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); UI != UE; ++UI) {
1144 continue; // Skip dead uses.
1145 OS << Indent << " [" << UI->BeginOffset << "," << UI->EndOffset << ") "
1146 << "used by: " << *UI->U->getUser() << "\n";
1147 if (MemTransferInst *II = dyn_cast<MemTransferInst>(UI->U->getUser())) {
1148 const MemTransferOffsets &MTO = MemTransferInstData.lookup(II);
1150 if (!MTO.IsSplittable)
1151 IsDest = UI->BeginOffset == MTO.DestBegin;
1153 IsDest = MTO.DestBegin != 0u;
1154 OS << Indent << " (original " << (IsDest ? "dest" : "source") << ": "
1155 << "[" << (IsDest ? MTO.DestBegin : MTO.SourceBegin)
1156 << "," << (IsDest ? MTO.DestEnd : MTO.SourceEnd) << ")\n";
1161 void AllocaPartitioning::print(raw_ostream &OS) const {
1162 if (PointerEscapingInstr) {
1163 OS << "No partitioning for alloca: " << AI << "\n"
1164 << " A pointer to this alloca escaped by:\n"
1165 << " " << *PointerEscapingInstr << "\n";
1169 OS << "Partitioning of alloca: " << AI << "\n";
1171 for (const_iterator I = begin(), E = end(); I != E; ++I, ++Num) {
1177 void AllocaPartitioning::dump(const_iterator I) const { print(dbgs(), I); }
1178 void AllocaPartitioning::dump() const { print(dbgs()); }
1180 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1184 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
1186 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
1187 /// the loads and stores of an alloca instruction, as well as updating its
1188 /// debug information. This is used when a domtree is unavailable and thus
1189 /// mem2reg in its full form can't be used to handle promotion of allocas to
1191 class AllocaPromoter : public LoadAndStorePromoter {
1195 SmallVector<DbgDeclareInst *, 4> DDIs;
1196 SmallVector<DbgValueInst *, 4> DVIs;
1199 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
1200 AllocaInst &AI, DIBuilder &DIB)
1201 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
1203 void run(const SmallVectorImpl<Instruction*> &Insts) {
1204 // Remember which alloca we're promoting (for isInstInList).
1205 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
1206 for (Value::use_iterator UI = DebugNode->use_begin(),
1207 UE = DebugNode->use_end();
1209 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
1210 DDIs.push_back(DDI);
1211 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
1212 DVIs.push_back(DVI);
1215 LoadAndStorePromoter::run(Insts);
1216 AI.eraseFromParent();
1217 while (!DDIs.empty())
1218 DDIs.pop_back_val()->eraseFromParent();
1219 while (!DVIs.empty())
1220 DVIs.pop_back_val()->eraseFromParent();
1223 virtual bool isInstInList(Instruction *I,
1224 const SmallVectorImpl<Instruction*> &Insts) const {
1225 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1226 return LI->getOperand(0) == &AI;
1227 return cast<StoreInst>(I)->getPointerOperand() == &AI;
1230 virtual void updateDebugInfo(Instruction *Inst) const {
1231 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(),
1232 E = DDIs.end(); I != E; ++I) {
1233 DbgDeclareInst *DDI = *I;
1234 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
1235 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1236 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
1237 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1239 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(),
1240 E = DVIs.end(); I != E; ++I) {
1241 DbgValueInst *DVI = *I;
1243 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1244 // If an argument is zero extended then use argument directly. The ZExt
1245 // may be zapped by an optimization pass in future.
1246 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1247 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
1248 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1249 Arg = dyn_cast<Argument>(SExt->getOperand(0));
1251 Arg = SI->getOperand(0);
1252 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
1253 Arg = LI->getOperand(0);
1257 Instruction *DbgVal =
1258 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
1260 DbgVal->setDebugLoc(DVI->getDebugLoc());
1264 } // end anon namespace
1268 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
1270 /// This pass takes allocations which can be completely analyzed (that is, they
1271 /// don't escape) and tries to turn them into scalar SSA values. There are
1272 /// a few steps to this process.
1274 /// 1) It takes allocations of aggregates and analyzes the ways in which they
1275 /// are used to try to split them into smaller allocations, ideally of
1276 /// a single scalar data type. It will split up memcpy and memset accesses
1277 /// as necessary and try to isolate individual scalar accesses.
1278 /// 2) It will transform accesses into forms which are suitable for SSA value
1279 /// promotion. This can be replacing a memset with a scalar store of an
1280 /// integer value, or it can involve speculating operations on a PHI or
1281 /// select to be a PHI or select of the results.
1282 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
1283 /// onto insert and extract operations on a vector value, and convert them to
1284 /// this form. By doing so, it will enable promotion of vector aggregates to
1285 /// SSA vector values.
1286 class SROA : public FunctionPass {
1287 const bool RequiresDomTree;
1290 const DataLayout *TD;
1293 /// \brief Worklist of alloca instructions to simplify.
1295 /// Each alloca in the function is added to this. Each new alloca formed gets
1296 /// added to it as well to recursively simplify unless that alloca can be
1297 /// directly promoted. Finally, each time we rewrite a use of an alloca other
1298 /// the one being actively rewritten, we add it back onto the list if not
1299 /// already present to ensure it is re-visited.
1300 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
1302 /// \brief A collection of instructions to delete.
1303 /// We try to batch deletions to simplify code and make things a bit more
1305 SetVector<Instruction *, SmallVector<Instruction *, 8> > DeadInsts;
1307 /// \brief Post-promotion worklist.
1309 /// Sometimes we discover an alloca which has a high probability of becoming
1310 /// viable for SROA after a round of promotion takes place. In those cases,
1311 /// the alloca is enqueued here for re-processing.
1313 /// Note that we have to be very careful to clear allocas out of this list in
1314 /// the event they are deleted.
1315 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist;
1317 /// \brief A collection of alloca instructions we can directly promote.
1318 std::vector<AllocaInst *> PromotableAllocas;
1321 SROA(bool RequiresDomTree = true)
1322 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
1323 C(0), TD(0), DT(0) {
1324 initializeSROAPass(*PassRegistry::getPassRegistry());
1326 bool runOnFunction(Function &F);
1327 void getAnalysisUsage(AnalysisUsage &AU) const;
1329 const char *getPassName() const { return "SROA"; }
1333 friend class PHIOrSelectSpeculator;
1334 friend class AllocaPartitionRewriter;
1335 friend class AllocaPartitionVectorRewriter;
1337 bool rewriteAllocaPartition(AllocaInst &AI,
1338 AllocaPartitioning &P,
1339 AllocaPartitioning::iterator PI);
1340 bool splitAlloca(AllocaInst &AI, AllocaPartitioning &P);
1341 bool runOnAlloca(AllocaInst &AI);
1342 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
1343 bool promoteAllocas(Function &F);
1349 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
1350 return new SROA(RequiresDomTree);
1353 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
1355 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
1356 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
1360 /// \brief Visitor to speculate PHIs and Selects where possible.
1361 class PHIOrSelectSpeculator : public InstVisitor<PHIOrSelectSpeculator> {
1362 // Befriend the base class so it can delegate to private visit methods.
1363 friend class llvm::InstVisitor<PHIOrSelectSpeculator>;
1365 const DataLayout &TD;
1366 AllocaPartitioning &P;
1370 PHIOrSelectSpeculator(const DataLayout &TD, AllocaPartitioning &P, SROA &Pass)
1371 : TD(TD), P(P), Pass(Pass) {}
1373 /// \brief Visit the users of an alloca partition and rewrite them.
1374 void visitUsers(AllocaPartitioning::const_iterator PI) {
1375 // Note that we need to use an index here as the underlying vector of uses
1376 // may be grown during speculation. However, we never need to re-visit the
1377 // new uses, and so we can use the initial size bound.
1378 for (unsigned Idx = 0, Size = P.use_size(PI); Idx != Size; ++Idx) {
1379 const AllocaPartitioning::PartitionUse &PU = P.getUse(PI, Idx);
1381 continue; // Skip dead use.
1383 visit(cast<Instruction>(PU.U->getUser()));
1388 // By default, skip this instruction.
1389 void visitInstruction(Instruction &I) {}
1391 /// PHI instructions that use an alloca and are subsequently loaded can be
1392 /// rewritten to load both input pointers in the pred blocks and then PHI the
1393 /// results, allowing the load of the alloca to be promoted.
1395 /// %P2 = phi [i32* %Alloca, i32* %Other]
1396 /// %V = load i32* %P2
1398 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1400 /// %V2 = load i32* %Other
1402 /// %V = phi [i32 %V1, i32 %V2]
1404 /// We can do this to a select if its only uses are loads and if the operands
1405 /// to the select can be loaded unconditionally.
1407 /// FIXME: This should be hoisted into a generic utility, likely in
1408 /// Transforms/Util/Local.h
1409 bool isSafePHIToSpeculate(PHINode &PN, SmallVectorImpl<LoadInst *> &Loads) {
1410 // For now, we can only do this promotion if the load is in the same block
1411 // as the PHI, and if there are no stores between the phi and load.
1412 // TODO: Allow recursive phi users.
1413 // TODO: Allow stores.
1414 BasicBlock *BB = PN.getParent();
1415 unsigned MaxAlign = 0;
1416 for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end();
1418 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1419 if (LI == 0 || !LI->isSimple()) return false;
1421 // For now we only allow loads in the same block as the PHI. This is
1422 // a common case that happens when instcombine merges two loads through
1424 if (LI->getParent() != BB) return false;
1426 // Ensure that there are no instructions between the PHI and the load that
1428 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1429 if (BBI->mayWriteToMemory())
1432 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1433 Loads.push_back(LI);
1436 // We can only transform this if it is safe to push the loads into the
1437 // predecessor blocks. The only thing to watch out for is that we can't put
1438 // a possibly trapping load in the predecessor if it is a critical edge.
1439 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1440 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1441 Value *InVal = PN.getIncomingValue(Idx);
1443 // If the value is produced by the terminator of the predecessor (an
1444 // invoke) or it has side-effects, there is no valid place to put a load
1445 // in the predecessor.
1446 if (TI == InVal || TI->mayHaveSideEffects())
1449 // If the predecessor has a single successor, then the edge isn't
1451 if (TI->getNumSuccessors() == 1)
1454 // If this pointer is always safe to load, or if we can prove that there
1455 // is already a load in the block, then we can move the load to the pred
1457 if (InVal->isDereferenceablePointer() ||
1458 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, &TD))
1467 void visitPHINode(PHINode &PN) {
1468 DEBUG(dbgs() << " original: " << PN << "\n");
1470 SmallVector<LoadInst *, 4> Loads;
1471 if (!isSafePHIToSpeculate(PN, Loads))
1474 assert(!Loads.empty());
1476 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1477 IRBuilder<> PHIBuilder(&PN);
1478 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1479 PN.getName() + ".sroa.speculated");
1481 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1482 // matter which one we get and if any differ.
1483 LoadInst *SomeLoad = cast<LoadInst>(Loads.back());
1484 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1485 unsigned Align = SomeLoad->getAlignment();
1487 // Rewrite all loads of the PN to use the new PHI.
1489 LoadInst *LI = Loads.pop_back_val();
1490 LI->replaceAllUsesWith(NewPN);
1491 Pass.DeadInsts.insert(LI);
1492 } while (!Loads.empty());
1494 // Inject loads into all of the pred blocks.
1495 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1496 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1497 TerminatorInst *TI = Pred->getTerminator();
1498 Use *InUse = &PN.getOperandUse(PN.getOperandNumForIncomingValue(Idx));
1499 Value *InVal = PN.getIncomingValue(Idx);
1500 IRBuilder<> PredBuilder(TI);
1503 = PredBuilder.CreateLoad(InVal, (PN.getName() + ".sroa.speculate.load." +
1505 ++NumLoadsSpeculated;
1506 Load->setAlignment(Align);
1508 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1509 NewPN->addIncoming(Load, Pred);
1511 Instruction *Ptr = dyn_cast<Instruction>(InVal);
1513 // No uses to rewrite.
1516 // Try to lookup and rewrite any partition uses corresponding to this phi
1518 AllocaPartitioning::iterator PI
1519 = P.findPartitionForPHIOrSelectOperand(InUse);
1523 // Replace the Use in the PartitionUse for this operand with the Use
1525 AllocaPartitioning::use_iterator UI
1526 = P.findPartitionUseForPHIOrSelectOperand(InUse);
1527 assert(isa<PHINode>(*UI->U->getUser()));
1528 UI->U = &Load->getOperandUse(Load->getPointerOperandIndex());
1530 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1533 /// Select instructions that use an alloca and are subsequently loaded can be
1534 /// rewritten to load both input pointers and then select between the result,
1535 /// allowing the load of the alloca to be promoted.
1537 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1538 /// %V = load i32* %P2
1540 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1541 /// %V2 = load i32* %Other
1542 /// %V = select i1 %cond, i32 %V1, i32 %V2
1544 /// We can do this to a select if its only uses are loads and if the operand
1545 /// to the select can be loaded unconditionally.
1546 bool isSafeSelectToSpeculate(SelectInst &SI,
1547 SmallVectorImpl<LoadInst *> &Loads) {
1548 Value *TValue = SI.getTrueValue();
1549 Value *FValue = SI.getFalseValue();
1550 bool TDerefable = TValue->isDereferenceablePointer();
1551 bool FDerefable = FValue->isDereferenceablePointer();
1553 for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end();
1555 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1556 if (LI == 0 || !LI->isSimple()) return false;
1558 // Both operands to the select need to be dereferencable, either
1559 // absolutely (e.g. allocas) or at this point because we can see other
1561 if (!TDerefable && !isSafeToLoadUnconditionally(TValue, LI,
1562 LI->getAlignment(), &TD))
1564 if (!FDerefable && !isSafeToLoadUnconditionally(FValue, LI,
1565 LI->getAlignment(), &TD))
1567 Loads.push_back(LI);
1573 void visitSelectInst(SelectInst &SI) {
1574 DEBUG(dbgs() << " original: " << SI << "\n");
1575 IRBuilder<> IRB(&SI);
1577 // If the select isn't safe to speculate, just use simple logic to emit it.
1578 SmallVector<LoadInst *, 4> Loads;
1579 if (!isSafeSelectToSpeculate(SI, Loads))
1582 Use *Ops[2] = { &SI.getOperandUse(1), &SI.getOperandUse(2) };
1583 AllocaPartitioning::iterator PIs[2];
1584 AllocaPartitioning::PartitionUse PUs[2];
1585 for (unsigned i = 0, e = 2; i != e; ++i) {
1586 PIs[i] = P.findPartitionForPHIOrSelectOperand(Ops[i]);
1587 if (PIs[i] != P.end()) {
1588 // If the pointer is within the partitioning, remove the select from
1589 // its uses. We'll add in the new loads below.
1590 AllocaPartitioning::use_iterator UI
1591 = P.findPartitionUseForPHIOrSelectOperand(Ops[i]);
1593 // Clear out the use here so that the offsets into the use list remain
1594 // stable but this use is ignored when rewriting.
1599 Value *TV = SI.getTrueValue();
1600 Value *FV = SI.getFalseValue();
1601 // Replace the loads of the select with a select of two loads.
1602 while (!Loads.empty()) {
1603 LoadInst *LI = Loads.pop_back_val();
1605 IRB.SetInsertPoint(LI);
1607 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1609 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1610 NumLoadsSpeculated += 2;
1612 // Transfer alignment and TBAA info if present.
1613 TL->setAlignment(LI->getAlignment());
1614 FL->setAlignment(LI->getAlignment());
1615 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1616 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
1617 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
1620 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1621 LI->getName() + ".sroa.speculated");
1623 LoadInst *Loads[2] = { TL, FL };
1624 for (unsigned i = 0, e = 2; i != e; ++i) {
1625 if (PIs[i] != P.end()) {
1626 Use *LoadUse = &Loads[i]->getOperandUse(0);
1627 assert(PUs[i].U->get() == LoadUse->get());
1629 P.use_push_back(PIs[i], PUs[i]);
1633 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1634 LI->replaceAllUsesWith(V);
1635 Pass.DeadInsts.insert(LI);
1641 /// \brief Build a GEP out of a base pointer and indices.
1643 /// This will return the BasePtr if that is valid, or build a new GEP
1644 /// instruction using the IRBuilder if GEP-ing is needed.
1645 static Value *buildGEP(IRBuilder<> &IRB, Value *BasePtr,
1646 SmallVectorImpl<Value *> &Indices,
1647 const Twine &Prefix) {
1648 if (Indices.empty())
1651 // A single zero index is a no-op, so check for this and avoid building a GEP
1653 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1656 return IRB.CreateInBoundsGEP(BasePtr, Indices, Prefix + ".idx");
1659 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1660 /// TargetTy without changing the offset of the pointer.
1662 /// This routine assumes we've already established a properly offset GEP with
1663 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1664 /// zero-indices down through type layers until we find one the same as
1665 /// TargetTy. If we can't find one with the same type, we at least try to use
1666 /// one with the same size. If none of that works, we just produce the GEP as
1667 /// indicated by Indices to have the correct offset.
1668 static Value *getNaturalGEPWithType(IRBuilder<> &IRB, const DataLayout &TD,
1669 Value *BasePtr, Type *Ty, Type *TargetTy,
1670 SmallVectorImpl<Value *> &Indices,
1671 const Twine &Prefix) {
1673 return buildGEP(IRB, BasePtr, Indices, Prefix);
1675 // See if we can descend into a struct and locate a field with the correct
1677 unsigned NumLayers = 0;
1678 Type *ElementTy = Ty;
1680 if (ElementTy->isPointerTy())
1682 if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
1683 ElementTy = SeqTy->getElementType();
1684 // Note that we use the default address space as this index is over an
1685 // array or a vector, not a pointer.
1686 Indices.push_back(IRB.getInt(APInt(TD.getPointerSizeInBits(0), 0)));
1687 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1688 if (STy->element_begin() == STy->element_end())
1689 break; // Nothing left to descend into.
1690 ElementTy = *STy->element_begin();
1691 Indices.push_back(IRB.getInt32(0));
1696 } while (ElementTy != TargetTy);
1697 if (ElementTy != TargetTy)
1698 Indices.erase(Indices.end() - NumLayers, Indices.end());
1700 return buildGEP(IRB, BasePtr, Indices, Prefix);
1703 /// \brief Recursively compute indices for a natural GEP.
1705 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1706 /// element types adding appropriate indices for the GEP.
1707 static Value *getNaturalGEPRecursively(IRBuilder<> &IRB, const DataLayout &TD,
1708 Value *Ptr, Type *Ty, APInt &Offset,
1710 SmallVectorImpl<Value *> &Indices,
1711 const Twine &Prefix) {
1713 return getNaturalGEPWithType(IRB, TD, Ptr, Ty, TargetTy, Indices, Prefix);
1715 // We can't recurse through pointer types.
1716 if (Ty->isPointerTy())
1719 // We try to analyze GEPs over vectors here, but note that these GEPs are
1720 // extremely poorly defined currently. The long-term goal is to remove GEPing
1721 // over a vector from the IR completely.
1722 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1723 unsigned ElementSizeInBits = TD.getTypeSizeInBits(VecTy->getScalarType());
1724 if (ElementSizeInBits % 8)
1725 return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
1726 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1727 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1728 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1730 Offset -= NumSkippedElements * ElementSize;
1731 Indices.push_back(IRB.getInt(NumSkippedElements));
1732 return getNaturalGEPRecursively(IRB, TD, Ptr, VecTy->getElementType(),
1733 Offset, TargetTy, Indices, Prefix);
1736 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1737 Type *ElementTy = ArrTy->getElementType();
1738 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1739 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1740 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1743 Offset -= NumSkippedElements * ElementSize;
1744 Indices.push_back(IRB.getInt(NumSkippedElements));
1745 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1749 StructType *STy = dyn_cast<StructType>(Ty);
1753 const StructLayout *SL = TD.getStructLayout(STy);
1754 uint64_t StructOffset = Offset.getZExtValue();
1755 if (StructOffset >= SL->getSizeInBytes())
1757 unsigned Index = SL->getElementContainingOffset(StructOffset);
1758 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1759 Type *ElementTy = STy->getElementType(Index);
1760 if (Offset.uge(TD.getTypeAllocSize(ElementTy)))
1761 return 0; // The offset points into alignment padding.
1763 Indices.push_back(IRB.getInt32(Index));
1764 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1768 /// \brief Get a natural GEP from a base pointer to a particular offset and
1769 /// resulting in a particular type.
1771 /// The goal is to produce a "natural" looking GEP that works with the existing
1772 /// composite types to arrive at the appropriate offset and element type for
1773 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1774 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1775 /// Indices, and setting Ty to the result subtype.
1777 /// If no natural GEP can be constructed, this function returns null.
1778 static Value *getNaturalGEPWithOffset(IRBuilder<> &IRB, const DataLayout &TD,
1779 Value *Ptr, APInt Offset, Type *TargetTy,
1780 SmallVectorImpl<Value *> &Indices,
1781 const Twine &Prefix) {
1782 PointerType *Ty = cast<PointerType>(Ptr->getType());
1784 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1786 if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
1789 Type *ElementTy = Ty->getElementType();
1790 if (!ElementTy->isSized())
1791 return 0; // We can't GEP through an unsized element.
1792 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1793 if (ElementSize == 0)
1794 return 0; // Zero-length arrays can't help us build a natural GEP.
1795 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1797 Offset -= NumSkippedElements * ElementSize;
1798 Indices.push_back(IRB.getInt(NumSkippedElements));
1799 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1803 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1804 /// resulting pointer has PointerTy.
1806 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1807 /// and produces the pointer type desired. Where it cannot, it will try to use
1808 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1809 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1810 /// bitcast to the type.
1812 /// The strategy for finding the more natural GEPs is to peel off layers of the
1813 /// pointer, walking back through bit casts and GEPs, searching for a base
1814 /// pointer from which we can compute a natural GEP with the desired
1815 /// properties. The algorithm tries to fold as many constant indices into
1816 /// a single GEP as possible, thus making each GEP more independent of the
1817 /// surrounding code.
1818 static Value *getAdjustedPtr(IRBuilder<> &IRB, const DataLayout &TD,
1819 Value *Ptr, APInt Offset, Type *PointerTy,
1820 const Twine &Prefix) {
1821 // Even though we don't look through PHI nodes, we could be called on an
1822 // instruction in an unreachable block, which may be on a cycle.
1823 SmallPtrSet<Value *, 4> Visited;
1824 Visited.insert(Ptr);
1825 SmallVector<Value *, 4> Indices;
1827 // We may end up computing an offset pointer that has the wrong type. If we
1828 // never are able to compute one directly that has the correct type, we'll
1829 // fall back to it, so keep it around here.
1830 Value *OffsetPtr = 0;
1832 // Remember any i8 pointer we come across to re-use if we need to do a raw
1835 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1837 Type *TargetTy = PointerTy->getPointerElementType();
1840 // First fold any existing GEPs into the offset.
1841 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1842 APInt GEPOffset(Offset.getBitWidth(), 0);
1843 if (!GEP->accumulateConstantOffset(TD, GEPOffset))
1845 Offset += GEPOffset;
1846 Ptr = GEP->getPointerOperand();
1847 if (!Visited.insert(Ptr))
1851 // See if we can perform a natural GEP here.
1853 if (Value *P = getNaturalGEPWithOffset(IRB, TD, Ptr, Offset, TargetTy,
1855 if (P->getType() == PointerTy) {
1856 // Zap any offset pointer that we ended up computing in previous rounds.
1857 if (OffsetPtr && OffsetPtr->use_empty())
1858 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1859 I->eraseFromParent();
1867 // Stash this pointer if we've found an i8*.
1868 if (Ptr->getType()->isIntegerTy(8)) {
1870 Int8PtrOffset = Offset;
1873 // Peel off a layer of the pointer and update the offset appropriately.
1874 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1875 Ptr = cast<Operator>(Ptr)->getOperand(0);
1876 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1877 if (GA->mayBeOverridden())
1879 Ptr = GA->getAliasee();
1883 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1884 } while (Visited.insert(Ptr));
1888 Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
1889 Prefix + ".raw_cast");
1890 Int8PtrOffset = Offset;
1893 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1894 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1895 Prefix + ".raw_idx");
1899 // On the off chance we were targeting i8*, guard the bitcast here.
1900 if (Ptr->getType() != PointerTy)
1901 Ptr = IRB.CreateBitCast(Ptr, PointerTy, Prefix + ".cast");
1906 /// \brief Test whether we can convert a value from the old to the new type.
1908 /// This predicate should be used to guard calls to convertValue in order to
1909 /// ensure that we only try to convert viable values. The strategy is that we
1910 /// will peel off single element struct and array wrappings to get to an
1911 /// underlying value, and convert that value.
1912 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1915 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1917 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1920 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1921 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1923 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1931 /// \brief Generic routine to convert an SSA value to a value of a different
1934 /// This will try various different casting techniques, such as bitcasts,
1935 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1936 /// two types for viability with this routine.
1937 static Value *convertValue(const DataLayout &DL, IRBuilder<> &IRB, Value *V,
1939 assert(canConvertValue(DL, V->getType(), Ty) &&
1940 "Value not convertable to type");
1941 if (V->getType() == Ty)
1943 if (V->getType()->isIntegerTy() && Ty->isPointerTy())
1944 return IRB.CreateIntToPtr(V, Ty);
1945 if (V->getType()->isPointerTy() && Ty->isIntegerTy())
1946 return IRB.CreatePtrToInt(V, Ty);
1948 return IRB.CreateBitCast(V, Ty);
1951 /// \brief Test whether the given alloca partition can be promoted to a vector.
1953 /// This is a quick test to check whether we can rewrite a particular alloca
1954 /// partition (and its newly formed alloca) into a vector alloca with only
1955 /// whole-vector loads and stores such that it could be promoted to a vector
1956 /// SSA value. We only can ensure this for a limited set of operations, and we
1957 /// don't want to do the rewrites unless we are confident that the result will
1958 /// be promotable, so we have an early test here.
1959 static bool isVectorPromotionViable(const DataLayout &TD,
1961 AllocaPartitioning &P,
1962 uint64_t PartitionBeginOffset,
1963 uint64_t PartitionEndOffset,
1964 AllocaPartitioning::const_use_iterator I,
1965 AllocaPartitioning::const_use_iterator E) {
1966 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
1970 uint64_t ElementSize = TD.getTypeSizeInBits(Ty->getScalarType());
1972 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1973 // that aren't byte sized.
1974 if (ElementSize % 8)
1976 assert((TD.getTypeSizeInBits(Ty) % 8) == 0 &&
1977 "vector size not a multiple of element size?");
1980 for (; I != E; ++I) {
1982 continue; // Skip dead use.
1984 uint64_t BeginOffset = I->BeginOffset - PartitionBeginOffset;
1985 uint64_t BeginIndex = BeginOffset / ElementSize;
1986 if (BeginIndex * ElementSize != BeginOffset ||
1987 BeginIndex >= Ty->getNumElements())
1989 uint64_t EndOffset = I->EndOffset - PartitionBeginOffset;
1990 uint64_t EndIndex = EndOffset / ElementSize;
1991 if (EndIndex * ElementSize != EndOffset ||
1992 EndIndex > Ty->getNumElements())
1995 assert(EndIndex > BeginIndex && "Empty vector!");
1996 uint64_t NumElements = EndIndex - BeginIndex;
1998 = (NumElements == 1) ? Ty->getElementType()
1999 : VectorType::get(Ty->getElementType(), NumElements);
2001 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
2002 if (MI->isVolatile())
2004 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
2005 const AllocaPartitioning::MemTransferOffsets &MTO
2006 = P.getMemTransferOffsets(*MTI);
2007 if (!MTO.IsSplittable)
2010 } else if (I->U->get()->getType()->getPointerElementType()->isStructTy()) {
2011 // Disable vector promotion when there are loads or stores of an FCA.
2013 } else if (LoadInst *LI = dyn_cast<LoadInst>(I->U->getUser())) {
2014 if (LI->isVolatile())
2016 if (!canConvertValue(TD, PartitionTy, LI->getType()))
2018 } else if (StoreInst *SI = dyn_cast<StoreInst>(I->U->getUser())) {
2019 if (SI->isVolatile())
2021 if (!canConvertValue(TD, SI->getValueOperand()->getType(), PartitionTy))
2030 /// \brief Test whether the given alloca partition's integer operations can be
2031 /// widened to promotable ones.
2033 /// This is a quick test to check whether we can rewrite the integer loads and
2034 /// stores to a particular alloca into wider loads and stores and be able to
2035 /// promote the resulting alloca.
2036 static bool isIntegerWideningViable(const DataLayout &TD,
2038 uint64_t AllocBeginOffset,
2039 AllocaPartitioning &P,
2040 AllocaPartitioning::const_use_iterator I,
2041 AllocaPartitioning::const_use_iterator E) {
2042 uint64_t SizeInBits = TD.getTypeSizeInBits(AllocaTy);
2043 // Don't create integer types larger than the maximum bitwidth.
2044 if (SizeInBits > IntegerType::MAX_INT_BITS)
2047 // Don't try to handle allocas with bit-padding.
2048 if (SizeInBits != TD.getTypeStoreSizeInBits(AllocaTy))
2051 // We need to ensure that an integer type with the appropriate bitwidth can
2052 // be converted to the alloca type, whatever that is. We don't want to force
2053 // the alloca itself to have an integer type if there is a more suitable one.
2054 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
2055 if (!canConvertValue(TD, AllocaTy, IntTy) ||
2056 !canConvertValue(TD, IntTy, AllocaTy))
2059 uint64_t Size = TD.getTypeStoreSize(AllocaTy);
2061 // Check the uses to ensure the uses are (likely) promotable integer uses.
2062 // Also ensure that the alloca has a covering load or store. We don't want
2063 // to widen the integer operations only to fail to promote due to some other
2064 // unsplittable entry (which we may make splittable later).
2065 bool WholeAllocaOp = false;
2066 for (; I != E; ++I) {
2068 continue; // Skip dead use.
2070 uint64_t RelBegin = I->BeginOffset - AllocBeginOffset;
2071 uint64_t RelEnd = I->EndOffset - AllocBeginOffset;
2073 // We can't reasonably handle cases where the load or store extends past
2074 // the end of the aloca's type and into its padding.
2078 if (LoadInst *LI = dyn_cast<LoadInst>(I->U->getUser())) {
2079 if (LI->isVolatile())
2081 if (RelBegin == 0 && RelEnd == Size)
2082 WholeAllocaOp = true;
2083 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
2084 if (ITy->getBitWidth() < TD.getTypeStoreSizeInBits(ITy))
2088 // Non-integer loads need to be convertible from the alloca type so that
2089 // they are promotable.
2090 if (RelBegin != 0 || RelEnd != Size ||
2091 !canConvertValue(TD, AllocaTy, LI->getType()))
2093 } else if (StoreInst *SI = dyn_cast<StoreInst>(I->U->getUser())) {
2094 Type *ValueTy = SI->getValueOperand()->getType();
2095 if (SI->isVolatile())
2097 if (RelBegin == 0 && RelEnd == Size)
2098 WholeAllocaOp = true;
2099 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
2100 if (ITy->getBitWidth() < TD.getTypeStoreSizeInBits(ITy))
2104 // Non-integer stores need to be convertible to the alloca type so that
2105 // they are promotable.
2106 if (RelBegin != 0 || RelEnd != Size ||
2107 !canConvertValue(TD, ValueTy, AllocaTy))
2109 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
2110 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
2112 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
2113 const AllocaPartitioning::MemTransferOffsets &MTO
2114 = P.getMemTransferOffsets(*MTI);
2115 if (!MTO.IsSplittable)
2118 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->U->getUser())) {
2119 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2120 II->getIntrinsicID() != Intrinsic::lifetime_end)
2126 return WholeAllocaOp;
2129 static Value *extractInteger(const DataLayout &DL, IRBuilder<> &IRB, Value *V,
2130 IntegerType *Ty, uint64_t Offset,
2131 const Twine &Name) {
2132 DEBUG(dbgs() << " start: " << *V << "\n");
2133 IntegerType *IntTy = cast<IntegerType>(V->getType());
2134 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2135 "Element extends past full value");
2136 uint64_t ShAmt = 8*Offset;
2137 if (DL.isBigEndian())
2138 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2140 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2141 DEBUG(dbgs() << " shifted: " << *V << "\n");
2143 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2144 "Cannot extract to a larger integer!");
2146 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2147 DEBUG(dbgs() << " trunced: " << *V << "\n");
2152 static Value *insertInteger(const DataLayout &DL, IRBuilder<> &IRB, Value *Old,
2153 Value *V, uint64_t Offset, const Twine &Name) {
2154 IntegerType *IntTy = cast<IntegerType>(Old->getType());
2155 IntegerType *Ty = cast<IntegerType>(V->getType());
2156 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2157 "Cannot insert a larger integer!");
2158 DEBUG(dbgs() << " start: " << *V << "\n");
2160 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2161 DEBUG(dbgs() << " extended: " << *V << "\n");
2163 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2164 "Element store outside of alloca store");
2165 uint64_t ShAmt = 8*Offset;
2166 if (DL.isBigEndian())
2167 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2169 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2170 DEBUG(dbgs() << " shifted: " << *V << "\n");
2173 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2174 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2175 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2176 DEBUG(dbgs() << " masked: " << *Old << "\n");
2177 V = IRB.CreateOr(Old, V, Name + ".insert");
2178 DEBUG(dbgs() << " inserted: " << *V << "\n");
2183 static Value *extractVector(IRBuilder<> &IRB, Value *V,
2184 unsigned BeginIndex, unsigned EndIndex,
2185 const Twine &Name) {
2186 VectorType *VecTy = cast<VectorType>(V->getType());
2187 unsigned NumElements = EndIndex - BeginIndex;
2188 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2190 if (NumElements == VecTy->getNumElements())
2193 if (NumElements == 1) {
2194 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2196 DEBUG(dbgs() << " extract: " << *V << "\n");
2200 SmallVector<Constant*, 8> Mask;
2201 Mask.reserve(NumElements);
2202 for (unsigned i = BeginIndex; i != EndIndex; ++i)
2203 Mask.push_back(IRB.getInt32(i));
2204 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2205 ConstantVector::get(Mask),
2207 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2211 static Value *insertVector(IRBuilder<> &IRB, Value *Old, Value *V,
2212 unsigned BeginIndex, const Twine &Name) {
2213 VectorType *VecTy = cast<VectorType>(Old->getType());
2214 assert(VecTy && "Can only insert a vector into a vector");
2216 VectorType *Ty = dyn_cast<VectorType>(V->getType());
2218 // Single element to insert.
2219 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2221 DEBUG(dbgs() << " insert: " << *V << "\n");
2225 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
2226 "Too many elements!");
2227 if (Ty->getNumElements() == VecTy->getNumElements()) {
2228 assert(V->getType() == VecTy && "Vector type mismatch");
2231 unsigned EndIndex = BeginIndex + Ty->getNumElements();
2233 // When inserting a smaller vector into the larger to store, we first
2234 // use a shuffle vector to widen it with undef elements, and then
2235 // a second shuffle vector to select between the loaded vector and the
2237 SmallVector<Constant*, 8> Mask;
2238 Mask.reserve(VecTy->getNumElements());
2239 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2240 if (i >= BeginIndex && i < EndIndex)
2241 Mask.push_back(IRB.getInt32(i - BeginIndex));
2243 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2244 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2245 ConstantVector::get(Mask),
2247 DEBUG(dbgs() << " shuffle1: " << *V << "\n");
2250 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2251 if (i >= BeginIndex && i < EndIndex)
2252 Mask.push_back(IRB.getInt32(i));
2254 Mask.push_back(IRB.getInt32(i + VecTy->getNumElements()));
2255 V = IRB.CreateShuffleVector(V, Old, ConstantVector::get(Mask),
2257 DEBUG(dbgs() << " shuffle2: " << *V << "\n");
2262 /// \brief Visitor to rewrite instructions using a partition of an alloca to
2263 /// use a new alloca.
2265 /// Also implements the rewriting to vector-based accesses when the partition
2266 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2268 class AllocaPartitionRewriter : public InstVisitor<AllocaPartitionRewriter,
2270 // Befriend the base class so it can delegate to private visit methods.
2271 friend class llvm::InstVisitor<AllocaPartitionRewriter, bool>;
2273 const DataLayout &TD;
2274 AllocaPartitioning &P;
2276 AllocaInst &OldAI, &NewAI;
2277 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2280 // If we are rewriting an alloca partition which can be written as pure
2281 // vector operations, we stash extra information here. When VecTy is
2282 // non-null, we have some strict guarantees about the rewritten alloca:
2283 // - The new alloca is exactly the size of the vector type here.
2284 // - The accesses all either map to the entire vector or to a single
2286 // - The set of accessing instructions is only one of those handled above
2287 // in isVectorPromotionViable. Generally these are the same access kinds
2288 // which are promotable via mem2reg.
2291 uint64_t ElementSize;
2293 // This is a convenience and flag variable that will be null unless the new
2294 // alloca's integer operations should be widened to this integer type due to
2295 // passing isIntegerWideningViable above. If it is non-null, the desired
2296 // integer type will be stored here for easy access during rewriting.
2299 // The offset of the partition user currently being rewritten.
2300 uint64_t BeginOffset, EndOffset;
2302 Instruction *OldPtr;
2304 // The name prefix to use when rewriting instructions for this alloca.
2305 std::string NamePrefix;
2308 AllocaPartitionRewriter(const DataLayout &TD, AllocaPartitioning &P,
2309 AllocaPartitioning::iterator PI,
2310 SROA &Pass, AllocaInst &OldAI, AllocaInst &NewAI,
2311 uint64_t NewBeginOffset, uint64_t NewEndOffset)
2312 : TD(TD), P(P), Pass(Pass),
2313 OldAI(OldAI), NewAI(NewAI),
2314 NewAllocaBeginOffset(NewBeginOffset),
2315 NewAllocaEndOffset(NewEndOffset),
2316 NewAllocaTy(NewAI.getAllocatedType()),
2317 VecTy(), ElementTy(), ElementSize(), IntTy(),
2318 BeginOffset(), EndOffset() {
2321 /// \brief Visit the users of the alloca partition and rewrite them.
2322 bool visitUsers(AllocaPartitioning::const_use_iterator I,
2323 AllocaPartitioning::const_use_iterator E) {
2324 if (isVectorPromotionViable(TD, NewAI.getAllocatedType(), P,
2325 NewAllocaBeginOffset, NewAllocaEndOffset,
2328 VecTy = cast<VectorType>(NewAI.getAllocatedType());
2329 ElementTy = VecTy->getElementType();
2330 assert((TD.getTypeSizeInBits(VecTy->getScalarType()) % 8) == 0 &&
2331 "Only multiple-of-8 sized vector elements are viable");
2332 ElementSize = TD.getTypeSizeInBits(VecTy->getScalarType()) / 8;
2333 } else if (isIntegerWideningViable(TD, NewAI.getAllocatedType(),
2334 NewAllocaBeginOffset, P, I, E)) {
2335 IntTy = Type::getIntNTy(NewAI.getContext(),
2336 TD.getTypeSizeInBits(NewAI.getAllocatedType()));
2338 bool CanSROA = true;
2339 for (; I != E; ++I) {
2341 continue; // Skip dead uses.
2342 BeginOffset = I->BeginOffset;
2343 EndOffset = I->EndOffset;
2345 OldPtr = cast<Instruction>(I->U->get());
2346 NamePrefix = (Twine(NewAI.getName()) + "." + Twine(BeginOffset)).str();
2347 CanSROA &= visit(cast<Instruction>(I->U->getUser()));
2363 // Every instruction which can end up as a user must have a rewrite rule.
2364 bool visitInstruction(Instruction &I) {
2365 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2366 llvm_unreachable("No rewrite rule for this instruction!");
2369 Twine getName(const Twine &Suffix) {
2370 return NamePrefix + Suffix;
2373 Value *getAdjustedAllocaPtr(IRBuilder<> &IRB, Type *PointerTy) {
2374 assert(BeginOffset >= NewAllocaBeginOffset);
2375 APInt Offset(TD.getPointerSizeInBits(), BeginOffset - NewAllocaBeginOffset);
2376 return getAdjustedPtr(IRB, TD, &NewAI, Offset, PointerTy, getName(""));
2379 /// \brief Compute suitable alignment to access an offset into the new alloca.
2380 unsigned getOffsetAlign(uint64_t Offset) {
2381 unsigned NewAIAlign = NewAI.getAlignment();
2383 NewAIAlign = TD.getABITypeAlignment(NewAI.getAllocatedType());
2384 return MinAlign(NewAIAlign, Offset);
2387 /// \brief Compute suitable alignment to access this partition of the new
2389 unsigned getPartitionAlign() {
2390 return getOffsetAlign(BeginOffset - NewAllocaBeginOffset);
2393 /// \brief Compute suitable alignment to access a type at an offset of the
2396 /// \returns zero if the type's ABI alignment is a suitable alignment,
2397 /// otherwise returns the maximal suitable alignment.
2398 unsigned getOffsetTypeAlign(Type *Ty, uint64_t Offset) {
2399 unsigned Align = getOffsetAlign(Offset);
2400 return Align == TD.getABITypeAlignment(Ty) ? 0 : Align;
2403 /// \brief Compute suitable alignment to access a type at the beginning of
2404 /// this partition of the new alloca.
2406 /// See \c getOffsetTypeAlign for details; this routine delegates to it.
2407 unsigned getPartitionTypeAlign(Type *Ty) {
2408 return getOffsetTypeAlign(Ty, BeginOffset - NewAllocaBeginOffset);
2411 unsigned getIndex(uint64_t Offset) {
2412 assert(VecTy && "Can only call getIndex when rewriting a vector");
2413 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2414 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2415 uint32_t Index = RelOffset / ElementSize;
2416 assert(Index * ElementSize == RelOffset);
2420 void deleteIfTriviallyDead(Value *V) {
2421 Instruction *I = cast<Instruction>(V);
2422 if (isInstructionTriviallyDead(I))
2423 Pass.DeadInsts.insert(I);
2426 Value *rewriteVectorizedLoadInst(IRBuilder<> &IRB) {
2427 unsigned BeginIndex = getIndex(BeginOffset);
2428 unsigned EndIndex = getIndex(EndOffset);
2429 assert(EndIndex > BeginIndex && "Empty vector!");
2431 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2433 return extractVector(IRB, V, BeginIndex, EndIndex, getName(".vec"));
2436 Value *rewriteIntegerLoad(IRBuilder<> &IRB, LoadInst &LI) {
2437 assert(IntTy && "We cannot insert an integer to the alloca");
2438 assert(!LI.isVolatile());
2439 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2441 V = convertValue(TD, IRB, V, IntTy);
2442 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2443 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2444 if (Offset > 0 || EndOffset < NewAllocaEndOffset)
2445 V = extractInteger(TD, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2446 getName(".extract"));
2450 bool visitLoadInst(LoadInst &LI) {
2451 DEBUG(dbgs() << " original: " << LI << "\n");
2452 Value *OldOp = LI.getOperand(0);
2453 assert(OldOp == OldPtr);
2454 IRBuilder<> IRB(&LI);
2456 uint64_t Size = EndOffset - BeginOffset;
2457 bool IsSplitIntLoad = Size < TD.getTypeStoreSize(LI.getType());
2459 // If this memory access can be shown to *statically* extend outside the
2460 // bounds of the original allocation it's behavior is undefined. Rather
2461 // than trying to transform it, just replace it with undef.
2462 // FIXME: We should do something more clever for functions being
2463 // instrumented by asan.
2464 // FIXME: Eventually, once ASan and friends can flush out bugs here, this
2465 // should be transformed to a load of null making it unreachable.
2466 uint64_t OldAllocSize = TD.getTypeAllocSize(OldAI.getAllocatedType());
2467 if (TD.getTypeStoreSize(LI.getType()) > OldAllocSize) {
2468 LI.replaceAllUsesWith(UndefValue::get(LI.getType()));
2469 Pass.DeadInsts.insert(&LI);
2470 deleteIfTriviallyDead(OldOp);
2471 DEBUG(dbgs() << " to: undef!!\n");
2475 Type *TargetTy = IsSplitIntLoad ? Type::getIntNTy(LI.getContext(), Size * 8)
2477 bool IsPtrAdjusted = false;
2480 V = rewriteVectorizedLoadInst(IRB);
2481 } else if (IntTy && LI.getType()->isIntegerTy()) {
2482 V = rewriteIntegerLoad(IRB, LI);
2483 } else if (BeginOffset == NewAllocaBeginOffset &&
2484 canConvertValue(TD, NewAllocaTy, LI.getType())) {
2485 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2486 LI.isVolatile(), getName(".load"));
2488 Type *LTy = TargetTy->getPointerTo();
2489 V = IRB.CreateAlignedLoad(getAdjustedAllocaPtr(IRB, LTy),
2490 getPartitionTypeAlign(TargetTy),
2491 LI.isVolatile(), getName(".load"));
2492 IsPtrAdjusted = true;
2494 V = convertValue(TD, IRB, V, TargetTy);
2496 if (IsSplitIntLoad) {
2497 assert(!LI.isVolatile());
2498 assert(LI.getType()->isIntegerTy() &&
2499 "Only integer type loads and stores are split");
2500 assert(LI.getType()->getIntegerBitWidth() ==
2501 TD.getTypeStoreSizeInBits(LI.getType()) &&
2502 "Non-byte-multiple bit width");
2503 assert(LI.getType()->getIntegerBitWidth() ==
2504 TD.getTypeAllocSizeInBits(OldAI.getAllocatedType()) &&
2505 "Only alloca-wide loads can be split and recomposed");
2506 // Move the insertion point just past the load so that we can refer to it.
2507 IRB.SetInsertPoint(llvm::next(BasicBlock::iterator(&LI)));
2508 // Create a placeholder value with the same type as LI to use as the
2509 // basis for the new value. This allows us to replace the uses of LI with
2510 // the computed value, and then replace the placeholder with LI, leaving
2511 // LI only used for this computation.
2513 = new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2514 V = insertInteger(TD, IRB, Placeholder, V, BeginOffset,
2515 getName(".insert"));
2516 LI.replaceAllUsesWith(V);
2517 Placeholder->replaceAllUsesWith(&LI);
2520 LI.replaceAllUsesWith(V);
2523 Pass.DeadInsts.insert(&LI);
2524 deleteIfTriviallyDead(OldOp);
2525 DEBUG(dbgs() << " to: " << *V << "\n");
2526 return !LI.isVolatile() && !IsPtrAdjusted;
2529 bool rewriteVectorizedStoreInst(IRBuilder<> &IRB, Value *V,
2530 StoreInst &SI, Value *OldOp) {
2531 unsigned BeginIndex = getIndex(BeginOffset);
2532 unsigned EndIndex = getIndex(EndOffset);
2533 assert(EndIndex > BeginIndex && "Empty vector!");
2534 unsigned NumElements = EndIndex - BeginIndex;
2535 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2537 = (NumElements == 1) ? ElementTy
2538 : VectorType::get(ElementTy, NumElements);
2539 if (V->getType() != PartitionTy)
2540 V = convertValue(TD, IRB, V, PartitionTy);
2542 // Mix in the existing elements.
2543 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2545 V = insertVector(IRB, Old, V, BeginIndex, getName(".vec"));
2547 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2548 Pass.DeadInsts.insert(&SI);
2551 DEBUG(dbgs() << " to: " << *Store << "\n");
2555 bool rewriteIntegerStore(IRBuilder<> &IRB, Value *V, StoreInst &SI) {
2556 assert(IntTy && "We cannot extract an integer from the alloca");
2557 assert(!SI.isVolatile());
2558 if (TD.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2559 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2560 getName(".oldload"));
2561 Old = convertValue(TD, IRB, Old, IntTy);
2562 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2563 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2564 V = insertInteger(TD, IRB, Old, SI.getValueOperand(), Offset,
2565 getName(".insert"));
2567 V = convertValue(TD, IRB, V, NewAllocaTy);
2568 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2569 Pass.DeadInsts.insert(&SI);
2571 DEBUG(dbgs() << " to: " << *Store << "\n");
2575 bool visitStoreInst(StoreInst &SI) {
2576 DEBUG(dbgs() << " original: " << SI << "\n");
2577 Value *OldOp = SI.getOperand(1);
2578 assert(OldOp == OldPtr);
2579 IRBuilder<> IRB(&SI);
2581 Value *V = SI.getValueOperand();
2583 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2584 // alloca that should be re-examined after promoting this alloca.
2585 if (V->getType()->isPointerTy())
2586 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2587 Pass.PostPromotionWorklist.insert(AI);
2589 uint64_t Size = EndOffset - BeginOffset;
2590 if (Size < TD.getTypeStoreSize(V->getType())) {
2591 assert(!SI.isVolatile());
2592 assert(V->getType()->isIntegerTy() &&
2593 "Only integer type loads and stores are split");
2594 assert(V->getType()->getIntegerBitWidth() ==
2595 TD.getTypeStoreSizeInBits(V->getType()) &&
2596 "Non-byte-multiple bit width");
2597 assert(V->getType()->getIntegerBitWidth() ==
2598 TD.getTypeAllocSizeInBits(OldAI.getAllocatedType()) &&
2599 "Only alloca-wide stores can be split and recomposed");
2600 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), Size * 8);
2601 V = extractInteger(TD, IRB, V, NarrowTy, BeginOffset,
2602 getName(".extract"));
2606 return rewriteVectorizedStoreInst(IRB, V, SI, OldOp);
2607 if (IntTy && V->getType()->isIntegerTy())
2608 return rewriteIntegerStore(IRB, V, SI);
2611 if (BeginOffset == NewAllocaBeginOffset &&
2612 canConvertValue(TD, V->getType(), NewAllocaTy)) {
2613 V = convertValue(TD, IRB, V, NewAllocaTy);
2614 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2617 Value *NewPtr = getAdjustedAllocaPtr(IRB, V->getType()->getPointerTo());
2618 NewSI = IRB.CreateAlignedStore(V, NewPtr,
2619 getPartitionTypeAlign(V->getType()),
2623 Pass.DeadInsts.insert(&SI);
2624 deleteIfTriviallyDead(OldOp);
2626 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2627 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2630 /// \brief Compute an integer value from splatting an i8 across the given
2631 /// number of bytes.
2633 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2634 /// call this routine.
2635 /// FIXME: Heed the advice above.
2637 /// \param V The i8 value to splat.
2638 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2639 Value *getIntegerSplat(IRBuilder<> &IRB, Value *V, unsigned Size) {
2640 assert(Size > 0 && "Expected a positive number of bytes.");
2641 IntegerType *VTy = cast<IntegerType>(V->getType());
2642 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2646 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size*8);
2647 V = IRB.CreateMul(IRB.CreateZExt(V, SplatIntTy, getName(".zext")),
2648 ConstantExpr::getUDiv(
2649 Constant::getAllOnesValue(SplatIntTy),
2650 ConstantExpr::getZExt(
2651 Constant::getAllOnesValue(V->getType()),
2653 getName(".isplat"));
2657 /// \brief Compute a vector splat for a given element value.
2658 Value *getVectorSplat(IRBuilder<> &IRB, Value *V, unsigned NumElements) {
2659 V = IRB.CreateVectorSplat(NumElements, V, NamePrefix);
2660 DEBUG(dbgs() << " splat: " << *V << "\n");
2664 bool visitMemSetInst(MemSetInst &II) {
2665 DEBUG(dbgs() << " original: " << II << "\n");
2666 IRBuilder<> IRB(&II);
2667 assert(II.getRawDest() == OldPtr);
2669 // If the memset has a variable size, it cannot be split, just adjust the
2670 // pointer to the new alloca.
2671 if (!isa<Constant>(II.getLength())) {
2672 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2673 Type *CstTy = II.getAlignmentCst()->getType();
2674 II.setAlignment(ConstantInt::get(CstTy, getPartitionAlign()));
2676 deleteIfTriviallyDead(OldPtr);
2680 // Record this instruction for deletion.
2681 Pass.DeadInsts.insert(&II);
2683 Type *AllocaTy = NewAI.getAllocatedType();
2684 Type *ScalarTy = AllocaTy->getScalarType();
2686 // If this doesn't map cleanly onto the alloca type, and that type isn't
2687 // a single value type, just emit a memset.
2688 if (!VecTy && !IntTy &&
2689 (BeginOffset != NewAllocaBeginOffset ||
2690 EndOffset != NewAllocaEndOffset ||
2691 !AllocaTy->isSingleValueType() ||
2692 !TD.isLegalInteger(TD.getTypeSizeInBits(ScalarTy)) ||
2693 TD.getTypeSizeInBits(ScalarTy)%8 != 0)) {
2694 Type *SizeTy = II.getLength()->getType();
2695 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2697 = IRB.CreateMemSet(getAdjustedAllocaPtr(IRB,
2698 II.getRawDest()->getType()),
2699 II.getValue(), Size, getPartitionAlign(),
2702 DEBUG(dbgs() << " to: " << *New << "\n");
2706 // If we can represent this as a simple value, we have to build the actual
2707 // value to store, which requires expanding the byte present in memset to
2708 // a sensible representation for the alloca type. This is essentially
2709 // splatting the byte to a sufficiently wide integer, splatting it across
2710 // any desired vector width, and bitcasting to the final type.
2714 // If this is a memset of a vectorized alloca, insert it.
2715 assert(ElementTy == ScalarTy);
2717 unsigned BeginIndex = getIndex(BeginOffset);
2718 unsigned EndIndex = getIndex(EndOffset);
2719 assert(EndIndex > BeginIndex && "Empty vector!");
2720 unsigned NumElements = EndIndex - BeginIndex;
2721 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2723 Value *Splat = getIntegerSplat(IRB, II.getValue(),
2724 TD.getTypeSizeInBits(ElementTy)/8);
2725 Splat = convertValue(TD, IRB, Splat, ElementTy);
2726 if (NumElements > 1)
2727 Splat = getVectorSplat(IRB, Splat, NumElements);
2729 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2730 getName(".oldload"));
2731 V = insertVector(IRB, Old, Splat, BeginIndex, getName(".vec"));
2733 // If this is a memset on an alloca where we can widen stores, insert the
2735 assert(!II.isVolatile());
2737 uint64_t Size = EndOffset - BeginOffset;
2738 V = getIntegerSplat(IRB, II.getValue(), Size);
2740 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2741 EndOffset != NewAllocaBeginOffset)) {
2742 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2743 getName(".oldload"));
2744 Old = convertValue(TD, IRB, Old, IntTy);
2745 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2746 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2747 V = insertInteger(TD, IRB, Old, V, Offset, getName(".insert"));
2749 assert(V->getType() == IntTy &&
2750 "Wrong type for an alloca wide integer!");
2752 V = convertValue(TD, IRB, V, AllocaTy);
2754 // Established these invariants above.
2755 assert(BeginOffset == NewAllocaBeginOffset);
2756 assert(EndOffset == NewAllocaEndOffset);
2758 V = getIntegerSplat(IRB, II.getValue(),
2759 TD.getTypeSizeInBits(ScalarTy)/8);
2760 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2761 V = getVectorSplat(IRB, V, AllocaVecTy->getNumElements());
2763 V = convertValue(TD, IRB, V, AllocaTy);
2766 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2769 DEBUG(dbgs() << " to: " << *New << "\n");
2770 return !II.isVolatile();
2773 bool visitMemTransferInst(MemTransferInst &II) {
2774 // Rewriting of memory transfer instructions can be a bit tricky. We break
2775 // them into two categories: split intrinsics and unsplit intrinsics.
2777 DEBUG(dbgs() << " original: " << II << "\n");
2778 IRBuilder<> IRB(&II);
2780 assert(II.getRawSource() == OldPtr || II.getRawDest() == OldPtr);
2781 bool IsDest = II.getRawDest() == OldPtr;
2783 const AllocaPartitioning::MemTransferOffsets &MTO
2784 = P.getMemTransferOffsets(II);
2786 // Compute the relative offset within the transfer.
2787 unsigned IntPtrWidth = TD.getPointerSizeInBits();
2788 APInt RelOffset(IntPtrWidth, BeginOffset - (IsDest ? MTO.DestBegin
2789 : MTO.SourceBegin));
2791 unsigned Align = II.getAlignment();
2793 Align = MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(),
2794 MinAlign(II.getAlignment(), getPartitionAlign()));
2796 // For unsplit intrinsics, we simply modify the source and destination
2797 // pointers in place. This isn't just an optimization, it is a matter of
2798 // correctness. With unsplit intrinsics we may be dealing with transfers
2799 // within a single alloca before SROA ran, or with transfers that have
2800 // a variable length. We may also be dealing with memmove instead of
2801 // memcpy, and so simply updating the pointers is the necessary for us to
2802 // update both source and dest of a single call.
2803 if (!MTO.IsSplittable) {
2804 Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource();
2806 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2808 II.setSource(getAdjustedAllocaPtr(IRB, II.getRawSource()->getType()));
2810 Type *CstTy = II.getAlignmentCst()->getType();
2811 II.setAlignment(ConstantInt::get(CstTy, Align));
2813 DEBUG(dbgs() << " to: " << II << "\n");
2814 deleteIfTriviallyDead(OldOp);
2817 // For split transfer intrinsics we have an incredibly useful assurance:
2818 // the source and destination do not reside within the same alloca, and at
2819 // least one of them does not escape. This means that we can replace
2820 // memmove with memcpy, and we don't need to worry about all manner of
2821 // downsides to splitting and transforming the operations.
2823 // If this doesn't map cleanly onto the alloca type, and that type isn't
2824 // a single value type, just emit a memcpy.
2826 = !VecTy && !IntTy && (BeginOffset != NewAllocaBeginOffset ||
2827 EndOffset != NewAllocaEndOffset ||
2828 !NewAI.getAllocatedType()->isSingleValueType());
2830 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2831 // size hasn't been shrunk based on analysis of the viable range, this is
2833 if (EmitMemCpy && &OldAI == &NewAI) {
2834 uint64_t OrigBegin = IsDest ? MTO.DestBegin : MTO.SourceBegin;
2835 uint64_t OrigEnd = IsDest ? MTO.DestEnd : MTO.SourceEnd;
2836 // Ensure the start lines up.
2837 assert(BeginOffset == OrigBegin);
2840 // Rewrite the size as needed.
2841 if (EndOffset != OrigEnd)
2842 II.setLength(ConstantInt::get(II.getLength()->getType(),
2843 EndOffset - BeginOffset));
2846 // Record this instruction for deletion.
2847 Pass.DeadInsts.insert(&II);
2849 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2850 // alloca that should be re-examined after rewriting this instruction.
2851 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2853 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets()))
2854 Pass.Worklist.insert(AI);
2857 Type *OtherPtrTy = IsDest ? II.getRawSource()->getType()
2858 : II.getRawDest()->getType();
2860 // Compute the other pointer, folding as much as possible to produce
2861 // a single, simple GEP in most cases.
2862 OtherPtr = getAdjustedPtr(IRB, TD, OtherPtr, RelOffset, OtherPtrTy,
2863 getName("." + OtherPtr->getName()));
2866 = getAdjustedAllocaPtr(IRB, IsDest ? II.getRawDest()->getType()
2867 : II.getRawSource()->getType());
2868 Type *SizeTy = II.getLength()->getType();
2869 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2871 CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
2872 IsDest ? OtherPtr : OurPtr,
2873 Size, Align, II.isVolatile());
2875 DEBUG(dbgs() << " to: " << *New << "\n");
2879 // Note that we clamp the alignment to 1 here as a 0 alignment for a memcpy
2880 // is equivalent to 1, but that isn't true if we end up rewriting this as
2885 bool IsWholeAlloca = BeginOffset == NewAllocaBeginOffset &&
2886 EndOffset == NewAllocaEndOffset;
2887 uint64_t Size = EndOffset - BeginOffset;
2888 unsigned BeginIndex = VecTy ? getIndex(BeginOffset) : 0;
2889 unsigned EndIndex = VecTy ? getIndex(EndOffset) : 0;
2890 unsigned NumElements = EndIndex - BeginIndex;
2891 IntegerType *SubIntTy
2892 = IntTy ? Type::getIntNTy(IntTy->getContext(), Size*8) : 0;
2894 Type *OtherPtrTy = NewAI.getType();
2895 if (VecTy && !IsWholeAlloca) {
2896 if (NumElements == 1)
2897 OtherPtrTy = VecTy->getElementType();
2899 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2901 OtherPtrTy = OtherPtrTy->getPointerTo();
2902 } else if (IntTy && !IsWholeAlloca) {
2903 OtherPtrTy = SubIntTy->getPointerTo();
2906 Value *SrcPtr = getAdjustedPtr(IRB, TD, OtherPtr, RelOffset, OtherPtrTy,
2907 getName("." + OtherPtr->getName()));
2908 Value *DstPtr = &NewAI;
2910 std::swap(SrcPtr, DstPtr);
2913 if (VecTy && !IsWholeAlloca && !IsDest) {
2914 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2916 Src = extractVector(IRB, Src, BeginIndex, EndIndex, getName(".vec"));
2917 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2918 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2920 Src = convertValue(TD, IRB, Src, IntTy);
2921 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2922 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2923 Src = extractInteger(TD, IRB, Src, SubIntTy, Offset, getName(".extract"));
2925 Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(),
2926 getName(".copyload"));
2929 if (VecTy && !IsWholeAlloca && IsDest) {
2930 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2931 getName(".oldload"));
2932 Src = insertVector(IRB, Old, Src, BeginIndex, getName(".vec"));
2933 } else if (IntTy && !IsWholeAlloca && IsDest) {
2934 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2935 getName(".oldload"));
2936 Old = convertValue(TD, IRB, Old, IntTy);
2937 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2938 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2939 Src = insertInteger(TD, IRB, Old, Src, Offset, getName(".insert"));
2940 Src = convertValue(TD, IRB, Src, NewAllocaTy);
2943 StoreInst *Store = cast<StoreInst>(
2944 IRB.CreateAlignedStore(Src, DstPtr, Align, II.isVolatile()));
2946 DEBUG(dbgs() << " to: " << *Store << "\n");
2947 return !II.isVolatile();
2950 bool visitIntrinsicInst(IntrinsicInst &II) {
2951 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2952 II.getIntrinsicID() == Intrinsic::lifetime_end);
2953 DEBUG(dbgs() << " original: " << II << "\n");
2954 IRBuilder<> IRB(&II);
2955 assert(II.getArgOperand(1) == OldPtr);
2957 // Record this instruction for deletion.
2958 Pass.DeadInsts.insert(&II);
2961 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2962 EndOffset - BeginOffset);
2963 Value *Ptr = getAdjustedAllocaPtr(IRB, II.getArgOperand(1)->getType());
2965 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2966 New = IRB.CreateLifetimeStart(Ptr, Size);
2968 New = IRB.CreateLifetimeEnd(Ptr, Size);
2971 DEBUG(dbgs() << " to: " << *New << "\n");
2975 bool visitPHINode(PHINode &PN) {
2976 DEBUG(dbgs() << " original: " << PN << "\n");
2978 // We would like to compute a new pointer in only one place, but have it be
2979 // as local as possible to the PHI. To do that, we re-use the location of
2980 // the old pointer, which necessarily must be in the right position to
2981 // dominate the PHI.
2982 IRBuilder<> PtrBuilder(cast<Instruction>(OldPtr));
2984 Value *NewPtr = getAdjustedAllocaPtr(PtrBuilder, OldPtr->getType());
2985 // Replace the operands which were using the old pointer.
2986 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
2988 DEBUG(dbgs() << " to: " << PN << "\n");
2989 deleteIfTriviallyDead(OldPtr);
2993 bool visitSelectInst(SelectInst &SI) {
2994 DEBUG(dbgs() << " original: " << SI << "\n");
2995 IRBuilder<> IRB(&SI);
2997 // Find the operand we need to rewrite here.
2998 bool IsTrueVal = SI.getTrueValue() == OldPtr;
3000 assert(SI.getFalseValue() != OldPtr && "Pointer is both operands!");
3002 assert(SI.getFalseValue() == OldPtr && "Pointer isn't an operand!");
3004 Value *NewPtr = getAdjustedAllocaPtr(IRB, OldPtr->getType());
3005 SI.setOperand(IsTrueVal ? 1 : 2, NewPtr);
3006 DEBUG(dbgs() << " to: " << SI << "\n");
3007 deleteIfTriviallyDead(OldPtr);
3015 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
3017 /// This pass aggressively rewrites all aggregate loads and stores on
3018 /// a particular pointer (or any pointer derived from it which we can identify)
3019 /// with scalar loads and stores.
3020 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
3021 // Befriend the base class so it can delegate to private visit methods.
3022 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
3024 const DataLayout &TD;
3026 /// Queue of pointer uses to analyze and potentially rewrite.
3027 SmallVector<Use *, 8> Queue;
3029 /// Set to prevent us from cycling with phi nodes and loops.
3030 SmallPtrSet<User *, 8> Visited;
3032 /// The current pointer use being rewritten. This is used to dig up the used
3033 /// value (as opposed to the user).
3037 AggLoadStoreRewriter(const DataLayout &TD) : TD(TD) {}
3039 /// Rewrite loads and stores through a pointer and all pointers derived from
3041 bool rewrite(Instruction &I) {
3042 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
3044 bool Changed = false;
3045 while (!Queue.empty()) {
3046 U = Queue.pop_back_val();
3047 Changed |= visit(cast<Instruction>(U->getUser()));
3053 /// Enqueue all the users of the given instruction for further processing.
3054 /// This uses a set to de-duplicate users.
3055 void enqueueUsers(Instruction &I) {
3056 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
3058 if (Visited.insert(*UI))
3059 Queue.push_back(&UI.getUse());
3062 // Conservative default is to not rewrite anything.
3063 bool visitInstruction(Instruction &I) { return false; }
3065 /// \brief Generic recursive split emission class.
3066 template <typename Derived>
3069 /// The builder used to form new instructions.
3071 /// The indices which to be used with insert- or extractvalue to select the
3072 /// appropriate value within the aggregate.
3073 SmallVector<unsigned, 4> Indices;
3074 /// The indices to a GEP instruction which will move Ptr to the correct slot
3075 /// within the aggregate.
3076 SmallVector<Value *, 4> GEPIndices;
3077 /// The base pointer of the original op, used as a base for GEPing the
3078 /// split operations.
3081 /// Initialize the splitter with an insertion point, Ptr and start with a
3082 /// single zero GEP index.
3083 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
3084 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
3087 /// \brief Generic recursive split emission routine.
3089 /// This method recursively splits an aggregate op (load or store) into
3090 /// scalar or vector ops. It splits recursively until it hits a single value
3091 /// and emits that single value operation via the template argument.
3093 /// The logic of this routine relies on GEPs and insertvalue and
3094 /// extractvalue all operating with the same fundamental index list, merely
3095 /// formatted differently (GEPs need actual values).
3097 /// \param Ty The type being split recursively into smaller ops.
3098 /// \param Agg The aggregate value being built up or stored, depending on
3099 /// whether this is splitting a load or a store respectively.
3100 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3101 if (Ty->isSingleValueType())
3102 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
3104 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3105 unsigned OldSize = Indices.size();
3107 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3109 assert(Indices.size() == OldSize && "Did not return to the old size");
3110 Indices.push_back(Idx);
3111 GEPIndices.push_back(IRB.getInt32(Idx));
3112 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3113 GEPIndices.pop_back();
3119 if (StructType *STy = dyn_cast<StructType>(Ty)) {
3120 unsigned OldSize = Indices.size();
3122 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3124 assert(Indices.size() == OldSize && "Did not return to the old size");
3125 Indices.push_back(Idx);
3126 GEPIndices.push_back(IRB.getInt32(Idx));
3127 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3128 GEPIndices.pop_back();
3134 llvm_unreachable("Only arrays and structs are aggregate loadable types");
3138 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3139 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3140 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
3142 /// Emit a leaf load of a single value. This is called at the leaves of the
3143 /// recursive emission to actually load values.
3144 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3145 assert(Ty->isSingleValueType());
3146 // Load the single value and insert it using the indices.
3147 Value *GEP = IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep");
3148 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
3149 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3150 DEBUG(dbgs() << " to: " << *Load << "\n");
3154 bool visitLoadInst(LoadInst &LI) {
3155 assert(LI.getPointerOperand() == *U);
3156 if (!LI.isSimple() || LI.getType()->isSingleValueType())
3159 // We have an aggregate being loaded, split it apart.
3160 DEBUG(dbgs() << " original: " << LI << "\n");
3161 LoadOpSplitter Splitter(&LI, *U);
3162 Value *V = UndefValue::get(LI.getType());
3163 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3164 LI.replaceAllUsesWith(V);
3165 LI.eraseFromParent();
3169 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3170 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3171 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
3173 /// Emit a leaf store of a single value. This is called at the leaves of the
3174 /// recursive emission to actually produce stores.
3175 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3176 assert(Ty->isSingleValueType());
3177 // Extract the single value and store it using the indices.
3178 Value *Store = IRB.CreateStore(
3179 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
3180 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
3182 DEBUG(dbgs() << " to: " << *Store << "\n");
3186 bool visitStoreInst(StoreInst &SI) {
3187 if (!SI.isSimple() || SI.getPointerOperand() != *U)
3189 Value *V = SI.getValueOperand();
3190 if (V->getType()->isSingleValueType())
3193 // We have an aggregate being stored, split it apart.
3194 DEBUG(dbgs() << " original: " << SI << "\n");
3195 StoreOpSplitter Splitter(&SI, *U);
3196 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3197 SI.eraseFromParent();
3201 bool visitBitCastInst(BitCastInst &BC) {
3206 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3211 bool visitPHINode(PHINode &PN) {
3216 bool visitSelectInst(SelectInst &SI) {
3223 /// \brief Strip aggregate type wrapping.
3225 /// This removes no-op aggregate types wrapping an underlying type. It will
3226 /// strip as many layers of types as it can without changing either the type
3227 /// size or the allocated size.
3228 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
3229 if (Ty->isSingleValueType())
3232 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3233 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
3236 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3237 InnerTy = ArrTy->getElementType();
3238 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3239 const StructLayout *SL = DL.getStructLayout(STy);
3240 unsigned Index = SL->getElementContainingOffset(0);
3241 InnerTy = STy->getElementType(Index);
3246 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3247 TypeSize > DL.getTypeSizeInBits(InnerTy))
3250 return stripAggregateTypeWrapping(DL, InnerTy);
3253 /// \brief Try to find a partition of the aggregate type passed in for a given
3254 /// offset and size.
3256 /// This recurses through the aggregate type and tries to compute a subtype
3257 /// based on the offset and size. When the offset and size span a sub-section
3258 /// of an array, it will even compute a new array type for that sub-section,
3259 /// and the same for structs.
3261 /// Note that this routine is very strict and tries to find a partition of the
3262 /// type which produces the *exact* right offset and size. It is not forgiving
3263 /// when the size or offset cause either end of type-based partition to be off.
3264 /// Also, this is a best-effort routine. It is reasonable to give up and not
3265 /// return a type if necessary.
3266 static Type *getTypePartition(const DataLayout &TD, Type *Ty,
3267 uint64_t Offset, uint64_t Size) {
3268 if (Offset == 0 && TD.getTypeAllocSize(Ty) == Size)
3269 return stripAggregateTypeWrapping(TD, Ty);
3270 if (Offset > TD.getTypeAllocSize(Ty) ||
3271 (TD.getTypeAllocSize(Ty) - Offset) < Size)
3274 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3275 // We can't partition pointers...
3276 if (SeqTy->isPointerTy())
3279 Type *ElementTy = SeqTy->getElementType();
3280 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
3281 uint64_t NumSkippedElements = Offset / ElementSize;
3282 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy))
3283 if (NumSkippedElements >= ArrTy->getNumElements())
3285 if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy))
3286 if (NumSkippedElements >= VecTy->getNumElements())
3288 Offset -= NumSkippedElements * ElementSize;
3290 // First check if we need to recurse.
3291 if (Offset > 0 || Size < ElementSize) {
3292 // Bail if the partition ends in a different array element.
3293 if ((Offset + Size) > ElementSize)
3295 // Recurse through the element type trying to peel off offset bytes.
3296 return getTypePartition(TD, ElementTy, Offset, Size);
3298 assert(Offset == 0);
3300 if (Size == ElementSize)
3301 return stripAggregateTypeWrapping(TD, ElementTy);
3302 assert(Size > ElementSize);
3303 uint64_t NumElements = Size / ElementSize;
3304 if (NumElements * ElementSize != Size)
3306 return ArrayType::get(ElementTy, NumElements);
3309 StructType *STy = dyn_cast<StructType>(Ty);
3313 const StructLayout *SL = TD.getStructLayout(STy);
3314 if (Offset >= SL->getSizeInBytes())
3316 uint64_t EndOffset = Offset + Size;
3317 if (EndOffset > SL->getSizeInBytes())
3320 unsigned Index = SL->getElementContainingOffset(Offset);
3321 Offset -= SL->getElementOffset(Index);
3323 Type *ElementTy = STy->getElementType(Index);
3324 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
3325 if (Offset >= ElementSize)
3326 return 0; // The offset points into alignment padding.
3328 // See if any partition must be contained by the element.
3329 if (Offset > 0 || Size < ElementSize) {
3330 if ((Offset + Size) > ElementSize)
3332 return getTypePartition(TD, ElementTy, Offset, Size);
3334 assert(Offset == 0);
3336 if (Size == ElementSize)
3337 return stripAggregateTypeWrapping(TD, ElementTy);
3339 StructType::element_iterator EI = STy->element_begin() + Index,
3340 EE = STy->element_end();
3341 if (EndOffset < SL->getSizeInBytes()) {
3342 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3343 if (Index == EndIndex)
3344 return 0; // Within a single element and its padding.
3346 // Don't try to form "natural" types if the elements don't line up with the
3348 // FIXME: We could potentially recurse down through the last element in the
3349 // sub-struct to find a natural end point.
3350 if (SL->getElementOffset(EndIndex) != EndOffset)
3353 assert(Index < EndIndex);
3354 EE = STy->element_begin() + EndIndex;
3357 // Try to build up a sub-structure.
3358 StructType *SubTy = StructType::get(STy->getContext(), makeArrayRef(EI, EE),
3360 const StructLayout *SubSL = TD.getStructLayout(SubTy);
3361 if (Size != SubSL->getSizeInBytes())
3362 return 0; // The sub-struct doesn't have quite the size needed.
3367 /// \brief Rewrite an alloca partition's users.
3369 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3370 /// to rewrite uses of an alloca partition to be conducive for SSA value
3371 /// promotion. If the partition needs a new, more refined alloca, this will
3372 /// build that new alloca, preserving as much type information as possible, and
3373 /// rewrite the uses of the old alloca to point at the new one and have the
3374 /// appropriate new offsets. It also evaluates how successful the rewrite was
3375 /// at enabling promotion and if it was successful queues the alloca to be
3377 bool SROA::rewriteAllocaPartition(AllocaInst &AI,
3378 AllocaPartitioning &P,
3379 AllocaPartitioning::iterator PI) {
3380 uint64_t AllocaSize = PI->EndOffset - PI->BeginOffset;
3381 bool IsLive = false;
3382 for (AllocaPartitioning::use_iterator UI = P.use_begin(PI),
3384 UI != UE && !IsLive; ++UI)
3388 return false; // No live uses left of this partition.
3390 DEBUG(dbgs() << "Speculating PHIs and selects in partition "
3391 << "[" << PI->BeginOffset << "," << PI->EndOffset << ")\n");
3393 PHIOrSelectSpeculator Speculator(*TD, P, *this);
3394 DEBUG(dbgs() << " speculating ");
3395 DEBUG(P.print(dbgs(), PI, ""));
3396 Speculator.visitUsers(PI);
3398 // Try to compute a friendly type for this partition of the alloca. This
3399 // won't always succeed, in which case we fall back to a legal integer type
3400 // or an i8 array of an appropriate size.
3402 if (Type *PartitionTy = P.getCommonType(PI))
3403 if (TD->getTypeAllocSize(PartitionTy) >= AllocaSize)
3404 AllocaTy = PartitionTy;
3406 if (Type *PartitionTy = getTypePartition(*TD, AI.getAllocatedType(),
3407 PI->BeginOffset, AllocaSize))
3408 AllocaTy = PartitionTy;
3410 (AllocaTy->isArrayTy() &&
3411 AllocaTy->getArrayElementType()->isIntegerTy())) &&
3412 TD->isLegalInteger(AllocaSize * 8))
3413 AllocaTy = Type::getIntNTy(*C, AllocaSize * 8);
3415 AllocaTy = ArrayType::get(Type::getInt8Ty(*C), AllocaSize);
3416 assert(TD->getTypeAllocSize(AllocaTy) >= AllocaSize);
3418 // Check for the case where we're going to rewrite to a new alloca of the
3419 // exact same type as the original, and with the same access offsets. In that
3420 // case, re-use the existing alloca, but still run through the rewriter to
3421 // perform phi and select speculation.
3423 if (AllocaTy == AI.getAllocatedType()) {
3424 assert(PI->BeginOffset == 0 &&
3425 "Non-zero begin offset but same alloca type");
3426 assert(PI == P.begin() && "Begin offset is zero on later partition");
3429 unsigned Alignment = AI.getAlignment();
3431 // The minimum alignment which users can rely on when the explicit
3432 // alignment is omitted or zero is that required by the ABI for this
3434 Alignment = TD->getABITypeAlignment(AI.getAllocatedType());
3436 Alignment = MinAlign(Alignment, PI->BeginOffset);
3437 // If we will get at least this much alignment from the type alone, leave
3438 // the alloca's alignment unconstrained.
3439 if (Alignment <= TD->getABITypeAlignment(AllocaTy))
3441 NewAI = new AllocaInst(AllocaTy, 0, Alignment,
3442 AI.getName() + ".sroa." + Twine(PI - P.begin()),
3447 DEBUG(dbgs() << "Rewriting alloca partition "
3448 << "[" << PI->BeginOffset << "," << PI->EndOffset << ") to: "
3451 // Track the high watermark of the post-promotion worklist. We will reset it
3452 // to this point if the alloca is not in fact scheduled for promotion.
3453 unsigned PPWOldSize = PostPromotionWorklist.size();
3455 AllocaPartitionRewriter Rewriter(*TD, P, PI, *this, AI, *NewAI,
3456 PI->BeginOffset, PI->EndOffset);
3457 DEBUG(dbgs() << " rewriting ");
3458 DEBUG(P.print(dbgs(), PI, ""));
3459 bool Promotable = Rewriter.visitUsers(P.use_begin(PI), P.use_end(PI));
3461 DEBUG(dbgs() << " and queuing for promotion\n");
3462 PromotableAllocas.push_back(NewAI);
3463 } else if (NewAI != &AI) {
3464 // If we can't promote the alloca, iterate on it to check for new
3465 // refinements exposed by splitting the current alloca. Don't iterate on an
3466 // alloca which didn't actually change and didn't get promoted.
3467 Worklist.insert(NewAI);
3470 // Drop any post-promotion work items if promotion didn't happen.
3472 while (PostPromotionWorklist.size() > PPWOldSize)
3473 PostPromotionWorklist.pop_back();
3478 /// \brief Walks the partitioning of an alloca rewriting uses of each partition.
3479 bool SROA::splitAlloca(AllocaInst &AI, AllocaPartitioning &P) {
3480 bool Changed = false;
3481 for (AllocaPartitioning::iterator PI = P.begin(), PE = P.end(); PI != PE;
3483 Changed |= rewriteAllocaPartition(AI, P, PI);
3488 /// \brief Analyze an alloca for SROA.
3490 /// This analyzes the alloca to ensure we can reason about it, builds
3491 /// a partitioning of the alloca, and then hands it off to be split and
3492 /// rewritten as needed.
3493 bool SROA::runOnAlloca(AllocaInst &AI) {
3494 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3495 ++NumAllocasAnalyzed;
3497 // Special case dead allocas, as they're trivial.
3498 if (AI.use_empty()) {
3499 AI.eraseFromParent();
3503 // Skip alloca forms that this analysis can't handle.
3504 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3505 TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
3508 bool Changed = false;
3510 // First, split any FCA loads and stores touching this alloca to promote
3511 // better splitting and promotion opportunities.
3512 AggLoadStoreRewriter AggRewriter(*TD);
3513 Changed |= AggRewriter.rewrite(AI);
3515 // Build the partition set using a recursive instruction-visiting builder.
3516 AllocaPartitioning P(*TD, AI);
3517 DEBUG(P.print(dbgs()));
3521 // Delete all the dead users of this alloca before splitting and rewriting it.
3522 for (AllocaPartitioning::dead_user_iterator DI = P.dead_user_begin(),
3523 DE = P.dead_user_end();
3526 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3527 DeadInsts.insert(*DI);
3529 for (AllocaPartitioning::dead_op_iterator DO = P.dead_op_begin(),
3530 DE = P.dead_op_end();
3533 // Clobber the use with an undef value.
3534 **DO = UndefValue::get(OldV->getType());
3535 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3536 if (isInstructionTriviallyDead(OldI)) {
3538 DeadInsts.insert(OldI);
3542 // No partitions to split. Leave the dead alloca for a later pass to clean up.
3543 if (P.begin() == P.end())
3546 return splitAlloca(AI, P) || Changed;
3549 /// \brief Delete the dead instructions accumulated in this run.
3551 /// Recursively deletes the dead instructions we've accumulated. This is done
3552 /// at the very end to maximize locality of the recursive delete and to
3553 /// minimize the problems of invalidated instruction pointers as such pointers
3554 /// are used heavily in the intermediate stages of the algorithm.
3556 /// We also record the alloca instructions deleted here so that they aren't
3557 /// subsequently handed to mem2reg to promote.
3558 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3559 while (!DeadInsts.empty()) {
3560 Instruction *I = DeadInsts.pop_back_val();
3561 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3563 I->replaceAllUsesWith(UndefValue::get(I->getType()));
3565 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
3566 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
3567 // Zero out the operand and see if it becomes trivially dead.
3569 if (isInstructionTriviallyDead(U))
3570 DeadInsts.insert(U);
3573 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3574 DeletedAllocas.insert(AI);
3577 I->eraseFromParent();
3581 /// \brief Promote the allocas, using the best available technique.
3583 /// This attempts to promote whatever allocas have been identified as viable in
3584 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3585 /// If there is a domtree available, we attempt to promote using the full power
3586 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3587 /// based on the SSAUpdater utilities. This function returns whether any
3588 /// promotion occurred.
3589 bool SROA::promoteAllocas(Function &F) {
3590 if (PromotableAllocas.empty())
3593 NumPromoted += PromotableAllocas.size();
3595 if (DT && !ForceSSAUpdater) {
3596 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3597 PromoteMemToReg(PromotableAllocas, *DT);
3598 PromotableAllocas.clear();
3602 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3604 DIBuilder DIB(*F.getParent());
3605 SmallVector<Instruction*, 64> Insts;
3607 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3608 AllocaInst *AI = PromotableAllocas[Idx];
3609 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
3611 Instruction *I = cast<Instruction>(*UI++);
3612 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3613 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3614 // leading to them) here. Eventually it should use them to optimize the
3615 // scalar values produced.
3616 if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
3617 assert(onlyUsedByLifetimeMarkers(I) &&
3618 "Found a bitcast used outside of a lifetime marker.");
3619 while (!I->use_empty())
3620 cast<Instruction>(*I->use_begin())->eraseFromParent();
3621 I->eraseFromParent();
3624 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3625 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3626 II->getIntrinsicID() == Intrinsic::lifetime_end);
3627 II->eraseFromParent();
3633 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3637 PromotableAllocas.clear();
3642 /// \brief A predicate to test whether an alloca belongs to a set.
3643 class IsAllocaInSet {
3644 typedef SmallPtrSet<AllocaInst *, 4> SetType;
3648 typedef AllocaInst *argument_type;
3650 IsAllocaInSet(const SetType &Set) : Set(Set) {}
3651 bool operator()(AllocaInst *AI) const { return Set.count(AI); }
3655 bool SROA::runOnFunction(Function &F) {
3656 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3657 C = &F.getContext();
3658 TD = getAnalysisIfAvailable<DataLayout>();
3660 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3663 DT = getAnalysisIfAvailable<DominatorTree>();
3665 BasicBlock &EntryBB = F.getEntryBlock();
3666 for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end());
3668 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3669 Worklist.insert(AI);
3671 bool Changed = false;
3672 // A set of deleted alloca instruction pointers which should be removed from
3673 // the list of promotable allocas.
3674 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3677 while (!Worklist.empty()) {
3678 Changed |= runOnAlloca(*Worklist.pop_back_val());
3679 deleteDeadInstructions(DeletedAllocas);
3681 // Remove the deleted allocas from various lists so that we don't try to
3682 // continue processing them.
3683 if (!DeletedAllocas.empty()) {
3684 Worklist.remove_if(IsAllocaInSet(DeletedAllocas));
3685 PostPromotionWorklist.remove_if(IsAllocaInSet(DeletedAllocas));
3686 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3687 PromotableAllocas.end(),
3688 IsAllocaInSet(DeletedAllocas)),
3689 PromotableAllocas.end());
3690 DeletedAllocas.clear();
3694 Changed |= promoteAllocas(F);
3696 Worklist = PostPromotionWorklist;
3697 PostPromotionWorklist.clear();
3698 } while (!Worklist.empty());
3703 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3704 if (RequiresDomTree)
3705 AU.addRequired<DominatorTree>();
3706 AU.setPreservesCFG();