use if we need to do a raw
1741 Value *Int8Ptr = nullptr;
1742 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1744 Type *TargetTy = PointerTy->getPointerElementType();
1747 // First fold any existing GEPs into the offset.
1748 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1749 APInt GEPOffset(Offset.getBitWidth(), 0);
1750 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1752 Offset += GEPOffset;
1753 Ptr = GEP->getPointerOperand();
1754 if (!Visited.insert(Ptr).second)
1758 // See if we can perform a natural GEP here.
1760 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1761 Indices, NamePrefix)) {
1762 // If we have a new natural pointer at the offset, clear out any old
1763 // offset pointer we computed. Unless it is the base pointer or
1764 // a non-instruction, we built a GEP we don't need. Zap it.
1765 if (OffsetPtr && OffsetPtr != OffsetBasePtr)
1766 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr)) {
1767 assert(I->use_empty() && "Built a GEP with uses some how!");
1768 I->eraseFromParent();
1771 OffsetBasePtr = Ptr;
1772 // If we also found a pointer of the right type, we're done.
1773 if (P->getType() == PointerTy)
1777 // Stash this pointer if we've found an i8*.
1778 if (Ptr->getType()->isIntegerTy(8)) {
1780 Int8PtrOffset = Offset;
1783 // Peel off a layer of the pointer and update the offset appropriately.
1784 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1785 Ptr = cast<Operator>(Ptr)->getOperand(0);
1786 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1787 if (GA->mayBeOverridden())
1789 Ptr = GA->getAliasee();
1793 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1794 } while (Visited.insert(Ptr).second);
1798 Int8Ptr = IRB.CreateBitCast(
1799 Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1800 NamePrefix + "sroa_raw_cast");
1801 Int8PtrOffset = Offset;
1804 OffsetPtr = Int8PtrOffset == 0
1806 : IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1807 NamePrefix + "sroa_raw_idx");
1811 // On the off chance we were targeting i8*, guard the bitcast here.
1812 if (Ptr->getType() != PointerTy)
1813 Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
1818 /// \brief Compute the adjusted alignment for a load or store from an offset.
1819 static unsigned getAdjustedAlignment(Instruction *I, uint64_t Offset,
1820 const DataLayout &DL) {
1823 if (auto *LI = dyn_cast<LoadInst>(I)) {
1824 Alignment = LI->getAlignment();
1826 } else if (auto *SI = dyn_cast<StoreInst>(I)) {
1827 Alignment = SI->getAlignment();
1828 Ty = SI->getValueOperand()->getType();
1830 llvm_unreachable("Only loads and stores are allowed!");
1834 Alignment = DL.getABITypeAlignment(Ty);
1836 return MinAlign(Alignment, Offset);
1839 /// \brief Test whether we can convert a value from the old to the new type.
1841 /// This predicate should be used to guard calls to convertValue in order to
1842 /// ensure that we only try to convert viable values. The strategy is that we
1843 /// will peel off single element struct and array wrappings to get to an
1844 /// underlying value, and convert that value.
1845 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1848 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1849 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1850 if (NewITy->getBitWidth() >= OldITy->getBitWidth())
1852 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1854 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1857 // We can convert pointers to integers and vice-versa. Same for vectors
1858 // of pointers and integers.
1859 OldTy = OldTy->getScalarType();
1860 NewTy = NewTy->getScalarType();
1861 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1862 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1864 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1872 /// \brief Generic routine to convert an SSA value to a value of a different
1875 /// This will try various different casting techniques, such as bitcasts,
1876 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1877 /// two types for viability with this routine.
1878 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1880 Type *OldTy = V->getType();
1881 assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1886 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1887 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1888 if (NewITy->getBitWidth() > OldITy->getBitWidth())
1889 return IRB.CreateZExt(V, NewITy);
1891 // See if we need inttoptr for this type pair. A cast involving both scalars
1892 // and vectors requires and additional bitcast.
1893 if (OldTy->getScalarType()->isIntegerTy() &&
1894 NewTy->getScalarType()->isPointerTy()) {
1895 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1896 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1897 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1900 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1901 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1902 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1905 return IRB.CreateIntToPtr(V, NewTy);
1908 // See if we need ptrtoint for this type pair. A cast involving both scalars
1909 // and vectors requires and additional bitcast.
1910 if (OldTy->getScalarType()->isPointerTy() &&
1911 NewTy->getScalarType()->isIntegerTy()) {
1912 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1913 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1914 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1917 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1918 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1919 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1922 return IRB.CreatePtrToInt(V, NewTy);
1925 return IRB.CreateBitCast(V, NewTy);
1928 /// \brief Test whether the given slice use can be promoted to a vector.
1930 /// This function is called to test each entry in a partioning which is slated
1931 /// for a single slice.
1932 static bool isVectorPromotionViableForSlice(AllocaSlices::Partition &P,
1933 const Slice &S, VectorType *Ty,
1934 uint64_t ElementSize,
1935 const DataLayout &DL) {
1936 // First validate the slice offsets.
1937 uint64_t BeginOffset =
1938 std::max(S.beginOffset(), P.beginOffset()) - P.beginOffset();
1939 uint64_t BeginIndex = BeginOffset / ElementSize;
1940 if (BeginIndex * ElementSize != BeginOffset ||
1941 BeginIndex >= Ty->getNumElements())
1943 uint64_t EndOffset =
1944 std::min(S.endOffset(), P.endOffset()) - P.beginOffset();
1945 uint64_t EndIndex = EndOffset / ElementSize;
1946 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1949 assert(EndIndex > BeginIndex && "Empty vector!");
1950 uint64_t NumElements = EndIndex - BeginIndex;
1951 Type *SliceTy = (NumElements == 1)
1952 ? Ty->getElementType()
1953 : VectorType::get(Ty->getElementType(), NumElements);
1956 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1958 Use *U = S.getUse();
1960 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1961 if (MI->isVolatile())
1963 if (!S.isSplittable())
1964 return false; // Skip any unsplittable intrinsics.
1965 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1966 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1967 II->getIntrinsicID() != Intrinsic::lifetime_end)
1969 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1970 // Disable vector promotion when there are loads or stores of an FCA.
1972 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1973 if (LI->isVolatile())
1975 Type *LTy = LI->getType();
1976 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1977 assert(LTy->isIntegerTy());
1980 if (!canConvertValue(DL, SliceTy, LTy))
1982 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1983 if (SI->isVolatile())
1985 Type *STy = SI->getValueOperand()->getType();
1986 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1987 assert(STy->isIntegerTy());
1990 if (!canConvertValue(DL, STy, SliceTy))
1999 /// \brief Test whether the given alloca partitioning and range of slices can be
2000 /// promoted to a vector.
2002 /// This is a quick test to check whether we can rewrite a particular alloca
2003 /// partition (and its newly formed alloca) into a vector alloca with only
2004 /// whole-vector loads and stores such that it could be promoted to a vector
2005 /// SSA value. We only can ensure this for a limited set of operations, and we
2006 /// don't want to do the rewrites unless we are confident that the result will
2007 /// be promotable, so we have an early test here.
2008 static VectorType *isVectorPromotionViable(AllocaSlices::Partition &P,
2009 const DataLayout &DL) {
2010 // Collect the candidate types for vector-based promotion. Also track whether
2011 // we have different element types.
2012 SmallVector<VectorType *, 4> CandidateTys;
2013 Type *CommonEltTy = nullptr;
2014 bool HaveCommonEltTy = true;
2015 auto CheckCandidateType = [&](Type *Ty) {
2016 if (auto *VTy = dyn_cast<VectorType>(Ty)) {
2017 CandidateTys.push_back(VTy);
2019 CommonEltTy = VTy->getElementType();
2020 else if (CommonEltTy != VTy->getElementType())
2021 HaveCommonEltTy = false;
2024 // Consider any loads or stores that are the exact size of the slice.
2025 for (const Slice &S : P)
2026 if (S.beginOffset() == P.beginOffset() &&
2027 S.endOffset() == P.endOffset()) {
2028 if (auto *LI = dyn_cast<LoadInst>(S.getUse()->getUser()))
2029 CheckCandidateType(LI->getType());
2030 else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser()))
2031 CheckCandidateType(SI->getValueOperand()->getType());
2034 // If we didn't find a vector type, nothing to do here.
2035 if (CandidateTys.empty())
2038 // Remove non-integer vector types if we had multiple common element types.
2039 // FIXME: It'd be nice to replace them with integer vector types, but we can't
2040 // do that until all the backends are known to produce good code for all
2041 // integer vector types.
2042 if (!HaveCommonEltTy) {
2043 CandidateTys.erase(std::remove_if(CandidateTys.begin(), CandidateTys.end(),
2044 [](VectorType *VTy) {
2045 return !VTy->getElementType()->isIntegerTy();
2047 CandidateTys.end());
2049 // If there were no integer vector types, give up.
2050 if (CandidateTys.empty())
2053 // Rank the remaining candidate vector types. This is easy because we know
2054 // they're all integer vectors. We sort by ascending number of elements.
2055 auto RankVectorTypes = [&DL](VectorType *RHSTy, VectorType *LHSTy) {
2056 assert(DL.getTypeSizeInBits(RHSTy) == DL.getTypeSizeInBits(LHSTy) &&
2057 "Cannot have vector types of different sizes!");
2058 assert(RHSTy->getElementType()->isIntegerTy() &&
2059 "All non-integer types eliminated!");
2060 assert(LHSTy->getElementType()->isIntegerTy() &&
2061 "All non-integer types eliminated!");
2062 return RHSTy->getNumElements() < LHSTy->getNumElements();
2064 std::sort(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes);
2066 std::unique(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes),
2067 CandidateTys.end());
2069 // The only way to have the same element type in every vector type is to
2070 // have the same vector type. Check that and remove all but one.
2072 for (VectorType *VTy : CandidateTys) {
2073 assert(VTy->getElementType() == CommonEltTy &&
2074 "Unaccounted for element type!");
2075 assert(VTy == CandidateTys[0] &&
2076 "Different vector types with the same element type!");
2079 CandidateTys.resize(1);
2082 // Try each vector type, and return the one which works.
2083 auto CheckVectorTypeForPromotion = [&](VectorType *VTy) {
2084 uint64_t ElementSize = DL.getTypeSizeInBits(VTy->getElementType());
2086 // While the definition of LLVM vectors is bitpacked, we don't support sizes
2087 // that aren't byte sized.
2088 if (ElementSize % 8)
2090 assert((DL.getTypeSizeInBits(VTy) % 8) == 0 &&
2091 "vector size not a multiple of element size?");
2094 for (const Slice &S : P)
2095 if (!isVectorPromotionViableForSlice(P, S, VTy, ElementSize, DL))
2098 for (const Slice *S : P.splitSliceTails())
2099 if (!isVectorPromotionViableForSlice(P, *S, VTy, ElementSize, DL))
2104 for (VectorType *VTy : CandidateTys)
2105 if (CheckVectorTypeForPromotion(VTy))
2111 /// \brief Test whether a slice of an alloca is valid for integer widening.
2113 /// This implements the necessary checking for the \c isIntegerWideningViable
2114 /// test below on a single slice of the alloca.
2115 static bool isIntegerWideningViableForSlice(const Slice &S,
2116 uint64_t AllocBeginOffset,
2118 const DataLayout &DL,
2119 bool &WholeAllocaOp) {
2120 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
2122 uint64_t RelBegin = S.beginOffset() - AllocBeginOffset;
2123 uint64_t RelEnd = S.endOffset() - AllocBeginOffset;
2125 // We can't reasonably handle cases where the load or store extends past
2126 // the end of the aloca's type and into its padding.
2130 Use *U = S.getUse();
2132 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
2133 if (LI->isVolatile())
2135 // Note that we don't count vector loads or stores as whole-alloca
2136 // operations which enable integer widening because we would prefer to use
2137 // vector widening instead.
2138 if (!isa<VectorType>(LI->getType()) && RelBegin == 0 && RelEnd == Size)
2139 WholeAllocaOp = true;
2140 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
2141 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
2143 } else if (RelBegin != 0 || RelEnd != Size ||
2144 !canConvertValue(DL, AllocaTy, LI->getType())) {
2145 // Non-integer loads need to be convertible from the alloca type so that
2146 // they are promotable.
2149 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
2150 Type *ValueTy = SI->getValueOperand()->getType();
2151 if (SI->isVolatile())
2153 // Note that we don't count vector loads or stores as whole-alloca
2154 // operations which enable integer widening because we would prefer to use
2155 // vector widening instead.
2156 if (!isa<VectorType>(ValueTy) && RelBegin == 0 && RelEnd == Size)
2157 WholeAllocaOp = true;
2158 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
2159 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
2161 } else if (RelBegin != 0 || RelEnd != Size ||
2162 !canConvertValue(DL, ValueTy, AllocaTy)) {
2163 // Non-integer stores need to be convertible to the alloca type so that
2164 // they are promotable.
2167 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
2168 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
2170 if (!S.isSplittable())
2171 return false; // Skip any unsplittable intrinsics.
2172 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
2173 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2174 II->getIntrinsicID() != Intrinsic::lifetime_end)
2183 /// \brief Test whether the given alloca partition's integer operations can be
2184 /// widened to promotable ones.
2186 /// This is a quick test to check whether we can rewrite the integer loads and
2187 /// stores to a particular alloca into wider loads and stores and be able to
2188 /// promote the resulting alloca.
2189 static bool isIntegerWideningViable(AllocaSlices::Partition &P, Type *AllocaTy,
2190 const DataLayout &DL) {
2191 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
2192 // Don't create integer types larger than the maximum bitwidth.
2193 if (SizeInBits > IntegerType::MAX_INT_BITS)
2196 // Don't try to handle allocas with bit-padding.
2197 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
2200 // We need to ensure that an integer type with the appropriate bitwidth can
2201 // be converted to the alloca type, whatever that is. We don't want to force
2202 // the alloca itself to have an integer type if there is a more suitable one.
2203 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
2204 if (!canConvertValue(DL, AllocaTy, IntTy) ||
2205 !canConvertValue(DL, IntTy, AllocaTy))
2208 // While examining uses, we ensure that the alloca has a covering load or
2209 // store. We don't want to widen the integer operations only to fail to
2210 // promote due to some other unsplittable entry (which we may make splittable
2211 // later). However, if there are only splittable uses, go ahead and assume
2212 // that we cover the alloca.
2213 // FIXME: We shouldn't consider split slices that happen to start in the
2214 // partition here...
2215 bool WholeAllocaOp =
2216 P.begin() != P.end() ? false : DL.isLegalInteger(SizeInBits);
2218 for (const Slice &S : P)
2219 if (!isIntegerWideningViableForSlice(S, P.beginOffset(), AllocaTy, DL,
2223 for (const Slice *S : P.splitSliceTails())
2224 if (!isIntegerWideningViableForSlice(*S, P.beginOffset(), AllocaTy, DL,
2228 return WholeAllocaOp;
2231 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
2232 IntegerType *Ty, uint64_t Offset,
2233 const Twine &Name) {
2234 DEBUG(dbgs() << " start: " << *V << "\n");
2235 IntegerType *IntTy = cast<IntegerType>(V->getType());
2236 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2237 "Element extends past full value");
2238 uint64_t ShAmt = 8 * Offset;
2239 if (DL.isBigEndian())
2240 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2242 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2243 DEBUG(dbgs() << " shifted: " << *V << "\n");
2245 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2246 "Cannot extract to a larger integer!");
2248 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2249 DEBUG(dbgs() << " trunced: " << *V << "\n");
2254 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
2255 Value *V, uint64_t Offset, const Twine &Name) {
2256 IntegerType *IntTy = cast<IntegerType>(Old->getType());
2257 IntegerType *Ty = cast<IntegerType>(V->getType());
2258 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2259 "Cannot insert a larger integer!");
2260 DEBUG(dbgs() << " start: " << *V << "\n");
2262 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2263 DEBUG(dbgs() << " extended: " << *V << "\n");
2265 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2266 "Element store outside of alloca store");
2267 uint64_t ShAmt = 8 * Offset;
2268 if (DL.isBigEndian())
2269 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2271 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2272 DEBUG(dbgs() << " shifted: " << *V << "\n");
2275 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2276 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2277 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2278 DEBUG(dbgs() << " masked: " << *Old << "\n");
2279 V = IRB.CreateOr(Old, V, Name + ".insert");
2280 DEBUG(dbgs() << " inserted: " << *V << "\n");
2285 static Value *extractVector(IRBuilderTy &IRB, Value *V, unsigned BeginIndex,
2286 unsigned EndIndex, const Twine &Name) {
2287 VectorType *VecTy = cast<VectorType>(V->getType());
2288 unsigned NumElements = EndIndex - BeginIndex;
2289 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2291 if (NumElements == VecTy->getNumElements())
2294 if (NumElements == 1) {
2295 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2297 DEBUG(dbgs() << " extract: " << *V << "\n");
2301 SmallVector<Constant *, 8> Mask;
2302 Mask.reserve(NumElements);
2303 for (unsigned i = BeginIndex; i != EndIndex; ++i)
2304 Mask.push_back(IRB.getInt32(i));
2305 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2306 ConstantVector::get(Mask), Name + ".extract");
2307 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2311 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
2312 unsigned BeginIndex, const Twine &Name) {
2313 VectorType *VecTy = cast<VectorType>(Old->getType());
2314 assert(VecTy && "Can only insert a vector into a vector");
2316 VectorType *Ty = dyn_cast<VectorType>(V->getType());
2318 // Single element to insert.
2319 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2321 DEBUG(dbgs() << " insert: " << *V << "\n");
2325 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
2326 "Too many elements!");
2327 if (Ty->getNumElements() == VecTy->getNumElements()) {
2328 assert(V->getType() == VecTy && "Vector type mismatch");
2331 unsigned EndIndex = BeginIndex + Ty->getNumElements();
2333 // When inserting a smaller vector into the larger to store, we first
2334 // use a shuffle vector to widen it with undef elements, and then
2335 // a second shuffle vector to select between the loaded vector and the
2337 SmallVector<Constant *, 8> Mask;
2338 Mask.reserve(VecTy->getNumElements());
2339 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2340 if (i >= BeginIndex && i < EndIndex)
2341 Mask.push_back(IRB.getInt32(i - BeginIndex));
2343 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2344 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2345 ConstantVector::get(Mask), Name + ".expand");
2346 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2349 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2350 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
2352 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
2354 DEBUG(dbgs() << " blend: " << *V << "\n");
2359 /// \brief Visitor to rewrite instructions using p particular slice of an alloca
2360 /// to use a new alloca.
2362 /// Also implements the rewriting to vector-based accesses when the partition
2363 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2365 class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
2366 // Befriend the base class so it can delegate to private visit methods.
2367 friend class llvm::InstVisitor<AllocaSliceRewriter, bool>;
2368 typedef llvm::InstVisitor<AllocaSliceRewriter, bool> Base;
2370 const DataLayout &DL;
2373 AllocaInst &OldAI, &NewAI;
2374 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2377 // This is a convenience and flag variable that will be null unless the new
2378 // alloca's integer operations should be widened to this integer type due to
2379 // passing isIntegerWideningViable above. If it is non-null, the desired
2380 // integer type will be stored here for easy access during rewriting.
2383 // If we are rewriting an alloca partition which can be written as pure
2384 // vector operations, we stash extra information here. When VecTy is
2385 // non-null, we have some strict guarantees about the rewritten alloca:
2386 // - The new alloca is exactly the size of the vector type here.
2387 // - The accesses all either map to the entire vector or to a single
2389 // - The set of accessing instructions is only one of those handled above
2390 // in isVectorPromotionViable. Generally these are the same access kinds
2391 // which are promotable via mem2reg.
2394 uint64_t ElementSize;
2396 // The original offset of the slice currently being rewritten relative to
2397 // the original alloca.
2398 uint64_t BeginOffset, EndOffset;
2399 // The new offsets of the slice currently being rewritten relative to the
2401 uint64_t NewBeginOffset, NewEndOffset;
2407 Instruction *OldPtr;
2409 // Track post-rewrite users which are PHI nodes and Selects.
2410 SmallPtrSetImpl<PHINode *> &PHIUsers;
2411 SmallPtrSetImpl<SelectInst *> &SelectUsers;
2413 // Utility IR builder, whose name prefix is setup for each visited use, and
2414 // the insertion point is set to point to the user.
2418 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &AS, SROA &Pass,
2419 AllocaInst &OldAI, AllocaInst &NewAI,
2420 uint64_t NewAllocaBeginOffset,
2421 uint64_t NewAllocaEndOffset, bool IsIntegerPromotable,
2422 VectorType *PromotableVecTy,
2423 SmallPtrSetImpl<PHINode *> &PHIUsers,
2424 SmallPtrSetImpl<SelectInst *> &SelectUsers)
2425 : DL(DL), AS(AS), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2426 NewAllocaBeginOffset(NewAllocaBeginOffset),
2427 NewAllocaEndOffset(NewAllocaEndOffset),
2428 NewAllocaTy(NewAI.getAllocatedType()),
2429 IntTy(IsIntegerPromotable
2432 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2434 VecTy(PromotableVecTy),
2435 ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2436 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
2437 BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
2438 OldPtr(), PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2439 IRB(NewAI.getContext(), ConstantFolder()) {
2441 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
2442 "Only multiple-of-8 sized vector elements are viable");
2445 assert((!IntTy && !VecTy) || (IntTy && !VecTy) || (!IntTy && VecTy));
2448 bool visit(AllocaSlices::const_iterator I) {
2449 bool CanSROA = true;
2450 BeginOffset = I->beginOffset();
2451 EndOffset = I->endOffset();
2452 IsSplittable = I->isSplittable();
2454 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2455 DEBUG(dbgs() << " rewriting " << (IsSplit ? "split " : ""));
2456 DEBUG(AS.printSlice(dbgs(), I, ""));
2457 DEBUG(dbgs() << "\n");
2459 // Compute the intersecting offset range.
2460 assert(BeginOffset < NewAllocaEndOffset);
2461 assert(EndOffset > NewAllocaBeginOffset);
2462 NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2463 NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2465 SliceSize = NewEndOffset - NewBeginOffset;
2467 OldUse = I->getUse();
2468 OldPtr = cast<Instruction>(OldUse->get());
2470 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2471 IRB.SetInsertPoint(OldUserI);
2472 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2473 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2475 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2482 // Make sure the other visit overloads are visible.
2485 // Every instruction which can end up as a user must have a rewrite rule.
2486 bool visitInstruction(Instruction &I) {
2487 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2488 llvm_unreachable("No rewrite rule for this instruction!");
2491 Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2492 // Note that the offset computation can use BeginOffset or NewBeginOffset
2493 // interchangeably for unsplit slices.
2494 assert(IsSplit || BeginOffset == NewBeginOffset);
2495 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2498 StringRef OldName = OldPtr->getName();
2499 // Skip through the last '.sroa.' component of the name.
2500 size_t LastSROAPrefix = OldName.rfind(".sroa.");
2501 if (LastSROAPrefix != StringRef::npos) {
2502 OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2503 // Look for an SROA slice index.
2504 size_t IndexEnd = OldName.find_first_not_of("0123456789");
2505 if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2506 // Strip the index and look for the offset.
2507 OldName = OldName.substr(IndexEnd + 1);
2508 size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2509 if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2510 // Strip the offset.
2511 OldName = OldName.substr(OffsetEnd + 1);
2514 // Strip any SROA suffixes as well.
2515 OldName = OldName.substr(0, OldName.find(".sroa_"));
2518 return getAdjustedPtr(IRB, DL, &NewAI,
2519 APInt(DL.getPointerSizeInBits(), Offset), PointerTy,
2521 Twine(OldName) + "."
2528 /// \brief Compute suitable alignment to access this slice of the *new*
2531 /// You can optionally pass a type to this routine and if that type's ABI
2532 /// alignment is itself suitable, this will return zero.
2533 unsigned getSliceAlign(Type *Ty = nullptr) {
2534 unsigned NewAIAlign = NewAI.getAlignment();
2536 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2538 MinAlign(NewAIAlign, NewBeginOffset - NewAllocaBeginOffset);
2539 return (Ty && Align == DL.getABITypeAlignment(Ty)) ? 0 : Align;
2542 unsigned getIndex(uint64_t Offset) {
2543 assert(VecTy && "Can only call getIndex when rewriting a vector");
2544 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2545 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2546 uint32_t Index = RelOffset / ElementSize;
2547 assert(Index * ElementSize == RelOffset);
2551 void deleteIfTriviallyDead(Value *V) {
2552 Instruction *I = cast<Instruction>(V);
2553 if (isInstructionTriviallyDead(I))
2554 Pass.DeadInsts.insert(I);
2557 Value *rewriteVectorizedLoadInst() {
2558 unsigned BeginIndex = getIndex(NewBeginOffset);
2559 unsigned EndIndex = getIndex(NewEndOffset);
2560 assert(EndIndex > BeginIndex && "Empty vector!");
2562 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2563 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2566 Value *rewriteIntegerLoad(LoadInst &LI) {
2567 assert(IntTy && "We cannot insert an integer to the alloca");
2568 assert(!LI.isVolatile());
2569 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2570 V = convertValue(DL, IRB, V, IntTy);
2571 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2572 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2573 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset)
2574 V = extractInteger(DL, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2579 bool visitLoadInst(LoadInst &LI) {
2580 DEBUG(dbgs() << " original: " << LI << "\n");
2581 Value *OldOp = LI.getOperand(0);
2582 assert(OldOp == OldPtr);
2584 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2586 bool IsPtrAdjusted = false;
2589 V = rewriteVectorizedLoadInst();
2590 } else if (IntTy && LI.getType()->isIntegerTy()) {
2591 V = rewriteIntegerLoad(LI);
2592 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2593 canConvertValue(DL, NewAllocaTy, LI.getType())) {
2594 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), LI.isVolatile(),
2597 Type *LTy = TargetTy->getPointerTo();
2598 V = IRB.CreateAlignedLoad(getNewAllocaSlicePtr(IRB, LTy),
2599 getSliceAlign(TargetTy), LI.isVolatile(),
2601 IsPtrAdjusted = true;
2603 V = convertValue(DL, IRB, V, TargetTy);
2606 assert(!LI.isVolatile());
2607 assert(LI.getType()->isIntegerTy() &&
2608 "Only integer type loads and stores are split");
2609 assert(SliceSize < DL.getTypeStoreSize(LI.getType()) &&
2610 "Split load isn't smaller than original load");
2611 assert(LI.getType()->getIntegerBitWidth() ==
2612 DL.getTypeStoreSizeInBits(LI.getType()) &&
2613 "Non-byte-multiple bit width");
2614 // Move the insertion point just past the load so that we can refer to it.
2615 IRB.SetInsertPoint(std::next(BasicBlock::iterator(&LI)));
2616 // Create a placeholder value with the same type as LI to use as the
2617 // basis for the new value. This allows us to replace the uses of LI with
2618 // the computed value, and then replace the placeholder with LI, leaving
2619 // LI only used for this computation.
2620 Value *Placeholder =
2621 new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2622 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset - BeginOffset,
2624 LI.replaceAllUsesWith(V);
2625 Placeholder->replaceAllUsesWith(&LI);
2628 LI.replaceAllUsesWith(V);
2631 Pass.DeadInsts.insert(&LI);
2632 deleteIfTriviallyDead(OldOp);
2633 DEBUG(dbgs() << " to: " << *V << "\n");
2634 return !LI.isVolatile() && !IsPtrAdjusted;
2637 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp) {
2638 if (V->getType() != VecTy) {
2639 unsigned BeginIndex = getIndex(NewBeginOffset);
2640 unsigned EndIndex = getIndex(NewEndOffset);
2641 assert(EndIndex > BeginIndex && "Empty vector!");
2642 unsigned NumElements = EndIndex - BeginIndex;
2643 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2644 Type *SliceTy = (NumElements == 1)
2646 : VectorType::get(ElementTy, NumElements);
2647 if (V->getType() != SliceTy)
2648 V = convertValue(DL, IRB, V, SliceTy);
2650 // Mix in the existing elements.
2651 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2652 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2654 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2655 Pass.DeadInsts.insert(&SI);
2658 DEBUG(dbgs() << " to: " << *Store << "\n");
2662 bool rewriteIntegerStore(Value *V, StoreInst &SI) {
2663 assert(IntTy && "We cannot extract an integer from the alloca");
2664 assert(!SI.isVolatile());
2665 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2667 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2668 Old = convertValue(DL, IRB, Old, IntTy);
2669 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2670 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2671 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset, "insert");
2673 V = convertValue(DL, IRB, V, NewAllocaTy);
2674 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2675 Pass.DeadInsts.insert(&SI);
2677 DEBUG(dbgs() << " to: " << *Store << "\n");
2681 bool visitStoreInst(StoreInst &SI) {
2682 DEBUG(dbgs() << " original: " << SI << "\n");
2683 Value *OldOp = SI.getOperand(1);
2684 assert(OldOp == OldPtr);
2686 Value *V = SI.getValueOperand();
2688 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2689 // alloca that should be re-examined after promoting this alloca.
2690 if (V->getType()->isPointerTy())
2691 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2692 Pass.PostPromotionWorklist.insert(AI);
2694 if (SliceSize < DL.getTypeStoreSize(V->getType())) {
2695 assert(!SI.isVolatile());
2696 assert(V->getType()->isIntegerTy() &&
2697 "Only integer type loads and stores are split");
2698 assert(V->getType()->getIntegerBitWidth() ==
2699 DL.getTypeStoreSizeInBits(V->getType()) &&
2700 "Non-byte-multiple bit width");
2701 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2702 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset - BeginOffset,
2707 return rewriteVectorizedStoreInst(V, SI, OldOp);
2708 if (IntTy && V->getType()->isIntegerTy())
2709 return rewriteIntegerStore(V, SI);
2712 if (NewBeginOffset == NewAllocaBeginOffset &&
2713 NewEndOffset == NewAllocaEndOffset &&
2714 canConvertValue(DL, V->getType(), NewAllocaTy)) {
2715 V = convertValue(DL, IRB, V, NewAllocaTy);
2716 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2719 Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo());
2720 NewSI = IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(V->getType()),
2724 Pass.DeadInsts.insert(&SI);
2725 deleteIfTriviallyDead(OldOp);
2727 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2728 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2731 /// \brief Compute an integer value from splatting an i8 across the given
2732 /// number of bytes.
2734 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2735 /// call this routine.
2736 /// FIXME: Heed the advice above.
2738 /// \param V The i8 value to splat.
2739 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2740 Value *getIntegerSplat(Value *V, unsigned Size) {
2741 assert(Size > 0 && "Expected a positive number of bytes.");
2742 IntegerType *VTy = cast<IntegerType>(V->getType());
2743 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2747 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size * 8);
2749 IRB.CreateZExt(V, SplatIntTy, "zext"),
2750 ConstantExpr::getUDiv(
2751 Constant::getAllOnesValue(SplatIntTy),
2752 ConstantExpr::getZExt(Constant::getAllOnesValue(V->getType()),
2758 /// \brief Compute a vector splat for a given element value.
2759 Value *getVectorSplat(Value *V, unsigned NumElements) {
2760 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2761 DEBUG(dbgs() << " splat: " << *V << "\n");
2765 bool visitMemSetInst(MemSetInst &II) {
2766 DEBUG(dbgs() << " original: " << II << "\n");
2767 assert(II.getRawDest() == OldPtr);
2769 // If the memset has a variable size, it cannot be split, just adjust the
2770 // pointer to the new alloca.
2771 if (!isa<Constant>(II.getLength())) {
2773 assert(NewBeginOffset == BeginOffset);
2774 II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2775 Type *CstTy = II.getAlignmentCst()->getType();
2776 II.setAlignment(ConstantInt::get(CstTy, getSliceAlign()));
2778 deleteIfTriviallyDead(OldPtr);
2782 // Record this instruction for deletion.
2783 Pass.DeadInsts.insert(&II);
2785 Type *AllocaTy = NewAI.getAllocatedType();
2786 Type *ScalarTy = AllocaTy->getScalarType();
2788 // If this doesn't map cleanly onto the alloca type, and that type isn't
2789 // a single value type, just emit a memset.
2790 if (!VecTy && !IntTy &&
2791 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2792 SliceSize != DL.getTypeStoreSize(AllocaTy) ||
2793 !AllocaTy->isSingleValueType() ||
2794 !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2795 DL.getTypeSizeInBits(ScalarTy) % 8 != 0)) {
2796 Type *SizeTy = II.getLength()->getType();
2797 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2798 CallInst *New = IRB.CreateMemSet(
2799 getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2800 getSliceAlign(), II.isVolatile());
2802 DEBUG(dbgs() << " to: " << *New << "\n");
2806 // If we can represent this as a simple value, we have to build the actual
2807 // value to store, which requires expanding the byte present in memset to
2808 // a sensible representation for the alloca type. This is essentially
2809 // splatting the byte to a sufficiently wide integer, splatting it across
2810 // any desired vector width, and bitcasting to the final type.
2814 // If this is a memset of a vectorized alloca, insert it.
2815 assert(ElementTy == ScalarTy);
2817 unsigned BeginIndex = getIndex(NewBeginOffset);
2818 unsigned EndIndex = getIndex(NewEndOffset);
2819 assert(EndIndex > BeginIndex && "Empty vector!");
2820 unsigned NumElements = EndIndex - BeginIndex;
2821 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2824 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2825 Splat = convertValue(DL, IRB, Splat, ElementTy);
2826 if (NumElements > 1)
2827 Splat = getVectorSplat(Splat, NumElements);
2830 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2831 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2833 // If this is a memset on an alloca where we can widen stores, insert the
2835 assert(!II.isVolatile());
2837 uint64_t Size = NewEndOffset - NewBeginOffset;
2838 V = getIntegerSplat(II.getValue(), Size);
2840 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2841 EndOffset != NewAllocaBeginOffset)) {
2843 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2844 Old = convertValue(DL, IRB, Old, IntTy);
2845 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2846 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2848 assert(V->getType() == IntTy &&
2849 "Wrong type for an alloca wide integer!");
2851 V = convertValue(DL, IRB, V, AllocaTy);
2853 // Established these invariants above.
2854 assert(NewBeginOffset == NewAllocaBeginOffset);
2855 assert(NewEndOffset == NewAllocaEndOffset);
2857 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2858 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2859 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2861 V = convertValue(DL, IRB, V, AllocaTy);
2864 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2867 DEBUG(dbgs() << " to: " << *New << "\n");
2868 return !II.isVolatile();
2871 bool visitMemTransferInst(MemTransferInst &II) {
2872 // Rewriting of memory transfer instructions can be a bit tricky. We break
2873 // them into two categories: split intrinsics and unsplit intrinsics.
2875 DEBUG(dbgs() << " original: " << II << "\n");
2877 bool IsDest = &II.getRawDestUse() == OldUse;
2878 assert((IsDest && II.getRawDest() == OldPtr) ||
2879 (!IsDest && II.getRawSource() == OldPtr));
2881 unsigned SliceAlign = getSliceAlign();
2883 // For unsplit intrinsics, we simply modify the source and destination
2884 // pointers in place. This isn't just an optimization, it is a matter of
2885 // correctness. With unsplit intrinsics we may be dealing with transfers
2886 // within a single alloca before SROA ran, or with transfers that have
2887 // a variable length. We may also be dealing with memmove instead of
2888 // memcpy, and so simply updating the pointers is the necessary for us to
2889 // update both source and dest of a single call.
2890 if (!IsSplittable) {
2891 Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2893 II.setDest(AdjustedPtr);
2895 II.setSource(AdjustedPtr);
2897 if (II.getAlignment() > SliceAlign) {
2898 Type *CstTy = II.getAlignmentCst()->getType();
2900 ConstantInt::get(CstTy, MinAlign(II.getAlignment(), SliceAlign)));
2903 DEBUG(dbgs() << " to: " << II << "\n");
2904 deleteIfTriviallyDead(OldPtr);
2907 // For split transfer intrinsics we have an incredibly useful assurance:
2908 // the source and destination do not reside within the same alloca, and at
2909 // least one of them does not escape. This means that we can replace
2910 // memmove with memcpy, and we don't need to worry about all manner of
2911 // downsides to splitting and transforming the operations.
2913 // If this doesn't map cleanly onto the alloca type, and that type isn't
2914 // a single value type, just emit a memcpy.
2917 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2918 SliceSize != DL.getTypeStoreSize(NewAI.getAllocatedType()) ||
2919 !NewAI.getAllocatedType()->isSingleValueType());
2921 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2922 // size hasn't been shrunk based on analysis of the viable range, this is
2924 if (EmitMemCpy && &OldAI == &NewAI) {
2925 // Ensure the start lines up.
2926 assert(NewBeginOffset == BeginOffset);
2928 // Rewrite the size as needed.
2929 if (NewEndOffset != EndOffset)
2930 II.setLength(ConstantInt::get(II.getLength()->getType(),
2931 NewEndOffset - NewBeginOffset));
2934 // Record this instruction for deletion.
2935 Pass.DeadInsts.insert(&II);
2937 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2938 // alloca that should be re-examined after rewriting this instruction.
2939 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2940 if (AllocaInst *AI =
2941 dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2942 assert(AI != &OldAI && AI != &NewAI &&
2943 "Splittable transfers cannot reach the same alloca on both ends.");
2944 Pass.Worklist.insert(AI);
2947 Type *OtherPtrTy = OtherPtr->getType();
2948 unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2950 // Compute the relative offset for the other pointer within the transfer.
2951 unsigned IntPtrWidth = DL.getPointerSizeInBits(OtherAS);
2952 APInt OtherOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2953 unsigned OtherAlign = MinAlign(II.getAlignment() ? II.getAlignment() : 1,
2954 OtherOffset.zextOrTrunc(64).getZExtValue());
2957 // Compute the other pointer, folding as much as possible to produce
2958 // a single, simple GEP in most cases.
2959 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2960 OtherPtr->getName() + ".");
2962 Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2963 Type *SizeTy = II.getLength()->getType();
2964 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2966 CallInst *New = IRB.CreateMemCpy(
2967 IsDest ? OurPtr : OtherPtr, IsDest ? OtherPtr : OurPtr, Size,
2968 MinAlign(SliceAlign, OtherAlign), II.isVolatile());
2970 DEBUG(dbgs() << " to: " << *New << "\n");
2974 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2975 NewEndOffset == NewAllocaEndOffset;
2976 uint64_t Size = NewEndOffset - NewBeginOffset;
2977 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2978 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2979 unsigned NumElements = EndIndex - BeginIndex;
2980 IntegerType *SubIntTy =
2981 IntTy ? Type::getIntNTy(IntTy->getContext(), Size * 8) : nullptr;
2983 // Reset the other pointer type to match the register type we're going to
2984 // use, but using the address space of the original other pointer.
2985 if (VecTy && !IsWholeAlloca) {
2986 if (NumElements == 1)
2987 OtherPtrTy = VecTy->getElementType();
2989 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2991 OtherPtrTy = OtherPtrTy->getPointerTo(OtherAS);
2992 } else if (IntTy && !IsWholeAlloca) {
2993 OtherPtrTy = SubIntTy->getPointerTo(OtherAS);
2995 OtherPtrTy = NewAllocaTy->getPointerTo(OtherAS);
2998 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2999 OtherPtr->getName() + ".");
3000 unsigned SrcAlign = OtherAlign;
3001 Value *DstPtr = &NewAI;
3002 unsigned DstAlign = SliceAlign;
3004 std::swap(SrcPtr, DstPtr);
3005 std::swap(SrcAlign, DstAlign);
3009 if (VecTy && !IsWholeAlloca && !IsDest) {
3010 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
3011 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
3012 } else if (IntTy && !IsWholeAlloca && !IsDest) {
3013 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
3014 Src = convertValue(DL, IRB, Src, IntTy);
3015 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
3016 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
3019 IRB.CreateAlignedLoad(SrcPtr, SrcAlign, II.isVolatile(), "copyload");
3022 if (VecTy && !IsWholeAlloca && IsDest) {
3024 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
3025 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
3026 } else if (IntTy && !IsWholeAlloca && IsDest) {
3028 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
3029 Old = convertValue(DL, IRB, Old, IntTy);
3030 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
3031 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
3032 Src = convertValue(DL, IRB, Src, NewAllocaTy);
3035 StoreInst *Store = cast<StoreInst>(
3036 IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
3038 DEBUG(dbgs() << " to: " << *Store << "\n");
3039 return !II.isVolatile();
3042 bool visitIntrinsicInst(IntrinsicInst &II) {
3043 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
3044 II.getIntrinsicID() == Intrinsic::lifetime_end);
3045 DEBUG(dbgs() << " original: " << II << "\n");
3046 assert(II.getArgOperand(1) == OldPtr);
3048 // Record this instruction for deletion.
3049 Pass.DeadInsts.insert(&II);
3052 ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
3053 NewEndOffset - NewBeginOffset);
3054 Value *Ptr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3056 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
3057 New = IRB.CreateLifetimeStart(Ptr, Size);
3059 New = IRB.CreateLifetimeEnd(Ptr, Size);
3062 DEBUG(dbgs() << " to: " << *New << "\n");
3066 bool visitPHINode(PHINode &PN) {
3067 DEBUG(dbgs() << " original: " << PN << "\n");
3068 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
3069 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
3071 // We would like to compute a new pointer in only one place, but have it be
3072 // as local as possible to the PHI. To do that, we re-use the location of
3073 // the old pointer, which necessarily must be in the right position to
3074 // dominate the PHI.
3075 IRBuilderTy PtrBuilder(IRB);
3076 if (isa<PHINode>(OldPtr))
3077 PtrBuilder.SetInsertPoint(OldPtr->getParent()->getFirstInsertionPt());
3079 PtrBuilder.SetInsertPoint(OldPtr);
3080 PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
3082 Value *NewPtr = getNewAllocaSlicePtr(PtrBuilder, OldPtr->getType());
3083 // Replace the operands which were using the old pointer.
3084 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
3086 DEBUG(dbgs() << " to: " << PN << "\n");
3087 deleteIfTriviallyDead(OldPtr);
3089 // PHIs can't be promoted on their own, but often can be speculated. We
3090 // check the speculation outside of the rewriter so that we see the
3091 // fully-rewritten alloca.
3092 PHIUsers.insert(&PN);
3096 bool visitSelectInst(SelectInst &SI) {
3097 DEBUG(dbgs() << " original: " << SI << "\n");
3098 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
3099 "Pointer isn't an operand!");
3100 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
3101 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
3103 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3104 // Replace the operands which were using the old pointer.
3105 if (SI.getOperand(1) == OldPtr)
3106 SI.setOperand(1, NewPtr);
3107 if (SI.getOperand(2) == OldPtr)
3108 SI.setOperand(2, NewPtr);
3110 DEBUG(dbgs() << " to: " << SI << "\n");
3111 deleteIfTriviallyDead(OldPtr);
3113 // Selects can't be promoted on their own, but often can be speculated. We
3114 // check the speculation outside of the rewriter so that we see the
3115 // fully-rewritten alloca.
3116 SelectUsers.insert(&SI);
3123 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
3125 /// This pass aggressively rewrites all aggregate loads and stores on
3126 /// a particular pointer (or any pointer derived from it which we can identify)
3127 /// with scalar loads and stores.
3128 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
3129 // Befriend the base class so it can delegate to private visit methods.
3130 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
3132 const DataLayout &DL;
3134 /// Queue of pointer uses to analyze and potentially rewrite.
3135 SmallVector<Use *, 8> Queue;
3137 /// Set to prevent us from cycling with phi nodes and loops.
3138 SmallPtrSet<User *, 8> Visited;
3140 /// The current pointer use being rewritten. This is used to dig up the used
3141 /// value (as opposed to the user).
3145 AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
3147 /// Rewrite loads and stores through a pointer and all pointers derived from
3149 bool rewrite(Instruction &I) {
3150 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
3152 bool Changed = false;
3153 while (!Queue.empty()) {
3154 U = Queue.pop_back_val();
3155 Changed |= visit(cast<Instruction>(U->getUser()));
3161 /// Enqueue all the users of the given instruction for further processing.
3162 /// This uses a set to de-duplicate users.
3163 void enqueueUsers(Instruction &I) {
3164 for (Use &U : I.uses())
3165 if (Visited.insert(U.getUser()).second)
3166 Queue.push_back(&U);
3169 // Conservative default is to not rewrite anything.
3170 bool visitInstruction(Instruction &I) { return false; }
3172 /// \brief Generic recursive split emission class.
3173 template <typename Derived> class OpSplitter {
3175 /// The builder used to form new instructions.
3177 /// The indices which to be used with insert- or extractvalue to select the
3178 /// appropriate value within the aggregate.
3179 SmallVector<unsigned, 4> Indices;
3180 /// The indices to a GEP instruction which will move Ptr to the correct slot
3181 /// within the aggregate.
3182 SmallVector<Value *, 4> GEPIndices;
3183 /// The base pointer of the original op, used as a base for GEPing the
3184 /// split operations.
3187 /// Initialize the splitter with an insertion point, Ptr and start with a
3188 /// single zero GEP index.
3189 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
3190 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
3193 /// \brief Generic recursive split emission routine.
3195 /// This method recursively splits an aggregate op (load or store) into
3196 /// scalar or vector ops. It splits recursively until it hits a single value
3197 /// and emits that single value operation via the template argument.
3199 /// The logic of this routine relies on GEPs and insertvalue and
3200 /// extractvalue all operating with the same fundamental index list, merely
3201 /// formatted differently (GEPs need actual values).
3203 /// \param Ty The type being split recursively into smaller ops.
3204 /// \param Agg The aggregate value being built up or stored, depending on
3205 /// whether this is splitting a load or a store respectively.
3206 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3207 if (Ty->isSingleValueType())
3208 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
3210 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3211 unsigned OldSize = Indices.size();
3213 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3215 assert(Indices.size() == OldSize && "Did not return to the old size");
3216 Indices.push_back(Idx);
3217 GEPIndices.push_back(IRB.getInt32(Idx));
3218 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3219 GEPIndices.pop_back();
3225 if (StructType *STy = dyn_cast<StructType>(Ty)) {
3226 unsigned OldSize = Indices.size();
3228 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3230 assert(Indices.size() == OldSize && "Did not return to the old size");
3231 Indices.push_back(Idx);
3232 GEPIndices.push_back(IRB.getInt32(Idx));
3233 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3234 GEPIndices.pop_back();
3240 llvm_unreachable("Only arrays and structs are aggregate loadable types");
3244 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3245 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3246 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
3248 /// Emit a leaf load of a single value. This is called at the leaves of the
3249 /// recursive emission to actually load values.
3250 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3251 assert(Ty->isSingleValueType());
3252 // Load the single value and insert it using the indices.
3253 Value *GEP = IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep");
3254 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
3255 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3256 DEBUG(dbgs() << " to: " << *Load << "\n");
3260 bool visitLoadInst(LoadInst &LI) {
3261 assert(LI.getPointerOperand() == *U);
3262 if (!LI.isSimple() || LI.getType()->isSingleValueType())
3265 // We have an aggregate being loaded, split it apart.
3266 DEBUG(dbgs() << " original: " << LI << "\n");
3267 LoadOpSplitter Splitter(&LI, *U);
3268 Value *V = UndefValue::get(LI.getType());
3269 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3270 LI.replaceAllUsesWith(V);
3271 LI.eraseFromParent();
3275 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3276 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3277 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
3279 /// Emit a leaf store of a single value. This is called at the leaves of the
3280 /// recursive emission to actually produce stores.
3281 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3282 assert(Ty->isSingleValueType());
3283 // Extract the single value and store it using the indices.
3284 Value *Store = IRB.CreateStore(
3285 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
3286 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
3288 DEBUG(dbgs() << " to: " << *Store << "\n");
3292 bool visitStoreInst(StoreInst &SI) {
3293 if (!SI.isSimple() || SI.getPointerOperand() != *U)
3295 Value *V = SI.getValueOperand();
3296 if (V->getType()->isSingleValueType())
3299 // We have an aggregate being stored, split it apart.
3300 DEBUG(dbgs() << " original: " << SI << "\n");
3301 StoreOpSplitter Splitter(&SI, *U);
3302 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3303 SI.eraseFromParent();
3307 bool visitBitCastInst(BitCastInst &BC) {
3312 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3317 bool visitPHINode(PHINode &PN) {
3322 bool visitSelectInst(SelectInst &SI) {
3329 /// \brief Strip aggregate type wrapping.
3331 /// This removes no-op aggregate types wrapping an underlying type. It will
3332 /// strip as many layers of types as it can without changing either the type
3333 /// size or the allocated size.
3334 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
3335 if (Ty->isSingleValueType())
3338 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3339 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
3342 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3343 InnerTy = ArrTy->getElementType();
3344 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3345 const StructLayout *SL = DL.getStructLayout(STy);
3346 unsigned Index = SL->getElementContainingOffset(0);
3347 InnerTy = STy->getElementType(Index);
3352 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3353 TypeSize > DL.getTypeSizeInBits(InnerTy))
3356 return stripAggregateTypeWrapping(DL, InnerTy);
3359 /// \brief Try to find a partition of the aggregate type passed in for a given
3360 /// offset and size.
3362 /// This recurses through the aggregate type and tries to compute a subtype
3363 /// based on the offset and size. When the offset and size span a sub-section
3364 /// of an array, it will even compute a new array type for that sub-section,
3365 /// and the same for structs.
3367 /// Note that this routine is very strict and tries to find a partition of the
3368 /// type which produces the *exact* right offset and size. It is not forgiving
3369 /// when the size or offset cause either end of type-based partition to be off.
3370 /// Also, this is a best-effort routine. It is reasonable to give up and not
3371 /// return a type if necessary.
3372 static Type *getTypePartition(const DataLayout &DL, Type *Ty, uint64_t Offset,
3374 if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
3375 return stripAggregateTypeWrapping(DL, Ty);
3376 if (Offset > DL.getTypeAllocSize(Ty) ||
3377 (DL.getTypeAllocSize(Ty) - Offset) < Size)
3380 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3381 // We can't partition pointers...
3382 if (SeqTy->isPointerTy())
3385 Type *ElementTy = SeqTy->getElementType();
3386 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3387 uint64_t NumSkippedElements = Offset / ElementSize;
3388 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) {
3389 if (NumSkippedElements >= ArrTy->getNumElements())
3391 } else if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) {
3392 if (NumSkippedElements >= VecTy->getNumElements())
3395 Offset -= NumSkippedElements * ElementSize;
3397 // First check if we need to recurse.
3398 if (Offset > 0 || Size < ElementSize) {
3399 // Bail if the partition ends in a different array element.
3400 if ((Offset + Size) > ElementSize)
3402 // Recurse through the element type trying to peel off offset bytes.
3403 return getTypePartition(DL, ElementTy, Offset, Size);
3405 assert(Offset == 0);
3407 if (Size == ElementSize)
3408 return stripAggregateTypeWrapping(DL, ElementTy);
3409 assert(Size > ElementSize);
3410 uint64_t NumElements = Size / ElementSize;
3411 if (NumElements * ElementSize != Size)
3413 return ArrayType::get(ElementTy, NumElements);
3416 StructType *STy = dyn_cast<StructType>(Ty);
3420 const StructLayout *SL = DL.getStructLayout(STy);
3421 if (Offset >= SL->getSizeInBytes())
3423 uint64_t EndOffset = Offset + Size;
3424 if (EndOffset > SL->getSizeInBytes())
3427 unsigned Index = SL->getElementContainingOffset(Offset);
3428 Offset -= SL->getElementOffset(Index);
3430 Type *ElementTy = STy->getElementType(Index);
3431 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3432 if (Offset >= ElementSize)
3433 return nullptr; // The offset points into alignment padding.
3435 // See if any partition must be contained by the element.
3436 if (Offset > 0 || Size < ElementSize) {
3437 if ((Offset + Size) > ElementSize)
3439 return getTypePartition(DL, ElementTy, Offset, Size);
3441 assert(Offset == 0);
3443 if (Size == ElementSize)
3444 return stripAggregateTypeWrapping(DL, ElementTy);
3446 StructType::element_iterator EI = STy->element_begin() + Index,
3447 EE = STy->element_end();
3448 if (EndOffset < SL->getSizeInBytes()) {
3449 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3450 if (Index == EndIndex)
3451 return nullptr; // Within a single element and its padding.
3453 // Don't try to form "natural" types if the elements don't line up with the
3455 // FIXME: We could potentially recurse down through the last element in the
3456 // sub-struct to find a natural end point.
3457 if (SL->getElementOffset(EndIndex) != EndOffset)
3460 assert(Index < EndIndex);
3461 EE = STy->element_begin() + EndIndex;
3464 // Try to build up a sub-structure.
3466 StructType::get(STy->getContext(), makeArrayRef(EI, EE), STy->isPacked());
3467 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3468 if (Size != SubSL->getSizeInBytes())
3469 return nullptr; // The sub-struct doesn't have quite the size needed.
3474 /// \brief Pre-split loads and stores to simplify rewriting.
3476 /// We want to break up the splittable load+store pairs as much as
3477 /// possible. This is important to do as a preprocessing step, as once we
3478 /// start rewriting the accesses to partitions of the alloca we lose the
3479 /// necessary information to correctly split apart paired loads and stores
3480 /// which both point into this alloca. The case to consider is something like
3483 /// %a = alloca [12 x i8]
3484 /// %gep1 = getelementptr [12 x i8]* %a, i32 0, i32 0
3485 /// %gep2 = getelementptr [12 x i8]* %a, i32 0, i32 4
3486 /// %gep3 = getelementptr [12 x i8]* %a, i32 0, i32 8
3487 /// %iptr1 = bitcast i8* %gep1 to i64*
3488 /// %iptr2 = bitcast i8* %gep2 to i64*
3489 /// %fptr1 = bitcast i8* %gep1 to float*
3490 /// %fptr2 = bitcast i8* %gep2 to float*
3491 /// %fptr3 = bitcast i8* %gep3 to float*
3492 /// store float 0.0, float* %fptr1
3493 /// store float 1.0, float* %fptr2
3494 /// %v = load i64* %iptr1
3495 /// store i64 %v, i64* %iptr2
3496 /// %f1 = load float* %fptr2
3497 /// %f2 = load float* %fptr3
3499 /// Here we want to form 3 partitions of the alloca, each 4 bytes large, and
3500 /// promote everything so we recover the 2 SSA values that should have been
3501 /// there all along.
3503 /// \returns true if any changes are made.
3504 bool SROA::presplitLoadsAndStores(AllocaInst &AI, AllocaSlices &AS) {
3505 DEBUG(dbgs() << "Pre-splitting loads and stores\n");
3507 // Track the loads and stores which are candidates for pre-splitting here, in
3508 // the order they first appear during the partition scan. These give stable
3509 // iteration order and a basis for tracking which loads and stores we
3511 SmallVector<LoadInst *, 4> Loads;
3512 SmallVector<StoreInst *, 4> Stores;
3514 // We need to accumulate the splits required of each load or store where we
3515 // can find them via a direct lookup. This is important to cross-check loads
3516 // and stores against each other. We also track the slice so that we can kill
3517 // all the slices that end up split.
3518 struct SplitOffsets {
3520 std::vector<uint64_t> Splits;
3522 SmallDenseMap<Instruction *, SplitOffsets, 8> SplitOffsetsMap;
3524 // Track loads out of this alloca which cannot, for any reason, be pre-split.
3525 // This is important as we also cannot pre-split stores of those loads!
3526 // FIXME: This is all pretty gross. It means that we can be more aggressive
3527 // in pre-splitting when the load feeding the store happens to come from
3528 // a separate alloca. Put another way, the effectiveness of SROA would be
3529 // decreased by a frontend which just concatenated all of its local allocas
3530 // into one big flat alloca. But defeating such patterns is exactly the job
3531 // SROA is tasked with! Sadly, to not have this discrepancy we would have
3532 // change store pre-splitting to actually force pre-splitting of the load
3533 // that feeds it *and all stores*. That makes pre-splitting much harder, but
3534 // maybe it would make it more principled?
3535 SmallPtrSet<LoadInst *, 8> UnsplittableLoads;
3537 DEBUG(dbgs() << " Searching for candidate loads and stores\n");
3538 for (auto &P : AS.partitions()) {
3539 for (Slice &S : P) {
3540 Instruction *I = cast<Instruction>(S.getUse()->getUser());
3541 if (!S.isSplittable() ||S.endOffset() <= P.endOffset()) {
3542 // If this was a load we have to track that it can't participate in any
3544 if (auto *LI = dyn_cast<LoadInst>(I))
3545 UnsplittableLoads.insert(LI);
3548 assert(P.endOffset() > S.beginOffset() &&
3549 "Empty or backwards partition!");
3551 // Determine if this is a pre-splittable slice.
3552 if (auto *LI = dyn_cast<LoadInst>(I)) {
3553 assert(!LI->isVolatile() && "Cannot split volatile loads!");
3555 // The load must be used exclusively to store into other pointers for
3556 // us to be able to arbitrarily pre-split it. The stores must also be
3557 // simple to avoid changing semantics.
3558 auto IsLoadSimplyStored = [](LoadInst *LI) {
3559 for (User *LU : LI->users()) {
3560 auto *SI = dyn_cast<StoreInst>(LU);
3561 if (!SI || !SI->isSimple())
3566 if (!IsLoadSimplyStored(LI)) {
3567 UnsplittableLoads.insert(LI);
3571 Loads.push_back(LI);
3572 } else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser())) {
3574 S.getUse() != &SI->getOperandUse(SI->getPointerOperandIndex()))
3576 auto *StoredLoad = dyn_cast<LoadInst>(SI->getValueOperand());
3577 if (!StoredLoad || !StoredLoad->isSimple())
3579 assert(!SI->isVolatile() && "Cannot split volatile stores!");
3581 Stores.push_back(SI);
3583 // Other uses cannot be pre-split.
3587 // Record the initial split.
3588 DEBUG(dbgs() << " Candidate: " << *I << "\n");
3589 auto &Offsets = SplitOffsetsMap[I];
3590 assert(Offsets.Splits.empty() &&
3591 "Should not have splits the first time we see an instruction!");
3593 Offsets.Splits.push_back(P.endOffset() - S.beginOffset());
3596 // Now scan the already split slices, and add a split for any of them which
3597 // we're going to pre-split.
3598 for (Slice *S : P.splitSliceTails()) {
3599 auto SplitOffsetsMapI =
3600 SplitOffsetsMap.find(cast<Instruction>(S->getUse()->getUser()));
3601 if (SplitOffsetsMapI == SplitOffsetsMap.end())
3603 auto &Offsets = SplitOffsetsMapI->second;
3605 assert(Offsets.S == S && "Found a mismatched slice!");
3606 assert(!Offsets.Splits.empty() &&
3607 "Cannot have an empty set of splits on the second partition!");
3608 assert(Offsets.Splits.back() ==
3609 P.beginOffset() - Offsets.S->beginOffset() &&
3610 "Previous split does not end where this one begins!");
3612 // Record each split. The last partition's end isn't needed as the size
3613 // of the slice dictates that.
3614 if (S->endOffset() > P.endOffset())
3615 Offsets.Splits.push_back(P.endOffset() - Offsets.S->beginOffset());
3619 // We may have split loads where some of their stores are split stores. For
3620 // such loads and stores, we can only pre-split them if their splits exactly
3621 // match relative to their starting offset. We have to verify this prior to
3624 std::remove_if(Stores.begin(), Stores.end(),
3625 [&UnsplittableLoads, &SplitOffsetsMap](StoreInst *SI) {
3626 // Lookup the load we are storing in our map of split
3628 auto *LI = cast<LoadInst>(SI->getValueOperand());
3629 // If it was completely unsplittable, then we're done,
3630 // and this store can't be pre-split.
3631 if (UnsplittableLoads.count(LI))
3634 auto LoadOffsetsI = SplitOffsetsMap.find(LI);
3635 if (LoadOffsetsI == SplitOffsetsMap.end())
3636 return false; // Unrelated loads are definitely safe.
3637 auto &LoadOffsets = LoadOffsetsI->second;
3639 // Now lookup the store's offsets.
3640 auto &StoreOffsets = SplitOffsetsMap[SI];
3642 // If the relative offsets of each split in the load and
3643 // store match exactly, then we can split them and we
3644 // don't need to remove them here.
3645 if (LoadOffsets.Splits == StoreOffsets.Splits)
3649 << " Mismatched splits for load and store:\n"
3650 << " " << *LI << "\n"
3651 << " " << *SI << "\n");
3653 // We've found a store and load that we need to split
3654 // with mismatched relative splits. Just give up on them
3655 // and remove both instructions from our list of
3657 UnsplittableLoads.insert(LI);
3661 // Now we have to go *back* through all te stores, because a later store may
3662 // have caused an earlier store's load to become unsplittable and if it is
3663 // unsplittable for the later store, then we can't rely on it being split in
3664 // the earlier store either.
3665 Stores.erase(std::remove_if(Stores.begin(), Stores.end(),
3666 [&UnsplittableLoads](StoreInst *SI) {
3668 cast<LoadInst>(SI->getValueOperand());
3669 return UnsplittableLoads.count(LI);
3672 // Once we've established all the loads that can't be split for some reason,
3673 // filter any that made it into our list out.
3674 Loads.erase(std::remove_if(Loads.begin(), Loads.end(),
3675 [&UnsplittableLoads](LoadInst *LI) {
3676 return UnsplittableLoads.count(LI);
3681 // If no loads or stores are left, there is no pre-splitting to be done for
3683 if (Loads.empty() && Stores.empty())
3686 // From here on, we can't fail and will be building new accesses, so rig up
3688 IRBuilderTy IRB(&AI);
3690 // Collect the new slices which we will merge into the alloca slices.
3691 SmallVector<Slice, 4> NewSlices;
3693 // Track any allocas we end up splitting loads and stores for so we iterate
3695 SmallPtrSet<AllocaInst *, 4> ResplitPromotableAllocas;
3697 // At this point, we have collected all of the loads and stores we can
3698 // pre-split, and the specific splits needed for them. We actually do the
3699 // splitting in a specific order in order to handle when one of the loads in
3700 // the value operand to one of the stores.
3702 // First, we rewrite all of the split loads, and just accumulate each split
3703 // load in a parallel structure. We also build the slices for them and append
3704 // them to the alloca slices.
3705 SmallDenseMap<LoadInst *, std::vector<LoadInst *>, 1> SplitLoadsMap;
3706 std::vector<LoadInst *> SplitLoads;
3707 const DataLayout &DL = AI.getModule()->getDataLayout();
3708 for (LoadInst *LI : Loads) {
3711 IntegerType *Ty = cast<IntegerType>(LI->getType());
3712 uint64_t LoadSize = Ty->getBitWidth() / 8;
3713 assert(LoadSize > 0 && "Cannot have a zero-sized integer load!");
3715 auto &Offsets = SplitOffsetsMap[LI];
3716 assert(LoadSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3717 "Slice size should always match load size exactly!");
3718 uint64_t BaseOffset = Offsets.S->beginOffset();
3719 assert(BaseOffset + LoadSize > BaseOffset &&
3720 "Cannot represent alloca access size using 64-bit integers!");
3722 Instruction *BasePtr = cast<Instruction>(LI->getPointerOperand());
3723 IRB.SetInsertPoint(BasicBlock::iterator(LI));
3725 DEBUG(dbgs() << " Splitting load: " << *LI << "\n");
3727 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3728 int Idx = 0, Size = Offsets.Splits.size();
3730 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3731 auto *PartPtrTy = PartTy->getPointerTo(LI->getPointerAddressSpace());
3732 LoadInst *PLoad = IRB.CreateAlignedLoad(
3733 getAdjustedPtr(IRB, DL, BasePtr,
3734 APInt(DL.getPointerSizeInBits(), PartOffset),
3735 PartPtrTy, BasePtr->getName() + "."),
3736 getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3739 // Append this load onto the list of split loads so we can find it later
3740 // to rewrite the stores.
3741 SplitLoads.push_back(PLoad);
3743 // Now build a new slice for the alloca.
3744 NewSlices.push_back(
3745 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3746 &PLoad->getOperandUse(PLoad->getPointerOperandIndex()),
3747 /*IsSplittable*/ false));
3748 DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3749 << ", " << NewSlices.back().endOffset() << "): " << *PLoad
3752 // See if we've handled all the splits.
3756 // Setup the next partition.
3757 PartOffset = Offsets.Splits[Idx];
3759 PartSize = (Idx < Size ? Offsets.Splits[Idx] : LoadSize) - PartOffset;
3762 // Now that we have the split loads, do the slow walk over all uses of the
3763 // load and rewrite them as split stores, or save the split loads to use
3764 // below if the store is going to be split there anyways.
3765 bool DeferredStores = false;
3766 for (User *LU : LI->users()) {
3767 StoreInst *SI = cast<StoreInst>(LU);
3768 if (!Stores.empty() && SplitOffsetsMap.count(SI)) {
3769 DeferredStores = true;
3770 DEBUG(dbgs() << " Deferred splitting of store: " << *SI << "\n");
3774 Value *StoreBasePtr = SI->getPointerOperand();
3775 IRB.SetInsertPoint(BasicBlock::iterator(SI));
3777 DEBUG(dbgs() << " Splitting store of load: " << *SI << "\n");
3779 for (int Idx = 0, Size = SplitLoads.size(); Idx < Size; ++Idx) {
3780 LoadInst *PLoad = SplitLoads[Idx];
3781 uint64_t PartOffset = Idx == 0 ? 0 : Offsets.Splits[Idx - 1];
3783 PLoad->getType()->getPointerTo(SI->getPointerAddressSpace());
3785 StoreInst *PStore = IRB.CreateAlignedStore(
3786 PLoad, getAdjustedPtr(IRB, DL, StoreBasePtr,
3787 APInt(DL.getPointerSizeInBits(), PartOffset),
3788 PartPtrTy, StoreBasePtr->getName() + "."),
3789 getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3791 DEBUG(dbgs() << " +" << PartOffset << ":" << *PStore << "\n");
3794 // We want to immediately iterate on any allocas impacted by splitting
3795 // this store, and we have to track any promotable alloca (indicated by
3796 // a direct store) as needing to be resplit because it is no longer
3798 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(StoreBasePtr)) {
3799 ResplitPromotableAllocas.insert(OtherAI);
3800 Worklist.insert(OtherAI);
3801 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3802 StoreBasePtr->stripInBoundsOffsets())) {
3803 Worklist.insert(OtherAI);
3806 // Mark the original store as dead.
3807 DeadInsts.insert(SI);
3810 // Save the split loads if there are deferred stores among the users.
3812 SplitLoadsMap.insert(std::make_pair(LI, std::move(SplitLoads)));
3814 // Mark the original load as dead and kill the original slice.
3815 DeadInsts.insert(LI);
3819 // Second, we rewrite all of the split stores. At this point, we know that
3820 // all loads from this alloca have been split already. For stores of such
3821 // loads, we can simply look up the pre-existing split loads. For stores of
3822 // other loads, we split those loads first and then write split stores of
3824 for (StoreInst *SI : Stores) {
3825 auto *LI = cast<LoadInst>(SI->getValueOperand());
3826 IntegerType *Ty = cast<IntegerType>(LI->getType());
3827 uint64_t StoreSize = Ty->getBitWidth() / 8;
3828 assert(StoreSize > 0 && "Cannot have a zero-sized integer store!");
3830 auto &Offsets = SplitOffsetsMap[SI];
3831 assert(StoreSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3832 "Slice size should always match load size exactly!");
3833 uint64_t BaseOffset = Offsets.S->beginOffset();
3834 assert(BaseOffset + StoreSize > BaseOffset &&
3835 "Cannot represent alloca access size using 64-bit integers!");
3837 Value *LoadBasePtr = LI->getPointerOperand();
3838 Instruction *StoreBasePtr = cast<Instruction>(SI->getPointerOperand());
3840 DEBUG(dbgs() << " Splitting store: " << *SI << "\n");
3842 // Check whether we have an already split load.
3843 auto SplitLoadsMapI = SplitLoadsMap.find(LI);
3844 std::vector<LoadInst *> *SplitLoads = nullptr;
3845 if (SplitLoadsMapI != SplitLoadsMap.end()) {
3846 SplitLoads = &SplitLoadsMapI->second;
3847 assert(SplitLoads->size() == Offsets.Splits.size() + 1 &&
3848 "Too few split loads for the number of splits in the store!");
3850 DEBUG(dbgs() << " of load: " << *LI << "\n");
3853 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3854 int Idx = 0, Size = Offsets.Splits.size();
3856 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3857 auto *PartPtrTy = PartTy->getPointerTo(SI->getPointerAddressSpace());
3859 // Either lookup a split load or create one.
3862 PLoad = (*SplitLoads)[Idx];
3864 IRB.SetInsertPoint(BasicBlock::iterator(LI));
3865 PLoad = IRB.CreateAlignedLoad(
3866 getAdjustedPtr(IRB, DL, LoadBasePtr,
3867 APInt(DL.getPointerSizeInBits(), PartOffset),
3868 PartPtrTy, LoadBasePtr->getName() + "."),
3869 getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3873 // And store this partition.
3874 IRB.SetInsertPoint(BasicBlock::iterator(SI));
3875 StoreInst *PStore = IRB.CreateAlignedStore(
3876 PLoad, getAdjustedPtr(IRB, DL, StoreBasePtr,
3877 APInt(DL.getPointerSizeInBits(), PartOffset),
3878 PartPtrTy, StoreBasePtr->getName() + "."),
3879 getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3881 // Now build a new slice for the alloca.
3882 NewSlices.push_back(
3883 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3884 &PStore->getOperandUse(PStore->getPointerOperandIndex()),
3885 /*IsSplittable*/ false));
3886 DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3887 << ", " << NewSlices.back().endOffset() << "): " << *PStore
3890 DEBUG(dbgs() << " of split load: " << *PLoad << "\n");
3893 // See if we've finished all the splits.
3897 // Setup the next partition.
3898 PartOffset = Offsets.Splits[Idx];
3900 PartSize = (Idx < Size ? Offsets.Splits[Idx] : StoreSize) - PartOffset;
3903 // We want to immediately iterate on any allocas impacted by splitting
3904 // this load, which is only relevant if it isn't a load of this alloca and
3905 // thus we didn't already split the loads above. We also have to keep track
3906 // of any promotable allocas we split loads on as they can no longer be
3909 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(LoadBasePtr)) {
3910 assert(OtherAI != &AI && "We can't re-split our own alloca!");
3911 ResplitPromotableAllocas.insert(OtherAI);
3912 Worklist.insert(OtherAI);
3913 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3914 LoadBasePtr->stripInBoundsOffsets())) {
3915 assert(OtherAI != &AI && "We can't re-split our own alloca!");
3916 Worklist.insert(OtherAI);
3920 // Mark the original store as dead now that we've split it up and kill its
3921 // slice. Note that we leave the original load in place unless this store
3922 // was its ownly use. It may in turn be split up if it is an alloca load
3923 // for some other alloca, but it may be a normal load. This may introduce
3924 // redundant loads, but where those can be merged the rest of the optimizer
3925 // should handle the merging, and this uncovers SSA splits which is more
3926 // important. In practice, the original loads will almost always be fully
3927 // split and removed eventually, and the splits will be merged by any
3928 // trivial CSE, including instcombine.
3929 if (LI->hasOneUse()) {
3930 assert(*LI->user_begin() == SI && "Single use isn't this store!");
3931 DeadInsts.insert(LI);
3933 DeadInsts.insert(SI);
3937 // Remove the killed slices that have ben pre-split.
3938 AS.erase(std::remove_if(AS.begin(), AS.end(), [](const Slice &S) {
3942 // Insert our new slices. This will sort and merge them into the sorted
3944 AS.insert(NewSlices);
3946 DEBUG(dbgs() << " Pre-split slices:\n");
3948 for (auto I = AS.begin(), E = AS.end(); I != E; ++I)
3949 DEBUG(AS.print(dbgs(), I, " "));
3952 // Finally, don't try to promote any allocas that new require re-splitting.
3953 // They have already been added to the worklist above.
3954 PromotableAllocas.erase(
3956 PromotableAllocas.begin(), PromotableAllocas.end(),
3957 [&](AllocaInst *AI) { return ResplitPromotableAllocas.count(AI); }),
3958 PromotableAllocas.end());
3963 /// \brief Rewrite an alloca partition's users.
3965 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3966 /// to rewrite uses of an alloca partition to be conducive for SSA value
3967 /// promotion. If the partition needs a new, more refined alloca, this will
3968 /// build that new alloca, preserving as much type information as possible, and
3969 /// rewrite the uses of the old alloca to point at the new one and have the
3970 /// appropriate new offsets. It also evaluates how successful the rewrite was
3971 /// at enabling promotion and if it was successful queues the alloca to be
3973 AllocaInst *SROA::rewritePartition(AllocaInst &AI, AllocaSlices &AS,
3974 AllocaSlices::Partition &P) {
3975 // Try to compute a friendly type for this partition of the alloca. This
3976 // won't always succeed, in which case we fall back to a legal integer type
3977 // or an i8 array of an appropriate size.
3978 Type *SliceTy = nullptr;
3979 const DataLayout &DL = AI.getModule()->getDataLayout();
3980 if (Type *CommonUseTy = findCommonType(P.begin(), P.end(), P.endOffset()))
3981 if (DL.getTypeAllocSize(CommonUseTy) >= P.size())
3982 SliceTy = CommonUseTy;
3984 if (Type *TypePartitionTy = getTypePartition(DL, AI.getAllocatedType(),
3985 P.beginOffset(), P.size()))
3986 SliceTy = TypePartitionTy;
3987 if ((!SliceTy || (SliceTy->isArrayTy() &&
3988 SliceTy->getArrayElementType()->isIntegerTy())) &&
3989 DL.isLegalInteger(P.size() * 8))
3990 SliceTy = Type::getIntNTy(*C, P.size() * 8);
3992 SliceTy = ArrayType::get(Type::getInt8Ty(*C), P.size());
3993 assert(DL.getTypeAllocSize(SliceTy) >= P.size());
3995 bool IsIntegerPromotable = isIntegerWideningViable(P, SliceTy, DL);
3998 IsIntegerPromotable ? nullptr : isVectorPromotionViable(P, DL);
4002 // Check for the case where we're going to rewrite to a new alloca of the
4003 // exact same type as the original, and with the same access offsets. In that
4004 // case, re-use the existing alloca, but still run through the rewriter to
4005 // perform phi and select speculation.
4007 if (SliceTy == AI.getAllocatedType()) {
4008 assert(P.beginOffset() == 0 &&
4009 "Non-zero begin offset but same alloca type");
4011 // FIXME: We should be able to bail at this point with "nothing changed".
4012 // FIXME: We might want to defer PHI speculation until after here.
4013 // FIXME: return nullptr;
4015 unsigned Alignment = AI.getAlignment();
4017 // The minimum alignment which users can rely on when the explicit
4018 // alignment is omitted or zero is that required by the ABI for this
4020 Alignment = DL.getABITypeAlignment(AI.getAllocatedType());
4022 Alignment = MinAlign(Alignment, P.beginOffset());
4023 // If we will get at least this much alignment from the type alone, leave
4024 // the alloca's alignment unconstrained.
4025 if (Alignment <= DL.getABITypeAlignment(SliceTy))
4027 NewAI = new AllocaInst(
4028 SliceTy, nullptr, Alignment,
4029 AI.getName() + ".sroa." + Twine(P.begin() - AS.begin()), &AI);
4033 DEBUG(dbgs() << "Rewriting alloca partition "
4034 << "[" << P.beginOffset() << "," << P.endOffset()
4035 << ") to: " << *NewAI << "\n");
4037 // Track the high watermark on the worklist as it is only relevant for
4038 // promoted allocas. We will reset it to this point if the alloca is not in
4039 // fact scheduled for promotion.
4040 unsigned PPWOldSize = PostPromotionWorklist.size();
4041 unsigned NumUses = 0;
4042 SmallPtrSet<PHINode *, 8> PHIUsers;
4043 SmallPtrSet<SelectInst *, 8> SelectUsers;
4045 AllocaSliceRewriter Rewriter(DL, AS, *this, AI, *NewAI, P.beginOffset(),
4046 P.endOffset(), IsIntegerPromotable, VecTy,
4047 PHIUsers, SelectUsers);
4048 bool Promotable = true;
4049 for (Slice *S : P.splitSliceTails()) {
4050 Promotable &= Rewriter.visit(S);
4053 for (Slice &S : P) {
4054 Promotable &= Rewriter.visit(&S);
4058 NumAllocaPartitionUses += NumUses;
4059 MaxUsesPerAllocaPartition =
4060 std::max<unsigned>(NumUses, MaxUsesPerAllocaPartition);
4062 // Now that we've processed all the slices in the new partition, check if any
4063 // PHIs or Selects would block promotion.
4064 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
4067 if (!isSafePHIToSpeculate(**I)) {
4070 SelectUsers.clear();
4073 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
4074 E = SelectUsers.end();
4076 if (!isSafeSelectToSpeculate(**I)) {
4079 SelectUsers.clear();
4084 if (PHIUsers.empty() && SelectUsers.empty()) {
4085 // Promote the alloca.
4086 PromotableAllocas.push_back(NewAI);
4088 // If we have either PHIs or Selects to speculate, add them to those
4089 // worklists and re-queue the new alloca so that we promote in on the
4091 for (PHINode *PHIUser : PHIUsers)
4092 SpeculatablePHIs.insert(PHIUser);
4093 for (SelectInst *SelectUser : SelectUsers)
4094 SpeculatableSelects.insert(SelectUser);
4095 Worklist.insert(NewAI);
4098 // If we can't promote the alloca, iterate on it to check for new
4099 // refinements exposed by splitting the current alloca. Don't iterate on an
4100 // alloca which didn't actually change and didn't get promoted.
4102 Worklist.insert(NewAI);
4104 // Drop any post-promotion work items if promotion didn't happen.
4105 while (PostPromotionWorklist.size() > PPWOldSize)
4106 PostPromotionWorklist.pop_back();
4112 /// \brief Walks the slices of an alloca and form partitions based on them,
4113 /// rewriting each of their uses.
4114 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &AS) {
4115 if (AS.begin() == AS.end())
4118 unsigned NumPartitions = 0;
4119 bool Changed = false;
4120 const DataLayout &DL = AI.getModule()->getDataLayout();
4122 // First try to pre-split loads and stores.
4123 Changed |= presplitLoadsAndStores(AI, AS);
4125 // Now that we have identified any pre-splitting opportunities, mark any
4126 // splittable (non-whole-alloca) loads and stores as unsplittable. If we fail
4127 // to split these during pre-splitting, we want to force them to be
4128 // rewritten into a partition.
4129 bool IsSorted = true;
4130 for (Slice &S : AS) {
4131 if (!S.isSplittable())
4133 // FIXME: We currently leave whole-alloca splittable loads and stores. This
4134 // used to be the only splittable loads and stores and we need to be
4135 // confident that the above handling of splittable loads and stores is
4136 // completely sufficient before we forcibly disable the remaining handling.
4137 if (S.beginOffset() == 0 &&
4138 S.endOffset() >= DL.getTypeAllocSize(AI.getAllocatedType()))
4140 if (isa<LoadInst>(S.getUse()->getUser()) ||
4141 isa<StoreInst>(S.getUse()->getUser())) {
4142 S.makeUnsplittable();
4147 std::sort(AS.begin(), AS.end());
4149 /// \brief Describes the allocas introduced by rewritePartition
4150 /// in order to migrate the debug info.
4155 Piece(AllocaInst *AI, uint64_t O, uint64_t S)
4156 : Alloca(AI), Offset(O), Size(S) {}
4158 SmallVector<Piece, 4> Pieces;
4160 // Rewrite each partition.
4161 for (auto &P : AS.partitions()) {
4162 if (AllocaInst *NewAI = rewritePartition(AI, AS, P)) {
4165 uint64_t SizeOfByte = 8;
4166 uint64_t AllocaSize = DL.getTypeSizeInBits(NewAI->getAllocatedType());
4167 // Don't include any padding.
4168 uint64_t Size = std::min(AllocaSize, P.size() * SizeOfByte);
4169 Pieces.push_back(Piece(NewAI, P.beginOffset() * SizeOfByte, Size));
4175 NumAllocaPartitions += NumPartitions;
4176 MaxPartitionsPerAlloca =
4177 std::max<unsigned>(NumPartitions, MaxPartitionsPerAlloca);
4179 // Migrate debug information from the old alloca to the new alloca(s)
4180 // and the individial partitions.
4181 if (DbgDeclareInst *DbgDecl = FindAllocaDbgDeclare(&AI)) {
4182 DIVariable Var(DbgDecl->getVariable());
4183 DIExpression Expr(DbgDecl->getExpression());
4184 DIBuilder DIB(*AI.getParent()->getParent()->getParent(),
4185 /*AllowUnresolved*/ false);
4186 bool IsSplit = Pieces.size() > 1;
4187 for (auto Piece : Pieces) {
4188 // Create a piece expression describing the new partition or reuse AI's
4189 // expression if there is only one partition.
4190 DIExpression PieceExpr = Expr;
4191 if (IsSplit || Expr.isBitPiece()) {
4192 // If this alloca is already a scalar replacement of a larger aggregate,
4193 // Piece.Offset describes the offset inside the scalar.
4194 uint64_t Offset = Expr.isBitPiece() ? Expr.getBitPieceOffset() : 0;
4195 uint64_t Start = Offset + Piece.Offset;
4196 uint64_t Size = Piece.Size;
4197 if (Expr.isBitPiece()) {
4198 uint64_t AbsEnd = Expr.getBitPieceOffset() + Expr.getBitPieceSize();
4199 if (Start >= AbsEnd)
4200 // No need to describe a SROAed padding.
4202 Size = std::min(Size, AbsEnd - Start);
4204 PieceExpr = DIB.createBitPieceExpression(Start, Size);
4207 // Remove any existing dbg.declare intrinsic describing the same alloca.
4208 if (DbgDeclareInst *OldDDI = FindAllocaDbgDeclare(Piece.Alloca))
4209 OldDDI->eraseFromParent();
4211 auto *NewDDI = DIB.insertDeclare(Piece.Alloca, Var, PieceExpr, &AI);
4212 NewDDI->setDebugLoc(DbgDecl->getDebugLoc());
4218 /// \brief Clobber a use with undef, deleting the used value if it becomes dead.
4219 void SROA::clobberUse(Use &U) {
4221 // Replace the use with an undef value.
4222 U = UndefValue::get(OldV->getType());
4224 // Check for this making an instruction dead. We have to garbage collect
4225 // all the dead instructions to ensure the uses of any alloca end up being
4227 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
4228 if (isInstructionTriviallyDead(OldI)) {
4229 DeadInsts.insert(OldI);
4233 /// \brief Analyze an alloca for SROA.
4235 /// This analyzes the alloca to ensure we can reason about it, builds
4236 /// the slices of the alloca, and then hands it off to be split and
4237 /// rewritten as needed.
4238 bool SROA::runOnAlloca(AllocaInst &AI) {
4239 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
4240 ++NumAllocasAnalyzed;
4242 // Special case dead allocas, as they're trivial.
4243 if (AI.use_empty()) {
4244 AI.eraseFromParent();
4247 const DataLayout &DL = AI.getModule()->getDataLayout();
4249 // Skip alloca forms that this analysis can't handle.
4250 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
4251 DL.getTypeAllocSize(AI.getAllocatedType()) == 0)
4254 bool Changed = false;
4256 // First, split any FCA loads and stores touching this alloca to promote
4257 // better splitting and promotion opportunities.
4258 AggLoadStoreRewriter AggRewriter(DL);
4259 Changed |= AggRewriter.rewrite(AI);
4261 // Build the slices using a recursive instruction-visiting builder.
4262 AllocaSlices AS(DL, AI);
4263 DEBUG(AS.print(dbgs()));
4267 // Delete all the dead users of this alloca before splitting and rewriting it.
4268 for (Instruction *DeadUser : AS.getDeadUsers()) {
4269 // Free up everything used by this instruction.
4270 for (Use &DeadOp : DeadUser->operands())
4273 // Now replace the uses of this instruction.
4274 DeadUser->replaceAllUsesWith(UndefValue::get(DeadUser->getType()));
4276 // And mark it for deletion.
4277 DeadInsts.insert(DeadUser);
4280 for (Use *DeadOp : AS.getDeadOperands()) {
4281 clobberUse(*DeadOp);
4285 // No slices to split. Leave the dead alloca for a later pass to clean up.
4286 if (AS.begin() == AS.end())
4289 Changed |= splitAlloca(AI, AS);
4291 DEBUG(dbgs() << " Speculating PHIs\n");
4292 while (!SpeculatablePHIs.empty())
4293 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
4295 DEBUG(dbgs() << " Speculating Selects\n");
4296 while (!SpeculatableSelects.empty())
4297 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
4302 /// \brief Delete the dead instructions accumulated in this run.
4304 /// Recursively deletes the dead instructions we've accumulated. This is done
4305 /// at the very end to maximize locality of the recursive delete and to
4306 /// minimize the problems of invalidated instruction pointers as such pointers
4307 /// are used heavily in the intermediate stages of the algorithm.
4309 /// We also record the alloca instructions deleted here so that they aren't
4310 /// subsequently handed to mem2reg to promote.
4311 void SROA::deleteDeadInstructions(
4312 SmallPtrSetImpl<AllocaInst *> &DeletedAllocas) {
4313 while (!DeadInsts.empty()) {
4314 Instruction *I = DeadInsts.pop_back_val();
4315 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
4317 I->replaceAllUsesWith(UndefValue::get(I->getType()));
4319 for (Use &Operand : I->operands())
4320 if (Instruction *U = dyn_cast<Instruction>(Operand)) {
4321 // Zero out the operand and see if it becomes trivially dead.
4323 if (isInstructionTriviallyDead(U))
4324 DeadInsts.insert(U);
4327 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
4328 DeletedAllocas.insert(AI);
4329 if (DbgDeclareInst *DbgDecl = FindAllocaDbgDeclare(AI))
4330 DbgDecl->eraseFromParent();
4334 I->eraseFromParent();
4338 static void enqueueUsersInWorklist(Instruction &I,
4339 SmallVectorImpl<Instruction *> &Worklist,
4340 SmallPtrSetImpl<Instruction *> &Visited) {
4341 for (User *U : I.users())
4342 if (Visited.insert(cast<Instruction>(U)).second)
4343 Worklist.push_back(cast<Instruction>(U));
4346 /// \brief Promote the allocas, using the best available technique.
4348 /// This attempts to promote whatever allocas have been identified as viable in
4349 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
4350 /// If there is a domtree available, we attempt to promote using the full power
4351 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
4352 /// based on the SSAUpdater utilities. This function returns whether any
4353 /// promotion occurred.
4354 bool SROA::promoteAllocas(Function &F) {
4355 if (PromotableAllocas.empty())
4358 NumPromoted += PromotableAllocas.size();
4360 if (DT && !ForceSSAUpdater) {
4361 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
4362 PromoteMemToReg(PromotableAllocas, *DT, nullptr, AC);
4363 PromotableAllocas.clear();
4367 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
4369 DIBuilder DIB(*F.getParent(), /*AllowUnresolved*/ false);
4370 SmallVector<Instruction *, 64> Insts;
4372 // We need a worklist to walk the uses of each alloca.
4373 SmallVector<Instruction *, 8> Worklist;
4374 SmallPtrSet<Instruction *, 8> Visited;
4375 SmallVector<Instruction *, 32> DeadInsts;
4377 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
4378 AllocaInst *AI = PromotableAllocas[Idx];
4383 enqueueUsersInWorklist(*AI, Worklist, Visited);
4385 while (!Worklist.empty()) {
4386 Instruction *I = Worklist.pop_back_val();
4388 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
4389 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
4390 // leading to them) here. Eventually it should use them to optimize the
4391 // scalar values produced.
4392 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
4393 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
4394 II->getIntrinsicID() == Intrinsic::lifetime_end);
4395 II->eraseFromParent();
4399 // Push the loads and stores we find onto the list. SROA will already
4400 // have validated that all loads and stores are viable candidates for
4402 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4403 assert(LI->getType() == AI->getAllocatedType());
4404 Insts.push_back(LI);
4407 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
4408 assert(SI->getValueOperand()->getType() == AI->getAllocatedType());
4409 Insts.push_back(SI);
4413 // For everything else, we know that only no-op bitcasts and GEPs will
4414 // make it this far, just recurse through them and recall them for later
4416 DeadInsts.push_back(I);
4417 enqueueUsersInWorklist(*I, Worklist, Visited);
4419 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
4420 while (!DeadInsts.empty())
4421 DeadInsts.pop_back_val()->eraseFromParent();
4422 AI->eraseFromParent();
4425 PromotableAllocas.clear();
4429 bool SROA::runOnFunction(Function &F) {
4430 if (skipOptnoneFunction(F))
4433 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
4434 C = &F.getContext();
4435 DominatorTreeWrapperPass *DTWP =
4436 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
4437 DT = DTWP ? &DTWP->getDomTree() : nullptr;
4438 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
4440 BasicBlock &EntryBB = F.getEntryBlock();
4441 for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
4443 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
4444 Worklist.insert(AI);
4447 bool Changed = false;
4448 // A set of deleted alloca instruction pointers which should be removed from
4449 // the list of promotable allocas.
4450 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
4453 while (!Worklist.empty()) {
4454 Changed |= runOnAlloca(*Worklist.pop_back_val());
4455 deleteDeadInstructions(DeletedAllocas);
4457 // Remove the deleted allocas from various lists so that we don't try to
4458 // continue processing them.
4459 if (!DeletedAllocas.empty()) {
4460 auto IsInSet = [&](AllocaInst *AI) { return DeletedAllocas.count(AI); };
4461 Worklist.remove_if(IsInSet);
4462 PostPromotionWorklist.remove_if(IsInSet);
4463 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
4464 PromotableAllocas.end(),
4466 PromotableAllocas.end());
4467 DeletedAllocas.clear();
4471 Changed |= promoteAllocas(F);
4473 Worklist = PostPromotionWorklist;
4474 PostPromotionWorklist.clear();
4475 } while (!Worklist.empty());
4480 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
4481 AU.addRequired<AssumptionCacheTracker>();
4482 if (RequiresDomTree)
4483 AU.addRequired<DominatorTreeWrapperPass>();
4484 AU.setPreservesCFG();