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/AssumptionTracker.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 // Forward declare an iterator to befriend it.
241 class partition_iterator;
243 /// \brief A partition of the slices.
245 /// An ephemeral representation for a range of slices which can be viewed as
246 /// a partition of the alloca. This range represents a span of the alloca's
247 /// memory which cannot be split, and provides access to all of the slices
248 /// overlapping some part of the partition.
250 /// Objects of this type are produced by traversing the alloca's slices, but
251 /// are only ephemeral and not persistent.
254 friend class AllocaSlices;
255 friend class AllocaSlices::partition_iterator;
257 /// \brief The begining and ending offsets of the alloca for this partition.
258 uint64_t BeginOffset, EndOffset;
260 /// \brief The start end end iterators of this partition.
263 /// \brief A collection of split slice tails overlapping the partition.
264 SmallVector<Slice *, 4> SplitTails;
266 /// \brief Raw constructor builds an empty partition starting and ending at
267 /// the given iterator.
268 Partition(iterator SI) : SI(SI), SJ(SI) {}
271 /// \brief The start offset of this partition.
273 /// All of the contained slices start at or after this offset.
274 uint64_t beginOffset() const { return BeginOffset; }
276 /// \brief The end offset of this partition.
278 /// All of the contained slices end at or before this offset.
279 uint64_t endOffset() const { return EndOffset; }
281 /// \brief The size of the partition.
283 /// Note that this can never be zero.
284 uint64_t size() const {
285 assert(BeginOffset < EndOffset && "Partitions must span some bytes!");
286 return EndOffset - BeginOffset;
289 /// \brief Test whether this partition contains no slices, and merely spans
290 /// a region occupied by split slices.
291 bool empty() const { return SI == SJ; }
293 /// \name Iterate slices that start within the partition.
294 /// These may be splittable or unsplittable. They have a begin offset >= the
295 /// partition begin offset.
297 // FIXME: We should probably define a "concat_iterator" helper and use that
298 // to stitch together pointee_iterators over the split tails and the
299 // contiguous iterators of the partition. That would give a much nicer
300 // interface here. We could then additionally expose filtered iterators for
301 // split, unsplit, and unsplittable splices based on the usage patterns.
302 iterator begin() const { return SI; }
303 iterator end() const { return SJ; }
306 /// \brief Get the sequence of split slice tails.
308 /// These tails are of slices which start before this partition but are
309 /// split and overlap into the partition. We accumulate these while forming
311 ArrayRef<Slice *> splitSliceTails() const { return SplitTails; }
314 /// \brief An iterator over partitions of the alloca's slices.
316 /// This iterator implements the core algorithm for partitioning the alloca's
317 /// slices. It is a forward iterator as we don't support backtracking for
318 /// efficiency reasons, and re-use a single storage area to maintain the
319 /// current set of split slices.
321 /// It is templated on the slice iterator type to use so that it can operate
322 /// with either const or non-const slice iterators.
323 class partition_iterator
324 : public iterator_facade_base<partition_iterator,
325 std::forward_iterator_tag, Partition> {
326 friend class AllocaSlices;
328 /// \brief Most of the state for walking the partitions is held in a class
329 /// with a nice interface for examining them.
332 /// \brief We need to keep the end of the slices to know when to stop.
333 AllocaSlices::iterator SE;
335 /// \brief We also need to keep track of the maximum split end offset seen.
336 /// FIXME: Do we really?
337 uint64_t MaxSplitSliceEndOffset;
339 /// \brief Sets the partition to be empty at given iterator, and sets the
341 partition_iterator(AllocaSlices::iterator SI, AllocaSlices::iterator SE)
342 : P(SI), SE(SE), MaxSplitSliceEndOffset(0) {
343 // If not already at the end, advance our state to form the initial
349 /// \brief Advance the iterator to the next partition.
351 /// Requires that the iterator not be at the end of the slices.
353 assert((P.SI != SE || !P.SplitTails.empty()) &&
354 "Cannot advance past the end of the slices!");
356 // Clear out any split uses which have ended.
357 if (!P.SplitTails.empty()) {
358 if (P.EndOffset >= MaxSplitSliceEndOffset) {
359 // If we've finished all splits, this is easy.
360 P.SplitTails.clear();
361 MaxSplitSliceEndOffset = 0;
363 // Remove the uses which have ended in the prior partition. This
364 // cannot change the max split slice end because we just checked that
365 // the prior partition ended prior to that max.
368 P.SplitTails.begin(), P.SplitTails.end(),
369 [&](Slice *S) { return S->endOffset() <= P.EndOffset; }),
371 assert(std::any_of(P.SplitTails.begin(), P.SplitTails.end(),
373 return S->endOffset() == MaxSplitSliceEndOffset;
375 "Could not find the current max split slice offset!");
376 assert(std::all_of(P.SplitTails.begin(), P.SplitTails.end(),
378 return S->endOffset() <= MaxSplitSliceEndOffset;
380 "Max split slice end offset is not actually the max!");
384 // If P.SI is already at the end, then we've cleared the split tail and
385 // now have an end iterator.
387 assert(P.SplitTails.empty() && "Failed to clear the split slices!");
391 // If we had a non-empty partition previously, set up the state for
392 // subsequent partitions.
394 // Accumulate all the splittable slices which started in the old
395 // partition into the split list.
397 if (S.isSplittable() && S.endOffset() > P.EndOffset) {
398 P.SplitTails.push_back(&S);
399 MaxSplitSliceEndOffset =
400 std::max(S.endOffset(), MaxSplitSliceEndOffset);
403 // Start from the end of the previous partition.
406 // If P.SI is now at the end, we at most have a tail of split slices.
408 P.BeginOffset = P.EndOffset;
409 P.EndOffset = MaxSplitSliceEndOffset;
413 // If the we have split slices and the next slice is after a gap and is
414 // not splittable immediately form an empty partition for the split
415 // slices up until the next slice begins.
416 if (!P.SplitTails.empty() && P.SI->beginOffset() != P.EndOffset &&
417 !P.SI->isSplittable()) {
418 P.BeginOffset = P.EndOffset;
419 P.EndOffset = P.SI->beginOffset();
424 // OK, we need to consume new slices. Set the end offset based on the
425 // current slice, and step SJ past it. The beginning offset of the
426 // parttion is the beginning offset of the next slice unless we have
427 // pre-existing split slices that are continuing, in which case we begin
428 // at the prior end offset.
429 P.BeginOffset = P.SplitTails.empty() ? P.SI->beginOffset() : P.EndOffset;
430 P.EndOffset = P.SI->endOffset();
433 // There are two strategies to form a partition based on whether the
434 // partition starts with an unsplittable slice or a splittable slice.
435 if (!P.SI->isSplittable()) {
436 // When we're forming an unsplittable region, it must always start at
437 // the first slice and will extend through its end.
438 assert(P.BeginOffset == P.SI->beginOffset());
440 // Form a partition including all of the overlapping slices with this
441 // unsplittable slice.
442 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
443 if (!P.SJ->isSplittable())
444 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
448 // We have a partition across a set of overlapping unsplittable
453 // If we're starting with a splittable slice, then we need to form
454 // a synthetic partition spanning it and any other overlapping splittable
456 assert(P.SI->isSplittable() && "Forming a splittable partition!");
458 // Collect all of the overlapping splittable slices.
459 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset &&
460 P.SJ->isSplittable()) {
461 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
465 // Back upiP.EndOffset if we ended the span early when encountering an
466 // unsplittable slice. This synthesizes the early end offset of
467 // a partition spanning only splittable slices.
468 if (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
469 assert(!P.SJ->isSplittable());
470 P.EndOffset = P.SJ->beginOffset();
475 bool operator==(const partition_iterator &RHS) const {
476 assert(SE == RHS.SE &&
477 "End iterators don't match between compared partition iterators!");
479 // The observed positions of partitions is marked by the P.SI iterator and
480 // the emptyness of the split slices. The latter is only relevant when
481 // P.SI == SE, as the end iterator will additionally have an empty split
482 // slices list, but the prior may have the same P.SI and a tail of split
484 if (P.SI == RHS.P.SI &&
485 P.SplitTails.empty() == RHS.P.SplitTails.empty()) {
486 assert(P.SJ == RHS.P.SJ &&
487 "Same set of slices formed two different sized partitions!");
488 assert(P.SplitTails.size() == RHS.P.SplitTails.size() &&
489 "Same slice position with differently sized non-empty split "
496 partition_iterator &operator++() {
501 Partition &operator*() { return P; }
504 /// \brief A forward range over the partitions of the alloca's slices.
506 /// This accesses an iterator range over the partitions of the alloca's
507 /// slices. It computes these partitions on the fly based on the overlapping
508 /// offsets of the slices and the ability to split them. It will visit "empty"
509 /// partitions to cover regions of the alloca only accessed via split
511 iterator_range<partition_iterator> partitions() {
512 return make_range(partition_iterator(begin(), end()),
513 partition_iterator(end(), end()));
516 /// \brief Access the dead users for this alloca.
517 ArrayRef<Instruction *> getDeadUsers() const { return DeadUsers; }
519 /// \brief Access the dead operands referring to this alloca.
521 /// These are operands which have cannot actually be used to refer to the
522 /// alloca as they are outside its range and the user doesn't correct for
523 /// that. These mostly consist of PHI node inputs and the like which we just
524 /// need to replace with undef.
525 ArrayRef<Use *> getDeadOperands() const { return DeadOperands; }
527 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
528 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
529 void printSlice(raw_ostream &OS, const_iterator I,
530 StringRef Indent = " ") const;
531 void printUse(raw_ostream &OS, const_iterator I,
532 StringRef Indent = " ") const;
533 void print(raw_ostream &OS) const;
534 void dump(const_iterator I) const;
539 template <typename DerivedT, typename RetT = void> class BuilderBase;
541 friend class AllocaSlices::SliceBuilder;
543 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
544 /// \brief Handle to alloca instruction to simplify method interfaces.
548 /// \brief The instruction responsible for this alloca not having a known set
551 /// When an instruction (potentially) escapes the pointer to the alloca, we
552 /// store a pointer to that here and abort trying to form slices of the
553 /// alloca. This will be null if the alloca slices are analyzed successfully.
554 Instruction *PointerEscapingInstr;
556 /// \brief The slices of the alloca.
558 /// We store a vector of the slices formed by uses of the alloca here. This
559 /// vector is sorted by increasing begin offset, and then the unsplittable
560 /// slices before the splittable ones. See the Slice inner class for more
562 SmallVector<Slice, 8> Slices;
564 /// \brief Instructions which will become dead if we rewrite the alloca.
566 /// Note that these are not separated by slice. This is because we expect an
567 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
568 /// all these instructions can simply be removed and replaced with undef as
569 /// they come from outside of the allocated space.
570 SmallVector<Instruction *, 8> DeadUsers;
572 /// \brief Operands which will become dead if we rewrite the alloca.
574 /// These are operands that in their particular use can be replaced with
575 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
576 /// to PHI nodes and the like. They aren't entirely dead (there might be
577 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
578 /// want to swap this particular input for undef to simplify the use lists of
580 SmallVector<Use *, 8> DeadOperands;
584 static Value *foldSelectInst(SelectInst &SI) {
585 // If the condition being selected on is a constant or the same value is
586 // being selected between, fold the select. Yes this does (rarely) happen
588 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
589 return SI.getOperand(1 + CI->isZero());
590 if (SI.getOperand(1) == SI.getOperand(2))
591 return SI.getOperand(1);
596 /// \brief A helper that folds a PHI node or a select.
597 static Value *foldPHINodeOrSelectInst(Instruction &I) {
598 if (PHINode *PN = dyn_cast<PHINode>(&I)) {
599 // If PN merges together the same value, return that value.
600 return PN->hasConstantValue();
602 return foldSelectInst(cast<SelectInst>(I));
605 /// \brief Builder for the alloca slices.
607 /// This class builds a set of alloca slices by recursively visiting the uses
608 /// of an alloca and making a slice for each load and store at each offset.
609 class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
610 friend class PtrUseVisitor<SliceBuilder>;
611 friend class InstVisitor<SliceBuilder>;
612 typedef PtrUseVisitor<SliceBuilder> Base;
614 const uint64_t AllocSize;
617 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
618 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
620 /// \brief Set to de-duplicate dead instructions found in the use walk.
621 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
624 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &AS)
625 : PtrUseVisitor<SliceBuilder>(DL),
626 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), AS(AS) {}
629 void markAsDead(Instruction &I) {
630 if (VisitedDeadInsts.insert(&I).second)
631 AS.DeadUsers.push_back(&I);
634 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
635 bool IsSplittable = false) {
636 // Completely skip uses which have a zero size or start either before or
637 // past the end of the allocation.
638 if (Size == 0 || Offset.uge(AllocSize)) {
639 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
640 << " which has zero size or starts outside of the "
641 << AllocSize << " byte alloca:\n"
642 << " alloca: " << AS.AI << "\n"
643 << " use: " << I << "\n");
644 return markAsDead(I);
647 uint64_t BeginOffset = Offset.getZExtValue();
648 uint64_t EndOffset = BeginOffset + Size;
650 // Clamp the end offset to the end of the allocation. Note that this is
651 // formulated to handle even the case where "BeginOffset + Size" overflows.
652 // This may appear superficially to be something we could ignore entirely,
653 // but that is not so! There may be widened loads or PHI-node uses where
654 // some instructions are dead but not others. We can't completely ignore
655 // them, and so have to record at least the information here.
656 assert(AllocSize >= BeginOffset); // Established above.
657 if (Size > AllocSize - BeginOffset) {
658 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
659 << " to remain within the " << AllocSize << " byte alloca:\n"
660 << " alloca: " << AS.AI << "\n"
661 << " use: " << I << "\n");
662 EndOffset = AllocSize;
665 AS.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
668 void visitBitCastInst(BitCastInst &BC) {
670 return markAsDead(BC);
672 return Base::visitBitCastInst(BC);
675 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
676 if (GEPI.use_empty())
677 return markAsDead(GEPI);
679 if (SROAStrictInbounds && GEPI.isInBounds()) {
680 // FIXME: This is a manually un-factored variant of the basic code inside
681 // of GEPs with checking of the inbounds invariant specified in the
682 // langref in a very strict sense. If we ever want to enable
683 // SROAStrictInbounds, this code should be factored cleanly into
684 // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
685 // by writing out the code here where we have tho underlying allocation
686 // size readily available.
687 APInt GEPOffset = Offset;
688 for (gep_type_iterator GTI = gep_type_begin(GEPI),
689 GTE = gep_type_end(GEPI);
691 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
695 // Handle a struct index, which adds its field offset to the pointer.
696 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
697 unsigned ElementIdx = OpC->getZExtValue();
698 const StructLayout *SL = DL.getStructLayout(STy);
700 APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
702 // For array or vector indices, scale the index by the size of the
704 APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
705 GEPOffset += Index * APInt(Offset.getBitWidth(),
706 DL.getTypeAllocSize(GTI.getIndexedType()));
709 // If this index has computed an intermediate pointer which is not
710 // inbounds, then the result of the GEP is a poison value and we can
711 // delete it and all uses.
712 if (GEPOffset.ugt(AllocSize))
713 return markAsDead(GEPI);
717 return Base::visitGetElementPtrInst(GEPI);
720 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
721 uint64_t Size, bool IsVolatile) {
722 // We allow splitting of loads and stores where the type is an integer type
723 // and cover the entire alloca. This prevents us from splitting over
725 // FIXME: In the great blue eventually, we should eagerly split all integer
726 // loads and stores, and then have a separate step that merges adjacent
727 // alloca partitions into a single partition suitable for integer widening.
728 // Or we should skip the merge step and rely on GVN and other passes to
729 // merge adjacent loads and stores that survive mem2reg.
731 Ty->isIntegerTy() && !IsVolatile && Offset == 0 && Size >= AllocSize;
733 insertUse(I, Offset, Size, IsSplittable);
736 void visitLoadInst(LoadInst &LI) {
737 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
738 "All simple FCA loads should have been pre-split");
741 return PI.setAborted(&LI);
743 uint64_t Size = DL.getTypeStoreSize(LI.getType());
744 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
747 void visitStoreInst(StoreInst &SI) {
748 Value *ValOp = SI.getValueOperand();
750 return PI.setEscapedAndAborted(&SI);
752 return PI.setAborted(&SI);
754 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
756 // If this memory access can be shown to *statically* extend outside the
757 // bounds of of the allocation, it's behavior is undefined, so simply
758 // ignore it. Note that this is more strict than the generic clamping
759 // behavior of insertUse. We also try to handle cases which might run the
761 // FIXME: We should instead consider the pointer to have escaped if this
762 // function is being instrumented for addressing bugs or race conditions.
763 if (Size > AllocSize || Offset.ugt(AllocSize - Size)) {
764 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
765 << " which extends past the end of the " << AllocSize
767 << " alloca: " << AS.AI << "\n"
768 << " use: " << SI << "\n");
769 return markAsDead(SI);
772 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
773 "All simple FCA stores should have been pre-split");
774 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
777 void visitMemSetInst(MemSetInst &II) {
778 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
779 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
780 if ((Length && Length->getValue() == 0) ||
781 (IsOffsetKnown && Offset.uge(AllocSize)))
782 // Zero-length mem transfer intrinsics can be ignored entirely.
783 return markAsDead(II);
786 return PI.setAborted(&II);
788 insertUse(II, Offset, Length ? Length->getLimitedValue()
789 : AllocSize - Offset.getLimitedValue(),
793 void visitMemTransferInst(MemTransferInst &II) {
794 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
795 if (Length && Length->getValue() == 0)
796 // Zero-length mem transfer intrinsics can be ignored entirely.
797 return markAsDead(II);
799 // Because we can visit these intrinsics twice, also check to see if the
800 // first time marked this instruction as dead. If so, skip it.
801 if (VisitedDeadInsts.count(&II))
805 return PI.setAborted(&II);
807 // This side of the transfer is completely out-of-bounds, and so we can
808 // nuke the entire transfer. However, we also need to nuke the other side
809 // if already added to our partitions.
810 // FIXME: Yet another place we really should bypass this when
811 // instrumenting for ASan.
812 if (Offset.uge(AllocSize)) {
813 SmallDenseMap<Instruction *, unsigned>::iterator MTPI =
814 MemTransferSliceMap.find(&II);
815 if (MTPI != MemTransferSliceMap.end())
816 AS.Slices[MTPI->second].kill();
817 return markAsDead(II);
820 uint64_t RawOffset = Offset.getLimitedValue();
821 uint64_t Size = Length ? Length->getLimitedValue() : AllocSize - RawOffset;
823 // Check for the special case where the same exact value is used for both
825 if (*U == II.getRawDest() && *U == II.getRawSource()) {
826 // For non-volatile transfers this is a no-op.
827 if (!II.isVolatile())
828 return markAsDead(II);
830 return insertUse(II, Offset, Size, /*IsSplittable=*/false);
833 // If we have seen both source and destination for a mem transfer, then
834 // they both point to the same alloca.
836 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
837 std::tie(MTPI, Inserted) =
838 MemTransferSliceMap.insert(std::make_pair(&II, AS.Slices.size()));
839 unsigned PrevIdx = MTPI->second;
841 Slice &PrevP = AS.Slices[PrevIdx];
843 // Check if the begin offsets match and this is a non-volatile transfer.
844 // In that case, we can completely elide the transfer.
845 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
847 return markAsDead(II);
850 // Otherwise we have an offset transfer within the same alloca. We can't
852 PrevP.makeUnsplittable();
855 // Insert the use now that we've fixed up the splittable nature.
856 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
858 // Check that we ended up with a valid index in the map.
859 assert(AS.Slices[PrevIdx].getUse()->getUser() == &II &&
860 "Map index doesn't point back to a slice with this user.");
863 // Disable SRoA for any intrinsics except for lifetime invariants.
864 // FIXME: What about debug intrinsics? This matches old behavior, but
865 // doesn't make sense.
866 void visitIntrinsicInst(IntrinsicInst &II) {
868 return PI.setAborted(&II);
870 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
871 II.getIntrinsicID() == Intrinsic::lifetime_end) {
872 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
873 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
874 Length->getLimitedValue());
875 insertUse(II, Offset, Size, true);
879 Base::visitIntrinsicInst(II);
882 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
883 // We consider any PHI or select that results in a direct load or store of
884 // the same offset to be a viable use for slicing purposes. These uses
885 // are considered unsplittable and the size is the maximum loaded or stored
887 SmallPtrSet<Instruction *, 4> Visited;
888 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
889 Visited.insert(Root);
890 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
891 // If there are no loads or stores, the access is dead. We mark that as
892 // a size zero access.
895 Instruction *I, *UsedI;
896 std::tie(UsedI, I) = Uses.pop_back_val();
898 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
899 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
902 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
903 Value *Op = SI->getOperand(0);
906 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
910 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
911 if (!GEP->hasAllZeroIndices())
913 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
914 !isa<SelectInst>(I)) {
918 for (User *U : I->users())
919 if (Visited.insert(cast<Instruction>(U)).second)
920 Uses.push_back(std::make_pair(I, cast<Instruction>(U)));
921 } while (!Uses.empty());
926 void visitPHINodeOrSelectInst(Instruction &I) {
927 assert(isa<PHINode>(I) || isa<SelectInst>(I));
929 return markAsDead(I);
931 // TODO: We could use SimplifyInstruction here to fold PHINodes and
932 // SelectInsts. However, doing so requires to change the current
933 // dead-operand-tracking mechanism. For instance, suppose neither loading
934 // from %U nor %other traps. Then "load (select undef, %U, %other)" does not
935 // trap either. However, if we simply replace %U with undef using the
936 // current dead-operand-tracking mechanism, "load (select undef, undef,
937 // %other)" may trap because the select may return the first operand
939 if (Value *Result = foldPHINodeOrSelectInst(I)) {
941 // If the result of the constant fold will be the pointer, recurse
942 // through the PHI/select as if we had RAUW'ed it.
945 // Otherwise the operand to the PHI/select is dead, and we can replace
947 AS.DeadOperands.push_back(U);
953 return PI.setAborted(&I);
955 // See if we already have computed info on this node.
956 uint64_t &Size = PHIOrSelectSizes[&I];
958 // This is a new PHI/Select, check for an unsafe use of it.
959 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&I, Size))
960 return PI.setAborted(UnsafeI);
963 // For PHI and select operands outside the alloca, we can't nuke the entire
964 // phi or select -- the other side might still be relevant, so we special
965 // case them here and use a separate structure to track the operands
966 // themselves which should be replaced with undef.
967 // FIXME: This should instead be escaped in the event we're instrumenting
968 // for address sanitization.
969 if (Offset.uge(AllocSize)) {
970 AS.DeadOperands.push_back(U);
974 insertUse(I, Offset, Size);
977 void visitPHINode(PHINode &PN) { visitPHINodeOrSelectInst(PN); }
979 void visitSelectInst(SelectInst &SI) { visitPHINodeOrSelectInst(SI); }
981 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
982 void visitInstruction(Instruction &I) { PI.setAborted(&I); }
985 AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
987 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
990 PointerEscapingInstr(nullptr) {
991 SliceBuilder PB(DL, AI, *this);
992 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
993 if (PtrI.isEscaped() || PtrI.isAborted()) {
994 // FIXME: We should sink the escape vs. abort info into the caller nicely,
995 // possibly by just storing the PtrInfo in the AllocaSlices.
996 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
997 : PtrI.getAbortingInst();
998 assert(PointerEscapingInstr && "Did not track a bad instruction");
1002 Slices.erase(std::remove_if(Slices.begin(), Slices.end(),
1003 [](const Slice &S) {
1008 #if __cplusplus >= 201103L && !defined(NDEBUG)
1009 if (SROARandomShuffleSlices) {
1010 std::mt19937 MT(static_cast<unsigned>(sys::TimeValue::now().msec()));
1011 std::shuffle(Slices.begin(), Slices.end(), MT);
1015 // Sort the uses. This arranges for the offsets to be in ascending order,
1016 // and the sizes to be in descending order.
1017 std::sort(Slices.begin(), Slices.end());
1020 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1022 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
1023 StringRef Indent) const {
1024 printSlice(OS, I, Indent);
1025 printUse(OS, I, Indent);
1028 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
1029 StringRef Indent) const {
1030 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
1031 << " slice #" << (I - begin())
1032 << (I->isSplittable() ? " (splittable)" : "") << "\n";
1035 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
1036 StringRef Indent) const {
1037 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
1040 void AllocaSlices::print(raw_ostream &OS) const {
1041 if (PointerEscapingInstr) {
1042 OS << "Can't analyze slices for alloca: " << AI << "\n"
1043 << " A pointer to this alloca escaped by:\n"
1044 << " " << *PointerEscapingInstr << "\n";
1048 OS << "Slices of alloca: " << AI << "\n";
1049 for (const_iterator I = begin(), E = end(); I != E; ++I)
1053 LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
1056 LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
1058 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1061 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
1063 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
1064 /// the loads and stores of an alloca instruction, as well as updating its
1065 /// debug information. This is used when a domtree is unavailable and thus
1066 /// mem2reg in its full form can't be used to handle promotion of allocas to
1068 class AllocaPromoter : public LoadAndStorePromoter {
1072 SmallVector<DbgDeclareInst *, 4> DDIs;
1073 SmallVector<DbgValueInst *, 4> DVIs;
1076 AllocaPromoter(const SmallVectorImpl<Instruction *> &Insts, SSAUpdater &S,
1077 AllocaInst &AI, DIBuilder &DIB)
1078 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
1080 void run(const SmallVectorImpl<Instruction *> &Insts) {
1081 // Retain the debug information attached to the alloca for use when
1082 // rewriting loads and stores.
1083 if (auto *L = LocalAsMetadata::getIfExists(&AI)) {
1084 if (auto *DebugNode = MetadataAsValue::getIfExists(AI.getContext(), L)) {
1085 for (User *U : DebugNode->users())
1086 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
1087 DDIs.push_back(DDI);
1088 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(U))
1089 DVIs.push_back(DVI);
1093 LoadAndStorePromoter::run(Insts);
1095 // While we have the debug information, clear it off of the alloca. The
1096 // caller takes care of deleting the alloca.
1097 while (!DDIs.empty())
1098 DDIs.pop_back_val()->eraseFromParent();
1099 while (!DVIs.empty())
1100 DVIs.pop_back_val()->eraseFromParent();
1104 isInstInList(Instruction *I,
1105 const SmallVectorImpl<Instruction *> &Insts) const override {
1107 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1108 Ptr = LI->getOperand(0);
1110 Ptr = cast<StoreInst>(I)->getPointerOperand();
1112 // Only used to detect cycles, which will be rare and quickly found as
1113 // we're walking up a chain of defs rather than down through uses.
1114 SmallPtrSet<Value *, 4> Visited;
1120 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr))
1121 Ptr = BCI->getOperand(0);
1122 else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr))
1123 Ptr = GEPI->getPointerOperand();
1127 } while (Visited.insert(Ptr).second);
1132 void updateDebugInfo(Instruction *Inst) const override {
1133 for (DbgDeclareInst *DDI : DDIs)
1134 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
1135 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1136 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
1137 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1138 for (DbgValueInst *DVI : DVIs) {
1139 Value *Arg = nullptr;
1140 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1141 // If an argument is zero extended then use argument directly. The ZExt
1142 // may be zapped by an optimization pass in future.
1143 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1144 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
1145 else if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1146 Arg = dyn_cast<Argument>(SExt->getOperand(0));
1148 Arg = SI->getValueOperand();
1149 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
1150 Arg = LI->getPointerOperand();
1154 Instruction *DbgVal =
1155 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
1156 DIExpression(DVI->getExpression()), Inst);
1157 DbgVal->setDebugLoc(DVI->getDebugLoc());
1161 } // end anon namespace
1164 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
1166 /// This pass takes allocations which can be completely analyzed (that is, they
1167 /// don't escape) and tries to turn them into scalar SSA values. There are
1168 /// a few steps to this process.
1170 /// 1) It takes allocations of aggregates and analyzes the ways in which they
1171 /// are used to try to split them into smaller allocations, ideally of
1172 /// a single scalar data type. It will split up memcpy and memset accesses
1173 /// as necessary and try to isolate individual scalar accesses.
1174 /// 2) It will transform accesses into forms which are suitable for SSA value
1175 /// promotion. This can be replacing a memset with a scalar store of an
1176 /// integer value, or it can involve speculating operations on a PHI or
1177 /// select to be a PHI or select of the results.
1178 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
1179 /// onto insert and extract operations on a vector value, and convert them to
1180 /// this form. By doing so, it will enable promotion of vector aggregates to
1181 /// SSA vector values.
1182 class SROA : public FunctionPass {
1183 const bool RequiresDomTree;
1186 const DataLayout *DL;
1188 AssumptionTracker *AT;
1190 /// \brief Worklist of alloca instructions to simplify.
1192 /// Each alloca in the function is added to this. Each new alloca formed gets
1193 /// added to it as well to recursively simplify unless that alloca can be
1194 /// directly promoted. Finally, each time we rewrite a use of an alloca other
1195 /// the one being actively rewritten, we add it back onto the list if not
1196 /// already present to ensure it is re-visited.
1197 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16>> Worklist;
1199 /// \brief A collection of instructions to delete.
1200 /// We try to batch deletions to simplify code and make things a bit more
1202 SetVector<Instruction *, SmallVector<Instruction *, 8>> DeadInsts;
1204 /// \brief Post-promotion worklist.
1206 /// Sometimes we discover an alloca which has a high probability of becoming
1207 /// viable for SROA after a round of promotion takes place. In those cases,
1208 /// the alloca is enqueued here for re-processing.
1210 /// Note that we have to be very careful to clear allocas out of this list in
1211 /// the event they are deleted.
1212 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16>> PostPromotionWorklist;
1214 /// \brief A collection of alloca instructions we can directly promote.
1215 std::vector<AllocaInst *> PromotableAllocas;
1217 /// \brief A worklist of PHIs to speculate prior to promoting allocas.
1219 /// All of these PHIs have been checked for the safety of speculation and by
1220 /// being speculated will allow promoting allocas currently in the promotable
1222 SetVector<PHINode *, SmallVector<PHINode *, 2>> SpeculatablePHIs;
1224 /// \brief A worklist of select instructions to speculate prior to promoting
1227 /// All of these select instructions have been checked for the safety of
1228 /// speculation and by being speculated will allow promoting allocas
1229 /// currently in the promotable queue.
1230 SetVector<SelectInst *, SmallVector<SelectInst *, 2>> SpeculatableSelects;
1233 SROA(bool RequiresDomTree = true)
1234 : FunctionPass(ID), RequiresDomTree(RequiresDomTree), C(nullptr),
1235 DL(nullptr), DT(nullptr) {
1236 initializeSROAPass(*PassRegistry::getPassRegistry());
1238 bool runOnFunction(Function &F) override;
1239 void getAnalysisUsage(AnalysisUsage &AU) const override;
1241 const char *getPassName() const override { return "SROA"; }
1245 friend class PHIOrSelectSpeculator;
1246 friend class AllocaSliceRewriter;
1248 bool rewritePartition(AllocaInst &AI, AllocaSlices &AS,
1249 AllocaSlices::Partition &P);
1250 bool splitAlloca(AllocaInst &AI, AllocaSlices &AS);
1251 bool runOnAlloca(AllocaInst &AI);
1252 void clobberUse(Use &U);
1253 void deleteDeadInstructions(SmallPtrSetImpl<AllocaInst *> &DeletedAllocas);
1254 bool promoteAllocas(Function &F);
1260 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
1261 return new SROA(RequiresDomTree);
1264 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates", false,
1266 INITIALIZE_PASS_DEPENDENCY(AssumptionTracker)
1267 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1268 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates", false,
1271 /// Walk the range of a partitioning looking for a common type to cover this
1272 /// sequence of slices.
1273 static Type *findCommonType(AllocaSlices::const_iterator B,
1274 AllocaSlices::const_iterator E,
1275 uint64_t EndOffset) {
1277 bool TyIsCommon = true;
1278 IntegerType *ITy = nullptr;
1280 // Note that we need to look at *every* alloca slice's Use to ensure we
1281 // always get consistent results regardless of the order of slices.
1282 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1283 Use *U = I->getUse();
1284 if (isa<IntrinsicInst>(*U->getUser()))
1286 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1289 Type *UserTy = nullptr;
1290 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1291 UserTy = LI->getType();
1292 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1293 UserTy = SI->getValueOperand()->getType();
1296 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1297 // If the type is larger than the partition, skip it. We only encounter
1298 // this for split integer operations where we want to use the type of the
1299 // entity causing the split. Also skip if the type is not a byte width
1301 if (UserITy->getBitWidth() % 8 != 0 ||
1302 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1305 // Track the largest bitwidth integer type used in this way in case there
1306 // is no common type.
1307 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1311 // To avoid depending on the order of slices, Ty and TyIsCommon must not
1312 // depend on types skipped above.
1313 if (!UserTy || (Ty && Ty != UserTy))
1314 TyIsCommon = false; // Give up on anything but an iN type.
1319 return TyIsCommon ? Ty : ITy;
1322 /// PHI instructions that use an alloca and are subsequently loaded can be
1323 /// rewritten to load both input pointers in the pred blocks and then PHI the
1324 /// results, allowing the load of the alloca to be promoted.
1326 /// %P2 = phi [i32* %Alloca, i32* %Other]
1327 /// %V = load i32* %P2
1329 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1331 /// %V2 = load i32* %Other
1333 /// %V = phi [i32 %V1, i32 %V2]
1335 /// We can do this to a select if its only uses are loads and if the operands
1336 /// to the select can be loaded unconditionally.
1338 /// FIXME: This should be hoisted into a generic utility, likely in
1339 /// Transforms/Util/Local.h
1340 static bool isSafePHIToSpeculate(PHINode &PN, const DataLayout *DL = nullptr) {
1341 // For now, we can only do this promotion if the load is in the same block
1342 // as the PHI, and if there are no stores between the phi and load.
1343 // TODO: Allow recursive phi users.
1344 // TODO: Allow stores.
1345 BasicBlock *BB = PN.getParent();
1346 unsigned MaxAlign = 0;
1347 bool HaveLoad = false;
1348 for (User *U : PN.users()) {
1349 LoadInst *LI = dyn_cast<LoadInst>(U);
1350 if (!LI || !LI->isSimple())
1353 // For now we only allow loads in the same block as the PHI. This is
1354 // a common case that happens when instcombine merges two loads through
1356 if (LI->getParent() != BB)
1359 // Ensure that there are no instructions between the PHI and the load that
1361 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1362 if (BBI->mayWriteToMemory())
1365 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1372 // We can only transform this if it is safe to push the loads into the
1373 // predecessor blocks. The only thing to watch out for is that we can't put
1374 // a possibly trapping load in the predecessor if it is a critical edge.
1375 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1376 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1377 Value *InVal = PN.getIncomingValue(Idx);
1379 // If the value is produced by the terminator of the predecessor (an
1380 // invoke) or it has side-effects, there is no valid place to put a load
1381 // in the predecessor.
1382 if (TI == InVal || TI->mayHaveSideEffects())
1385 // If the predecessor has a single successor, then the edge isn't
1387 if (TI->getNumSuccessors() == 1)
1390 // If this pointer is always safe to load, or if we can prove that there
1391 // is already a load in the block, then we can move the load to the pred
1393 if (InVal->isDereferenceablePointer(DL) ||
1394 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, DL))
1403 static void speculatePHINodeLoads(PHINode &PN) {
1404 DEBUG(dbgs() << " original: " << PN << "\n");
1406 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1407 IRBuilderTy PHIBuilder(&PN);
1408 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1409 PN.getName() + ".sroa.speculated");
1411 // Get the AA tags and alignment to use from one of the loads. It doesn't
1412 // matter which one we get and if any differ.
1413 LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1416 SomeLoad->getAAMetadata(AATags);
1417 unsigned Align = SomeLoad->getAlignment();
1419 // Rewrite all loads of the PN to use the new PHI.
1420 while (!PN.use_empty()) {
1421 LoadInst *LI = cast<LoadInst>(PN.user_back());
1422 LI->replaceAllUsesWith(NewPN);
1423 LI->eraseFromParent();
1426 // Inject loads into all of the pred blocks.
1427 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1428 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1429 TerminatorInst *TI = Pred->getTerminator();
1430 Value *InVal = PN.getIncomingValue(Idx);
1431 IRBuilderTy PredBuilder(TI);
1433 LoadInst *Load = PredBuilder.CreateLoad(
1434 InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1435 ++NumLoadsSpeculated;
1436 Load->setAlignment(Align);
1438 Load->setAAMetadata(AATags);
1439 NewPN->addIncoming(Load, Pred);
1442 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1443 PN.eraseFromParent();
1446 /// Select instructions that use an alloca and are subsequently loaded can be
1447 /// rewritten to load both input pointers and then select between the result,
1448 /// allowing the load of the alloca to be promoted.
1450 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1451 /// %V = load i32* %P2
1453 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1454 /// %V2 = load i32* %Other
1455 /// %V = select i1 %cond, i32 %V1, i32 %V2
1457 /// We can do this to a select if its only uses are loads and if the operand
1458 /// to the select can be loaded unconditionally.
1459 static bool isSafeSelectToSpeculate(SelectInst &SI,
1460 const DataLayout *DL = nullptr) {
1461 Value *TValue = SI.getTrueValue();
1462 Value *FValue = SI.getFalseValue();
1463 bool TDerefable = TValue->isDereferenceablePointer(DL);
1464 bool FDerefable = FValue->isDereferenceablePointer(DL);
1466 for (User *U : SI.users()) {
1467 LoadInst *LI = dyn_cast<LoadInst>(U);
1468 if (!LI || !LI->isSimple())
1471 // Both operands to the select need to be dereferencable, either
1472 // absolutely (e.g. allocas) or at this point because we can see other
1475 !isSafeToLoadUnconditionally(TValue, LI, LI->getAlignment(), DL))
1478 !isSafeToLoadUnconditionally(FValue, LI, LI->getAlignment(), DL))
1485 static void speculateSelectInstLoads(SelectInst &SI) {
1486 DEBUG(dbgs() << " original: " << SI << "\n");
1488 IRBuilderTy IRB(&SI);
1489 Value *TV = SI.getTrueValue();
1490 Value *FV = SI.getFalseValue();
1491 // Replace the loads of the select with a select of two loads.
1492 while (!SI.use_empty()) {
1493 LoadInst *LI = cast<LoadInst>(SI.user_back());
1494 assert(LI->isSimple() && "We only speculate simple loads");
1496 IRB.SetInsertPoint(LI);
1498 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1500 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1501 NumLoadsSpeculated += 2;
1503 // Transfer alignment and AA info if present.
1504 TL->setAlignment(LI->getAlignment());
1505 FL->setAlignment(LI->getAlignment());
1508 LI->getAAMetadata(Tags);
1510 TL->setAAMetadata(Tags);
1511 FL->setAAMetadata(Tags);
1514 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1515 LI->getName() + ".sroa.speculated");
1517 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1518 LI->replaceAllUsesWith(V);
1519 LI->eraseFromParent();
1521 SI.eraseFromParent();
1524 /// \brief Build a GEP out of a base pointer and indices.
1526 /// This will return the BasePtr if that is valid, or build a new GEP
1527 /// instruction using the IRBuilder if GEP-ing is needed.
1528 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1529 SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1530 if (Indices.empty())
1533 // A single zero index is a no-op, so check for this and avoid building a GEP
1535 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1538 return IRB.CreateInBoundsGEP(BasePtr, Indices, NamePrefix + "sroa_idx");
1541 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1542 /// TargetTy without changing the offset of the pointer.
1544 /// This routine assumes we've already established a properly offset GEP with
1545 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1546 /// zero-indices down through type layers until we find one the same as
1547 /// TargetTy. If we can't find one with the same type, we at least try to use
1548 /// one with the same size. If none of that works, we just produce the GEP as
1549 /// indicated by Indices to have the correct offset.
1550 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1551 Value *BasePtr, Type *Ty, Type *TargetTy,
1552 SmallVectorImpl<Value *> &Indices,
1555 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1557 // Pointer size to use for the indices.
1558 unsigned PtrSize = DL.getPointerTypeSizeInBits(BasePtr->getType());
1560 // See if we can descend into a struct and locate a field with the correct
1562 unsigned NumLayers = 0;
1563 Type *ElementTy = Ty;
1565 if (ElementTy->isPointerTy())
1568 if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1569 ElementTy = ArrayTy->getElementType();
1570 Indices.push_back(IRB.getIntN(PtrSize, 0));
1571 } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1572 ElementTy = VectorTy->getElementType();
1573 Indices.push_back(IRB.getInt32(0));
1574 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1575 if (STy->element_begin() == STy->element_end())
1576 break; // Nothing left to descend into.
1577 ElementTy = *STy->element_begin();
1578 Indices.push_back(IRB.getInt32(0));
1583 } while (ElementTy != TargetTy);
1584 if (ElementTy != TargetTy)
1585 Indices.erase(Indices.end() - NumLayers, Indices.end());
1587 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1590 /// \brief Recursively compute indices for a natural GEP.
1592 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1593 /// element types adding appropriate indices for the GEP.
1594 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1595 Value *Ptr, Type *Ty, APInt &Offset,
1597 SmallVectorImpl<Value *> &Indices,
1600 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices,
1603 // We can't recurse through pointer types.
1604 if (Ty->isPointerTy())
1607 // We try to analyze GEPs over vectors here, but note that these GEPs are
1608 // extremely poorly defined currently. The long-term goal is to remove GEPing
1609 // over a vector from the IR completely.
1610 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1611 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1612 if (ElementSizeInBits % 8 != 0) {
1613 // GEPs over non-multiple of 8 size vector elements are invalid.
1616 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1617 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1618 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1620 Offset -= NumSkippedElements * ElementSize;
1621 Indices.push_back(IRB.getInt(NumSkippedElements));
1622 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1623 Offset, TargetTy, Indices, NamePrefix);
1626 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1627 Type *ElementTy = ArrTy->getElementType();
1628 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1629 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1630 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1633 Offset -= NumSkippedElements * ElementSize;
1634 Indices.push_back(IRB.getInt(NumSkippedElements));
1635 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1636 Indices, NamePrefix);
1639 StructType *STy = dyn_cast<StructType>(Ty);
1643 const StructLayout *SL = DL.getStructLayout(STy);
1644 uint64_t StructOffset = Offset.getZExtValue();
1645 if (StructOffset >= SL->getSizeInBytes())
1647 unsigned Index = SL->getElementContainingOffset(StructOffset);
1648 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1649 Type *ElementTy = STy->getElementType(Index);
1650 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1651 return nullptr; // The offset points into alignment padding.
1653 Indices.push_back(IRB.getInt32(Index));
1654 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1655 Indices, NamePrefix);
1658 /// \brief Get a natural GEP from a base pointer to a particular offset and
1659 /// resulting in a particular type.
1661 /// The goal is to produce a "natural" looking GEP that works with the existing
1662 /// composite types to arrive at the appropriate offset and element type for
1663 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1664 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1665 /// Indices, and setting Ty to the result subtype.
1667 /// If no natural GEP can be constructed, this function returns null.
1668 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1669 Value *Ptr, APInt Offset, Type *TargetTy,
1670 SmallVectorImpl<Value *> &Indices,
1672 PointerType *Ty = cast<PointerType>(Ptr->getType());
1674 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1676 if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1679 Type *ElementTy = Ty->getElementType();
1680 if (!ElementTy->isSized())
1681 return nullptr; // We can't GEP through an unsized element.
1682 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1683 if (ElementSize == 0)
1684 return nullptr; // Zero-length arrays can't help us build a natural GEP.
1685 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1687 Offset -= NumSkippedElements * ElementSize;
1688 Indices.push_back(IRB.getInt(NumSkippedElements));
1689 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1690 Indices, NamePrefix);
1693 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1694 /// resulting pointer has PointerTy.
1696 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1697 /// and produces the pointer type desired. Where it cannot, it will try to use
1698 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1699 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1700 /// bitcast to the type.
1702 /// The strategy for finding the more natural GEPs is to peel off layers of the
1703 /// pointer, walking back through bit casts and GEPs, searching for a base
1704 /// pointer from which we can compute a natural GEP with the desired
1705 /// properties. The algorithm tries to fold as many constant indices into
1706 /// a single GEP as possible, thus making each GEP more independent of the
1707 /// surrounding code.
1708 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1709 APInt Offset, Type *PointerTy, Twine NamePrefix) {
1710 // Even though we don't look through PHI nodes, we could be called on an
1711 // instruction in an unreachable block, which may be on a cycle.
1712 SmallPtrSet<Value *, 4> Visited;
1713 Visited.insert(Ptr);
1714 SmallVector<Value *, 4> Indices;
1716 // We may end up computing an offset pointer that has the wrong type. If we
1717 // never are able to compute one directly that has the correct type, we'll
1718 // fall back to it, so keep it around here.
1719 Value *OffsetPtr = nullptr;
1721 // Remember any i8 pointer we come across to re-use if we need to do a raw
1723 Value *Int8Ptr = nullptr;
1724 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1726 Type *TargetTy = PointerTy->getPointerElementType();
1729 // First fold any existing GEPs into the offset.
1730 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1731 APInt GEPOffset(Offset.getBitWidth(), 0);
1732 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1734 Offset += GEPOffset;
1735 Ptr = GEP->getPointerOperand();
1736 if (!Visited.insert(Ptr).second)
1740 // See if we can perform a natural GEP here.
1742 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1743 Indices, NamePrefix)) {
1744 if (P->getType() == PointerTy) {
1745 // Zap any offset pointer that we ended up computing in previous rounds.
1746 if (OffsetPtr && OffsetPtr->use_empty())
1747 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1748 I->eraseFromParent();
1756 // Stash this pointer if we've found an i8*.
1757 if (Ptr->getType()->isIntegerTy(8)) {
1759 Int8PtrOffset = Offset;
1762 // Peel off a layer of the pointer and update the offset appropriately.
1763 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1764 Ptr = cast<Operator>(Ptr)->getOperand(0);
1765 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1766 if (GA->mayBeOverridden())
1768 Ptr = GA->getAliasee();
1772 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1773 } while (Visited.insert(Ptr).second);
1777 Int8Ptr = IRB.CreateBitCast(
1778 Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1779 NamePrefix + "sroa_raw_cast");
1780 Int8PtrOffset = Offset;
1783 OffsetPtr = Int8PtrOffset == 0
1785 : IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1786 NamePrefix + "sroa_raw_idx");
1790 // On the off chance we were targeting i8*, guard the bitcast here.
1791 if (Ptr->getType() != PointerTy)
1792 Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
1797 /// \brief Test whether we can convert a value from the old to the new type.
1799 /// This predicate should be used to guard calls to convertValue in order to
1800 /// ensure that we only try to convert viable values. The strategy is that we
1801 /// will peel off single element struct and array wrappings to get to an
1802 /// underlying value, and convert that value.
1803 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1806 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1807 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1808 if (NewITy->getBitWidth() >= OldITy->getBitWidth())
1810 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1812 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1815 // We can convert pointers to integers and vice-versa. Same for vectors
1816 // of pointers and integers.
1817 OldTy = OldTy->getScalarType();
1818 NewTy = NewTy->getScalarType();
1819 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1820 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1822 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1830 /// \brief Generic routine to convert an SSA value to a value of a different
1833 /// This will try various different casting techniques, such as bitcasts,
1834 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1835 /// two types for viability with this routine.
1836 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1838 Type *OldTy = V->getType();
1839 assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1844 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1845 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1846 if (NewITy->getBitWidth() > OldITy->getBitWidth())
1847 return IRB.CreateZExt(V, NewITy);
1849 // See if we need inttoptr for this type pair. A cast involving both scalars
1850 // and vectors requires and additional bitcast.
1851 if (OldTy->getScalarType()->isIntegerTy() &&
1852 NewTy->getScalarType()->isPointerTy()) {
1853 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1854 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1855 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1858 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1859 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1860 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1863 return IRB.CreateIntToPtr(V, NewTy);
1866 // See if we need ptrtoint for this type pair. A cast involving both scalars
1867 // and vectors requires and additional bitcast.
1868 if (OldTy->getScalarType()->isPointerTy() &&
1869 NewTy->getScalarType()->isIntegerTy()) {
1870 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1871 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1872 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1875 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1876 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1877 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1880 return IRB.CreatePtrToInt(V, NewTy);
1883 return IRB.CreateBitCast(V, NewTy);
1886 /// \brief Test whether the given slice use can be promoted to a vector.
1888 /// This function is called to test each entry in a partioning which is slated
1889 /// for a single slice.
1890 static bool isVectorPromotionViableForSlice(AllocaSlices::Partition &P,
1891 const Slice &S, VectorType *Ty,
1892 uint64_t ElementSize,
1893 const DataLayout &DL) {
1894 // First validate the slice offsets.
1895 uint64_t BeginOffset =
1896 std::max(S.beginOffset(), P.beginOffset()) - P.beginOffset();
1897 uint64_t BeginIndex = BeginOffset / ElementSize;
1898 if (BeginIndex * ElementSize != BeginOffset ||
1899 BeginIndex >= Ty->getNumElements())
1901 uint64_t EndOffset =
1902 std::min(S.endOffset(), P.endOffset()) - P.beginOffset();
1903 uint64_t EndIndex = EndOffset / ElementSize;
1904 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1907 assert(EndIndex > BeginIndex && "Empty vector!");
1908 uint64_t NumElements = EndIndex - BeginIndex;
1909 Type *SliceTy = (NumElements == 1)
1910 ? Ty->getElementType()
1911 : VectorType::get(Ty->getElementType(), NumElements);
1914 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1916 Use *U = S.getUse();
1918 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1919 if (MI->isVolatile())
1921 if (!S.isSplittable())
1922 return false; // Skip any unsplittable intrinsics.
1923 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1924 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1925 II->getIntrinsicID() != Intrinsic::lifetime_end)
1927 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1928 // Disable vector promotion when there are loads or stores of an FCA.
1930 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1931 if (LI->isVolatile())
1933 Type *LTy = LI->getType();
1934 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1935 assert(LTy->isIntegerTy());
1938 if (!canConvertValue(DL, SliceTy, LTy))
1940 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1941 if (SI->isVolatile())
1943 Type *STy = SI->getValueOperand()->getType();
1944 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1945 assert(STy->isIntegerTy());
1948 if (!canConvertValue(DL, STy, SliceTy))
1957 /// \brief Test whether the given alloca partitioning and range of slices can be
1958 /// promoted to a vector.
1960 /// This is a quick test to check whether we can rewrite a particular alloca
1961 /// partition (and its newly formed alloca) into a vector alloca with only
1962 /// whole-vector loads and stores such that it could be promoted to a vector
1963 /// SSA value. We only can ensure this for a limited set of operations, and we
1964 /// don't want to do the rewrites unless we are confident that the result will
1965 /// be promotable, so we have an early test here.
1966 static VectorType *isVectorPromotionViable(AllocaSlices::Partition &P,
1967 const DataLayout &DL) {
1968 // Collect the candidate types for vector-based promotion. Also track whether
1969 // we have different element types.
1970 SmallVector<VectorType *, 4> CandidateTys;
1971 Type *CommonEltTy = nullptr;
1972 bool HaveCommonEltTy = true;
1973 auto CheckCandidateType = [&](Type *Ty) {
1974 if (auto *VTy = dyn_cast<VectorType>(Ty)) {
1975 CandidateTys.push_back(VTy);
1977 CommonEltTy = VTy->getElementType();
1978 else if (CommonEltTy != VTy->getElementType())
1979 HaveCommonEltTy = false;
1982 // Consider any loads or stores that are the exact size of the slice.
1983 for (const Slice &S : P)
1984 if (S.beginOffset() == P.beginOffset() &&
1985 S.endOffset() == P.endOffset()) {
1986 if (auto *LI = dyn_cast<LoadInst>(S.getUse()->getUser()))
1987 CheckCandidateType(LI->getType());
1988 else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser()))
1989 CheckCandidateType(SI->getValueOperand()->getType());
1992 // If we didn't find a vector type, nothing to do here.
1993 if (CandidateTys.empty())
1996 // Remove non-integer vector types if we had multiple common element types.
1997 // FIXME: It'd be nice to replace them with integer vector types, but we can't
1998 // do that until all the backends are known to produce good code for all
1999 // integer vector types.
2000 if (!HaveCommonEltTy) {
2001 CandidateTys.erase(std::remove_if(CandidateTys.begin(), CandidateTys.end(),
2002 [](VectorType *VTy) {
2003 return !VTy->getElementType()->isIntegerTy();
2005 CandidateTys.end());
2007 // If there were no integer vector types, give up.
2008 if (CandidateTys.empty())
2011 // Rank the remaining candidate vector types. This is easy because we know
2012 // they're all integer vectors. We sort by ascending number of elements.
2013 auto RankVectorTypes = [&DL](VectorType *RHSTy, VectorType *LHSTy) {
2014 assert(DL.getTypeSizeInBits(RHSTy) == DL.getTypeSizeInBits(LHSTy) &&
2015 "Cannot have vector types of different sizes!");
2016 assert(RHSTy->getElementType()->isIntegerTy() &&
2017 "All non-integer types eliminated!");
2018 assert(LHSTy->getElementType()->isIntegerTy() &&
2019 "All non-integer types eliminated!");
2020 return RHSTy->getNumElements() < LHSTy->getNumElements();
2022 std::sort(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes);
2024 std::unique(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes),
2025 CandidateTys.end());
2027 // The only way to have the same element type in every vector type is to
2028 // have the same vector type. Check that and remove all but one.
2030 for (VectorType *VTy : CandidateTys) {
2031 assert(VTy->getElementType() == CommonEltTy &&
2032 "Unaccounted for element type!");
2033 assert(VTy == CandidateTys[0] &&
2034 "Different vector types with the same element type!");
2037 CandidateTys.resize(1);
2040 // Try each vector type, and return the one which works.
2041 auto CheckVectorTypeForPromotion = [&](VectorType *VTy) {
2042 uint64_t ElementSize = DL.getTypeSizeInBits(VTy->getElementType());
2044 // While the definition of LLVM vectors is bitpacked, we don't support sizes
2045 // that aren't byte sized.
2046 if (ElementSize % 8)
2048 assert((DL.getTypeSizeInBits(VTy) % 8) == 0 &&
2049 "vector size not a multiple of element size?");
2052 for (const Slice &S : P)
2053 if (!isVectorPromotionViableForSlice(P, S, VTy, ElementSize, DL))
2056 for (const Slice *S : P.splitSliceTails())
2057 if (!isVectorPromotionViableForSlice(P, *S, VTy, ElementSize, DL))
2062 for (VectorType *VTy : CandidateTys)
2063 if (CheckVectorTypeForPromotion(VTy))
2069 /// \brief Test whether a slice of an alloca is valid for integer widening.
2071 /// This implements the necessary checking for the \c isIntegerWideningViable
2072 /// test below on a single slice of the alloca.
2073 static bool isIntegerWideningViableForSlice(const Slice &S,
2074 uint64_t AllocBeginOffset,
2076 const DataLayout &DL,
2077 bool &WholeAllocaOp) {
2078 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
2080 uint64_t RelBegin = S.beginOffset() - AllocBeginOffset;
2081 uint64_t RelEnd = S.endOffset() - AllocBeginOffset;
2083 // We can't reasonably handle cases where the load or store extends past
2084 // the end of the aloca's type and into its padding.
2088 Use *U = S.getUse();
2090 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
2091 if (LI->isVolatile())
2093 // Note that we don't count vector loads or stores as whole-alloca
2094 // operations which enable integer widening because we would prefer to use
2095 // vector widening instead.
2096 if (!isa<VectorType>(LI->getType()) && RelBegin == 0 && RelEnd == Size)
2097 WholeAllocaOp = true;
2098 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
2099 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
2101 } else if (RelBegin != 0 || RelEnd != Size ||
2102 !canConvertValue(DL, AllocaTy, LI->getType())) {
2103 // Non-integer loads need to be convertible from the alloca type so that
2104 // they are promotable.
2107 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
2108 Type *ValueTy = SI->getValueOperand()->getType();
2109 if (SI->isVolatile())
2111 // Note that we don't count vector loads or stores as whole-alloca
2112 // operations which enable integer widening because we would prefer to use
2113 // vector widening instead.
2114 if (!isa<VectorType>(ValueTy) && RelBegin == 0 && RelEnd == Size)
2115 WholeAllocaOp = true;
2116 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
2117 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
2119 } else if (RelBegin != 0 || RelEnd != Size ||
2120 !canConvertValue(DL, ValueTy, AllocaTy)) {
2121 // Non-integer stores need to be convertible to the alloca type so that
2122 // they are promotable.
2125 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
2126 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
2128 if (!S.isSplittable())
2129 return false; // Skip any unsplittable intrinsics.
2130 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
2131 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2132 II->getIntrinsicID() != Intrinsic::lifetime_end)
2141 /// \brief Test whether the given alloca partition's integer operations can be
2142 /// widened to promotable ones.
2144 /// This is a quick test to check whether we can rewrite the integer loads and
2145 /// stores to a particular alloca into wider loads and stores and be able to
2146 /// promote the resulting alloca.
2147 static bool isIntegerWideningViable(AllocaSlices::Partition &P, Type *AllocaTy,
2148 const DataLayout &DL) {
2149 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
2150 // Don't create integer types larger than the maximum bitwidth.
2151 if (SizeInBits > IntegerType::MAX_INT_BITS)
2154 // Don't try to handle allocas with bit-padding.
2155 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
2158 // We need to ensure that an integer type with the appropriate bitwidth can
2159 // be converted to the alloca type, whatever that is. We don't want to force
2160 // the alloca itself to have an integer type if there is a more suitable one.
2161 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
2162 if (!canConvertValue(DL, AllocaTy, IntTy) ||
2163 !canConvertValue(DL, IntTy, AllocaTy))
2166 // While examining uses, we ensure that the alloca has a covering load or
2167 // store. We don't want to widen the integer operations only to fail to
2168 // promote due to some other unsplittable entry (which we may make splittable
2169 // later). However, if there are only splittable uses, go ahead and assume
2170 // that we cover the alloca.
2171 // FIXME: We shouldn't consider split slices that happen to start in the
2172 // partition here...
2173 bool WholeAllocaOp =
2174 P.begin() != P.end() ? false : DL.isLegalInteger(SizeInBits);
2176 for (const Slice &S : P)
2177 if (!isIntegerWideningViableForSlice(S, P.beginOffset(), AllocaTy, DL,
2181 for (const Slice *S : P.splitSliceTails())
2182 if (!isIntegerWideningViableForSlice(*S, P.beginOffset(), AllocaTy, DL,
2186 return WholeAllocaOp;
2189 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
2190 IntegerType *Ty, uint64_t Offset,
2191 const Twine &Name) {
2192 DEBUG(dbgs() << " start: " << *V << "\n");
2193 IntegerType *IntTy = cast<IntegerType>(V->getType());
2194 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2195 "Element extends past full value");
2196 uint64_t ShAmt = 8 * Offset;
2197 if (DL.isBigEndian())
2198 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2200 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2201 DEBUG(dbgs() << " shifted: " << *V << "\n");
2203 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2204 "Cannot extract to a larger integer!");
2206 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2207 DEBUG(dbgs() << " trunced: " << *V << "\n");
2212 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
2213 Value *V, uint64_t Offset, const Twine &Name) {
2214 IntegerType *IntTy = cast<IntegerType>(Old->getType());
2215 IntegerType *Ty = cast<IntegerType>(V->getType());
2216 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2217 "Cannot insert a larger integer!");
2218 DEBUG(dbgs() << " start: " << *V << "\n");
2220 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2221 DEBUG(dbgs() << " extended: " << *V << "\n");
2223 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2224 "Element store outside of alloca store");
2225 uint64_t ShAmt = 8 * Offset;
2226 if (DL.isBigEndian())
2227 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2229 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2230 DEBUG(dbgs() << " shifted: " << *V << "\n");
2233 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2234 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2235 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2236 DEBUG(dbgs() << " masked: " << *Old << "\n");
2237 V = IRB.CreateOr(Old, V, Name + ".insert");
2238 DEBUG(dbgs() << " inserted: " << *V << "\n");
2243 static Value *extractVector(IRBuilderTy &IRB, Value *V, unsigned BeginIndex,
2244 unsigned EndIndex, const Twine &Name) {
2245 VectorType *VecTy = cast<VectorType>(V->getType());
2246 unsigned NumElements = EndIndex - BeginIndex;
2247 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2249 if (NumElements == VecTy->getNumElements())
2252 if (NumElements == 1) {
2253 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2255 DEBUG(dbgs() << " extract: " << *V << "\n");
2259 SmallVector<Constant *, 8> Mask;
2260 Mask.reserve(NumElements);
2261 for (unsigned i = BeginIndex; i != EndIndex; ++i)
2262 Mask.push_back(IRB.getInt32(i));
2263 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2264 ConstantVector::get(Mask), Name + ".extract");
2265 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2269 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
2270 unsigned BeginIndex, const Twine &Name) {
2271 VectorType *VecTy = cast<VectorType>(Old->getType());
2272 assert(VecTy && "Can only insert a vector into a vector");
2274 VectorType *Ty = dyn_cast<VectorType>(V->getType());
2276 // Single element to insert.
2277 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2279 DEBUG(dbgs() << " insert: " << *V << "\n");
2283 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
2284 "Too many elements!");
2285 if (Ty->getNumElements() == VecTy->getNumElements()) {
2286 assert(V->getType() == VecTy && "Vector type mismatch");
2289 unsigned EndIndex = BeginIndex + Ty->getNumElements();
2291 // When inserting a smaller vector into the larger to store, we first
2292 // use a shuffle vector to widen it with undef elements, and then
2293 // a second shuffle vector to select between the loaded vector and the
2295 SmallVector<Constant *, 8> Mask;
2296 Mask.reserve(VecTy->getNumElements());
2297 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2298 if (i >= BeginIndex && i < EndIndex)
2299 Mask.push_back(IRB.getInt32(i - BeginIndex));
2301 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2302 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2303 ConstantVector::get(Mask), Name + ".expand");
2304 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2307 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2308 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
2310 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
2312 DEBUG(dbgs() << " blend: " << *V << "\n");
2317 /// \brief Visitor to rewrite instructions using p particular slice of an alloca
2318 /// to use a new alloca.
2320 /// Also implements the rewriting to vector-based accesses when the partition
2321 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2323 class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
2324 // Befriend the base class so it can delegate to private visit methods.
2325 friend class llvm::InstVisitor<AllocaSliceRewriter, bool>;
2326 typedef llvm::InstVisitor<AllocaSliceRewriter, bool> Base;
2328 const DataLayout &DL;
2331 AllocaInst &OldAI, &NewAI;
2332 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2335 // This is a convenience and flag variable that will be null unless the new
2336 // alloca's integer operations should be widened to this integer type due to
2337 // passing isIntegerWideningViable above. If it is non-null, the desired
2338 // integer type will be stored here for easy access during rewriting.
2341 // If we are rewriting an alloca partition which can be written as pure
2342 // vector operations, we stash extra information here. When VecTy is
2343 // non-null, we have some strict guarantees about the rewritten alloca:
2344 // - The new alloca is exactly the size of the vector type here.
2345 // - The accesses all either map to the entire vector or to a single
2347 // - The set of accessing instructions is only one of those handled above
2348 // in isVectorPromotionViable. Generally these are the same access kinds
2349 // which are promotable via mem2reg.
2352 uint64_t ElementSize;
2354 // The original offset of the slice currently being rewritten relative to
2355 // the original alloca.
2356 uint64_t BeginOffset, EndOffset;
2357 // The new offsets of the slice currently being rewritten relative to the
2359 uint64_t NewBeginOffset, NewEndOffset;
2365 Instruction *OldPtr;
2367 // Track post-rewrite users which are PHI nodes and Selects.
2368 SmallPtrSetImpl<PHINode *> &PHIUsers;
2369 SmallPtrSetImpl<SelectInst *> &SelectUsers;
2371 // Utility IR builder, whose name prefix is setup for each visited use, and
2372 // the insertion point is set to point to the user.
2376 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &AS, SROA &Pass,
2377 AllocaInst &OldAI, AllocaInst &NewAI,
2378 uint64_t NewAllocaBeginOffset,
2379 uint64_t NewAllocaEndOffset, bool IsIntegerPromotable,
2380 VectorType *PromotableVecTy,
2381 SmallPtrSetImpl<PHINode *> &PHIUsers,
2382 SmallPtrSetImpl<SelectInst *> &SelectUsers)
2383 : DL(DL), AS(AS), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2384 NewAllocaBeginOffset(NewAllocaBeginOffset),
2385 NewAllocaEndOffset(NewAllocaEndOffset),
2386 NewAllocaTy(NewAI.getAllocatedType()),
2387 IntTy(IsIntegerPromotable
2390 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2392 VecTy(PromotableVecTy),
2393 ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2394 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
2395 BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
2396 OldPtr(), PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2397 IRB(NewAI.getContext(), ConstantFolder()) {
2399 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
2400 "Only multiple-of-8 sized vector elements are viable");
2403 assert((!IntTy && !VecTy) || (IntTy && !VecTy) || (!IntTy && VecTy));
2406 bool visit(AllocaSlices::const_iterator I) {
2407 bool CanSROA = true;
2408 BeginOffset = I->beginOffset();
2409 EndOffset = I->endOffset();
2410 IsSplittable = I->isSplittable();
2412 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2413 DEBUG(dbgs() << " rewriting " << (IsSplit ? "split " : ""));
2414 DEBUG(AS.printSlice(dbgs(), I, ""));
2416 // Compute the intersecting offset range.
2417 assert(BeginOffset < NewAllocaEndOffset);
2418 assert(EndOffset > NewAllocaBeginOffset);
2419 NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2420 NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2422 SliceSize = NewEndOffset - NewBeginOffset;
2424 OldUse = I->getUse();
2425 OldPtr = cast<Instruction>(OldUse->get());
2427 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2428 IRB.SetInsertPoint(OldUserI);
2429 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2430 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2432 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2439 // Make sure the other visit overloads are visible.
2442 // Every instruction which can end up as a user must have a rewrite rule.
2443 bool visitInstruction(Instruction &I) {
2444 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2445 llvm_unreachable("No rewrite rule for this instruction!");
2448 Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2449 // Note that the offset computation can use BeginOffset or NewBeginOffset
2450 // interchangeably for unsplit slices.
2451 assert(IsSplit || BeginOffset == NewBeginOffset);
2452 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2455 StringRef OldName = OldPtr->getName();
2456 // Skip through the last '.sroa.' component of the name.
2457 size_t LastSROAPrefix = OldName.rfind(".sroa.");
2458 if (LastSROAPrefix != StringRef::npos) {
2459 OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2460 // Look for an SROA slice index.
2461 size_t IndexEnd = OldName.find_first_not_of("0123456789");
2462 if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2463 // Strip the index and look for the offset.
2464 OldName = OldName.substr(IndexEnd + 1);
2465 size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2466 if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2467 // Strip the offset.
2468 OldName = OldName.substr(OffsetEnd + 1);
2471 // Strip any SROA suffixes as well.
2472 OldName = OldName.substr(0, OldName.find(".sroa_"));
2475 return getAdjustedPtr(IRB, DL, &NewAI,
2476 APInt(DL.getPointerSizeInBits(), Offset), PointerTy,
2478 Twine(OldName) + "."
2485 /// \brief Compute suitable alignment to access this slice of the *new*
2488 /// You can optionally pass a type to this routine and if that type's ABI
2489 /// alignment is itself suitable, this will return zero.
2490 unsigned getSliceAlign(Type *Ty = nullptr) {
2491 unsigned NewAIAlign = NewAI.getAlignment();
2493 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2495 MinAlign(NewAIAlign, NewBeginOffset - NewAllocaBeginOffset);
2496 return (Ty && Align == DL.getABITypeAlignment(Ty)) ? 0 : Align;
2499 unsigned getIndex(uint64_t Offset) {
2500 assert(VecTy && "Can only call getIndex when rewriting a vector");
2501 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2502 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2503 uint32_t Index = RelOffset / ElementSize;
2504 assert(Index * ElementSize == RelOffset);
2508 void deleteIfTriviallyDead(Value *V) {
2509 Instruction *I = cast<Instruction>(V);
2510 if (isInstructionTriviallyDead(I))
2511 Pass.DeadInsts.insert(I);
2514 Value *rewriteVectorizedLoadInst() {
2515 unsigned BeginIndex = getIndex(NewBeginOffset);
2516 unsigned EndIndex = getIndex(NewEndOffset);
2517 assert(EndIndex > BeginIndex && "Empty vector!");
2519 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2520 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2523 Value *rewriteIntegerLoad(LoadInst &LI) {
2524 assert(IntTy && "We cannot insert an integer to the alloca");
2525 assert(!LI.isVolatile());
2526 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2527 V = convertValue(DL, IRB, V, IntTy);
2528 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2529 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2530 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset)
2531 V = extractInteger(DL, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2536 bool visitLoadInst(LoadInst &LI) {
2537 DEBUG(dbgs() << " original: " << LI << "\n");
2538 Value *OldOp = LI.getOperand(0);
2539 assert(OldOp == OldPtr);
2541 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2543 bool IsPtrAdjusted = false;
2546 V = rewriteVectorizedLoadInst();
2547 } else if (IntTy && LI.getType()->isIntegerTy()) {
2548 V = rewriteIntegerLoad(LI);
2549 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2550 canConvertValue(DL, NewAllocaTy, LI.getType())) {
2551 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), LI.isVolatile(),
2554 Type *LTy = TargetTy->getPointerTo();
2555 V = IRB.CreateAlignedLoad(getNewAllocaSlicePtr(IRB, LTy),
2556 getSliceAlign(TargetTy), LI.isVolatile(),
2558 IsPtrAdjusted = true;
2560 V = convertValue(DL, IRB, V, TargetTy);
2563 assert(!LI.isVolatile());
2564 assert(LI.getType()->isIntegerTy() &&
2565 "Only integer type loads and stores are split");
2566 assert(SliceSize < DL.getTypeStoreSize(LI.getType()) &&
2567 "Split load isn't smaller than original load");
2568 assert(LI.getType()->getIntegerBitWidth() ==
2569 DL.getTypeStoreSizeInBits(LI.getType()) &&
2570 "Non-byte-multiple bit width");
2571 // Move the insertion point just past the load so that we can refer to it.
2572 IRB.SetInsertPoint(std::next(BasicBlock::iterator(&LI)));
2573 // Create a placeholder value with the same type as LI to use as the
2574 // basis for the new value. This allows us to replace the uses of LI with
2575 // the computed value, and then replace the placeholder with LI, leaving
2576 // LI only used for this computation.
2577 Value *Placeholder =
2578 new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2579 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset, "insert");
2580 LI.replaceAllUsesWith(V);
2581 Placeholder->replaceAllUsesWith(&LI);
2584 LI.replaceAllUsesWith(V);
2587 Pass.DeadInsts.insert(&LI);
2588 deleteIfTriviallyDead(OldOp);
2589 DEBUG(dbgs() << " to: " << *V << "\n");
2590 return !LI.isVolatile() && !IsPtrAdjusted;
2593 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp) {
2594 if (V->getType() != VecTy) {
2595 unsigned BeginIndex = getIndex(NewBeginOffset);
2596 unsigned EndIndex = getIndex(NewEndOffset);
2597 assert(EndIndex > BeginIndex && "Empty vector!");
2598 unsigned NumElements = EndIndex - BeginIndex;
2599 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2600 Type *SliceTy = (NumElements == 1)
2602 : VectorType::get(ElementTy, NumElements);
2603 if (V->getType() != SliceTy)
2604 V = convertValue(DL, IRB, V, SliceTy);
2606 // Mix in the existing elements.
2607 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2608 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2610 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2611 Pass.DeadInsts.insert(&SI);
2614 DEBUG(dbgs() << " to: " << *Store << "\n");
2618 bool rewriteIntegerStore(Value *V, StoreInst &SI) {
2619 assert(IntTy && "We cannot extract an integer from the alloca");
2620 assert(!SI.isVolatile());
2621 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2623 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2624 Old = convertValue(DL, IRB, Old, IntTy);
2625 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2626 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2627 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset, "insert");
2629 V = convertValue(DL, IRB, V, NewAllocaTy);
2630 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2631 Pass.DeadInsts.insert(&SI);
2633 DEBUG(dbgs() << " to: " << *Store << "\n");
2637 bool visitStoreInst(StoreInst &SI) {
2638 DEBUG(dbgs() << " original: " << SI << "\n");
2639 Value *OldOp = SI.getOperand(1);
2640 assert(OldOp == OldPtr);
2642 Value *V = SI.getValueOperand();
2644 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2645 // alloca that should be re-examined after promoting this alloca.
2646 if (V->getType()->isPointerTy())
2647 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2648 Pass.PostPromotionWorklist.insert(AI);
2650 if (SliceSize < DL.getTypeStoreSize(V->getType())) {
2651 assert(!SI.isVolatile());
2652 assert(V->getType()->isIntegerTy() &&
2653 "Only integer type loads and stores are split");
2654 assert(V->getType()->getIntegerBitWidth() ==
2655 DL.getTypeStoreSizeInBits(V->getType()) &&
2656 "Non-byte-multiple bit width");
2657 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2658 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset, "extract");
2662 return rewriteVectorizedStoreInst(V, SI, OldOp);
2663 if (IntTy && V->getType()->isIntegerTy())
2664 return rewriteIntegerStore(V, SI);
2667 if (NewBeginOffset == NewAllocaBeginOffset &&
2668 NewEndOffset == NewAllocaEndOffset &&
2669 canConvertValue(DL, V->getType(), NewAllocaTy)) {
2670 V = convertValue(DL, IRB, V, NewAllocaTy);
2671 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2674 Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo());
2675 NewSI = IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(V->getType()),
2679 Pass.DeadInsts.insert(&SI);
2680 deleteIfTriviallyDead(OldOp);
2682 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2683 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2686 /// \brief Compute an integer value from splatting an i8 across the given
2687 /// number of bytes.
2689 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2690 /// call this routine.
2691 /// FIXME: Heed the advice above.
2693 /// \param V The i8 value to splat.
2694 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2695 Value *getIntegerSplat(Value *V, unsigned Size) {
2696 assert(Size > 0 && "Expected a positive number of bytes.");
2697 IntegerType *VTy = cast<IntegerType>(V->getType());
2698 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2702 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size * 8);
2704 IRB.CreateZExt(V, SplatIntTy, "zext"),
2705 ConstantExpr::getUDiv(
2706 Constant::getAllOnesValue(SplatIntTy),
2707 ConstantExpr::getZExt(Constant::getAllOnesValue(V->getType()),
2713 /// \brief Compute a vector splat for a given element value.
2714 Value *getVectorSplat(Value *V, unsigned NumElements) {
2715 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2716 DEBUG(dbgs() << " splat: " << *V << "\n");
2720 bool visitMemSetInst(MemSetInst &II) {
2721 DEBUG(dbgs() << " original: " << II << "\n");
2722 assert(II.getRawDest() == OldPtr);
2724 // If the memset has a variable size, it cannot be split, just adjust the
2725 // pointer to the new alloca.
2726 if (!isa<Constant>(II.getLength())) {
2728 assert(NewBeginOffset == BeginOffset);
2729 II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2730 Type *CstTy = II.getAlignmentCst()->getType();
2731 II.setAlignment(ConstantInt::get(CstTy, getSliceAlign()));
2733 deleteIfTriviallyDead(OldPtr);
2737 // Record this instruction for deletion.
2738 Pass.DeadInsts.insert(&II);
2740 Type *AllocaTy = NewAI.getAllocatedType();
2741 Type *ScalarTy = AllocaTy->getScalarType();
2743 // If this doesn't map cleanly onto the alloca type, and that type isn't
2744 // a single value type, just emit a memset.
2745 if (!VecTy && !IntTy &&
2746 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2747 SliceSize != DL.getTypeStoreSize(AllocaTy) ||
2748 !AllocaTy->isSingleValueType() ||
2749 !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2750 DL.getTypeSizeInBits(ScalarTy) % 8 != 0)) {
2751 Type *SizeTy = II.getLength()->getType();
2752 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2753 CallInst *New = IRB.CreateMemSet(
2754 getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2755 getSliceAlign(), II.isVolatile());
2757 DEBUG(dbgs() << " to: " << *New << "\n");
2761 // If we can represent this as a simple value, we have to build the actual
2762 // value to store, which requires expanding the byte present in memset to
2763 // a sensible representation for the alloca type. This is essentially
2764 // splatting the byte to a sufficiently wide integer, splatting it across
2765 // any desired vector width, and bitcasting to the final type.
2769 // If this is a memset of a vectorized alloca, insert it.
2770 assert(ElementTy == ScalarTy);
2772 unsigned BeginIndex = getIndex(NewBeginOffset);
2773 unsigned EndIndex = getIndex(NewEndOffset);
2774 assert(EndIndex > BeginIndex && "Empty vector!");
2775 unsigned NumElements = EndIndex - BeginIndex;
2776 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2779 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2780 Splat = convertValue(DL, IRB, Splat, ElementTy);
2781 if (NumElements > 1)
2782 Splat = getVectorSplat(Splat, NumElements);
2785 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2786 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2788 // If this is a memset on an alloca where we can widen stores, insert the
2790 assert(!II.isVolatile());
2792 uint64_t Size = NewEndOffset - NewBeginOffset;
2793 V = getIntegerSplat(II.getValue(), Size);
2795 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2796 EndOffset != NewAllocaBeginOffset)) {
2798 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2799 Old = convertValue(DL, IRB, Old, IntTy);
2800 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2801 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2803 assert(V->getType() == IntTy &&
2804 "Wrong type for an alloca wide integer!");
2806 V = convertValue(DL, IRB, V, AllocaTy);
2808 // Established these invariants above.
2809 assert(NewBeginOffset == NewAllocaBeginOffset);
2810 assert(NewEndOffset == NewAllocaEndOffset);
2812 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2813 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2814 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2816 V = convertValue(DL, IRB, V, AllocaTy);
2819 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2822 DEBUG(dbgs() << " to: " << *New << "\n");
2823 return !II.isVolatile();
2826 bool visitMemTransferInst(MemTransferInst &II) {
2827 // Rewriting of memory transfer instructions can be a bit tricky. We break
2828 // them into two categories: split intrinsics and unsplit intrinsics.
2830 DEBUG(dbgs() << " original: " << II << "\n");
2832 bool IsDest = &II.getRawDestUse() == OldUse;
2833 assert((IsDest && II.getRawDest() == OldPtr) ||
2834 (!IsDest && II.getRawSource() == OldPtr));
2836 unsigned SliceAlign = getSliceAlign();
2838 // For unsplit intrinsics, we simply modify the source and destination
2839 // pointers in place. This isn't just an optimization, it is a matter of
2840 // correctness. With unsplit intrinsics we may be dealing with transfers
2841 // within a single alloca before SROA ran, or with transfers that have
2842 // a variable length. We may also be dealing with memmove instead of
2843 // memcpy, and so simply updating the pointers is the necessary for us to
2844 // update both source and dest of a single call.
2845 if (!IsSplittable) {
2846 Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2848 II.setDest(AdjustedPtr);
2850 II.setSource(AdjustedPtr);
2852 if (II.getAlignment() > SliceAlign) {
2853 Type *CstTy = II.getAlignmentCst()->getType();
2855 ConstantInt::get(CstTy, MinAlign(II.getAlignment(), SliceAlign)));
2858 DEBUG(dbgs() << " to: " << II << "\n");
2859 deleteIfTriviallyDead(OldPtr);
2862 // For split transfer intrinsics we have an incredibly useful assurance:
2863 // the source and destination do not reside within the same alloca, and at
2864 // least one of them does not escape. This means that we can replace
2865 // memmove with memcpy, and we don't need to worry about all manner of
2866 // downsides to splitting and transforming the operations.
2868 // If this doesn't map cleanly onto the alloca type, and that type isn't
2869 // a single value type, just emit a memcpy.
2872 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2873 SliceSize != DL.getTypeStoreSize(NewAI.getAllocatedType()) ||
2874 !NewAI.getAllocatedType()->isSingleValueType());
2876 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2877 // size hasn't been shrunk based on analysis of the viable range, this is
2879 if (EmitMemCpy && &OldAI == &NewAI) {
2880 // Ensure the start lines up.
2881 assert(NewBeginOffset == BeginOffset);
2883 // Rewrite the size as needed.
2884 if (NewEndOffset != EndOffset)
2885 II.setLength(ConstantInt::get(II.getLength()->getType(),
2886 NewEndOffset - NewBeginOffset));
2889 // Record this instruction for deletion.
2890 Pass.DeadInsts.insert(&II);
2892 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2893 // alloca that should be re-examined after rewriting this instruction.
2894 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2895 if (AllocaInst *AI =
2896 dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2897 assert(AI != &OldAI && AI != &NewAI &&
2898 "Splittable transfers cannot reach the same alloca on both ends.");
2899 Pass.Worklist.insert(AI);
2902 Type *OtherPtrTy = OtherPtr->getType();
2903 unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2905 // Compute the relative offset for the other pointer within the transfer.
2906 unsigned IntPtrWidth = DL.getPointerSizeInBits(OtherAS);
2907 APInt OtherOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2908 unsigned OtherAlign = MinAlign(II.getAlignment() ? II.getAlignment() : 1,
2909 OtherOffset.zextOrTrunc(64).getZExtValue());
2912 // Compute the other pointer, folding as much as possible to produce
2913 // a single, simple GEP in most cases.
2914 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2915 OtherPtr->getName() + ".");
2917 Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2918 Type *SizeTy = II.getLength()->getType();
2919 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2921 CallInst *New = IRB.CreateMemCpy(
2922 IsDest ? OurPtr : OtherPtr, IsDest ? OtherPtr : OurPtr, Size,
2923 MinAlign(SliceAlign, OtherAlign), II.isVolatile());
2925 DEBUG(dbgs() << " to: " << *New << "\n");
2929 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2930 NewEndOffset == NewAllocaEndOffset;
2931 uint64_t Size = NewEndOffset - NewBeginOffset;
2932 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2933 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2934 unsigned NumElements = EndIndex - BeginIndex;
2935 IntegerType *SubIntTy =
2936 IntTy ? Type::getIntNTy(IntTy->getContext(), Size * 8) : nullptr;
2938 // Reset the other pointer type to match the register type we're going to
2939 // use, but using the address space of the original other pointer.
2940 if (VecTy && !IsWholeAlloca) {
2941 if (NumElements == 1)
2942 OtherPtrTy = VecTy->getElementType();
2944 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2946 OtherPtrTy = OtherPtrTy->getPointerTo(OtherAS);
2947 } else if (IntTy && !IsWholeAlloca) {
2948 OtherPtrTy = SubIntTy->getPointerTo(OtherAS);
2950 OtherPtrTy = NewAllocaTy->getPointerTo(OtherAS);
2953 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2954 OtherPtr->getName() + ".");
2955 unsigned SrcAlign = OtherAlign;
2956 Value *DstPtr = &NewAI;
2957 unsigned DstAlign = SliceAlign;
2959 std::swap(SrcPtr, DstPtr);
2960 std::swap(SrcAlign, DstAlign);
2964 if (VecTy && !IsWholeAlloca && !IsDest) {
2965 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2966 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2967 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2968 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2969 Src = convertValue(DL, IRB, Src, IntTy);
2970 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2971 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2974 IRB.CreateAlignedLoad(SrcPtr, SrcAlign, II.isVolatile(), "copyload");
2977 if (VecTy && !IsWholeAlloca && IsDest) {
2979 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2980 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2981 } else if (IntTy && !IsWholeAlloca && IsDest) {
2983 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2984 Old = convertValue(DL, IRB, Old, IntTy);
2985 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2986 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
2987 Src = convertValue(DL, IRB, Src, NewAllocaTy);
2990 StoreInst *Store = cast<StoreInst>(
2991 IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
2993 DEBUG(dbgs() << " to: " << *Store << "\n");
2994 return !II.isVolatile();
2997 bool visitIntrinsicInst(IntrinsicInst &II) {
2998 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2999 II.getIntrinsicID() == Intrinsic::lifetime_end);
3000 DEBUG(dbgs() << " original: " << II << "\n");
3001 assert(II.getArgOperand(1) == OldPtr);
3003 // Record this instruction for deletion.
3004 Pass.DeadInsts.insert(&II);
3007 ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
3008 NewEndOffset - NewBeginOffset);
3009 Value *Ptr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3011 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
3012 New = IRB.CreateLifetimeStart(Ptr, Size);
3014 New = IRB.CreateLifetimeEnd(Ptr, Size);
3017 DEBUG(dbgs() << " to: " << *New << "\n");
3021 bool visitPHINode(PHINode &PN) {
3022 DEBUG(dbgs() << " original: " << PN << "\n");
3023 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
3024 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
3026 // We would like to compute a new pointer in only one place, but have it be
3027 // as local as possible to the PHI. To do that, we re-use the location of
3028 // the old pointer, which necessarily must be in the right position to
3029 // dominate the PHI.
3030 IRBuilderTy PtrBuilder(IRB);
3031 if (isa<PHINode>(OldPtr))
3032 PtrBuilder.SetInsertPoint(OldPtr->getParent()->getFirstInsertionPt());
3034 PtrBuilder.SetInsertPoint(OldPtr);
3035 PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
3037 Value *NewPtr = getNewAllocaSlicePtr(PtrBuilder, OldPtr->getType());
3038 // Replace the operands which were using the old pointer.
3039 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
3041 DEBUG(dbgs() << " to: " << PN << "\n");
3042 deleteIfTriviallyDead(OldPtr);
3044 // PHIs can't be promoted on their own, but often can be speculated. We
3045 // check the speculation outside of the rewriter so that we see the
3046 // fully-rewritten alloca.
3047 PHIUsers.insert(&PN);
3051 bool visitSelectInst(SelectInst &SI) {
3052 DEBUG(dbgs() << " original: " << SI << "\n");
3053 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
3054 "Pointer isn't an operand!");
3055 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
3056 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
3058 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3059 // Replace the operands which were using the old pointer.
3060 if (SI.getOperand(1) == OldPtr)
3061 SI.setOperand(1, NewPtr);
3062 if (SI.getOperand(2) == OldPtr)
3063 SI.setOperand(2, NewPtr);
3065 DEBUG(dbgs() << " to: " << SI << "\n");
3066 deleteIfTriviallyDead(OldPtr);
3068 // Selects can't be promoted on their own, but often can be speculated. We
3069 // check the speculation outside of the rewriter so that we see the
3070 // fully-rewritten alloca.
3071 SelectUsers.insert(&SI);
3078 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
3080 /// This pass aggressively rewrites all aggregate loads and stores on
3081 /// a particular pointer (or any pointer derived from it which we can identify)
3082 /// with scalar loads and stores.
3083 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
3084 // Befriend the base class so it can delegate to private visit methods.
3085 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
3087 const DataLayout &DL;
3089 /// Queue of pointer uses to analyze and potentially rewrite.
3090 SmallVector<Use *, 8> Queue;
3092 /// Set to prevent us from cycling with phi nodes and loops.
3093 SmallPtrSet<User *, 8> Visited;
3095 /// The current pointer use being rewritten. This is used to dig up the used
3096 /// value (as opposed to the user).
3100 AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
3102 /// Rewrite loads and stores through a pointer and all pointers derived from
3104 bool rewrite(Instruction &I) {
3105 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
3107 bool Changed = false;
3108 while (!Queue.empty()) {
3109 U = Queue.pop_back_val();
3110 Changed |= visit(cast<Instruction>(U->getUser()));
3116 /// Enqueue all the users of the given instruction for further processing.
3117 /// This uses a set to de-duplicate users.
3118 void enqueueUsers(Instruction &I) {
3119 for (Use &U : I.uses())
3120 if (Visited.insert(U.getUser()).second)
3121 Queue.push_back(&U);
3124 // Conservative default is to not rewrite anything.
3125 bool visitInstruction(Instruction &I) { return false; }
3127 /// \brief Generic recursive split emission class.
3128 template <typename Derived> class OpSplitter {
3130 /// The builder used to form new instructions.
3132 /// The indices which to be used with insert- or extractvalue to select the
3133 /// appropriate value within the aggregate.
3134 SmallVector<unsigned, 4> Indices;
3135 /// The indices to a GEP instruction which will move Ptr to the correct slot
3136 /// within the aggregate.
3137 SmallVector<Value *, 4> GEPIndices;
3138 /// The base pointer of the original op, used as a base for GEPing the
3139 /// split operations.
3142 /// Initialize the splitter with an insertion point, Ptr and start with a
3143 /// single zero GEP index.
3144 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
3145 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
3148 /// \brief Generic recursive split emission routine.
3150 /// This method recursively splits an aggregate op (load or store) into
3151 /// scalar or vector ops. It splits recursively until it hits a single value
3152 /// and emits that single value operation via the template argument.
3154 /// The logic of this routine relies on GEPs and insertvalue and
3155 /// extractvalue all operating with the same fundamental index list, merely
3156 /// formatted differently (GEPs need actual values).
3158 /// \param Ty The type being split recursively into smaller ops.
3159 /// \param Agg The aggregate value being built up or stored, depending on
3160 /// whether this is splitting a load or a store respectively.
3161 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3162 if (Ty->isSingleValueType())
3163 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
3165 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3166 unsigned OldSize = Indices.size();
3168 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3170 assert(Indices.size() == OldSize && "Did not return to the old size");
3171 Indices.push_back(Idx);
3172 GEPIndices.push_back(IRB.getInt32(Idx));
3173 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3174 GEPIndices.pop_back();
3180 if (StructType *STy = dyn_cast<StructType>(Ty)) {
3181 unsigned OldSize = Indices.size();
3183 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3185 assert(Indices.size() == OldSize && "Did not return to the old size");
3186 Indices.push_back(Idx);
3187 GEPIndices.push_back(IRB.getInt32(Idx));
3188 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3189 GEPIndices.pop_back();
3195 llvm_unreachable("Only arrays and structs are aggregate loadable types");
3199 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3200 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3201 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
3203 /// Emit a leaf load of a single value. This is called at the leaves of the
3204 /// recursive emission to actually load values.
3205 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3206 assert(Ty->isSingleValueType());
3207 // Load the single value and insert it using the indices.
3208 Value *GEP = IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep");
3209 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
3210 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3211 DEBUG(dbgs() << " to: " << *Load << "\n");
3215 bool visitLoadInst(LoadInst &LI) {
3216 assert(LI.getPointerOperand() == *U);
3217 if (!LI.isSimple() || LI.getType()->isSingleValueType())
3220 // We have an aggregate being loaded, split it apart.
3221 DEBUG(dbgs() << " original: " << LI << "\n");
3222 LoadOpSplitter Splitter(&LI, *U);
3223 Value *V = UndefValue::get(LI.getType());
3224 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3225 LI.replaceAllUsesWith(V);
3226 LI.eraseFromParent();
3230 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3231 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3232 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
3234 /// Emit a leaf store of a single value. This is called at the leaves of the
3235 /// recursive emission to actually produce stores.
3236 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3237 assert(Ty->isSingleValueType());
3238 // Extract the single value and store it using the indices.
3239 Value *Store = IRB.CreateStore(
3240 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
3241 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
3243 DEBUG(dbgs() << " to: " << *Store << "\n");
3247 bool visitStoreInst(StoreInst &SI) {
3248 if (!SI.isSimple() || SI.getPointerOperand() != *U)
3250 Value *V = SI.getValueOperand();
3251 if (V->getType()->isSingleValueType())
3254 // We have an aggregate being stored, split it apart.
3255 DEBUG(dbgs() << " original: " << SI << "\n");
3256 StoreOpSplitter Splitter(&SI, *U);
3257 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3258 SI.eraseFromParent();
3262 bool visitBitCastInst(BitCastInst &BC) {
3267 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3272 bool visitPHINode(PHINode &PN) {
3277 bool visitSelectInst(SelectInst &SI) {
3284 /// \brief Strip aggregate type wrapping.
3286 /// This removes no-op aggregate types wrapping an underlying type. It will
3287 /// strip as many layers of types as it can without changing either the type
3288 /// size or the allocated size.
3289 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
3290 if (Ty->isSingleValueType())
3293 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3294 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
3297 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3298 InnerTy = ArrTy->getElementType();
3299 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3300 const StructLayout *SL = DL.getStructLayout(STy);
3301 unsigned Index = SL->getElementContainingOffset(0);
3302 InnerTy = STy->getElementType(Index);
3307 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3308 TypeSize > DL.getTypeSizeInBits(InnerTy))
3311 return stripAggregateTypeWrapping(DL, InnerTy);
3314 /// \brief Try to find a partition of the aggregate type passed in for a given
3315 /// offset and size.
3317 /// This recurses through the aggregate type and tries to compute a subtype
3318 /// based on the offset and size. When the offset and size span a sub-section
3319 /// of an array, it will even compute a new array type for that sub-section,
3320 /// and the same for structs.
3322 /// Note that this routine is very strict and tries to find a partition of the
3323 /// type which produces the *exact* right offset and size. It is not forgiving
3324 /// when the size or offset cause either end of type-based partition to be off.
3325 /// Also, this is a best-effort routine. It is reasonable to give up and not
3326 /// return a type if necessary.
3327 static Type *getTypePartition(const DataLayout &DL, Type *Ty, uint64_t Offset,
3329 if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
3330 return stripAggregateTypeWrapping(DL, Ty);
3331 if (Offset > DL.getTypeAllocSize(Ty) ||
3332 (DL.getTypeAllocSize(Ty) - Offset) < Size)
3335 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3336 // We can't partition pointers...
3337 if (SeqTy->isPointerTy())
3340 Type *ElementTy = SeqTy->getElementType();
3341 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3342 uint64_t NumSkippedElements = Offset / ElementSize;
3343 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) {
3344 if (NumSkippedElements >= ArrTy->getNumElements())
3346 } else if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) {
3347 if (NumSkippedElements >= VecTy->getNumElements())
3350 Offset -= NumSkippedElements * ElementSize;
3352 // First check if we need to recurse.
3353 if (Offset > 0 || Size < ElementSize) {
3354 // Bail if the partition ends in a different array element.
3355 if ((Offset + Size) > ElementSize)
3357 // Recurse through the element type trying to peel off offset bytes.
3358 return getTypePartition(DL, ElementTy, Offset, Size);
3360 assert(Offset == 0);
3362 if (Size == ElementSize)
3363 return stripAggregateTypeWrapping(DL, ElementTy);
3364 assert(Size > ElementSize);
3365 uint64_t NumElements = Size / ElementSize;
3366 if (NumElements * ElementSize != Size)
3368 return ArrayType::get(ElementTy, NumElements);
3371 StructType *STy = dyn_cast<StructType>(Ty);
3375 const StructLayout *SL = DL.getStructLayout(STy);
3376 if (Offset >= SL->getSizeInBytes())
3378 uint64_t EndOffset = Offset + Size;
3379 if (EndOffset > SL->getSizeInBytes())
3382 unsigned Index = SL->getElementContainingOffset(Offset);
3383 Offset -= SL->getElementOffset(Index);
3385 Type *ElementTy = STy->getElementType(Index);
3386 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3387 if (Offset >= ElementSize)
3388 return nullptr; // The offset points into alignment padding.
3390 // See if any partition must be contained by the element.
3391 if (Offset > 0 || Size < ElementSize) {
3392 if ((Offset + Size) > ElementSize)
3394 return getTypePartition(DL, ElementTy, Offset, Size);
3396 assert(Offset == 0);
3398 if (Size == ElementSize)
3399 return stripAggregateTypeWrapping(DL, ElementTy);
3401 StructType::element_iterator EI = STy->element_begin() + Index,
3402 EE = STy->element_end();
3403 if (EndOffset < SL->getSizeInBytes()) {
3404 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3405 if (Index == EndIndex)
3406 return nullptr; // Within a single element and its padding.
3408 // Don't try to form "natural" types if the elements don't line up with the
3410 // FIXME: We could potentially recurse down through the last element in the
3411 // sub-struct to find a natural end point.
3412 if (SL->getElementOffset(EndIndex) != EndOffset)
3415 assert(Index < EndIndex);
3416 EE = STy->element_begin() + EndIndex;
3419 // Try to build up a sub-structure.
3421 StructType::get(STy->getContext(), makeArrayRef(EI, EE), STy->isPacked());
3422 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3423 if (Size != SubSL->getSizeInBytes())
3424 return nullptr; // The sub-struct doesn't have quite the size needed.
3429 /// \brief Rewrite an alloca partition's users.
3431 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3432 /// to rewrite uses of an alloca partition to be conducive for SSA value
3433 /// promotion. If the partition needs a new, more refined alloca, this will
3434 /// build that new alloca, preserving as much type information as possible, and
3435 /// rewrite the uses of the old alloca to point at the new one and have the
3436 /// appropriate new offsets. It also evaluates how successful the rewrite was
3437 /// at enabling promotion and if it was successful queues the alloca to be
3439 bool SROA::rewritePartition(AllocaInst &AI, AllocaSlices &AS,
3440 AllocaSlices::Partition &P) {
3441 // Try to compute a friendly type for this partition of the alloca. This
3442 // won't always succeed, in which case we fall back to a legal integer type
3443 // or an i8 array of an appropriate size.
3444 Type *SliceTy = nullptr;
3445 if (Type *CommonUseTy = findCommonType(P.begin(), P.end(), P.endOffset()))
3446 if (DL->getTypeAllocSize(CommonUseTy) >= P.size())
3447 SliceTy = CommonUseTy;
3449 if (Type *TypePartitionTy = getTypePartition(*DL, AI.getAllocatedType(),
3450 P.beginOffset(), P.size()))
3451 SliceTy = TypePartitionTy;
3452 if ((!SliceTy || (SliceTy->isArrayTy() &&
3453 SliceTy->getArrayElementType()->isIntegerTy())) &&
3454 DL->isLegalInteger(P.size() * 8))
3455 SliceTy = Type::getIntNTy(*C, P.size() * 8);
3457 SliceTy = ArrayType::get(Type::getInt8Ty(*C), P.size());
3458 assert(DL->getTypeAllocSize(SliceTy) >= P.size());
3460 bool IsIntegerPromotable = isIntegerWideningViable(P, SliceTy, *DL);
3463 IsIntegerPromotable ? nullptr : isVectorPromotionViable(P, *DL);
3467 // Check for the case where we're going to rewrite to a new alloca of the
3468 // exact same type as the original, and with the same access offsets. In that
3469 // case, re-use the existing alloca, but still run through the rewriter to
3470 // perform phi and select speculation.
3472 if (SliceTy == AI.getAllocatedType()) {
3473 assert(P.beginOffset() == 0 &&
3474 "Non-zero begin offset but same alloca type");
3476 // FIXME: We should be able to bail at this point with "nothing changed".
3477 // FIXME: We might want to defer PHI speculation until after here.
3479 unsigned Alignment = AI.getAlignment();
3481 // The minimum alignment which users can rely on when the explicit
3482 // alignment is omitted or zero is that required by the ABI for this
3484 Alignment = DL->getABITypeAlignment(AI.getAllocatedType());
3486 Alignment = MinAlign(Alignment, P.beginOffset());
3487 // If we will get at least this much alignment from the type alone, leave
3488 // the alloca's alignment unconstrained.
3489 if (Alignment <= DL->getABITypeAlignment(SliceTy))
3491 NewAI = new AllocaInst(
3492 SliceTy, nullptr, Alignment,
3493 AI.getName() + ".sroa." + Twine(P.begin() - AS.begin()), &AI);
3497 DEBUG(dbgs() << "Rewriting alloca partition "
3498 << "[" << P.beginOffset() << "," << P.endOffset()
3499 << ") to: " << *NewAI << "\n");
3501 // Track the high watermark on the worklist as it is only relevant for
3502 // promoted allocas. We will reset it to this point if the alloca is not in
3503 // fact scheduled for promotion.
3504 unsigned PPWOldSize = PostPromotionWorklist.size();
3505 unsigned NumUses = 0;
3506 SmallPtrSet<PHINode *, 8> PHIUsers;
3507 SmallPtrSet<SelectInst *, 8> SelectUsers;
3509 AllocaSliceRewriter Rewriter(*DL, AS, *this, AI, *NewAI, P.beginOffset(),
3510 P.endOffset(), IsIntegerPromotable, VecTy,
3511 PHIUsers, SelectUsers);
3512 bool Promotable = true;
3513 for (Slice *S : P.splitSliceTails()) {
3514 Promotable &= Rewriter.visit(S);
3517 for (Slice &S : P) {
3518 Promotable &= Rewriter.visit(&S);
3522 NumAllocaPartitionUses += NumUses;
3523 MaxUsesPerAllocaPartition =
3524 std::max<unsigned>(NumUses, MaxUsesPerAllocaPartition);
3526 // Now that we've processed all the slices in the new partition, check if any
3527 // PHIs or Selects would block promotion.
3528 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
3531 if (!isSafePHIToSpeculate(**I, DL)) {
3534 SelectUsers.clear();
3537 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
3538 E = SelectUsers.end();
3540 if (!isSafeSelectToSpeculate(**I, DL)) {
3543 SelectUsers.clear();
3548 if (PHIUsers.empty() && SelectUsers.empty()) {
3549 // Promote the alloca.
3550 PromotableAllocas.push_back(NewAI);
3552 // If we have either PHIs or Selects to speculate, add them to those
3553 // worklists and re-queue the new alloca so that we promote in on the
3555 for (PHINode *PHIUser : PHIUsers)
3556 SpeculatablePHIs.insert(PHIUser);
3557 for (SelectInst *SelectUser : SelectUsers)
3558 SpeculatableSelects.insert(SelectUser);
3559 Worklist.insert(NewAI);
3562 // If we can't promote the alloca, iterate on it to check for new
3563 // refinements exposed by splitting the current alloca. Don't iterate on an
3564 // alloca which didn't actually change and didn't get promoted.
3566 Worklist.insert(NewAI);
3568 // Drop any post-promotion work items if promotion didn't happen.
3569 while (PostPromotionWorklist.size() > PPWOldSize)
3570 PostPromotionWorklist.pop_back();
3576 /// \brief Walks the slices of an alloca and form partitions based on them,
3577 /// rewriting each of their uses.
3578 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &AS) {
3579 if (AS.begin() == AS.end())
3582 unsigned NumPartitions = 0;
3583 bool Changed = false;
3585 // Rewrite each parttion.
3586 for (auto &P : AS.partitions()) {
3587 Changed |= rewritePartition(AI, AS, P);
3591 NumAllocaPartitions += NumPartitions;
3592 MaxPartitionsPerAlloca =
3593 std::max<unsigned>(NumPartitions, MaxPartitionsPerAlloca);
3598 /// \brief Clobber a use with undef, deleting the used value if it becomes dead.
3599 void SROA::clobberUse(Use &U) {
3601 // Replace the use with an undef value.
3602 U = UndefValue::get(OldV->getType());
3604 // Check for this making an instruction dead. We have to garbage collect
3605 // all the dead instructions to ensure the uses of any alloca end up being
3607 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3608 if (isInstructionTriviallyDead(OldI)) {
3609 DeadInsts.insert(OldI);
3613 /// \brief Analyze an alloca for SROA.
3615 /// This analyzes the alloca to ensure we can reason about it, builds
3616 /// the slices of the alloca, and then hands it off to be split and
3617 /// rewritten as needed.
3618 bool SROA::runOnAlloca(AllocaInst &AI) {
3619 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3620 ++NumAllocasAnalyzed;
3622 // Special case dead allocas, as they're trivial.
3623 if (AI.use_empty()) {
3624 AI.eraseFromParent();
3628 // Skip alloca forms that this analysis can't handle.
3629 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3630 DL->getTypeAllocSize(AI.getAllocatedType()) == 0)
3633 bool Changed = false;
3635 // First, split any FCA loads and stores touching this alloca to promote
3636 // better splitting and promotion opportunities.
3637 AggLoadStoreRewriter AggRewriter(*DL);
3638 Changed |= AggRewriter.rewrite(AI);
3640 // Build the slices using a recursive instruction-visiting builder.
3641 AllocaSlices AS(*DL, AI);
3642 DEBUG(AS.print(dbgs()));
3646 // Delete all the dead users of this alloca before splitting and rewriting it.
3647 for (Instruction *DeadUser : AS.getDeadUsers()) {
3648 // Free up everything used by this instruction.
3649 for (Use &DeadOp : DeadUser->operands())
3652 // Now replace the uses of this instruction.
3653 DeadUser->replaceAllUsesWith(UndefValue::get(DeadUser->getType()));
3655 // And mark it for deletion.
3656 DeadInsts.insert(DeadUser);
3659 for (Use *DeadOp : AS.getDeadOperands()) {
3660 clobberUse(*DeadOp);
3664 // No slices to split. Leave the dead alloca for a later pass to clean up.
3665 if (AS.begin() == AS.end())
3668 Changed |= splitAlloca(AI, AS);
3670 DEBUG(dbgs() << " Speculating PHIs\n");
3671 while (!SpeculatablePHIs.empty())
3672 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
3674 DEBUG(dbgs() << " Speculating Selects\n");
3675 while (!SpeculatableSelects.empty())
3676 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
3681 /// \brief Delete the dead instructions accumulated in this run.
3683 /// Recursively deletes the dead instructions we've accumulated. This is done
3684 /// at the very end to maximize locality of the recursive delete and to
3685 /// minimize the problems of invalidated instruction pointers as such pointers
3686 /// are used heavily in the intermediate stages of the algorithm.
3688 /// We also record the alloca instructions deleted here so that they aren't
3689 /// subsequently handed to mem2reg to promote.
3690 void SROA::deleteDeadInstructions(
3691 SmallPtrSetImpl<AllocaInst *> &DeletedAllocas) {
3692 while (!DeadInsts.empty()) {
3693 Instruction *I = DeadInsts.pop_back_val();
3694 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3696 I->replaceAllUsesWith(UndefValue::get(I->getType()));
3698 for (Use &Operand : I->operands())
3699 if (Instruction *U = dyn_cast<Instruction>(Operand)) {
3700 // Zero out the operand and see if it becomes trivially dead.
3702 if (isInstructionTriviallyDead(U))
3703 DeadInsts.insert(U);
3706 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3707 DeletedAllocas.insert(AI);
3710 I->eraseFromParent();
3714 static void enqueueUsersInWorklist(Instruction &I,
3715 SmallVectorImpl<Instruction *> &Worklist,
3716 SmallPtrSetImpl<Instruction *> &Visited) {
3717 for (User *U : I.users())
3718 if (Visited.insert(cast<Instruction>(U)).second)
3719 Worklist.push_back(cast<Instruction>(U));
3722 /// \brief Promote the allocas, using the best available technique.
3724 /// This attempts to promote whatever allocas have been identified as viable in
3725 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3726 /// If there is a domtree available, we attempt to promote using the full power
3727 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3728 /// based on the SSAUpdater utilities. This function returns whether any
3729 /// promotion occurred.
3730 bool SROA::promoteAllocas(Function &F) {
3731 if (PromotableAllocas.empty())
3734 NumPromoted += PromotableAllocas.size();
3736 if (DT && !ForceSSAUpdater) {
3737 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3738 PromoteMemToReg(PromotableAllocas, *DT, nullptr, AT);
3739 PromotableAllocas.clear();
3743 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3745 DIBuilder DIB(*F.getParent(), /*AllowUnresolved*/ false);
3746 SmallVector<Instruction *, 64> Insts;
3748 // We need a worklist to walk the uses of each alloca.
3749 SmallVector<Instruction *, 8> Worklist;
3750 SmallPtrSet<Instruction *, 8> Visited;
3751 SmallVector<Instruction *, 32> DeadInsts;
3753 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3754 AllocaInst *AI = PromotableAllocas[Idx];
3759 enqueueUsersInWorklist(*AI, Worklist, Visited);
3761 while (!Worklist.empty()) {
3762 Instruction *I = Worklist.pop_back_val();
3764 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3765 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3766 // leading to them) here. Eventually it should use them to optimize the
3767 // scalar values produced.
3768 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3769 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3770 II->getIntrinsicID() == Intrinsic::lifetime_end);
3771 II->eraseFromParent();
3775 // Push the loads and stores we find onto the list. SROA will already
3776 // have validated that all loads and stores are viable candidates for
3778 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3779 assert(LI->getType() == AI->getAllocatedType());
3780 Insts.push_back(LI);
3783 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3784 assert(SI->getValueOperand()->getType() == AI->getAllocatedType());
3785 Insts.push_back(SI);
3789 // For everything else, we know that only no-op bitcasts and GEPs will
3790 // make it this far, just recurse through them and recall them for later
3792 DeadInsts.push_back(I);
3793 enqueueUsersInWorklist(*I, Worklist, Visited);
3795 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3796 while (!DeadInsts.empty())
3797 DeadInsts.pop_back_val()->eraseFromParent();
3798 AI->eraseFromParent();
3801 PromotableAllocas.clear();
3805 bool SROA::runOnFunction(Function &F) {
3806 if (skipOptnoneFunction(F))
3809 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3810 C = &F.getContext();
3811 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
3813 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3816 DL = &DLP->getDataLayout();
3817 DominatorTreeWrapperPass *DTWP =
3818 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
3819 DT = DTWP ? &DTWP->getDomTree() : nullptr;
3820 AT = &getAnalysis<AssumptionTracker>();
3822 BasicBlock &EntryBB = F.getEntryBlock();
3823 for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
3825 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3826 Worklist.insert(AI);
3828 bool Changed = false;
3829 // A set of deleted alloca instruction pointers which should be removed from
3830 // the list of promotable allocas.
3831 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3834 while (!Worklist.empty()) {
3835 Changed |= runOnAlloca(*Worklist.pop_back_val());
3836 deleteDeadInstructions(DeletedAllocas);
3838 // Remove the deleted allocas from various lists so that we don't try to
3839 // continue processing them.
3840 if (!DeletedAllocas.empty()) {
3841 auto IsInSet = [&](AllocaInst *AI) { return DeletedAllocas.count(AI); };
3842 Worklist.remove_if(IsInSet);
3843 PostPromotionWorklist.remove_if(IsInSet);
3844 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3845 PromotableAllocas.end(),
3847 PromotableAllocas.end());
3848 DeletedAllocas.clear();
3852 Changed |= promoteAllocas(F);
3854 Worklist = PostPromotionWorklist;
3855 PostPromotionWorklist.clear();
3856 } while (!Worklist.empty());
3861 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3862 AU.addRequired<AssumptionTracker>();
3863 if (RequiresDomTree)
3864 AU.addRequired<DominatorTreeWrapperPass>();
3865 AU.setPreservesCFG();