1 //===- SROA.cpp - Scalar Replacement Of Aggregates ------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 /// This transformation implements the well known scalar replacement of
11 /// aggregates transformation. It tries to identify promotable elements of an
12 /// aggregate alloca, and promote them to registers. It will also try to
13 /// convert uses of an element (or set of elements) of an alloca into a vector
14 /// or bitfield-style integer scalar if appropriate.
16 /// It works to do this with minimal slicing of the alloca so that regions
17 /// which are merely transferred in and out of external memory remain unchanged
18 /// and are not decomposed to scalar code.
20 /// Because this also performs alloca promotion, it can be thought of as also
21 /// serving the purpose of SSA formation. The algorithm iterates on the
22 /// function until all opportunities for promotion have been realized.
24 //===----------------------------------------------------------------------===//
26 #define DEBUG_TYPE "sroa"
27 #include "llvm/Transforms/Scalar.h"
28 #include "llvm/ADT/STLExtras.h"
29 #include "llvm/ADT/SetVector.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/Analysis/Dominators.h"
33 #include "llvm/Analysis/Loads.h"
34 #include "llvm/Analysis/PtrUseVisitor.h"
35 #include "llvm/Analysis/ValueTracking.h"
36 #include "llvm/DIBuilder.h"
37 #include "llvm/DebugInfo.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/DerivedTypes.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/IRBuilder.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/IntrinsicInst.h"
45 #include "llvm/IR/LLVMContext.h"
46 #include "llvm/IR/Operator.h"
47 #include "llvm/InstVisitor.h"
48 #include "llvm/Pass.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/raw_ostream.h"
54 #include "llvm/Transforms/Utils/Local.h"
55 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
56 #include "llvm/Transforms/Utils/SSAUpdater.h"
59 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
60 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
61 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
62 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
63 STATISTIC(NumDeleted, "Number of instructions deleted");
64 STATISTIC(NumVectorized, "Number of vectorized aggregates");
66 /// Hidden option to force the pass to not use DomTree and mem2reg, instead
67 /// forming SSA values through the SSAUpdater infrastructure.
69 ForceSSAUpdater("force-ssa-updater", cl::init(false), cl::Hidden);
72 /// \brief Alloca partitioning representation.
74 /// This class represents a partitioning of an alloca into slices, and
75 /// information about the nature of uses of each slice of the alloca. The goal
76 /// is that this information is sufficient to decide if and how to split the
77 /// alloca apart and replace slices with scalars. It is also intended that this
78 /// structure can capture the relevant information needed both to decide about
79 /// and to enact these transformations.
80 class AllocaPartitioning {
82 /// \brief A common base class for representing a half-open byte range.
84 /// \brief The beginning offset of the range.
87 /// \brief The ending offset, not included in the range.
90 ByteRange() : BeginOffset(), EndOffset() {}
91 ByteRange(uint64_t BeginOffset, uint64_t EndOffset)
92 : BeginOffset(BeginOffset), EndOffset(EndOffset) {}
94 /// \brief Support for ordering ranges.
96 /// This provides an ordering over ranges such that start offsets are
97 /// always increasing, and within equal start offsets, the end offsets are
98 /// decreasing. Thus the spanning range comes first in a cluster with the
99 /// same start position.
100 bool operator<(const ByteRange &RHS) const {
101 if (BeginOffset < RHS.BeginOffset) return true;
102 if (BeginOffset > RHS.BeginOffset) return false;
103 if (EndOffset > RHS.EndOffset) return true;
107 /// \brief Support comparison with a single offset to allow binary searches.
108 friend bool operator<(const ByteRange &LHS, uint64_t RHSOffset) {
109 return LHS.BeginOffset < RHSOffset;
112 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
113 const ByteRange &RHS) {
114 return LHSOffset < RHS.BeginOffset;
117 bool operator==(const ByteRange &RHS) const {
118 return BeginOffset == RHS.BeginOffset && EndOffset == RHS.EndOffset;
120 bool operator!=(const ByteRange &RHS) const { return !operator==(RHS); }
123 /// \brief A partition of an alloca.
125 /// This structure represents a contiguous partition of the alloca. These are
126 /// formed by examining the uses of the alloca. During formation, they may
127 /// overlap but once an AllocaPartitioning is built, the Partitions within it
128 /// are all disjoint.
129 struct Partition : public ByteRange {
130 /// \brief Whether this partition is splittable into smaller partitions.
132 /// We flag partitions as splittable when they are formed entirely due to
133 /// accesses by trivially splittable operations such as memset and memcpy.
136 /// \brief Test whether a partition has been marked as dead.
137 bool isDead() const {
138 if (BeginOffset == UINT64_MAX) {
139 assert(EndOffset == UINT64_MAX);
145 /// \brief Kill a partition.
146 /// This is accomplished by setting both its beginning and end offset to
147 /// the maximum possible value.
149 assert(!isDead() && "He's Dead, Jim!");
150 BeginOffset = EndOffset = UINT64_MAX;
153 Partition() : ByteRange(), IsSplittable() {}
154 Partition(uint64_t BeginOffset, uint64_t EndOffset, bool IsSplittable)
155 : ByteRange(BeginOffset, EndOffset), IsSplittable(IsSplittable) {}
158 /// \brief A particular use of a partition of the alloca.
160 /// This structure is used to associate uses of a partition with it. They
161 /// mark the range of bytes which are referenced by a particular instruction,
162 /// and includes a handle to the user itself and the pointer value in use.
163 /// The bounds of these uses are determined by intersecting the bounds of the
164 /// memory use itself with a particular partition. As a consequence there is
165 /// intentionally overlap between various uses of the same partition.
166 struct PartitionUse : public ByteRange {
167 /// \brief The use in question. Provides access to both user and used value.
169 /// Note that this may be null if the partition use is *dead*, that is, it
170 /// should be ignored.
173 PartitionUse() : ByteRange(), U() {}
174 PartitionUse(uint64_t BeginOffset, uint64_t EndOffset, Use *U)
175 : ByteRange(BeginOffset, EndOffset), U(U) {}
178 /// \brief Construct a partitioning of a particular alloca.
180 /// Construction does most of the work for partitioning the alloca. This
181 /// performs the necessary walks of users and builds a partitioning from it.
182 AllocaPartitioning(const DataLayout &TD, AllocaInst &AI);
184 /// \brief Test whether a pointer to the allocation escapes our analysis.
186 /// If this is true, the partitioning is never fully built and should be
188 bool isEscaped() const { return PointerEscapingInstr; }
190 /// \brief Support for iterating over the partitions.
192 typedef SmallVectorImpl<Partition>::iterator iterator;
193 iterator begin() { return Partitions.begin(); }
194 iterator end() { return Partitions.end(); }
196 typedef SmallVectorImpl<Partition>::const_iterator const_iterator;
197 const_iterator begin() const { return Partitions.begin(); }
198 const_iterator end() const { return Partitions.end(); }
201 /// \brief Support for iterating over and manipulating a particular
202 /// partition's uses.
204 /// The iteration support provided for uses is more limited, but also
205 /// includes some manipulation routines to support rewriting the uses of
206 /// partitions during SROA.
208 typedef SmallVectorImpl<PartitionUse>::iterator use_iterator;
209 use_iterator use_begin(unsigned Idx) { return Uses[Idx].begin(); }
210 use_iterator use_begin(const_iterator I) { return Uses[I - begin()].begin(); }
211 use_iterator use_end(unsigned Idx) { return Uses[Idx].end(); }
212 use_iterator use_end(const_iterator I) { return Uses[I - begin()].end(); }
214 typedef SmallVectorImpl<PartitionUse>::const_iterator const_use_iterator;
215 const_use_iterator use_begin(unsigned Idx) const { return Uses[Idx].begin(); }
216 const_use_iterator use_begin(const_iterator I) const {
217 return Uses[I - begin()].begin();
219 const_use_iterator use_end(unsigned Idx) const { return Uses[Idx].end(); }
220 const_use_iterator use_end(const_iterator I) const {
221 return Uses[I - begin()].end();
224 unsigned use_size(unsigned Idx) const { return Uses[Idx].size(); }
225 unsigned use_size(const_iterator I) const { return Uses[I - begin()].size(); }
226 const PartitionUse &getUse(unsigned PIdx, unsigned UIdx) const {
227 return Uses[PIdx][UIdx];
229 const PartitionUse &getUse(const_iterator I, unsigned UIdx) const {
230 return Uses[I - begin()][UIdx];
233 void use_push_back(unsigned Idx, const PartitionUse &PU) {
234 Uses[Idx].push_back(PU);
236 void use_push_back(const_iterator I, const PartitionUse &PU) {
237 Uses[I - begin()].push_back(PU);
241 /// \brief Allow iterating the dead users for this alloca.
243 /// These are instructions which will never actually use the alloca as they
244 /// are outside the allocated range. They are safe to replace with undef and
247 typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator;
248 dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); }
249 dead_user_iterator dead_user_end() const { return DeadUsers.end(); }
252 /// \brief Allow iterating the dead expressions referring to this alloca.
254 /// These are operands which have cannot actually be used to refer to the
255 /// alloca as they are outside its range and the user doesn't correct for
256 /// that. These mostly consist of PHI node inputs and the like which we just
257 /// need to replace with undef.
259 typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator;
260 dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); }
261 dead_op_iterator dead_op_end() const { return DeadOperands.end(); }
264 /// \brief MemTransferInst auxiliary data.
265 /// This struct provides some auxiliary data about memory transfer
266 /// intrinsics such as memcpy and memmove. These intrinsics can use two
267 /// different ranges within the same alloca, and provide other challenges to
268 /// correctly represent. We stash extra data to help us untangle this
269 /// after the partitioning is complete.
270 struct MemTransferOffsets {
271 /// The destination begin and end offsets when the destination is within
272 /// this alloca. If the end offset is zero the destination is not within
274 uint64_t DestBegin, DestEnd;
276 /// The source begin and end offsets when the source is within this alloca.
277 /// If the end offset is zero, the source is not within this alloca.
278 uint64_t SourceBegin, SourceEnd;
280 /// Flag for whether an alloca is splittable.
283 MemTransferOffsets getMemTransferOffsets(MemTransferInst &II) const {
284 return MemTransferInstData.lookup(&II);
287 /// \brief Map from a PHI or select operand back to a partition.
289 /// When manipulating PHI nodes or selects, they can use more than one
290 /// partition of an alloca. We store a special mapping to allow finding the
291 /// partition referenced by each of these operands, if any.
292 iterator findPartitionForPHIOrSelectOperand(Use *U) {
293 SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt
294 = PHIOrSelectOpMap.find(U);
295 if (MapIt == PHIOrSelectOpMap.end())
298 return begin() + MapIt->second.first;
301 /// \brief Map from a PHI or select operand back to the specific use of
304 /// Similar to mapping these operands back to the partitions, this maps
305 /// directly to the use structure of that partition.
306 use_iterator findPartitionUseForPHIOrSelectOperand(Use *U) {
307 SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt
308 = PHIOrSelectOpMap.find(U);
309 assert(MapIt != PHIOrSelectOpMap.end());
310 return Uses[MapIt->second.first].begin() + MapIt->second.second;
313 /// \brief Compute a common type among the uses of a particular partition.
315 /// This routines walks all of the uses of a particular partition and tries
316 /// to find a common type between them. Untyped operations such as memset and
317 /// memcpy are ignored.
318 Type *getCommonType(iterator I) const;
320 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
321 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
322 void printUsers(raw_ostream &OS, const_iterator I,
323 StringRef Indent = " ") const;
324 void print(raw_ostream &OS) const;
325 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump(const_iterator I) const;
326 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump() const;
330 template <typename DerivedT, typename RetT = void> class BuilderBase;
331 class PartitionBuilder;
332 friend class AllocaPartitioning::PartitionBuilder;
334 friend class AllocaPartitioning::UseBuilder;
336 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
337 /// \brief Handle to alloca instruction to simplify method interfaces.
341 /// \brief The instruction responsible for this alloca having no partitioning.
343 /// When an instruction (potentially) escapes the pointer to the alloca, we
344 /// store a pointer to that here and abort trying to partition the alloca.
345 /// This will be null if the alloca is partitioned successfully.
346 Instruction *PointerEscapingInstr;
348 /// \brief The partitions of the alloca.
350 /// We store a vector of the partitions over the alloca here. This vector is
351 /// sorted by increasing begin offset, and then by decreasing end offset. See
352 /// the Partition inner class for more details. Initially (during
353 /// construction) there are overlaps, but we form a disjoint sequence of
354 /// partitions while finishing construction and a fully constructed object is
355 /// expected to always have this as a disjoint space.
356 SmallVector<Partition, 8> Partitions;
358 /// \brief The uses of the partitions.
360 /// This is essentially a mapping from each partition to a list of uses of
361 /// that partition. The mapping is done with a Uses vector that has the exact
362 /// same number of entries as the partition vector. Each entry is itself
363 /// a vector of the uses.
364 SmallVector<SmallVector<PartitionUse, 2>, 8> Uses;
366 /// \brief Instructions which will become dead if we rewrite the alloca.
368 /// Note that these are not separated by partition. This is because we expect
369 /// a partitioned alloca to be completely rewritten or not rewritten at all.
370 /// If rewritten, all these instructions can simply be removed and replaced
371 /// with undef as they come from outside of the allocated space.
372 SmallVector<Instruction *, 8> DeadUsers;
374 /// \brief Operands which will become dead if we rewrite the alloca.
376 /// These are operands that in their particular use can be replaced with
377 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
378 /// to PHI nodes and the like. They aren't entirely dead (there might be
379 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
380 /// want to swap this particular input for undef to simplify the use lists of
382 SmallVector<Use *, 8> DeadOperands;
384 /// \brief The underlying storage for auxiliary memcpy and memset info.
385 SmallDenseMap<MemTransferInst *, MemTransferOffsets, 4> MemTransferInstData;
387 /// \brief A side datastructure used when building up the partitions and uses.
389 /// This mapping is only really used during the initial building of the
390 /// partitioning so that we can retain information about PHI and select nodes
392 SmallDenseMap<Instruction *, std::pair<uint64_t, bool> > PHIOrSelectSizes;
394 /// \brief Auxiliary information for particular PHI or select operands.
395 SmallDenseMap<Use *, std::pair<unsigned, unsigned>, 4> PHIOrSelectOpMap;
397 /// \brief A utility routine called from the constructor.
399 /// This does what it says on the tin. It is the key of the alloca partition
400 /// splitting and merging. After it is called we have the desired disjoint
401 /// collection of partitions.
402 void splitAndMergePartitions();
406 static Value *foldSelectInst(SelectInst &SI) {
407 // If the condition being selected on is a constant or the same value is
408 // being selected between, fold the select. Yes this does (rarely) happen
410 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
411 return SI.getOperand(1+CI->isZero());
412 if (SI.getOperand(1) == SI.getOperand(2)) {
413 return SI.getOperand(1);
418 /// \brief Builder for the alloca partitioning.
420 /// This class builds an alloca partitioning by recursively visiting the uses
421 /// of an alloca and splitting the partitions for each load and store at each
423 class AllocaPartitioning::PartitionBuilder
424 : public PtrUseVisitor<PartitionBuilder> {
425 friend class PtrUseVisitor<PartitionBuilder>;
426 friend class InstVisitor<PartitionBuilder>;
427 typedef PtrUseVisitor<PartitionBuilder> Base;
429 const uint64_t AllocSize;
430 AllocaPartitioning &P;
432 SmallDenseMap<Instruction *, unsigned> MemTransferPartitionMap;
435 PartitionBuilder(const DataLayout &DL, AllocaInst &AI, AllocaPartitioning &P)
436 : PtrUseVisitor<PartitionBuilder>(DL),
437 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())),
441 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
442 bool IsSplittable = false) {
443 // Completely skip uses which have a zero size or start either before or
444 // past the end of the allocation.
445 if (Size == 0 || Offset.isNegative() || Offset.uge(AllocSize)) {
446 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
447 << " which has zero size or starts outside of the "
448 << AllocSize << " byte alloca:\n"
449 << " alloca: " << P.AI << "\n"
450 << " use: " << I << "\n");
454 uint64_t BeginOffset = Offset.getZExtValue();
455 uint64_t EndOffset = BeginOffset + Size;
457 // Clamp the end offset to the end of the allocation. Note that this is
458 // formulated to handle even the case where "BeginOffset + Size" overflows.
459 // NOTE! This may appear superficially to be something we could ignore
460 // entirely, but that is not so! There may be PHI-node uses where some
461 // instructions are dead but not others. We can't completely ignore the
462 // PHI node, and so have to record at least the information here.
463 assert(AllocSize >= BeginOffset); // Established above.
464 if (Size > AllocSize - BeginOffset) {
465 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
466 << " to remain within the " << AllocSize << " byte alloca:\n"
467 << " alloca: " << P.AI << "\n"
468 << " use: " << I << "\n");
469 EndOffset = AllocSize;
472 Partition New(BeginOffset, EndOffset, IsSplittable);
473 P.Partitions.push_back(New);
476 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
478 uint64_t Size = DL.getTypeStoreSize(Ty);
480 // If this memory access can be shown to *statically* extend outside the
481 // bounds of of the allocation, it's behavior is undefined, so simply
482 // ignore it. Note that this is more strict than the generic clamping
483 // behavior of insertUse. We also try to handle cases which might run the
485 // FIXME: We should instead consider the pointer to have escaped if this
486 // function is being instrumented for addressing bugs or race conditions.
487 if (Offset.isNegative() || Size > AllocSize ||
488 Offset.ugt(AllocSize - Size)) {
489 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte "
490 << (isa<LoadInst>(I) ? "load" : "store") << " @" << Offset
491 << " which extends past the end of the " << AllocSize
493 << " alloca: " << P.AI << "\n"
494 << " use: " << I << "\n");
498 // We allow splitting of loads and stores where the type is an integer type
499 // and which cover the entire alloca. Such integer loads and stores
500 // often require decomposition into fine grained loads and stores.
501 bool IsSplittable = false;
502 if (IntegerType *ITy = dyn_cast<IntegerType>(Ty))
503 IsSplittable = !IsVolatile && ITy->getBitWidth() == AllocSize*8;
505 insertUse(I, Offset, Size, IsSplittable);
508 void visitLoadInst(LoadInst &LI) {
509 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
510 "All simple FCA loads should have been pre-split");
513 return PI.setAborted(&LI);
515 return handleLoadOrStore(LI.getType(), LI, Offset, LI.isVolatile());
518 void visitStoreInst(StoreInst &SI) {
519 Value *ValOp = SI.getValueOperand();
521 return PI.setEscapedAndAborted(&SI);
523 return PI.setAborted(&SI);
525 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
526 "All simple FCA stores should have been pre-split");
527 handleLoadOrStore(ValOp->getType(), SI, Offset, SI.isVolatile());
531 void visitMemSetInst(MemSetInst &II) {
532 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
533 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
534 if ((Length && Length->getValue() == 0) ||
535 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
536 // Zero-length mem transfer intrinsics can be ignored entirely.
540 return PI.setAborted(&II);
542 insertUse(II, Offset,
543 Length ? Length->getLimitedValue()
544 : AllocSize - Offset.getLimitedValue(),
548 void visitMemTransferInst(MemTransferInst &II) {
549 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
550 if ((Length && Length->getValue() == 0) ||
551 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
552 // Zero-length mem transfer intrinsics can be ignored entirely.
556 return PI.setAborted(&II);
558 uint64_t RawOffset = Offset.getLimitedValue();
559 uint64_t Size = Length ? Length->getLimitedValue()
560 : AllocSize - RawOffset;
562 MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
564 // Only intrinsics with a constant length can be split.
565 Offsets.IsSplittable = Length;
567 if (*U == II.getRawDest()) {
568 Offsets.DestBegin = RawOffset;
569 Offsets.DestEnd = RawOffset + Size;
571 if (*U == II.getRawSource()) {
572 Offsets.SourceBegin = RawOffset;
573 Offsets.SourceEnd = RawOffset + Size;
576 // If we have set up end offsets for both the source and the destination,
577 // we have found both sides of this transfer pointing at the same alloca.
578 bool SeenBothEnds = Offsets.SourceEnd && Offsets.DestEnd;
579 if (SeenBothEnds && II.getRawDest() != II.getRawSource()) {
580 unsigned PrevIdx = MemTransferPartitionMap[&II];
582 // Check if the begin offsets match and this is a non-volatile transfer.
583 // In that case, we can completely elide the transfer.
584 if (!II.isVolatile() && Offsets.SourceBegin == Offsets.DestBegin) {
585 P.Partitions[PrevIdx].kill();
589 // Otherwise we have an offset transfer within the same alloca. We can't
591 P.Partitions[PrevIdx].IsSplittable = Offsets.IsSplittable = false;
592 } else if (SeenBothEnds) {
593 // Handle the case where this exact use provides both ends of the
595 assert(II.getRawDest() == II.getRawSource());
597 // For non-volatile transfers this is a no-op.
598 if (!II.isVolatile())
601 // Otherwise just suppress splitting.
602 Offsets.IsSplittable = false;
606 // Insert the use now that we've fixed up the splittable nature.
607 insertUse(II, Offset, Size, Offsets.IsSplittable);
609 // Setup the mapping from intrinsic to partition of we've not seen both
610 // ends of this transfer.
612 unsigned NewIdx = P.Partitions.size() - 1;
614 = MemTransferPartitionMap.insert(std::make_pair(&II, NewIdx)).second;
616 "Already have intrinsic in map but haven't seen both ends");
621 // Disable SRoA for any intrinsics except for lifetime invariants.
622 // FIXME: What about debug intrinsics? This matches old behavior, but
623 // doesn't make sense.
624 void visitIntrinsicInst(IntrinsicInst &II) {
626 return PI.setAborted(&II);
628 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
629 II.getIntrinsicID() == Intrinsic::lifetime_end) {
630 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
631 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
632 Length->getLimitedValue());
633 insertUse(II, Offset, Size, true);
637 Base::visitIntrinsicInst(II);
640 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
641 // We consider any PHI or select that results in a direct load or store of
642 // the same offset to be a viable use for partitioning purposes. These uses
643 // are considered unsplittable and the size is the maximum loaded or stored
645 SmallPtrSet<Instruction *, 4> Visited;
646 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
647 Visited.insert(Root);
648 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
649 // If there are no loads or stores, the access is dead. We mark that as
650 // a size zero access.
653 Instruction *I, *UsedI;
654 llvm::tie(UsedI, I) = Uses.pop_back_val();
656 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
657 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
660 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
661 Value *Op = SI->getOperand(0);
664 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
668 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
669 if (!GEP->hasAllZeroIndices())
671 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
672 !isa<SelectInst>(I)) {
676 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
678 if (Visited.insert(cast<Instruction>(*UI)))
679 Uses.push_back(std::make_pair(I, cast<Instruction>(*UI)));
680 } while (!Uses.empty());
685 void visitPHINode(PHINode &PN) {
689 return PI.setAborted(&PN);
691 // See if we already have computed info on this node.
692 std::pair<uint64_t, bool> &PHIInfo = P.PHIOrSelectSizes[&PN];
694 PHIInfo.second = true;
695 insertUse(PN, Offset, PHIInfo.first);
699 // Check for an unsafe use of the PHI node.
700 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&PN, PHIInfo.first))
701 return PI.setAborted(UnsafeI);
703 insertUse(PN, Offset, PHIInfo.first);
706 void visitSelectInst(SelectInst &SI) {
709 if (Value *Result = foldSelectInst(SI)) {
711 // If the result of the constant fold will be the pointer, recurse
712 // through the select as if we had RAUW'ed it.
718 return PI.setAborted(&SI);
720 // See if we already have computed info on this node.
721 std::pair<uint64_t, bool> &SelectInfo = P.PHIOrSelectSizes[&SI];
722 if (SelectInfo.first) {
723 SelectInfo.second = true;
724 insertUse(SI, Offset, SelectInfo.first);
728 // Check for an unsafe use of the PHI node.
729 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&SI, SelectInfo.first))
730 return PI.setAborted(UnsafeI);
732 insertUse(SI, Offset, SelectInfo.first);
735 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
736 void visitInstruction(Instruction &I) {
741 /// \brief Use adder for the alloca partitioning.
743 /// This class adds the uses of an alloca to all of the partitions which they
744 /// use. For splittable partitions, this can end up doing essentially a linear
745 /// walk of the partitions, but the number of steps remains bounded by the
746 /// total result instruction size:
747 /// - The number of partitions is a result of the number unsplittable
748 /// instructions using the alloca.
749 /// - The number of users of each partition is at worst the total number of
750 /// splittable instructions using the alloca.
751 /// Thus we will produce N * M instructions in the end, where N are the number
752 /// of unsplittable uses and M are the number of splittable. This visitor does
753 /// the exact same number of updates to the partitioning.
755 /// In the more common case, this visitor will leverage the fact that the
756 /// partition space is pre-sorted, and do a logarithmic search for the
757 /// partition needed, making the total visit a classical ((N + M) * log(N))
758 /// complexity operation.
759 class AllocaPartitioning::UseBuilder : public PtrUseVisitor<UseBuilder> {
760 friend class PtrUseVisitor<UseBuilder>;
761 friend class InstVisitor<UseBuilder>;
762 typedef PtrUseVisitor<UseBuilder> Base;
764 const uint64_t AllocSize;
765 AllocaPartitioning &P;
767 /// \brief Set to de-duplicate dead instructions found in the use walk.
768 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
771 UseBuilder(const DataLayout &TD, AllocaInst &AI, AllocaPartitioning &P)
772 : PtrUseVisitor<UseBuilder>(TD),
773 AllocSize(TD.getTypeAllocSize(AI.getAllocatedType())),
777 void markAsDead(Instruction &I) {
778 if (VisitedDeadInsts.insert(&I))
779 P.DeadUsers.push_back(&I);
782 void insertUse(Instruction &User, const APInt &Offset, uint64_t Size) {
783 // If the use has a zero size or extends outside of the allocation, record
784 // it as a dead use for elimination later.
785 if (Size == 0 || Offset.isNegative() || Offset.uge(AllocSize))
786 return markAsDead(User);
788 uint64_t BeginOffset = Offset.getZExtValue();
789 uint64_t EndOffset = BeginOffset + Size;
791 // Clamp the end offset to the end of the allocation. Note that this is
792 // formulated to handle even the case where "BeginOffset + Size" overflows.
793 assert(AllocSize >= BeginOffset); // Established above.
794 if (Size > AllocSize - BeginOffset)
795 EndOffset = AllocSize;
797 // NB: This only works if we have zero overlapping partitions.
798 iterator B = std::lower_bound(P.begin(), P.end(), BeginOffset);
799 if (B != P.begin() && llvm::prior(B)->EndOffset > BeginOffset)
801 for (iterator I = B, E = P.end(); I != E && I->BeginOffset < EndOffset;
803 PartitionUse NewPU(std::max(I->BeginOffset, BeginOffset),
804 std::min(I->EndOffset, EndOffset), U);
805 P.use_push_back(I, NewPU);
806 if (isa<PHINode>(U->getUser()) || isa<SelectInst>(U->getUser()))
807 P.PHIOrSelectOpMap[U]
808 = std::make_pair(I - P.begin(), P.Uses[I - P.begin()].size() - 1);
812 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset) {
813 uint64_t Size = DL.getTypeStoreSize(Ty);
815 // If this memory access can be shown to *statically* extend outside the
816 // bounds of of the allocation, it's behavior is undefined, so simply
817 // ignore it. Note that this is more strict than the generic clamping
818 // behavior of insertUse.
819 if (Offset.isNegative() || Size > AllocSize ||
820 Offset.ugt(AllocSize - Size))
821 return markAsDead(I);
823 insertUse(I, Offset, Size);
826 void visitBitCastInst(BitCastInst &BC) {
828 return markAsDead(BC);
830 return Base::visitBitCastInst(BC);
833 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
834 if (GEPI.use_empty())
835 return markAsDead(GEPI);
837 return Base::visitGetElementPtrInst(GEPI);
840 void visitLoadInst(LoadInst &LI) {
841 assert(IsOffsetKnown);
842 handleLoadOrStore(LI.getType(), LI, Offset);
845 void visitStoreInst(StoreInst &SI) {
846 assert(IsOffsetKnown);
847 handleLoadOrStore(SI.getOperand(0)->getType(), SI, Offset);
850 void visitMemSetInst(MemSetInst &II) {
851 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
852 if ((Length && Length->getValue() == 0) ||
853 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
854 return markAsDead(II);
856 assert(IsOffsetKnown);
857 insertUse(II, Offset, Length ? Length->getLimitedValue()
858 : AllocSize - Offset.getLimitedValue());
861 void visitMemTransferInst(MemTransferInst &II) {
862 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
863 if ((Length && Length->getValue() == 0) ||
864 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
865 return markAsDead(II);
867 assert(IsOffsetKnown);
868 uint64_t Size = Length ? Length->getLimitedValue()
869 : AllocSize - Offset.getLimitedValue();
871 MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
872 if (!II.isVolatile() && Offsets.DestEnd && Offsets.SourceEnd &&
873 Offsets.DestBegin == Offsets.SourceBegin)
874 return markAsDead(II); // Skip identity transfers without side-effects.
876 insertUse(II, Offset, Size);
879 void visitIntrinsicInst(IntrinsicInst &II) {
880 assert(IsOffsetKnown);
881 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
882 II.getIntrinsicID() == Intrinsic::lifetime_end);
884 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
885 insertUse(II, Offset, std::min(Length->getLimitedValue(),
886 AllocSize - Offset.getLimitedValue()));
889 void insertPHIOrSelect(Instruction &User, const APInt &Offset) {
890 uint64_t Size = P.PHIOrSelectSizes.lookup(&User).first;
892 // For PHI and select operands outside the alloca, we can't nuke the entire
893 // phi or select -- the other side might still be relevant, so we special
894 // case them here and use a separate structure to track the operands
895 // themselves which should be replaced with undef.
896 if ((Offset.isNegative() && Offset.uge(Size)) ||
897 (!Offset.isNegative() && Offset.uge(AllocSize))) {
898 P.DeadOperands.push_back(U);
902 insertUse(User, Offset, Size);
905 void visitPHINode(PHINode &PN) {
907 return markAsDead(PN);
909 assert(IsOffsetKnown);
910 insertPHIOrSelect(PN, Offset);
913 void visitSelectInst(SelectInst &SI) {
915 return markAsDead(SI);
917 if (Value *Result = foldSelectInst(SI)) {
919 // If the result of the constant fold will be the pointer, recurse
920 // through the select as if we had RAUW'ed it.
923 // Otherwise the operand to the select is dead, and we can replace it
925 P.DeadOperands.push_back(U);
930 assert(IsOffsetKnown);
931 insertPHIOrSelect(SI, Offset);
934 /// \brief Unreachable, we've already visited the alloca once.
935 void visitInstruction(Instruction &I) {
936 llvm_unreachable("Unhandled instruction in use builder.");
940 void AllocaPartitioning::splitAndMergePartitions() {
941 size_t NumDeadPartitions = 0;
943 // Track the range of splittable partitions that we pass when accumulating
944 // overlapping unsplittable partitions.
945 uint64_t SplitEndOffset = 0ull;
947 Partition New(0ull, 0ull, false);
949 for (unsigned i = 0, j = i, e = Partitions.size(); i != e; i = j) {
952 if (!Partitions[i].IsSplittable || New.BeginOffset == New.EndOffset) {
953 assert(New.BeginOffset == New.EndOffset);
956 assert(New.IsSplittable);
957 New.EndOffset = std::max(New.EndOffset, Partitions[i].EndOffset);
959 assert(New.BeginOffset != New.EndOffset);
961 // Scan the overlapping partitions.
962 while (j != e && New.EndOffset > Partitions[j].BeginOffset) {
963 // If the new partition we are forming is splittable, stop at the first
964 // unsplittable partition.
965 if (New.IsSplittable && !Partitions[j].IsSplittable)
968 // Grow the new partition to include any equally splittable range. 'j' is
969 // always equally splittable when New is splittable, but when New is not
970 // splittable, we may subsume some (or part of some) splitable partition
971 // without growing the new one.
972 if (New.IsSplittable == Partitions[j].IsSplittable) {
973 New.EndOffset = std::max(New.EndOffset, Partitions[j].EndOffset);
975 assert(!New.IsSplittable);
976 assert(Partitions[j].IsSplittable);
977 SplitEndOffset = std::max(SplitEndOffset, Partitions[j].EndOffset);
980 Partitions[j].kill();
985 // If the new partition is splittable, chop off the end as soon as the
986 // unsplittable subsequent partition starts and ensure we eventually cover
987 // the splittable area.
988 if (j != e && New.IsSplittable) {
989 SplitEndOffset = std::max(SplitEndOffset, New.EndOffset);
990 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
993 // Add the new partition if it differs from the original one and is
994 // non-empty. We can end up with an empty partition here if it was
995 // splittable but there is an unsplittable one that starts at the same
997 if (New != Partitions[i]) {
998 if (New.BeginOffset != New.EndOffset)
999 Partitions.push_back(New);
1000 // Mark the old one for removal.
1001 Partitions[i].kill();
1002 ++NumDeadPartitions;
1005 New.BeginOffset = New.EndOffset;
1006 if (!New.IsSplittable) {
1007 New.EndOffset = std::max(New.EndOffset, SplitEndOffset);
1008 if (j != e && !Partitions[j].IsSplittable)
1009 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
1010 New.IsSplittable = true;
1011 // If there is a trailing splittable partition which won't be fused into
1012 // the next splittable partition go ahead and add it onto the partitions
1014 if (New.BeginOffset < New.EndOffset &&
1015 (j == e || !Partitions[j].IsSplittable ||
1016 New.EndOffset < Partitions[j].BeginOffset)) {
1017 Partitions.push_back(New);
1018 New.BeginOffset = New.EndOffset = 0ull;
1023 // Re-sort the partitions now that they have been split and merged into
1024 // disjoint set of partitions. Also remove any of the dead partitions we've
1025 // replaced in the process.
1026 std::sort(Partitions.begin(), Partitions.end());
1027 if (NumDeadPartitions) {
1028 assert(Partitions.back().isDead());
1029 assert((ptrdiff_t)NumDeadPartitions ==
1030 std::count(Partitions.begin(), Partitions.end(), Partitions.back()));
1032 Partitions.erase(Partitions.end() - NumDeadPartitions, Partitions.end());
1035 AllocaPartitioning::AllocaPartitioning(const DataLayout &TD, AllocaInst &AI)
1037 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1040 PointerEscapingInstr(0) {
1041 PartitionBuilder PB(TD, AI, *this);
1042 PartitionBuilder::PtrInfo PtrI = PB.visitPtr(AI);
1043 if (PtrI.isEscaped() || PtrI.isAborted()) {
1044 // FIXME: We should sink the escape vs. abort info into the caller nicely,
1045 // possibly by just storing the PtrInfo in the AllocaPartitioning.
1046 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
1047 : PtrI.getAbortingInst();
1048 assert(PointerEscapingInstr && "Did not track a bad instruction");
1052 // Sort the uses. This arranges for the offsets to be in ascending order,
1053 // and the sizes to be in descending order.
1054 std::sort(Partitions.begin(), Partitions.end());
1056 // Remove any partitions from the back which are marked as dead.
1057 while (!Partitions.empty() && Partitions.back().isDead())
1058 Partitions.pop_back();
1060 if (Partitions.size() > 1) {
1061 // Intersect splittability for all partitions with equal offsets and sizes.
1062 // Then remove all but the first so that we have a sequence of non-equal but
1063 // potentially overlapping partitions.
1064 for (iterator I = Partitions.begin(), J = I, E = Partitions.end(); I != E;
1067 while (J != E && *I == *J) {
1068 I->IsSplittable &= J->IsSplittable;
1072 Partitions.erase(std::unique(Partitions.begin(), Partitions.end()),
1075 // Split splittable and merge unsplittable partitions into a disjoint set
1076 // of partitions over the used space of the allocation.
1077 splitAndMergePartitions();
1080 // Now build up the user lists for each of these disjoint partitions by
1081 // re-walking the recursive users of the alloca.
1082 Uses.resize(Partitions.size());
1083 UseBuilder UB(TD, AI, *this);
1084 PtrI = UB.visitPtr(AI);
1085 assert(!PtrI.isEscaped() && "Previously analyzed pointer now escapes!");
1086 assert(!PtrI.isAborted() && "Early aborted the visit of the pointer.");
1089 Type *AllocaPartitioning::getCommonType(iterator I) const {
1091 for (const_use_iterator UI = use_begin(I), UE = use_end(I); UI != UE; ++UI) {
1093 continue; // Skip dead uses.
1094 if (isa<IntrinsicInst>(*UI->U->getUser()))
1096 if (UI->BeginOffset != I->BeginOffset || UI->EndOffset != I->EndOffset)
1100 if (LoadInst *LI = dyn_cast<LoadInst>(UI->U->getUser())) {
1101 UserTy = LI->getType();
1102 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI->U->getUser())) {
1103 UserTy = SI->getValueOperand()->getType();
1105 return 0; // Bail if we have weird uses.
1108 if (IntegerType *ITy = dyn_cast<IntegerType>(UserTy)) {
1109 // If the type is larger than the partition, skip it. We only encounter
1110 // this for split integer operations where we want to use the type of the
1111 // entity causing the split.
1112 if (ITy->getBitWidth() > (I->EndOffset - I->BeginOffset)*8)
1115 // If we have found an integer type use covering the alloca, use that
1116 // regardless of the other types, as integers are often used for a "bucket
1121 if (Ty && Ty != UserTy)
1129 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1131 void AllocaPartitioning::print(raw_ostream &OS, const_iterator I,
1132 StringRef Indent) const {
1133 OS << Indent << "partition #" << (I - begin())
1134 << " [" << I->BeginOffset << "," << I->EndOffset << ")"
1135 << (I->IsSplittable ? " (splittable)" : "")
1136 << (Uses[I - begin()].empty() ? " (zero uses)" : "")
1140 void AllocaPartitioning::printUsers(raw_ostream &OS, const_iterator I,
1141 StringRef Indent) const {
1142 for (const_use_iterator UI = use_begin(I), UE = use_end(I);
1145 continue; // Skip dead uses.
1146 OS << Indent << " [" << UI->BeginOffset << "," << UI->EndOffset << ") "
1147 << "used by: " << *UI->U->getUser() << "\n";
1148 if (MemTransferInst *II = dyn_cast<MemTransferInst>(UI->U->getUser())) {
1149 const MemTransferOffsets &MTO = MemTransferInstData.lookup(II);
1151 if (!MTO.IsSplittable)
1152 IsDest = UI->BeginOffset == MTO.DestBegin;
1154 IsDest = MTO.DestBegin != 0u;
1155 OS << Indent << " (original " << (IsDest ? "dest" : "source") << ": "
1156 << "[" << (IsDest ? MTO.DestBegin : MTO.SourceBegin)
1157 << "," << (IsDest ? MTO.DestEnd : MTO.SourceEnd) << ")\n";
1162 void AllocaPartitioning::print(raw_ostream &OS) const {
1163 if (PointerEscapingInstr) {
1164 OS << "No partitioning for alloca: " << AI << "\n"
1165 << " A pointer to this alloca escaped by:\n"
1166 << " " << *PointerEscapingInstr << "\n";
1170 OS << "Partitioning of alloca: " << AI << "\n";
1172 for (const_iterator I = begin(), E = end(); I != E; ++I, ++Num) {
1178 void AllocaPartitioning::dump(const_iterator I) const { print(dbgs(), I); }
1179 void AllocaPartitioning::dump() const { print(dbgs()); }
1181 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1185 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
1187 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
1188 /// the loads and stores of an alloca instruction, as well as updating its
1189 /// debug information. This is used when a domtree is unavailable and thus
1190 /// mem2reg in its full form can't be used to handle promotion of allocas to
1192 class AllocaPromoter : public LoadAndStorePromoter {
1196 SmallVector<DbgDeclareInst *, 4> DDIs;
1197 SmallVector<DbgValueInst *, 4> DVIs;
1200 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
1201 AllocaInst &AI, DIBuilder &DIB)
1202 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
1204 void run(const SmallVectorImpl<Instruction*> &Insts) {
1205 // Remember which alloca we're promoting (for isInstInList).
1206 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
1207 for (Value::use_iterator UI = DebugNode->use_begin(),
1208 UE = DebugNode->use_end();
1210 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
1211 DDIs.push_back(DDI);
1212 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
1213 DVIs.push_back(DVI);
1216 LoadAndStorePromoter::run(Insts);
1217 AI.eraseFromParent();
1218 while (!DDIs.empty())
1219 DDIs.pop_back_val()->eraseFromParent();
1220 while (!DVIs.empty())
1221 DVIs.pop_back_val()->eraseFromParent();
1224 virtual bool isInstInList(Instruction *I,
1225 const SmallVectorImpl<Instruction*> &Insts) const {
1226 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1227 return LI->getOperand(0) == &AI;
1228 return cast<StoreInst>(I)->getPointerOperand() == &AI;
1231 virtual void updateDebugInfo(Instruction *Inst) const {
1232 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(),
1233 E = DDIs.end(); I != E; ++I) {
1234 DbgDeclareInst *DDI = *I;
1235 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
1236 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1237 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
1238 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1240 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(),
1241 E = DVIs.end(); I != E; ++I) {
1242 DbgValueInst *DVI = *I;
1244 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1245 // If an argument is zero extended then use argument directly. The ZExt
1246 // may be zapped by an optimization pass in future.
1247 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1248 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
1249 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1250 Arg = dyn_cast<Argument>(SExt->getOperand(0));
1252 Arg = SI->getOperand(0);
1253 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
1254 Arg = LI->getOperand(0);
1258 Instruction *DbgVal =
1259 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
1261 DbgVal->setDebugLoc(DVI->getDebugLoc());
1265 } // end anon namespace
1269 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
1271 /// This pass takes allocations which can be completely analyzed (that is, they
1272 /// don't escape) and tries to turn them into scalar SSA values. There are
1273 /// a few steps to this process.
1275 /// 1) It takes allocations of aggregates and analyzes the ways in which they
1276 /// are used to try to split them into smaller allocations, ideally of
1277 /// a single scalar data type. It will split up memcpy and memset accesses
1278 /// as necessary and try to isolate individual scalar accesses.
1279 /// 2) It will transform accesses into forms which are suitable for SSA value
1280 /// promotion. This can be replacing a memset with a scalar store of an
1281 /// integer value, or it can involve speculating operations on a PHI or
1282 /// select to be a PHI or select of the results.
1283 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
1284 /// onto insert and extract operations on a vector value, and convert them to
1285 /// this form. By doing so, it will enable promotion of vector aggregates to
1286 /// SSA vector values.
1287 class SROA : public FunctionPass {
1288 const bool RequiresDomTree;
1291 const DataLayout *TD;
1294 /// \brief Worklist of alloca instructions to simplify.
1296 /// Each alloca in the function is added to this. Each new alloca formed gets
1297 /// added to it as well to recursively simplify unless that alloca can be
1298 /// directly promoted. Finally, each time we rewrite a use of an alloca other
1299 /// the one being actively rewritten, we add it back onto the list if not
1300 /// already present to ensure it is re-visited.
1301 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
1303 /// \brief A collection of instructions to delete.
1304 /// We try to batch deletions to simplify code and make things a bit more
1306 SetVector<Instruction *, SmallVector<Instruction *, 8> > DeadInsts;
1308 /// \brief Post-promotion worklist.
1310 /// Sometimes we discover an alloca which has a high probability of becoming
1311 /// viable for SROA after a round of promotion takes place. In those cases,
1312 /// the alloca is enqueued here for re-processing.
1314 /// Note that we have to be very careful to clear allocas out of this list in
1315 /// the event they are deleted.
1316 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist;
1318 /// \brief A collection of alloca instructions we can directly promote.
1319 std::vector<AllocaInst *> PromotableAllocas;
1322 SROA(bool RequiresDomTree = true)
1323 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
1324 C(0), TD(0), DT(0) {
1325 initializeSROAPass(*PassRegistry::getPassRegistry());
1327 bool runOnFunction(Function &F);
1328 void getAnalysisUsage(AnalysisUsage &AU) const;
1330 const char *getPassName() const { return "SROA"; }
1334 friend class PHIOrSelectSpeculator;
1335 friend class AllocaPartitionRewriter;
1336 friend class AllocaPartitionVectorRewriter;
1338 bool rewriteAllocaPartition(AllocaInst &AI,
1339 AllocaPartitioning &P,
1340 AllocaPartitioning::iterator PI);
1341 bool splitAlloca(AllocaInst &AI, AllocaPartitioning &P);
1342 bool runOnAlloca(AllocaInst &AI);
1343 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
1344 bool promoteAllocas(Function &F);
1350 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
1351 return new SROA(RequiresDomTree);
1354 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
1356 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
1357 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
1361 /// \brief Visitor to speculate PHIs and Selects where possible.
1362 class PHIOrSelectSpeculator : public InstVisitor<PHIOrSelectSpeculator> {
1363 // Befriend the base class so it can delegate to private visit methods.
1364 friend class llvm::InstVisitor<PHIOrSelectSpeculator>;
1366 const DataLayout &TD;
1367 AllocaPartitioning &P;
1371 PHIOrSelectSpeculator(const DataLayout &TD, AllocaPartitioning &P, SROA &Pass)
1372 : TD(TD), P(P), Pass(Pass) {}
1374 /// \brief Visit the users of an alloca partition and rewrite them.
1375 void visitUsers(AllocaPartitioning::const_iterator PI) {
1376 // Note that we need to use an index here as the underlying vector of uses
1377 // may be grown during speculation. However, we never need to re-visit the
1378 // new uses, and so we can use the initial size bound.
1379 for (unsigned Idx = 0, Size = P.use_size(PI); Idx != Size; ++Idx) {
1380 const AllocaPartitioning::PartitionUse &PU = P.getUse(PI, Idx);
1382 continue; // Skip dead use.
1384 visit(cast<Instruction>(PU.U->getUser()));
1389 // By default, skip this instruction.
1390 void visitInstruction(Instruction &I) {}
1392 /// PHI instructions that use an alloca and are subsequently loaded can be
1393 /// rewritten to load both input pointers in the pred blocks and then PHI the
1394 /// results, allowing the load of the alloca to be promoted.
1396 /// %P2 = phi [i32* %Alloca, i32* %Other]
1397 /// %V = load i32* %P2
1399 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1401 /// %V2 = load i32* %Other
1403 /// %V = phi [i32 %V1, i32 %V2]
1405 /// We can do this to a select if its only uses are loads and if the operands
1406 /// to the select can be loaded unconditionally.
1408 /// FIXME: This should be hoisted into a generic utility, likely in
1409 /// Transforms/Util/Local.h
1410 bool isSafePHIToSpeculate(PHINode &PN, SmallVectorImpl<LoadInst *> &Loads) {
1411 // For now, we can only do this promotion if the load is in the same block
1412 // as the PHI, and if there are no stores between the phi and load.
1413 // TODO: Allow recursive phi users.
1414 // TODO: Allow stores.
1415 BasicBlock *BB = PN.getParent();
1416 unsigned MaxAlign = 0;
1417 for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end();
1419 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1420 if (LI == 0 || !LI->isSimple()) return false;
1422 // For now we only allow loads in the same block as the PHI. This is
1423 // a common case that happens when instcombine merges two loads through
1425 if (LI->getParent() != BB) return false;
1427 // Ensure that there are no instructions between the PHI and the load that
1429 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1430 if (BBI->mayWriteToMemory())
1433 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1434 Loads.push_back(LI);
1437 // We can only transform this if it is safe to push the loads into the
1438 // predecessor blocks. The only thing to watch out for is that we can't put
1439 // a possibly trapping load in the predecessor if it is a critical edge.
1440 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num;
1442 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1443 Value *InVal = PN.getIncomingValue(Idx);
1445 // If the value is produced by the terminator of the predecessor (an
1446 // invoke) or it has side-effects, there is no valid place to put a load
1447 // in the predecessor.
1448 if (TI == InVal || TI->mayHaveSideEffects())
1451 // If the predecessor has a single successor, then the edge isn't
1453 if (TI->getNumSuccessors() == 1)
1456 // If this pointer is always safe to load, or if we can prove that there
1457 // is already a load in the block, then we can move the load to the pred
1459 if (InVal->isDereferenceablePointer() ||
1460 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, &TD))
1469 void visitPHINode(PHINode &PN) {
1470 DEBUG(dbgs() << " original: " << PN << "\n");
1472 SmallVector<LoadInst *, 4> Loads;
1473 if (!isSafePHIToSpeculate(PN, Loads))
1476 assert(!Loads.empty());
1478 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1479 IRBuilder<> PHIBuilder(&PN);
1480 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1481 PN.getName() + ".sroa.speculated");
1483 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1484 // matter which one we get and if any differ.
1485 LoadInst *SomeLoad = cast<LoadInst>(Loads.back());
1486 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1487 unsigned Align = SomeLoad->getAlignment();
1489 // Rewrite all loads of the PN to use the new PHI.
1491 LoadInst *LI = Loads.pop_back_val();
1492 LI->replaceAllUsesWith(NewPN);
1493 Pass.DeadInsts.insert(LI);
1494 } while (!Loads.empty());
1496 // Inject loads into all of the pred blocks.
1497 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1498 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1499 TerminatorInst *TI = Pred->getTerminator();
1500 Use *InUse = &PN.getOperandUse(PN.getOperandNumForIncomingValue(Idx));
1501 Value *InVal = PN.getIncomingValue(Idx);
1502 IRBuilder<> PredBuilder(TI);
1505 = PredBuilder.CreateLoad(InVal, (PN.getName() + ".sroa.speculate.load." +
1507 ++NumLoadsSpeculated;
1508 Load->setAlignment(Align);
1510 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1511 NewPN->addIncoming(Load, Pred);
1513 Instruction *Ptr = dyn_cast<Instruction>(InVal);
1515 // No uses to rewrite.
1518 // Try to lookup and rewrite any partition uses corresponding to this phi
1520 AllocaPartitioning::iterator PI
1521 = P.findPartitionForPHIOrSelectOperand(InUse);
1525 // Replace the Use in the PartitionUse for this operand with the Use
1527 AllocaPartitioning::use_iterator UI
1528 = P.findPartitionUseForPHIOrSelectOperand(InUse);
1529 assert(isa<PHINode>(*UI->U->getUser()));
1530 UI->U = &Load->getOperandUse(Load->getPointerOperandIndex());
1532 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1535 /// Select instructions that use an alloca and are subsequently loaded can be
1536 /// rewritten to load both input pointers and then select between the result,
1537 /// allowing the load of the alloca to be promoted.
1539 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1540 /// %V = load i32* %P2
1542 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1543 /// %V2 = load i32* %Other
1544 /// %V = select i1 %cond, i32 %V1, i32 %V2
1546 /// We can do this to a select if its only uses are loads and if the operand
1547 /// to the select can be loaded unconditionally.
1548 bool isSafeSelectToSpeculate(SelectInst &SI,
1549 SmallVectorImpl<LoadInst *> &Loads) {
1550 Value *TValue = SI.getTrueValue();
1551 Value *FValue = SI.getFalseValue();
1552 bool TDerefable = TValue->isDereferenceablePointer();
1553 bool FDerefable = FValue->isDereferenceablePointer();
1555 for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end();
1557 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1558 if (LI == 0 || !LI->isSimple()) return false;
1560 // Both operands to the select need to be dereferencable, either
1561 // absolutely (e.g. allocas) or at this point because we can see other
1563 if (!TDerefable && !isSafeToLoadUnconditionally(TValue, LI,
1564 LI->getAlignment(), &TD))
1566 if (!FDerefable && !isSafeToLoadUnconditionally(FValue, LI,
1567 LI->getAlignment(), &TD))
1569 Loads.push_back(LI);
1575 void visitSelectInst(SelectInst &SI) {
1576 DEBUG(dbgs() << " original: " << SI << "\n");
1577 IRBuilder<> IRB(&SI);
1579 // If the select isn't safe to speculate, just use simple logic to emit it.
1580 SmallVector<LoadInst *, 4> Loads;
1581 if (!isSafeSelectToSpeculate(SI, Loads))
1584 Use *Ops[2] = { &SI.getOperandUse(1), &SI.getOperandUse(2) };
1585 AllocaPartitioning::iterator PIs[2];
1586 AllocaPartitioning::PartitionUse PUs[2];
1587 for (unsigned i = 0, e = 2; i != e; ++i) {
1588 PIs[i] = P.findPartitionForPHIOrSelectOperand(Ops[i]);
1589 if (PIs[i] != P.end()) {
1590 // If the pointer is within the partitioning, remove the select from
1591 // its uses. We'll add in the new loads below.
1592 AllocaPartitioning::use_iterator UI
1593 = P.findPartitionUseForPHIOrSelectOperand(Ops[i]);
1595 // Clear out the use here so that the offsets into the use list remain
1596 // stable but this use is ignored when rewriting.
1601 Value *TV = SI.getTrueValue();
1602 Value *FV = SI.getFalseValue();
1603 // Replace the loads of the select with a select of two loads.
1604 while (!Loads.empty()) {
1605 LoadInst *LI = Loads.pop_back_val();
1607 IRB.SetInsertPoint(LI);
1609 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1611 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1612 NumLoadsSpeculated += 2;
1614 // Transfer alignment and TBAA info if present.
1615 TL->setAlignment(LI->getAlignment());
1616 FL->setAlignment(LI->getAlignment());
1617 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1618 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
1619 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
1622 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1623 LI->getName() + ".sroa.speculated");
1625 LoadInst *Loads[2] = { TL, FL };
1626 for (unsigned i = 0, e = 2; i != e; ++i) {
1627 if (PIs[i] != P.end()) {
1628 Use *LoadUse = &Loads[i]->getOperandUse(0);
1629 assert(PUs[i].U->get() == LoadUse->get());
1631 P.use_push_back(PIs[i], PUs[i]);
1635 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1636 LI->replaceAllUsesWith(V);
1637 Pass.DeadInsts.insert(LI);
1643 /// \brief Build a GEP out of a base pointer and indices.
1645 /// This will return the BasePtr if that is valid, or build a new GEP
1646 /// instruction using the IRBuilder if GEP-ing is needed.
1647 static Value *buildGEP(IRBuilder<> &IRB, Value *BasePtr,
1648 SmallVectorImpl<Value *> &Indices,
1649 const Twine &Prefix) {
1650 if (Indices.empty())
1653 // A single zero index is a no-op, so check for this and avoid building a GEP
1655 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1658 return IRB.CreateInBoundsGEP(BasePtr, Indices, Prefix + ".idx");
1661 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1662 /// TargetTy without changing the offset of the pointer.
1664 /// This routine assumes we've already established a properly offset GEP with
1665 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1666 /// zero-indices down through type layers until we find one the same as
1667 /// TargetTy. If we can't find one with the same type, we at least try to use
1668 /// one with the same size. If none of that works, we just produce the GEP as
1669 /// indicated by Indices to have the correct offset.
1670 static Value *getNaturalGEPWithType(IRBuilder<> &IRB, const DataLayout &TD,
1671 Value *BasePtr, Type *Ty, Type *TargetTy,
1672 SmallVectorImpl<Value *> &Indices,
1673 const Twine &Prefix) {
1675 return buildGEP(IRB, BasePtr, Indices, Prefix);
1677 // See if we can descend into a struct and locate a field with the correct
1679 unsigned NumLayers = 0;
1680 Type *ElementTy = Ty;
1682 if (ElementTy->isPointerTy())
1684 if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
1685 ElementTy = SeqTy->getElementType();
1686 // Note that we use the default address space as this index is over an
1687 // array or a vector, not a pointer.
1688 Indices.push_back(IRB.getInt(APInt(TD.getPointerSizeInBits(0), 0)));
1689 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1690 if (STy->element_begin() == STy->element_end())
1691 break; // Nothing left to descend into.
1692 ElementTy = *STy->element_begin();
1693 Indices.push_back(IRB.getInt32(0));
1698 } while (ElementTy != TargetTy);
1699 if (ElementTy != TargetTy)
1700 Indices.erase(Indices.end() - NumLayers, Indices.end());
1702 return buildGEP(IRB, BasePtr, Indices, Prefix);
1705 /// \brief Recursively compute indices for a natural GEP.
1707 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1708 /// element types adding appropriate indices for the GEP.
1709 static Value *getNaturalGEPRecursively(IRBuilder<> &IRB, const DataLayout &TD,
1710 Value *Ptr, Type *Ty, APInt &Offset,
1712 SmallVectorImpl<Value *> &Indices,
1713 const Twine &Prefix) {
1715 return getNaturalGEPWithType(IRB, TD, Ptr, Ty, TargetTy, Indices, Prefix);
1717 // We can't recurse through pointer types.
1718 if (Ty->isPointerTy())
1721 // We try to analyze GEPs over vectors here, but note that these GEPs are
1722 // extremely poorly defined currently. The long-term goal is to remove GEPing
1723 // over a vector from the IR completely.
1724 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1725 unsigned ElementSizeInBits = TD.getTypeSizeInBits(VecTy->getScalarType());
1726 if (ElementSizeInBits % 8)
1727 return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
1728 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1729 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1730 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1732 Offset -= NumSkippedElements * ElementSize;
1733 Indices.push_back(IRB.getInt(NumSkippedElements));
1734 return getNaturalGEPRecursively(IRB, TD, Ptr, VecTy->getElementType(),
1735 Offset, TargetTy, Indices, Prefix);
1738 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1739 Type *ElementTy = ArrTy->getElementType();
1740 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1741 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1742 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1745 Offset -= NumSkippedElements * ElementSize;
1746 Indices.push_back(IRB.getInt(NumSkippedElements));
1747 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1751 StructType *STy = dyn_cast<StructType>(Ty);
1755 const StructLayout *SL = TD.getStructLayout(STy);
1756 uint64_t StructOffset = Offset.getZExtValue();
1757 if (StructOffset >= SL->getSizeInBytes())
1759 unsigned Index = SL->getElementContainingOffset(StructOffset);
1760 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1761 Type *ElementTy = STy->getElementType(Index);
1762 if (Offset.uge(TD.getTypeAllocSize(ElementTy)))
1763 return 0; // The offset points into alignment padding.
1765 Indices.push_back(IRB.getInt32(Index));
1766 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1770 /// \brief Get a natural GEP from a base pointer to a particular offset and
1771 /// resulting in a particular type.
1773 /// The goal is to produce a "natural" looking GEP that works with the existing
1774 /// composite types to arrive at the appropriate offset and element type for
1775 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1776 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1777 /// Indices, and setting Ty to the result subtype.
1779 /// If no natural GEP can be constructed, this function returns null.
1780 static Value *getNaturalGEPWithOffset(IRBuilder<> &IRB, const DataLayout &TD,
1781 Value *Ptr, APInt Offset, Type *TargetTy,
1782 SmallVectorImpl<Value *> &Indices,
1783 const Twine &Prefix) {
1784 PointerType *Ty = cast<PointerType>(Ptr->getType());
1786 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1788 if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
1791 Type *ElementTy = Ty->getElementType();
1792 if (!ElementTy->isSized())
1793 return 0; // We can't GEP through an unsized element.
1794 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1795 if (ElementSize == 0)
1796 return 0; // Zero-length arrays can't help us build a natural GEP.
1797 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1799 Offset -= NumSkippedElements * ElementSize;
1800 Indices.push_back(IRB.getInt(NumSkippedElements));
1801 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1805 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1806 /// resulting pointer has PointerTy.
1808 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1809 /// and produces the pointer type desired. Where it cannot, it will try to use
1810 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1811 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1812 /// bitcast to the type.
1814 /// The strategy for finding the more natural GEPs is to peel off layers of the
1815 /// pointer, walking back through bit casts and GEPs, searching for a base
1816 /// pointer from which we can compute a natural GEP with the desired
1817 /// properties. The algorithm tries to fold as many constant indices into
1818 /// a single GEP as possible, thus making each GEP more independent of the
1819 /// surrounding code.
1820 static Value *getAdjustedPtr(IRBuilder<> &IRB, const DataLayout &TD,
1821 Value *Ptr, APInt Offset, Type *PointerTy,
1822 const Twine &Prefix) {
1823 // Even though we don't look through PHI nodes, we could be called on an
1824 // instruction in an unreachable block, which may be on a cycle.
1825 SmallPtrSet<Value *, 4> Visited;
1826 Visited.insert(Ptr);
1827 SmallVector<Value *, 4> Indices;
1829 // We may end up computing an offset pointer that has the wrong type. If we
1830 // never are able to compute one directly that has the correct type, we'll
1831 // fall back to it, so keep it around here.
1832 Value *OffsetPtr = 0;
1834 // Remember any i8 pointer we come across to re-use if we need to do a raw
1837 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1839 Type *TargetTy = PointerTy->getPointerElementType();
1842 // First fold any existing GEPs into the offset.
1843 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1844 APInt GEPOffset(Offset.getBitWidth(), 0);
1845 if (!GEP->accumulateConstantOffset(TD, GEPOffset))
1847 Offset += GEPOffset;
1848 Ptr = GEP->getPointerOperand();
1849 if (!Visited.insert(Ptr))
1853 // See if we can perform a natural GEP here.
1855 if (Value *P = getNaturalGEPWithOffset(IRB, TD, Ptr, Offset, TargetTy,
1857 if (P->getType() == PointerTy) {
1858 // Zap any offset pointer that we ended up computing in previous rounds.
1859 if (OffsetPtr && OffsetPtr->use_empty())
1860 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1861 I->eraseFromParent();
1869 // Stash this pointer if we've found an i8*.
1870 if (Ptr->getType()->isIntegerTy(8)) {
1872 Int8PtrOffset = Offset;
1875 // Peel off a layer of the pointer and update the offset appropriately.
1876 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1877 Ptr = cast<Operator>(Ptr)->getOperand(0);
1878 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1879 if (GA->mayBeOverridden())
1881 Ptr = GA->getAliasee();
1885 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1886 } while (Visited.insert(Ptr));
1890 Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
1891 Prefix + ".raw_cast");
1892 Int8PtrOffset = Offset;
1895 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1896 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1897 Prefix + ".raw_idx");
1901 // On the off chance we were targeting i8*, guard the bitcast here.
1902 if (Ptr->getType() != PointerTy)
1903 Ptr = IRB.CreateBitCast(Ptr, PointerTy, Prefix + ".cast");
1908 /// \brief Test whether we can convert a value from the old to the new type.
1910 /// This predicate should be used to guard calls to convertValue in order to
1911 /// ensure that we only try to convert viable values. The strategy is that we
1912 /// will peel off single element struct and array wrappings to get to an
1913 /// underlying value, and convert that value.
1914 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1917 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1919 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1922 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1923 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1925 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1933 /// \brief Generic routine to convert an SSA value to a value of a different
1936 /// This will try various different casting techniques, such as bitcasts,
1937 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1938 /// two types for viability with this routine.
1939 static Value *convertValue(const DataLayout &DL, IRBuilder<> &IRB, Value *V,
1941 assert(canConvertValue(DL, V->getType(), Ty) &&
1942 "Value not convertable to type");
1943 if (V->getType() == Ty)
1945 if (V->getType()->isIntegerTy() && Ty->isPointerTy())
1946 return IRB.CreateIntToPtr(V, Ty);
1947 if (V->getType()->isPointerTy() && Ty->isIntegerTy())
1948 return IRB.CreatePtrToInt(V, Ty);
1950 return IRB.CreateBitCast(V, Ty);
1953 /// \brief Test whether the given alloca partition can be promoted to a vector.
1955 /// This is a quick test to check whether we can rewrite a particular alloca
1956 /// partition (and its newly formed alloca) into a vector alloca with only
1957 /// whole-vector loads and stores such that it could be promoted to a vector
1958 /// SSA value. We only can ensure this for a limited set of operations, and we
1959 /// don't want to do the rewrites unless we are confident that the result will
1960 /// be promotable, so we have an early test here.
1961 static bool isVectorPromotionViable(const DataLayout &TD,
1963 AllocaPartitioning &P,
1964 uint64_t PartitionBeginOffset,
1965 uint64_t PartitionEndOffset,
1966 AllocaPartitioning::const_use_iterator I,
1967 AllocaPartitioning::const_use_iterator E) {
1968 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
1972 uint64_t ElementSize = TD.getTypeSizeInBits(Ty->getScalarType());
1974 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1975 // that aren't byte sized.
1976 if (ElementSize % 8)
1978 assert((TD.getTypeSizeInBits(Ty) % 8) == 0 &&
1979 "vector size not a multiple of element size?");
1982 for (; I != E; ++I) {
1984 continue; // Skip dead use.
1986 uint64_t BeginOffset = I->BeginOffset - PartitionBeginOffset;
1987 uint64_t BeginIndex = BeginOffset / ElementSize;
1988 if (BeginIndex * ElementSize != BeginOffset ||
1989 BeginIndex >= Ty->getNumElements())
1991 uint64_t EndOffset = I->EndOffset - PartitionBeginOffset;
1992 uint64_t EndIndex = EndOffset / ElementSize;
1993 if (EndIndex * ElementSize != EndOffset ||
1994 EndIndex > Ty->getNumElements())
1997 assert(EndIndex > BeginIndex && "Empty vector!");
1998 uint64_t NumElements = EndIndex - BeginIndex;
2000 = (NumElements == 1) ? Ty->getElementType()
2001 : VectorType::get(Ty->getElementType(), NumElements);
2003 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
2004 if (MI->isVolatile())
2006 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
2007 const AllocaPartitioning::MemTransferOffsets &MTO
2008 = P.getMemTransferOffsets(*MTI);
2009 if (!MTO.IsSplittable)
2012 } else if (I->U->get()->getType()->getPointerElementType()->isStructTy()) {
2013 // Disable vector promotion when there are loads or stores of an FCA.
2015 } else if (LoadInst *LI = dyn_cast<LoadInst>(I->U->getUser())) {
2016 if (LI->isVolatile())
2018 if (!canConvertValue(TD, PartitionTy, LI->getType()))
2020 } else if (StoreInst *SI = dyn_cast<StoreInst>(I->U->getUser())) {
2021 if (SI->isVolatile())
2023 if (!canConvertValue(TD, SI->getValueOperand()->getType(), PartitionTy))
2032 /// \brief Test whether the given alloca partition's integer operations can be
2033 /// widened to promotable ones.
2035 /// This is a quick test to check whether we can rewrite the integer loads and
2036 /// stores to a particular alloca into wider loads and stores and be able to
2037 /// promote the resulting alloca.
2038 static bool isIntegerWideningViable(const DataLayout &TD,
2040 uint64_t AllocBeginOffset,
2041 AllocaPartitioning &P,
2042 AllocaPartitioning::const_use_iterator I,
2043 AllocaPartitioning::const_use_iterator E) {
2044 uint64_t SizeInBits = TD.getTypeSizeInBits(AllocaTy);
2045 // Don't create integer types larger than the maximum bitwidth.
2046 if (SizeInBits > IntegerType::MAX_INT_BITS)
2049 // Don't try to handle allocas with bit-padding.
2050 if (SizeInBits != TD.getTypeStoreSizeInBits(AllocaTy))
2053 // We need to ensure that an integer type with the appropriate bitwidth can
2054 // be converted to the alloca type, whatever that is. We don't want to force
2055 // the alloca itself to have an integer type if there is a more suitable one.
2056 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
2057 if (!canConvertValue(TD, AllocaTy, IntTy) ||
2058 !canConvertValue(TD, IntTy, AllocaTy))
2061 uint64_t Size = TD.getTypeStoreSize(AllocaTy);
2063 // Check the uses to ensure the uses are (likely) promotable integer uses.
2064 // Also ensure that the alloca has a covering load or store. We don't want
2065 // to widen the integer operations only to fail to promote due to some other
2066 // unsplittable entry (which we may make splittable later).
2067 bool WholeAllocaOp = false;
2068 for (; I != E; ++I) {
2070 continue; // Skip dead use.
2072 uint64_t RelBegin = I->BeginOffset - AllocBeginOffset;
2073 uint64_t RelEnd = I->EndOffset - AllocBeginOffset;
2075 // We can't reasonably handle cases where the load or store extends past
2076 // the end of the aloca's type and into its padding.
2080 if (LoadInst *LI = dyn_cast<LoadInst>(I->U->getUser())) {
2081 if (LI->isVolatile())
2083 if (RelBegin == 0 && RelEnd == Size)
2084 WholeAllocaOp = true;
2085 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
2086 if (ITy->getBitWidth() < TD.getTypeStoreSizeInBits(ITy))
2090 // Non-integer loads need to be convertible from the alloca type so that
2091 // they are promotable.
2092 if (RelBegin != 0 || RelEnd != Size ||
2093 !canConvertValue(TD, AllocaTy, LI->getType()))
2095 } else if (StoreInst *SI = dyn_cast<StoreInst>(I->U->getUser())) {
2096 Type *ValueTy = SI->getValueOperand()->getType();
2097 if (SI->isVolatile())
2099 if (RelBegin == 0 && RelEnd == Size)
2100 WholeAllocaOp = true;
2101 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
2102 if (ITy->getBitWidth() < TD.getTypeStoreSizeInBits(ITy))
2106 // Non-integer stores need to be convertible to the alloca type so that
2107 // they are promotable.
2108 if (RelBegin != 0 || RelEnd != Size ||
2109 !canConvertValue(TD, ValueTy, AllocaTy))
2111 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
2112 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
2114 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
2115 const AllocaPartitioning::MemTransferOffsets &MTO
2116 = P.getMemTransferOffsets(*MTI);
2117 if (!MTO.IsSplittable)
2120 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->U->getUser())) {
2121 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2122 II->getIntrinsicID() != Intrinsic::lifetime_end)
2128 return WholeAllocaOp;
2131 static Value *extractInteger(const DataLayout &DL, IRBuilder<> &IRB, Value *V,
2132 IntegerType *Ty, uint64_t Offset,
2133 const Twine &Name) {
2134 DEBUG(dbgs() << " start: " << *V << "\n");
2135 IntegerType *IntTy = cast<IntegerType>(V->getType());
2136 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2137 "Element extends past full value");
2138 uint64_t ShAmt = 8*Offset;
2139 if (DL.isBigEndian())
2140 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2142 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2143 DEBUG(dbgs() << " shifted: " << *V << "\n");
2145 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2146 "Cannot extract to a larger integer!");
2148 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2149 DEBUG(dbgs() << " trunced: " << *V << "\n");
2154 static Value *insertInteger(const DataLayout &DL, IRBuilder<> &IRB, Value *Old,
2155 Value *V, uint64_t Offset, const Twine &Name) {
2156 IntegerType *IntTy = cast<IntegerType>(Old->getType());
2157 IntegerType *Ty = cast<IntegerType>(V->getType());
2158 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2159 "Cannot insert a larger integer!");
2160 DEBUG(dbgs() << " start: " << *V << "\n");
2162 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2163 DEBUG(dbgs() << " extended: " << *V << "\n");
2165 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2166 "Element store outside of alloca store");
2167 uint64_t ShAmt = 8*Offset;
2168 if (DL.isBigEndian())
2169 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2171 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2172 DEBUG(dbgs() << " shifted: " << *V << "\n");
2175 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2176 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2177 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2178 DEBUG(dbgs() << " masked: " << *Old << "\n");
2179 V = IRB.CreateOr(Old, V, Name + ".insert");
2180 DEBUG(dbgs() << " inserted: " << *V << "\n");
2185 static Value *extractVector(IRBuilder<> &IRB, Value *V,
2186 unsigned BeginIndex, unsigned EndIndex,
2187 const Twine &Name) {
2188 VectorType *VecTy = cast<VectorType>(V->getType());
2189 unsigned NumElements = EndIndex - BeginIndex;
2190 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2192 if (NumElements == VecTy->getNumElements())
2195 if (NumElements == 1) {
2196 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2198 DEBUG(dbgs() << " extract: " << *V << "\n");
2202 SmallVector<Constant*, 8> Mask;
2203 Mask.reserve(NumElements);
2204 for (unsigned i = BeginIndex; i != EndIndex; ++i)
2205 Mask.push_back(IRB.getInt32(i));
2206 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2207 ConstantVector::get(Mask),
2209 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2213 static Value *insertVector(IRBuilder<> &IRB, Value *Old, Value *V,
2214 unsigned BeginIndex, const Twine &Name) {
2215 VectorType *VecTy = cast<VectorType>(Old->getType());
2216 assert(VecTy && "Can only insert a vector into a vector");
2218 VectorType *Ty = dyn_cast<VectorType>(V->getType());
2220 // Single element to insert.
2221 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2223 DEBUG(dbgs() << " insert: " << *V << "\n");
2227 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
2228 "Too many elements!");
2229 if (Ty->getNumElements() == VecTy->getNumElements()) {
2230 assert(V->getType() == VecTy && "Vector type mismatch");
2233 unsigned EndIndex = BeginIndex + Ty->getNumElements();
2235 // When inserting a smaller vector into the larger to store, we first
2236 // use a shuffle vector to widen it with undef elements, and then
2237 // a second shuffle vector to select between the loaded vector and the
2239 SmallVector<Constant*, 8> Mask;
2240 Mask.reserve(VecTy->getNumElements());
2241 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2242 if (i >= BeginIndex && i < EndIndex)
2243 Mask.push_back(IRB.getInt32(i - BeginIndex));
2245 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2246 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2247 ConstantVector::get(Mask),
2249 DEBUG(dbgs() << " shuffle1: " << *V << "\n");
2252 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2253 if (i >= BeginIndex && i < EndIndex)
2254 Mask.push_back(IRB.getInt32(i));
2256 Mask.push_back(IRB.getInt32(i + VecTy->getNumElements()));
2257 V = IRB.CreateShuffleVector(V, Old, ConstantVector::get(Mask),
2259 DEBUG(dbgs() << " shuffle2: " << *V << "\n");
2264 /// \brief Visitor to rewrite instructions using a partition of an alloca to
2265 /// use a new alloca.
2267 /// Also implements the rewriting to vector-based accesses when the partition
2268 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2270 class AllocaPartitionRewriter : public InstVisitor<AllocaPartitionRewriter,
2272 // Befriend the base class so it can delegate to private visit methods.
2273 friend class llvm::InstVisitor<AllocaPartitionRewriter, bool>;
2275 const DataLayout &TD;
2276 AllocaPartitioning &P;
2278 AllocaInst &OldAI, &NewAI;
2279 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2282 // If we are rewriting an alloca partition which can be written as pure
2283 // vector operations, we stash extra information here. When VecTy is
2284 // non-null, we have some strict guarantees about the rewritten alloca:
2285 // - The new alloca is exactly the size of the vector type here.
2286 // - The accesses all either map to the entire vector or to a single
2288 // - The set of accessing instructions is only one of those handled above
2289 // in isVectorPromotionViable. Generally these are the same access kinds
2290 // which are promotable via mem2reg.
2293 uint64_t ElementSize;
2295 // This is a convenience and flag variable that will be null unless the new
2296 // alloca's integer operations should be widened to this integer type due to
2297 // passing isIntegerWideningViable above. If it is non-null, the desired
2298 // integer type will be stored here for easy access during rewriting.
2301 // The offset of the partition user currently being rewritten.
2302 uint64_t BeginOffset, EndOffset;
2304 Instruction *OldPtr;
2306 // The name prefix to use when rewriting instructions for this alloca.
2307 std::string NamePrefix;
2310 AllocaPartitionRewriter(const DataLayout &TD, AllocaPartitioning &P,
2311 AllocaPartitioning::iterator PI,
2312 SROA &Pass, AllocaInst &OldAI, AllocaInst &NewAI,
2313 uint64_t NewBeginOffset, uint64_t NewEndOffset)
2314 : TD(TD), P(P), Pass(Pass),
2315 OldAI(OldAI), NewAI(NewAI),
2316 NewAllocaBeginOffset(NewBeginOffset),
2317 NewAllocaEndOffset(NewEndOffset),
2318 NewAllocaTy(NewAI.getAllocatedType()),
2319 VecTy(), ElementTy(), ElementSize(), IntTy(),
2320 BeginOffset(), EndOffset() {
2323 /// \brief Visit the users of the alloca partition and rewrite them.
2324 bool visitUsers(AllocaPartitioning::const_use_iterator I,
2325 AllocaPartitioning::const_use_iterator E) {
2326 if (isVectorPromotionViable(TD, NewAI.getAllocatedType(), P,
2327 NewAllocaBeginOffset, NewAllocaEndOffset,
2330 VecTy = cast<VectorType>(NewAI.getAllocatedType());
2331 ElementTy = VecTy->getElementType();
2332 assert((TD.getTypeSizeInBits(VecTy->getScalarType()) % 8) == 0 &&
2333 "Only multiple-of-8 sized vector elements are viable");
2334 ElementSize = TD.getTypeSizeInBits(VecTy->getScalarType()) / 8;
2335 } else if (isIntegerWideningViable(TD, NewAI.getAllocatedType(),
2336 NewAllocaBeginOffset, P, I, E)) {
2337 IntTy = Type::getIntNTy(NewAI.getContext(),
2338 TD.getTypeSizeInBits(NewAI.getAllocatedType()));
2340 bool CanSROA = true;
2341 for (; I != E; ++I) {
2343 continue; // Skip dead uses.
2344 BeginOffset = I->BeginOffset;
2345 EndOffset = I->EndOffset;
2347 OldPtr = cast<Instruction>(I->U->get());
2348 NamePrefix = (Twine(NewAI.getName()) + "." + Twine(BeginOffset)).str();
2349 CanSROA &= visit(cast<Instruction>(I->U->getUser()));
2365 // Every instruction which can end up as a user must have a rewrite rule.
2366 bool visitInstruction(Instruction &I) {
2367 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2368 llvm_unreachable("No rewrite rule for this instruction!");
2371 Twine getName(const Twine &Suffix) {
2372 return NamePrefix + Suffix;
2375 Value *getAdjustedAllocaPtr(IRBuilder<> &IRB, Type *PointerTy) {
2376 assert(BeginOffset >= NewAllocaBeginOffset);
2377 APInt Offset(TD.getPointerSizeInBits(), BeginOffset - NewAllocaBeginOffset);
2378 return getAdjustedPtr(IRB, TD, &NewAI, Offset, PointerTy, getName(""));
2381 /// \brief Compute suitable alignment to access an offset into the new alloca.
2382 unsigned getOffsetAlign(uint64_t Offset) {
2383 unsigned NewAIAlign = NewAI.getAlignment();
2385 NewAIAlign = TD.getABITypeAlignment(NewAI.getAllocatedType());
2386 return MinAlign(NewAIAlign, Offset);
2389 /// \brief Compute suitable alignment to access this partition of the new
2391 unsigned getPartitionAlign() {
2392 return getOffsetAlign(BeginOffset - NewAllocaBeginOffset);
2395 /// \brief Compute suitable alignment to access a type at an offset of the
2398 /// \returns zero if the type's ABI alignment is a suitable alignment,
2399 /// otherwise returns the maximal suitable alignment.
2400 unsigned getOffsetTypeAlign(Type *Ty, uint64_t Offset) {
2401 unsigned Align = getOffsetAlign(Offset);
2402 return Align == TD.getABITypeAlignment(Ty) ? 0 : Align;
2405 /// \brief Compute suitable alignment to access a type at the beginning of
2406 /// this partition of the new alloca.
2408 /// See \c getOffsetTypeAlign for details; this routine delegates to it.
2409 unsigned getPartitionTypeAlign(Type *Ty) {
2410 return getOffsetTypeAlign(Ty, BeginOffset - NewAllocaBeginOffset);
2413 unsigned getIndex(uint64_t Offset) {
2414 assert(VecTy && "Can only call getIndex when rewriting a vector");
2415 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2416 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2417 uint32_t Index = RelOffset / ElementSize;
2418 assert(Index * ElementSize == RelOffset);
2422 void deleteIfTriviallyDead(Value *V) {
2423 Instruction *I = cast<Instruction>(V);
2424 if (isInstructionTriviallyDead(I))
2425 Pass.DeadInsts.insert(I);
2428 Value *rewriteVectorizedLoadInst(IRBuilder<> &IRB) {
2429 unsigned BeginIndex = getIndex(BeginOffset);
2430 unsigned EndIndex = getIndex(EndOffset);
2431 assert(EndIndex > BeginIndex && "Empty vector!");
2433 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2435 return extractVector(IRB, V, BeginIndex, EndIndex, getName(".vec"));
2438 Value *rewriteIntegerLoad(IRBuilder<> &IRB, LoadInst &LI) {
2439 assert(IntTy && "We cannot insert an integer to the alloca");
2440 assert(!LI.isVolatile());
2441 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2443 V = convertValue(TD, IRB, V, IntTy);
2444 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2445 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2446 if (Offset > 0 || EndOffset < NewAllocaEndOffset)
2447 V = extractInteger(TD, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2448 getName(".extract"));
2452 bool visitLoadInst(LoadInst &LI) {
2453 DEBUG(dbgs() << " original: " << LI << "\n");
2454 Value *OldOp = LI.getOperand(0);
2455 assert(OldOp == OldPtr);
2456 IRBuilder<> IRB(&LI);
2458 uint64_t Size = EndOffset - BeginOffset;
2459 bool IsSplitIntLoad = Size < TD.getTypeStoreSize(LI.getType());
2461 // If this memory access can be shown to *statically* extend outside the
2462 // bounds of the original allocation it's behavior is undefined. Rather
2463 // than trying to transform it, just replace it with undef.
2464 // FIXME: We should do something more clever for functions being
2465 // instrumented by asan.
2466 // FIXME: Eventually, once ASan and friends can flush out bugs here, this
2467 // should be transformed to a load of null making it unreachable.
2468 uint64_t OldAllocSize = TD.getTypeAllocSize(OldAI.getAllocatedType());
2469 if (TD.getTypeStoreSize(LI.getType()) > OldAllocSize) {
2470 LI.replaceAllUsesWith(UndefValue::get(LI.getType()));
2471 Pass.DeadInsts.insert(&LI);
2472 deleteIfTriviallyDead(OldOp);
2473 DEBUG(dbgs() << " to: undef!!\n");
2477 Type *TargetTy = IsSplitIntLoad ? Type::getIntNTy(LI.getContext(), Size * 8)
2479 bool IsPtrAdjusted = false;
2482 V = rewriteVectorizedLoadInst(IRB);
2483 } else if (IntTy && LI.getType()->isIntegerTy()) {
2484 V = rewriteIntegerLoad(IRB, LI);
2485 } else if (BeginOffset == NewAllocaBeginOffset &&
2486 canConvertValue(TD, NewAllocaTy, LI.getType())) {
2487 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2488 LI.isVolatile(), getName(".load"));
2490 Type *LTy = TargetTy->getPointerTo();
2491 V = IRB.CreateAlignedLoad(getAdjustedAllocaPtr(IRB, LTy),
2492 getPartitionTypeAlign(TargetTy),
2493 LI.isVolatile(), getName(".load"));
2494 IsPtrAdjusted = true;
2496 V = convertValue(TD, IRB, V, TargetTy);
2498 if (IsSplitIntLoad) {
2499 assert(!LI.isVolatile());
2500 assert(LI.getType()->isIntegerTy() &&
2501 "Only integer type loads and stores are split");
2502 assert(LI.getType()->getIntegerBitWidth() ==
2503 TD.getTypeStoreSizeInBits(LI.getType()) &&
2504 "Non-byte-multiple bit width");
2505 assert(LI.getType()->getIntegerBitWidth() ==
2506 TD.getTypeAllocSizeInBits(OldAI.getAllocatedType()) &&
2507 "Only alloca-wide loads can be split and recomposed");
2508 // Move the insertion point just past the load so that we can refer to it.
2509 IRB.SetInsertPoint(llvm::next(BasicBlock::iterator(&LI)));
2510 // Create a placeholder value with the same type as LI to use as the
2511 // basis for the new value. This allows us to replace the uses of LI with
2512 // the computed value, and then replace the placeholder with LI, leaving
2513 // LI only used for this computation.
2515 = new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2516 V = insertInteger(TD, IRB, Placeholder, V, BeginOffset,
2517 getName(".insert"));
2518 LI.replaceAllUsesWith(V);
2519 Placeholder->replaceAllUsesWith(&LI);
2522 LI.replaceAllUsesWith(V);
2525 Pass.DeadInsts.insert(&LI);
2526 deleteIfTriviallyDead(OldOp);
2527 DEBUG(dbgs() << " to: " << *V << "\n");
2528 return !LI.isVolatile() && !IsPtrAdjusted;
2531 bool rewriteVectorizedStoreInst(IRBuilder<> &IRB, Value *V,
2532 StoreInst &SI, Value *OldOp) {
2533 unsigned BeginIndex = getIndex(BeginOffset);
2534 unsigned EndIndex = getIndex(EndOffset);
2535 assert(EndIndex > BeginIndex && "Empty vector!");
2536 unsigned NumElements = EndIndex - BeginIndex;
2537 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2539 = (NumElements == 1) ? ElementTy
2540 : VectorType::get(ElementTy, NumElements);
2541 if (V->getType() != PartitionTy)
2542 V = convertValue(TD, IRB, V, PartitionTy);
2544 // Mix in the existing elements.
2545 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2547 V = insertVector(IRB, Old, V, BeginIndex, getName(".vec"));
2549 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2550 Pass.DeadInsts.insert(&SI);
2553 DEBUG(dbgs() << " to: " << *Store << "\n");
2557 bool rewriteIntegerStore(IRBuilder<> &IRB, Value *V, StoreInst &SI) {
2558 assert(IntTy && "We cannot extract an integer from the alloca");
2559 assert(!SI.isVolatile());
2560 if (TD.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2561 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2562 getName(".oldload"));
2563 Old = convertValue(TD, IRB, Old, IntTy);
2564 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2565 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2566 V = insertInteger(TD, IRB, Old, SI.getValueOperand(), Offset,
2567 getName(".insert"));
2569 V = convertValue(TD, IRB, V, NewAllocaTy);
2570 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2571 Pass.DeadInsts.insert(&SI);
2573 DEBUG(dbgs() << " to: " << *Store << "\n");
2577 bool visitStoreInst(StoreInst &SI) {
2578 DEBUG(dbgs() << " original: " << SI << "\n");
2579 Value *OldOp = SI.getOperand(1);
2580 assert(OldOp == OldPtr);
2581 IRBuilder<> IRB(&SI);
2583 Value *V = SI.getValueOperand();
2585 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2586 // alloca that should be re-examined after promoting this alloca.
2587 if (V->getType()->isPointerTy())
2588 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2589 Pass.PostPromotionWorklist.insert(AI);
2591 uint64_t Size = EndOffset - BeginOffset;
2592 if (Size < TD.getTypeStoreSize(V->getType())) {
2593 assert(!SI.isVolatile());
2594 assert(V->getType()->isIntegerTy() &&
2595 "Only integer type loads and stores are split");
2596 assert(V->getType()->getIntegerBitWidth() ==
2597 TD.getTypeStoreSizeInBits(V->getType()) &&
2598 "Non-byte-multiple bit width");
2599 assert(V->getType()->getIntegerBitWidth() ==
2600 TD.getTypeAllocSizeInBits(OldAI.getAllocatedType()) &&
2601 "Only alloca-wide stores can be split and recomposed");
2602 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), Size * 8);
2603 V = extractInteger(TD, IRB, V, NarrowTy, BeginOffset,
2604 getName(".extract"));
2608 return rewriteVectorizedStoreInst(IRB, V, SI, OldOp);
2609 if (IntTy && V->getType()->isIntegerTy())
2610 return rewriteIntegerStore(IRB, V, SI);
2613 if (BeginOffset == NewAllocaBeginOffset &&
2614 canConvertValue(TD, V->getType(), NewAllocaTy)) {
2615 V = convertValue(TD, IRB, V, NewAllocaTy);
2616 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2619 Value *NewPtr = getAdjustedAllocaPtr(IRB, V->getType()->getPointerTo());
2620 NewSI = IRB.CreateAlignedStore(V, NewPtr,
2621 getPartitionTypeAlign(V->getType()),
2625 Pass.DeadInsts.insert(&SI);
2626 deleteIfTriviallyDead(OldOp);
2628 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2629 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2632 /// \brief Compute an integer value from splatting an i8 across the given
2633 /// number of bytes.
2635 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2636 /// call this routine.
2637 /// FIXME: Heed the advice above.
2639 /// \param V The i8 value to splat.
2640 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2641 Value *getIntegerSplat(IRBuilder<> &IRB, Value *V, unsigned Size) {
2642 assert(Size > 0 && "Expected a positive number of bytes.");
2643 IntegerType *VTy = cast<IntegerType>(V->getType());
2644 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2648 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size*8);
2649 V = IRB.CreateMul(IRB.CreateZExt(V, SplatIntTy, getName(".zext")),
2650 ConstantExpr::getUDiv(
2651 Constant::getAllOnesValue(SplatIntTy),
2652 ConstantExpr::getZExt(
2653 Constant::getAllOnesValue(V->getType()),
2655 getName(".isplat"));
2659 /// \brief Compute a vector splat for a given element value.
2660 Value *getVectorSplat(IRBuilder<> &IRB, Value *V, unsigned NumElements) {
2661 V = IRB.CreateVectorSplat(NumElements, V, NamePrefix);
2662 DEBUG(dbgs() << " splat: " << *V << "\n");
2666 bool visitMemSetInst(MemSetInst &II) {
2667 DEBUG(dbgs() << " original: " << II << "\n");
2668 IRBuilder<> IRB(&II);
2669 assert(II.getRawDest() == OldPtr);
2671 // If the memset has a variable size, it cannot be split, just adjust the
2672 // pointer to the new alloca.
2673 if (!isa<Constant>(II.getLength())) {
2674 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2675 Type *CstTy = II.getAlignmentCst()->getType();
2676 II.setAlignment(ConstantInt::get(CstTy, getPartitionAlign()));
2678 deleteIfTriviallyDead(OldPtr);
2682 // Record this instruction for deletion.
2683 Pass.DeadInsts.insert(&II);
2685 Type *AllocaTy = NewAI.getAllocatedType();
2686 Type *ScalarTy = AllocaTy->getScalarType();
2688 // If this doesn't map cleanly onto the alloca type, and that type isn't
2689 // a single value type, just emit a memset.
2690 if (!VecTy && !IntTy &&
2691 (BeginOffset != NewAllocaBeginOffset ||
2692 EndOffset != NewAllocaEndOffset ||
2693 !AllocaTy->isSingleValueType() ||
2694 !TD.isLegalInteger(TD.getTypeSizeInBits(ScalarTy)) ||
2695 TD.getTypeSizeInBits(ScalarTy)%8 != 0)) {
2696 Type *SizeTy = II.getLength()->getType();
2697 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2699 = IRB.CreateMemSet(getAdjustedAllocaPtr(IRB,
2700 II.getRawDest()->getType()),
2701 II.getValue(), Size, getPartitionAlign(),
2704 DEBUG(dbgs() << " to: " << *New << "\n");
2708 // If we can represent this as a simple value, we have to build the actual
2709 // value to store, which requires expanding the byte present in memset to
2710 // a sensible representation for the alloca type. This is essentially
2711 // splatting the byte to a sufficiently wide integer, splatting it across
2712 // any desired vector width, and bitcasting to the final type.
2716 // If this is a memset of a vectorized alloca, insert it.
2717 assert(ElementTy == ScalarTy);
2719 unsigned BeginIndex = getIndex(BeginOffset);
2720 unsigned EndIndex = getIndex(EndOffset);
2721 assert(EndIndex > BeginIndex && "Empty vector!");
2722 unsigned NumElements = EndIndex - BeginIndex;
2723 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2725 Value *Splat = getIntegerSplat(IRB, II.getValue(),
2726 TD.getTypeSizeInBits(ElementTy)/8);
2727 Splat = convertValue(TD, IRB, Splat, ElementTy);
2728 if (NumElements > 1)
2729 Splat = getVectorSplat(IRB, Splat, NumElements);
2731 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2732 getName(".oldload"));
2733 V = insertVector(IRB, Old, Splat, BeginIndex, getName(".vec"));
2735 // If this is a memset on an alloca where we can widen stores, insert the
2737 assert(!II.isVolatile());
2739 uint64_t Size = EndOffset - BeginOffset;
2740 V = getIntegerSplat(IRB, II.getValue(), Size);
2742 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2743 EndOffset != NewAllocaBeginOffset)) {
2744 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2745 getName(".oldload"));
2746 Old = convertValue(TD, IRB, Old, IntTy);
2747 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2748 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2749 V = insertInteger(TD, IRB, Old, V, Offset, getName(".insert"));
2751 assert(V->getType() == IntTy &&
2752 "Wrong type for an alloca wide integer!");
2754 V = convertValue(TD, IRB, V, AllocaTy);
2756 // Established these invariants above.
2757 assert(BeginOffset == NewAllocaBeginOffset);
2758 assert(EndOffset == NewAllocaEndOffset);
2760 V = getIntegerSplat(IRB, II.getValue(),
2761 TD.getTypeSizeInBits(ScalarTy)/8);
2762 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2763 V = getVectorSplat(IRB, V, AllocaVecTy->getNumElements());
2765 V = convertValue(TD, IRB, V, AllocaTy);
2768 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2771 DEBUG(dbgs() << " to: " << *New << "\n");
2772 return !II.isVolatile();
2775 bool visitMemTransferInst(MemTransferInst &II) {
2776 // Rewriting of memory transfer instructions can be a bit tricky. We break
2777 // them into two categories: split intrinsics and unsplit intrinsics.
2779 DEBUG(dbgs() << " original: " << II << "\n");
2780 IRBuilder<> IRB(&II);
2782 assert(II.getRawSource() == OldPtr || II.getRawDest() == OldPtr);
2783 bool IsDest = II.getRawDest() == OldPtr;
2785 const AllocaPartitioning::MemTransferOffsets &MTO
2786 = P.getMemTransferOffsets(II);
2788 // Compute the relative offset within the transfer.
2789 unsigned IntPtrWidth = TD.getPointerSizeInBits();
2790 APInt RelOffset(IntPtrWidth, BeginOffset - (IsDest ? MTO.DestBegin
2791 : MTO.SourceBegin));
2793 unsigned Align = II.getAlignment();
2795 Align = MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(),
2796 MinAlign(II.getAlignment(), getPartitionAlign()));
2798 // For unsplit intrinsics, we simply modify the source and destination
2799 // pointers in place. This isn't just an optimization, it is a matter of
2800 // correctness. With unsplit intrinsics we may be dealing with transfers
2801 // within a single alloca before SROA ran, or with transfers that have
2802 // a variable length. We may also be dealing with memmove instead of
2803 // memcpy, and so simply updating the pointers is the necessary for us to
2804 // update both source and dest of a single call.
2805 if (!MTO.IsSplittable) {
2806 Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource();
2808 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2810 II.setSource(getAdjustedAllocaPtr(IRB, II.getRawSource()->getType()));
2812 Type *CstTy = II.getAlignmentCst()->getType();
2813 II.setAlignment(ConstantInt::get(CstTy, Align));
2815 DEBUG(dbgs() << " to: " << II << "\n");
2816 deleteIfTriviallyDead(OldOp);
2819 // For split transfer intrinsics we have an incredibly useful assurance:
2820 // the source and destination do not reside within the same alloca, and at
2821 // least one of them does not escape. This means that we can replace
2822 // memmove with memcpy, and we don't need to worry about all manner of
2823 // downsides to splitting and transforming the operations.
2825 // If this doesn't map cleanly onto the alloca type, and that type isn't
2826 // a single value type, just emit a memcpy.
2828 = !VecTy && !IntTy && (BeginOffset != NewAllocaBeginOffset ||
2829 EndOffset != NewAllocaEndOffset ||
2830 !NewAI.getAllocatedType()->isSingleValueType());
2832 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2833 // size hasn't been shrunk based on analysis of the viable range, this is
2835 if (EmitMemCpy && &OldAI == &NewAI) {
2836 uint64_t OrigBegin = IsDest ? MTO.DestBegin : MTO.SourceBegin;
2837 uint64_t OrigEnd = IsDest ? MTO.DestEnd : MTO.SourceEnd;
2838 // Ensure the start lines up.
2839 assert(BeginOffset == OrigBegin);
2842 // Rewrite the size as needed.
2843 if (EndOffset != OrigEnd)
2844 II.setLength(ConstantInt::get(II.getLength()->getType(),
2845 EndOffset - BeginOffset));
2848 // Record this instruction for deletion.
2849 Pass.DeadInsts.insert(&II);
2851 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2852 // alloca that should be re-examined after rewriting this instruction.
2853 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2855 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets()))
2856 Pass.Worklist.insert(AI);
2859 Type *OtherPtrTy = IsDest ? II.getRawSource()->getType()
2860 : II.getRawDest()->getType();
2862 // Compute the other pointer, folding as much as possible to produce
2863 // a single, simple GEP in most cases.
2864 OtherPtr = getAdjustedPtr(IRB, TD, OtherPtr, RelOffset, OtherPtrTy,
2865 getName("." + OtherPtr->getName()));
2868 = getAdjustedAllocaPtr(IRB, IsDest ? II.getRawDest()->getType()
2869 : II.getRawSource()->getType());
2870 Type *SizeTy = II.getLength()->getType();
2871 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2873 CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
2874 IsDest ? OtherPtr : OurPtr,
2875 Size, Align, II.isVolatile());
2877 DEBUG(dbgs() << " to: " << *New << "\n");
2881 // Note that we clamp the alignment to 1 here as a 0 alignment for a memcpy
2882 // is equivalent to 1, but that isn't true if we end up rewriting this as
2887 bool IsWholeAlloca = BeginOffset == NewAllocaBeginOffset &&
2888 EndOffset == NewAllocaEndOffset;
2889 uint64_t Size = EndOffset - BeginOffset;
2890 unsigned BeginIndex = VecTy ? getIndex(BeginOffset) : 0;
2891 unsigned EndIndex = VecTy ? getIndex(EndOffset) : 0;
2892 unsigned NumElements = EndIndex - BeginIndex;
2893 IntegerType *SubIntTy
2894 = IntTy ? Type::getIntNTy(IntTy->getContext(), Size*8) : 0;
2896 Type *OtherPtrTy = NewAI.getType();
2897 if (VecTy && !IsWholeAlloca) {
2898 if (NumElements == 1)
2899 OtherPtrTy = VecTy->getElementType();
2901 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2903 OtherPtrTy = OtherPtrTy->getPointerTo();
2904 } else if (IntTy && !IsWholeAlloca) {
2905 OtherPtrTy = SubIntTy->getPointerTo();
2908 Value *SrcPtr = getAdjustedPtr(IRB, TD, OtherPtr, RelOffset, OtherPtrTy,
2909 getName("." + OtherPtr->getName()));
2910 Value *DstPtr = &NewAI;
2912 std::swap(SrcPtr, DstPtr);
2915 if (VecTy && !IsWholeAlloca && !IsDest) {
2916 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2918 Src = extractVector(IRB, Src, BeginIndex, EndIndex, getName(".vec"));
2919 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2920 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2922 Src = convertValue(TD, IRB, Src, IntTy);
2923 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2924 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2925 Src = extractInteger(TD, IRB, Src, SubIntTy, Offset, getName(".extract"));
2927 Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(),
2928 getName(".copyload"));
2931 if (VecTy && !IsWholeAlloca && IsDest) {
2932 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2933 getName(".oldload"));
2934 Src = insertVector(IRB, Old, Src, BeginIndex, getName(".vec"));
2935 } else if (IntTy && !IsWholeAlloca && IsDest) {
2936 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2937 getName(".oldload"));
2938 Old = convertValue(TD, IRB, Old, IntTy);
2939 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2940 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2941 Src = insertInteger(TD, IRB, Old, Src, Offset, getName(".insert"));
2942 Src = convertValue(TD, IRB, Src, NewAllocaTy);
2945 StoreInst *Store = cast<StoreInst>(
2946 IRB.CreateAlignedStore(Src, DstPtr, Align, II.isVolatile()));
2948 DEBUG(dbgs() << " to: " << *Store << "\n");
2949 return !II.isVolatile();
2952 bool visitIntrinsicInst(IntrinsicInst &II) {
2953 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2954 II.getIntrinsicID() == Intrinsic::lifetime_end);
2955 DEBUG(dbgs() << " original: " << II << "\n");
2956 IRBuilder<> IRB(&II);
2957 assert(II.getArgOperand(1) == OldPtr);
2959 // Record this instruction for deletion.
2960 Pass.DeadInsts.insert(&II);
2963 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2964 EndOffset - BeginOffset);
2965 Value *Ptr = getAdjustedAllocaPtr(IRB, II.getArgOperand(1)->getType());
2967 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2968 New = IRB.CreateLifetimeStart(Ptr, Size);
2970 New = IRB.CreateLifetimeEnd(Ptr, Size);
2973 DEBUG(dbgs() << " to: " << *New << "\n");
2977 bool visitPHINode(PHINode &PN) {
2978 DEBUG(dbgs() << " original: " << PN << "\n");
2980 // We would like to compute a new pointer in only one place, but have it be
2981 // as local as possible to the PHI. To do that, we re-use the location of
2982 // the old pointer, which necessarily must be in the right position to
2983 // dominate the PHI.
2984 IRBuilder<> PtrBuilder(cast<Instruction>(OldPtr));
2986 Value *NewPtr = getAdjustedAllocaPtr(PtrBuilder, OldPtr->getType());
2987 // Replace the operands which were using the old pointer.
2988 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
2990 DEBUG(dbgs() << " to: " << PN << "\n");
2991 deleteIfTriviallyDead(OldPtr);
2995 bool visitSelectInst(SelectInst &SI) {
2996 DEBUG(dbgs() << " original: " << SI << "\n");
2997 IRBuilder<> IRB(&SI);
2999 // Find the operand we need to rewrite here.
3000 bool IsTrueVal = SI.getTrueValue() == OldPtr;
3002 assert(SI.getFalseValue() != OldPtr && "Pointer is both operands!");
3004 assert(SI.getFalseValue() == OldPtr && "Pointer isn't an operand!");
3006 Value *NewPtr = getAdjustedAllocaPtr(IRB, OldPtr->getType());
3007 SI.setOperand(IsTrueVal ? 1 : 2, NewPtr);
3008 DEBUG(dbgs() << " to: " << SI << "\n");
3009 deleteIfTriviallyDead(OldPtr);
3017 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
3019 /// This pass aggressively rewrites all aggregate loads and stores on
3020 /// a particular pointer (or any pointer derived from it which we can identify)
3021 /// with scalar loads and stores.
3022 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
3023 // Befriend the base class so it can delegate to private visit methods.
3024 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
3026 const DataLayout &TD;
3028 /// Queue of pointer uses to analyze and potentially rewrite.
3029 SmallVector<Use *, 8> Queue;
3031 /// Set to prevent us from cycling with phi nodes and loops.
3032 SmallPtrSet<User *, 8> Visited;
3034 /// The current pointer use being rewritten. This is used to dig up the used
3035 /// value (as opposed to the user).
3039 AggLoadStoreRewriter(const DataLayout &TD) : TD(TD) {}
3041 /// Rewrite loads and stores through a pointer and all pointers derived from
3043 bool rewrite(Instruction &I) {
3044 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
3046 bool Changed = false;
3047 while (!Queue.empty()) {
3048 U = Queue.pop_back_val();
3049 Changed |= visit(cast<Instruction>(U->getUser()));
3055 /// Enqueue all the users of the given instruction for further processing.
3056 /// This uses a set to de-duplicate users.
3057 void enqueueUsers(Instruction &I) {
3058 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
3060 if (Visited.insert(*UI))
3061 Queue.push_back(&UI.getUse());
3064 // Conservative default is to not rewrite anything.
3065 bool visitInstruction(Instruction &I) { return false; }
3067 /// \brief Generic recursive split emission class.
3068 template <typename Derived>
3071 /// The builder used to form new instructions.
3073 /// The indices which to be used with insert- or extractvalue to select the
3074 /// appropriate value within the aggregate.
3075 SmallVector<unsigned, 4> Indices;
3076 /// The indices to a GEP instruction which will move Ptr to the correct slot
3077 /// within the aggregate.
3078 SmallVector<Value *, 4> GEPIndices;
3079 /// The base pointer of the original op, used as a base for GEPing the
3080 /// split operations.
3083 /// Initialize the splitter with an insertion point, Ptr and start with a
3084 /// single zero GEP index.
3085 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
3086 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
3089 /// \brief Generic recursive split emission routine.
3091 /// This method recursively splits an aggregate op (load or store) into
3092 /// scalar or vector ops. It splits recursively until it hits a single value
3093 /// and emits that single value operation via the template argument.
3095 /// The logic of this routine relies on GEPs and insertvalue and
3096 /// extractvalue all operating with the same fundamental index list, merely
3097 /// formatted differently (GEPs need actual values).
3099 /// \param Ty The type being split recursively into smaller ops.
3100 /// \param Agg The aggregate value being built up or stored, depending on
3101 /// whether this is splitting a load or a store respectively.
3102 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3103 if (Ty->isSingleValueType())
3104 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
3106 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3107 unsigned OldSize = Indices.size();
3109 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3111 assert(Indices.size() == OldSize && "Did not return to the old size");
3112 Indices.push_back(Idx);
3113 GEPIndices.push_back(IRB.getInt32(Idx));
3114 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3115 GEPIndices.pop_back();
3121 if (StructType *STy = dyn_cast<StructType>(Ty)) {
3122 unsigned OldSize = Indices.size();
3124 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3126 assert(Indices.size() == OldSize && "Did not return to the old size");
3127 Indices.push_back(Idx);
3128 GEPIndices.push_back(IRB.getInt32(Idx));
3129 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3130 GEPIndices.pop_back();
3136 llvm_unreachable("Only arrays and structs are aggregate loadable types");
3140 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3141 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3142 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
3144 /// Emit a leaf load of a single value. This is called at the leaves of the
3145 /// recursive emission to actually load values.
3146 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3147 assert(Ty->isSingleValueType());
3148 // Load the single value and insert it using the indices.
3149 Value *Load = IRB.CreateLoad(IRB.CreateInBoundsGEP(Ptr, GEPIndices,
3152 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3153 DEBUG(dbgs() << " to: " << *Load << "\n");
3157 bool visitLoadInst(LoadInst &LI) {
3158 assert(LI.getPointerOperand() == *U);
3159 if (!LI.isSimple() || LI.getType()->isSingleValueType())
3162 // We have an aggregate being loaded, split it apart.
3163 DEBUG(dbgs() << " original: " << LI << "\n");
3164 LoadOpSplitter Splitter(&LI, *U);
3165 Value *V = UndefValue::get(LI.getType());
3166 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3167 LI.replaceAllUsesWith(V);
3168 LI.eraseFromParent();
3172 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3173 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3174 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
3176 /// Emit a leaf store of a single value. This is called at the leaves of the
3177 /// recursive emission to actually produce stores.
3178 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3179 assert(Ty->isSingleValueType());
3180 // Extract the single value and store it using the indices.
3181 Value *Store = IRB.CreateStore(
3182 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
3183 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
3185 DEBUG(dbgs() << " to: " << *Store << "\n");
3189 bool visitStoreInst(StoreInst &SI) {
3190 if (!SI.isSimple() || SI.getPointerOperand() != *U)
3192 Value *V = SI.getValueOperand();
3193 if (V->getType()->isSingleValueType())
3196 // We have an aggregate being stored, split it apart.
3197 DEBUG(dbgs() << " original: " << SI << "\n");
3198 StoreOpSplitter Splitter(&SI, *U);
3199 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3200 SI.eraseFromParent();
3204 bool visitBitCastInst(BitCastInst &BC) {
3209 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3214 bool visitPHINode(PHINode &PN) {
3219 bool visitSelectInst(SelectInst &SI) {
3226 /// \brief Strip aggregate type wrapping.
3228 /// This removes no-op aggregate types wrapping an underlying type. It will
3229 /// strip as many layers of types as it can without changing either the type
3230 /// size or the allocated size.
3231 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
3232 if (Ty->isSingleValueType())
3235 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3236 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
3239 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3240 InnerTy = ArrTy->getElementType();
3241 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3242 const StructLayout *SL = DL.getStructLayout(STy);
3243 unsigned Index = SL->getElementContainingOffset(0);
3244 InnerTy = STy->getElementType(Index);
3249 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3250 TypeSize > DL.getTypeSizeInBits(InnerTy))
3253 return stripAggregateTypeWrapping(DL, InnerTy);
3256 /// \brief Try to find a partition of the aggregate type passed in for a given
3257 /// offset and size.
3259 /// This recurses through the aggregate type and tries to compute a subtype
3260 /// based on the offset and size. When the offset and size span a sub-section
3261 /// of an array, it will even compute a new array type for that sub-section,
3262 /// and the same for structs.
3264 /// Note that this routine is very strict and tries to find a partition of the
3265 /// type which produces the *exact* right offset and size. It is not forgiving
3266 /// when the size or offset cause either end of type-based partition to be off.
3267 /// Also, this is a best-effort routine. It is reasonable to give up and not
3268 /// return a type if necessary.
3269 static Type *getTypePartition(const DataLayout &TD, Type *Ty,
3270 uint64_t Offset, uint64_t Size) {
3271 if (Offset == 0 && TD.getTypeAllocSize(Ty) == Size)
3272 return stripAggregateTypeWrapping(TD, Ty);
3273 if (Offset > TD.getTypeAllocSize(Ty) ||
3274 (TD.getTypeAllocSize(Ty) - Offset) < Size)
3277 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3278 // We can't partition pointers...
3279 if (SeqTy->isPointerTy())
3282 Type *ElementTy = SeqTy->getElementType();
3283 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
3284 uint64_t NumSkippedElements = Offset / ElementSize;
3285 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy))
3286 if (NumSkippedElements >= ArrTy->getNumElements())
3288 if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy))
3289 if (NumSkippedElements >= VecTy->getNumElements())
3291 Offset -= NumSkippedElements * ElementSize;
3293 // First check if we need to recurse.
3294 if (Offset > 0 || Size < ElementSize) {
3295 // Bail if the partition ends in a different array element.
3296 if ((Offset + Size) > ElementSize)
3298 // Recurse through the element type trying to peel off offset bytes.
3299 return getTypePartition(TD, ElementTy, Offset, Size);
3301 assert(Offset == 0);
3303 if (Size == ElementSize)
3304 return stripAggregateTypeWrapping(TD, ElementTy);
3305 assert(Size > ElementSize);
3306 uint64_t NumElements = Size / ElementSize;
3307 if (NumElements * ElementSize != Size)
3309 return ArrayType::get(ElementTy, NumElements);
3312 StructType *STy = dyn_cast<StructType>(Ty);
3316 const StructLayout *SL = TD.getStructLayout(STy);
3317 if (Offset >= SL->getSizeInBytes())
3319 uint64_t EndOffset = Offset + Size;
3320 if (EndOffset > SL->getSizeInBytes())
3323 unsigned Index = SL->getElementContainingOffset(Offset);
3324 Offset -= SL->getElementOffset(Index);
3326 Type *ElementTy = STy->getElementType(Index);
3327 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
3328 if (Offset >= ElementSize)
3329 return 0; // The offset points into alignment padding.
3331 // See if any partition must be contained by the element.
3332 if (Offset > 0 || Size < ElementSize) {
3333 if ((Offset + Size) > ElementSize)
3335 return getTypePartition(TD, ElementTy, Offset, Size);
3337 assert(Offset == 0);
3339 if (Size == ElementSize)
3340 return stripAggregateTypeWrapping(TD, ElementTy);
3342 StructType::element_iterator EI = STy->element_begin() + Index,
3343 EE = STy->element_end();
3344 if (EndOffset < SL->getSizeInBytes()) {
3345 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3346 if (Index == EndIndex)
3347 return 0; // Within a single element and its padding.
3349 // Don't try to form "natural" types if the elements don't line up with the
3351 // FIXME: We could potentially recurse down through the last element in the
3352 // sub-struct to find a natural end point.
3353 if (SL->getElementOffset(EndIndex) != EndOffset)
3356 assert(Index < EndIndex);
3357 EE = STy->element_begin() + EndIndex;
3360 // Try to build up a sub-structure.
3361 StructType *SubTy = StructType::get(STy->getContext(), makeArrayRef(EI, EE),
3363 const StructLayout *SubSL = TD.getStructLayout(SubTy);
3364 if (Size != SubSL->getSizeInBytes())
3365 return 0; // The sub-struct doesn't have quite the size needed.
3370 /// \brief Rewrite an alloca partition's users.
3372 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3373 /// to rewrite uses of an alloca partition to be conducive for SSA value
3374 /// promotion. If the partition needs a new, more refined alloca, this will
3375 /// build that new alloca, preserving as much type information as possible, and
3376 /// rewrite the uses of the old alloca to point at the new one and have the
3377 /// appropriate new offsets. It also evaluates how successful the rewrite was
3378 /// at enabling promotion and if it was successful queues the alloca to be
3380 bool SROA::rewriteAllocaPartition(AllocaInst &AI,
3381 AllocaPartitioning &P,
3382 AllocaPartitioning::iterator PI) {
3383 uint64_t AllocaSize = PI->EndOffset - PI->BeginOffset;
3384 bool IsLive = false;
3385 for (AllocaPartitioning::use_iterator UI = P.use_begin(PI),
3387 UI != UE && !IsLive; ++UI)
3391 return false; // No live uses left of this partition.
3393 DEBUG(dbgs() << "Speculating PHIs and selects in partition "
3394 << "[" << PI->BeginOffset << "," << PI->EndOffset << ")\n");
3396 PHIOrSelectSpeculator Speculator(*TD, P, *this);
3397 DEBUG(dbgs() << " speculating ");
3398 DEBUG(P.print(dbgs(), PI, ""));
3399 Speculator.visitUsers(PI);
3401 // Try to compute a friendly type for this partition of the alloca. This
3402 // won't always succeed, in which case we fall back to a legal integer type
3403 // or an i8 array of an appropriate size.
3405 if (Type *PartitionTy = P.getCommonType(PI))
3406 if (TD->getTypeAllocSize(PartitionTy) >= AllocaSize)
3407 AllocaTy = PartitionTy;
3409 if (Type *PartitionTy = getTypePartition(*TD, AI.getAllocatedType(),
3410 PI->BeginOffset, AllocaSize))
3411 AllocaTy = PartitionTy;
3413 (AllocaTy->isArrayTy() &&
3414 AllocaTy->getArrayElementType()->isIntegerTy())) &&
3415 TD->isLegalInteger(AllocaSize * 8))
3416 AllocaTy = Type::getIntNTy(*C, AllocaSize * 8);
3418 AllocaTy = ArrayType::get(Type::getInt8Ty(*C), AllocaSize);
3419 assert(TD->getTypeAllocSize(AllocaTy) >= AllocaSize);
3421 // Check for the case where we're going to rewrite to a new alloca of the
3422 // exact same type as the original, and with the same access offsets. In that
3423 // case, re-use the existing alloca, but still run through the rewriter to
3424 // perform phi and select speculation.
3426 if (AllocaTy == AI.getAllocatedType()) {
3427 assert(PI->BeginOffset == 0 &&
3428 "Non-zero begin offset but same alloca type");
3429 assert(PI == P.begin() && "Begin offset is zero on later partition");
3432 unsigned Alignment = AI.getAlignment();
3434 // The minimum alignment which users can rely on when the explicit
3435 // alignment is omitted or zero is that required by the ABI for this
3437 Alignment = TD->getABITypeAlignment(AI.getAllocatedType());
3439 Alignment = MinAlign(Alignment, PI->BeginOffset);
3440 // If we will get at least this much alignment from the type alone, leave
3441 // the alloca's alignment unconstrained.
3442 if (Alignment <= TD->getABITypeAlignment(AllocaTy))
3444 NewAI = new AllocaInst(AllocaTy, 0, Alignment,
3445 AI.getName() + ".sroa." + Twine(PI - P.begin()),
3450 DEBUG(dbgs() << "Rewriting alloca partition "
3451 << "[" << PI->BeginOffset << "," << PI->EndOffset << ") to: "
3454 // Track the high watermark of the post-promotion worklist. We will reset it
3455 // to this point if the alloca is not in fact scheduled for promotion.
3456 unsigned PPWOldSize = PostPromotionWorklist.size();
3458 AllocaPartitionRewriter Rewriter(*TD, P, PI, *this, AI, *NewAI,
3459 PI->BeginOffset, PI->EndOffset);
3460 DEBUG(dbgs() << " rewriting ");
3461 DEBUG(P.print(dbgs(), PI, ""));
3462 bool Promotable = Rewriter.visitUsers(P.use_begin(PI), P.use_end(PI));
3464 DEBUG(dbgs() << " and queuing for promotion\n");
3465 PromotableAllocas.push_back(NewAI);
3466 } else if (NewAI != &AI) {
3467 // If we can't promote the alloca, iterate on it to check for new
3468 // refinements exposed by splitting the current alloca. Don't iterate on an
3469 // alloca which didn't actually change and didn't get promoted.
3470 Worklist.insert(NewAI);
3473 // Drop any post-promotion work items if promotion didn't happen.
3475 while (PostPromotionWorklist.size() > PPWOldSize)
3476 PostPromotionWorklist.pop_back();
3481 /// \brief Walks the partitioning of an alloca rewriting uses of each partition.
3482 bool SROA::splitAlloca(AllocaInst &AI, AllocaPartitioning &P) {
3483 bool Changed = false;
3484 for (AllocaPartitioning::iterator PI = P.begin(), PE = P.end(); PI != PE;
3486 Changed |= rewriteAllocaPartition(AI, P, PI);
3491 /// \brief Analyze an alloca for SROA.
3493 /// This analyzes the alloca to ensure we can reason about it, builds
3494 /// a partitioning of the alloca, and then hands it off to be split and
3495 /// rewritten as needed.
3496 bool SROA::runOnAlloca(AllocaInst &AI) {
3497 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3498 ++NumAllocasAnalyzed;
3500 // Special case dead allocas, as they're trivial.
3501 if (AI.use_empty()) {
3502 AI.eraseFromParent();
3506 // Skip alloca forms that this analysis can't handle.
3507 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3508 TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
3511 bool Changed = false;
3513 // First, split any FCA loads and stores touching this alloca to promote
3514 // better splitting and promotion opportunities.
3515 AggLoadStoreRewriter AggRewriter(*TD);
3516 Changed |= AggRewriter.rewrite(AI);
3518 // Build the partition set using a recursive instruction-visiting builder.
3519 AllocaPartitioning P(*TD, AI);
3520 DEBUG(P.print(dbgs()));
3524 // Delete all the dead users of this alloca before splitting and rewriting it.
3525 for (AllocaPartitioning::dead_user_iterator DI = P.dead_user_begin(),
3526 DE = P.dead_user_end();
3529 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3530 DeadInsts.insert(*DI);
3532 for (AllocaPartitioning::dead_op_iterator DO = P.dead_op_begin(),
3533 DE = P.dead_op_end();
3536 // Clobber the use with an undef value.
3537 **DO = UndefValue::get(OldV->getType());
3538 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3539 if (isInstructionTriviallyDead(OldI)) {
3541 DeadInsts.insert(OldI);
3545 // No partitions to split. Leave the dead alloca for a later pass to clean up.
3546 if (P.begin() == P.end())
3549 return splitAlloca(AI, P) || Changed;
3552 /// \brief Delete the dead instructions accumulated in this run.
3554 /// Recursively deletes the dead instructions we've accumulated. This is done
3555 /// at the very end to maximize locality of the recursive delete and to
3556 /// minimize the problems of invalidated instruction pointers as such pointers
3557 /// are used heavily in the intermediate stages of the algorithm.
3559 /// We also record the alloca instructions deleted here so that they aren't
3560 /// subsequently handed to mem2reg to promote.
3561 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3562 while (!DeadInsts.empty()) {
3563 Instruction *I = DeadInsts.pop_back_val();
3564 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3566 I->replaceAllUsesWith(UndefValue::get(I->getType()));
3568 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
3569 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
3570 // Zero out the operand and see if it becomes trivially dead.
3572 if (isInstructionTriviallyDead(U))
3573 DeadInsts.insert(U);
3576 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3577 DeletedAllocas.insert(AI);
3580 I->eraseFromParent();
3584 /// \brief Promote the allocas, using the best available technique.
3586 /// This attempts to promote whatever allocas have been identified as viable in
3587 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3588 /// If there is a domtree available, we attempt to promote using the full power
3589 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3590 /// based on the SSAUpdater utilities. This function returns whether any
3591 /// promotion occurred.
3592 bool SROA::promoteAllocas(Function &F) {
3593 if (PromotableAllocas.empty())
3596 NumPromoted += PromotableAllocas.size();
3598 if (DT && !ForceSSAUpdater) {
3599 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3600 PromoteMemToReg(PromotableAllocas, *DT);
3601 PromotableAllocas.clear();
3605 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3607 DIBuilder DIB(*F.getParent());
3608 SmallVector<Instruction*, 64> Insts;
3610 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3611 AllocaInst *AI = PromotableAllocas[Idx];
3612 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
3614 Instruction *I = cast<Instruction>(*UI++);
3615 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3616 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3617 // leading to them) here. Eventually it should use them to optimize the
3618 // scalar values produced.
3619 if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
3620 assert(onlyUsedByLifetimeMarkers(I) &&
3621 "Found a bitcast used outside of a lifetime marker.");
3622 while (!I->use_empty())
3623 cast<Instruction>(*I->use_begin())->eraseFromParent();
3624 I->eraseFromParent();
3627 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3628 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3629 II->getIntrinsicID() == Intrinsic::lifetime_end);
3630 II->eraseFromParent();
3636 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3640 PromotableAllocas.clear();
3645 /// \brief A predicate to test whether an alloca belongs to a set.
3646 class IsAllocaInSet {
3647 typedef SmallPtrSet<AllocaInst *, 4> SetType;
3651 typedef AllocaInst *argument_type;
3653 IsAllocaInSet(const SetType &Set) : Set(Set) {}
3654 bool operator()(AllocaInst *AI) const { return Set.count(AI); }
3658 bool SROA::runOnFunction(Function &F) {
3659 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3660 C = &F.getContext();
3661 TD = getAnalysisIfAvailable<DataLayout>();
3663 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3666 DT = getAnalysisIfAvailable<DominatorTree>();
3668 BasicBlock &EntryBB = F.getEntryBlock();
3669 for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end());
3671 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3672 Worklist.insert(AI);
3674 bool Changed = false;
3675 // A set of deleted alloca instruction pointers which should be removed from
3676 // the list of promotable allocas.
3677 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3680 while (!Worklist.empty()) {
3681 Changed |= runOnAlloca(*Worklist.pop_back_val());
3682 deleteDeadInstructions(DeletedAllocas);
3684 // Remove the deleted allocas from various lists so that we don't try to
3685 // continue processing them.
3686 if (!DeletedAllocas.empty()) {
3687 Worklist.remove_if(IsAllocaInSet(DeletedAllocas));
3688 PostPromotionWorklist.remove_if(IsAllocaInSet(DeletedAllocas));
3689 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3690 PromotableAllocas.end(),
3691 IsAllocaInSet(DeletedAllocas)),
3692 PromotableAllocas.end());
3693 DeletedAllocas.clear();
3697 Changed |= promoteAllocas(F);
3699 Worklist = PostPromotionWorklist;
3700 PostPromotionWorklist.clear();
3701 } while (!Worklist.empty());
3706 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3707 if (RequiresDomTree)
3708 AU.addRequired<DominatorTree>();
3709 AU.setPreservesCFG();