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/Constants.h"
29 #include "llvm/DIBuilder.h"
30 #include "llvm/DebugInfo.h"
31 #include "llvm/DerivedTypes.h"
32 #include "llvm/Function.h"
33 #include "llvm/IRBuilder.h"
34 #include "llvm/Instructions.h"
35 #include "llvm/IntrinsicInst.h"
36 #include "llvm/LLVMContext.h"
37 #include "llvm/Module.h"
38 #include "llvm/Operator.h"
39 #include "llvm/Pass.h"
40 #include "llvm/ADT/SetVector.h"
41 #include "llvm/ADT/SmallVector.h"
42 #include "llvm/ADT/Statistic.h"
43 #include "llvm/ADT/STLExtras.h"
44 #include "llvm/Analysis/Dominators.h"
45 #include "llvm/Analysis/Loads.h"
46 #include "llvm/Analysis/ValueTracking.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/ErrorHandling.h"
50 #include "llvm/Support/GetElementPtrTypeIterator.h"
51 #include "llvm/Support/InstVisitor.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/raw_ostream.h"
54 #include "llvm/Target/TargetData.h"
55 #include "llvm/Transforms/Utils/Local.h"
56 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
57 #include "llvm/Transforms/Utils/SSAUpdater.h"
60 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
61 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
62 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
63 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
64 STATISTIC(NumDeleted, "Number of instructions deleted");
65 STATISTIC(NumVectorized, "Number of vectorized aggregates");
67 /// Hidden option to force the pass to not use DomTree and mem2reg, instead
68 /// forming SSA values through the SSAUpdater infrastructure.
70 ForceSSAUpdater("force-ssa-updater", cl::init(false), cl::Hidden);
73 /// \brief Alloca partitioning representation.
75 /// This class represents a partitioning of an alloca into slices, and
76 /// information about the nature of uses of each slice of the alloca. The goal
77 /// is that this information is sufficient to decide if and how to split the
78 /// alloca apart and replace slices with scalars. It is also intended that this
79 /// structure can capture the relevant information needed both to decide about
80 /// and to enact these transformations.
81 class AllocaPartitioning {
83 /// \brief A common base class for representing a half-open byte range.
85 /// \brief The beginning offset of the range.
88 /// \brief The ending offset, not included in the range.
91 ByteRange() : BeginOffset(), EndOffset() {}
92 ByteRange(uint64_t BeginOffset, uint64_t EndOffset)
93 : BeginOffset(BeginOffset), EndOffset(EndOffset) {}
95 /// \brief Support for ordering ranges.
97 /// This provides an ordering over ranges such that start offsets are
98 /// always increasing, and within equal start offsets, the end offsets are
99 /// decreasing. Thus the spanning range comes first in a cluster with the
100 /// same start position.
101 bool operator<(const ByteRange &RHS) const {
102 if (BeginOffset < RHS.BeginOffset) return true;
103 if (BeginOffset > RHS.BeginOffset) return false;
104 if (EndOffset > RHS.EndOffset) return true;
108 /// \brief Support comparison with a single offset to allow binary searches.
109 friend bool operator<(const ByteRange &LHS, uint64_t RHSOffset) {
110 return LHS.BeginOffset < RHSOffset;
113 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
114 const ByteRange &RHS) {
115 return LHSOffset < RHS.BeginOffset;
118 bool operator==(const ByteRange &RHS) const {
119 return BeginOffset == RHS.BeginOffset && EndOffset == RHS.EndOffset;
121 bool operator!=(const ByteRange &RHS) const { return !operator==(RHS); }
124 /// \brief A partition of an alloca.
126 /// This structure represents a contiguous partition of the alloca. These are
127 /// formed by examining the uses of the alloca. During formation, they may
128 /// overlap but once an AllocaPartitioning is built, the Partitions within it
129 /// are all disjoint.
130 struct Partition : public ByteRange {
131 /// \brief Whether this partition is splittable into smaller partitions.
133 /// We flag partitions as splittable when they are formed entirely due to
134 /// accesses by trivially splittable operations such as memset and memcpy.
136 /// FIXME: At some point we should consider loads and stores of FCAs to be
137 /// splittable and eagerly split them into scalar values.
140 Partition() : ByteRange(), IsSplittable() {}
141 Partition(uint64_t BeginOffset, uint64_t EndOffset, bool IsSplittable)
142 : ByteRange(BeginOffset, EndOffset), IsSplittable(IsSplittable) {}
145 /// \brief A particular use of a partition of the alloca.
147 /// This structure is used to associate uses of a partition with it. They
148 /// mark the range of bytes which are referenced by a particular instruction,
149 /// and includes a handle to the user itself and the pointer value in use.
150 /// The bounds of these uses are determined by intersecting the bounds of the
151 /// memory use itself with a particular partition. As a consequence there is
152 /// intentionally overlap between various uses of the same partition.
153 struct PartitionUse : public ByteRange {
154 /// \brief The use in question. Provides access to both user and used value.
156 /// Note that this may be null if the partition use is *dead*, that is, it
157 /// should be ignored.
160 PartitionUse() : ByteRange(), U() {}
161 PartitionUse(uint64_t BeginOffset, uint64_t EndOffset, Use *U)
162 : ByteRange(BeginOffset, EndOffset), U(U) {}
165 /// \brief Construct a partitioning of a particular alloca.
167 /// Construction does most of the work for partitioning the alloca. This
168 /// performs the necessary walks of users and builds a partitioning from it.
169 AllocaPartitioning(const TargetData &TD, AllocaInst &AI);
171 /// \brief Test whether a pointer to the allocation escapes our analysis.
173 /// If this is true, the partitioning is never fully built and should be
175 bool isEscaped() const { return PointerEscapingInstr; }
177 /// \brief Support for iterating over the partitions.
179 typedef SmallVectorImpl<Partition>::iterator iterator;
180 iterator begin() { return Partitions.begin(); }
181 iterator end() { return Partitions.end(); }
183 typedef SmallVectorImpl<Partition>::const_iterator const_iterator;
184 const_iterator begin() const { return Partitions.begin(); }
185 const_iterator end() const { return Partitions.end(); }
188 /// \brief Support for iterating over and manipulating a particular
189 /// partition's uses.
191 /// The iteration support provided for uses is more limited, but also
192 /// includes some manipulation routines to support rewriting the uses of
193 /// partitions during SROA.
195 typedef SmallVectorImpl<PartitionUse>::iterator use_iterator;
196 use_iterator use_begin(unsigned Idx) { return Uses[Idx].begin(); }
197 use_iterator use_begin(const_iterator I) { return Uses[I - begin()].begin(); }
198 use_iterator use_end(unsigned Idx) { return Uses[Idx].end(); }
199 use_iterator use_end(const_iterator I) { return Uses[I - begin()].end(); }
201 typedef SmallVectorImpl<PartitionUse>::const_iterator const_use_iterator;
202 const_use_iterator use_begin(unsigned Idx) const { return Uses[Idx].begin(); }
203 const_use_iterator use_begin(const_iterator I) const {
204 return Uses[I - begin()].begin();
206 const_use_iterator use_end(unsigned Idx) const { return Uses[Idx].end(); }
207 const_use_iterator use_end(const_iterator I) const {
208 return Uses[I - begin()].end();
211 unsigned use_size(unsigned Idx) const { return Uses[Idx].size(); }
212 unsigned use_size(const_iterator I) const { return Uses[I - begin()].size(); }
213 const PartitionUse &getUse(unsigned PIdx, unsigned UIdx) const {
214 return Uses[PIdx][UIdx];
216 const PartitionUse &getUse(const_iterator I, unsigned UIdx) const {
217 return Uses[I - begin()][UIdx];
220 void use_push_back(unsigned Idx, const PartitionUse &PU) {
221 Uses[Idx].push_back(PU);
223 void use_push_back(const_iterator I, const PartitionUse &PU) {
224 Uses[I - begin()].push_back(PU);
228 /// \brief Allow iterating the dead users for this alloca.
230 /// These are instructions which will never actually use the alloca as they
231 /// are outside the allocated range. They are safe to replace with undef and
234 typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator;
235 dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); }
236 dead_user_iterator dead_user_end() const { return DeadUsers.end(); }
239 /// \brief Allow iterating the dead expressions referring to this alloca.
241 /// These are operands which have cannot actually be used to refer to the
242 /// alloca as they are outside its range and the user doesn't correct for
243 /// that. These mostly consist of PHI node inputs and the like which we just
244 /// need to replace with undef.
246 typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator;
247 dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); }
248 dead_op_iterator dead_op_end() const { return DeadOperands.end(); }
251 /// \brief MemTransferInst auxiliary data.
252 /// This struct provides some auxiliary data about memory transfer
253 /// intrinsics such as memcpy and memmove. These intrinsics can use two
254 /// different ranges within the same alloca, and provide other challenges to
255 /// correctly represent. We stash extra data to help us untangle this
256 /// after the partitioning is complete.
257 struct MemTransferOffsets {
258 uint64_t DestBegin, DestEnd;
259 uint64_t SourceBegin, SourceEnd;
262 MemTransferOffsets getMemTransferOffsets(MemTransferInst &II) const {
263 return MemTransferInstData.lookup(&II);
266 /// \brief Map from a PHI or select operand back to a partition.
268 /// When manipulating PHI nodes or selects, they can use more than one
269 /// partition of an alloca. We store a special mapping to allow finding the
270 /// partition referenced by each of these operands, if any.
271 iterator findPartitionForPHIOrSelectOperand(Use *U) {
272 SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt
273 = PHIOrSelectOpMap.find(U);
274 if (MapIt == PHIOrSelectOpMap.end())
277 return begin() + MapIt->second.first;
280 /// \brief Map from a PHI or select operand back to the specific use of
283 /// Similar to mapping these operands back to the partitions, this maps
284 /// directly to the use structure of that partition.
285 use_iterator findPartitionUseForPHIOrSelectOperand(Use *U) {
286 SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt
287 = PHIOrSelectOpMap.find(U);
288 assert(MapIt != PHIOrSelectOpMap.end());
289 return Uses[MapIt->second.first].begin() + MapIt->second.second;
292 /// \brief Compute a common type among the uses of a particular partition.
294 /// This routines walks all of the uses of a particular partition and tries
295 /// to find a common type between them. Untyped operations such as memset and
296 /// memcpy are ignored.
297 Type *getCommonType(iterator I) const;
299 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
300 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
301 void printUsers(raw_ostream &OS, const_iterator I,
302 StringRef Indent = " ") const;
303 void print(raw_ostream &OS) const;
304 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump(const_iterator I) const;
305 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump() const;
309 template <typename DerivedT, typename RetT = void> class BuilderBase;
310 class PartitionBuilder;
311 friend class AllocaPartitioning::PartitionBuilder;
313 friend class AllocaPartitioning::UseBuilder;
316 /// \brief Handle to alloca instruction to simplify method interfaces.
320 /// \brief The instruction responsible for this alloca having no partitioning.
322 /// When an instruction (potentially) escapes the pointer to the alloca, we
323 /// store a pointer to that here and abort trying to partition the alloca.
324 /// This will be null if the alloca is partitioned successfully.
325 Instruction *PointerEscapingInstr;
327 /// \brief The partitions of the alloca.
329 /// We store a vector of the partitions over the alloca here. This vector is
330 /// sorted by increasing begin offset, and then by decreasing end offset. See
331 /// the Partition inner class for more details. Initially (during
332 /// construction) there are overlaps, but we form a disjoint sequence of
333 /// partitions while finishing construction and a fully constructed object is
334 /// expected to always have this as a disjoint space.
335 SmallVector<Partition, 8> Partitions;
337 /// \brief The uses of the partitions.
339 /// This is essentially a mapping from each partition to a list of uses of
340 /// that partition. The mapping is done with a Uses vector that has the exact
341 /// same number of entries as the partition vector. Each entry is itself
342 /// a vector of the uses.
343 SmallVector<SmallVector<PartitionUse, 2>, 8> Uses;
345 /// \brief Instructions which will become dead if we rewrite the alloca.
347 /// Note that these are not separated by partition. This is because we expect
348 /// a partitioned alloca to be completely rewritten or not rewritten at all.
349 /// If rewritten, all these instructions can simply be removed and replaced
350 /// with undef as they come from outside of the allocated space.
351 SmallVector<Instruction *, 8> DeadUsers;
353 /// \brief Operands which will become dead if we rewrite the alloca.
355 /// These are operands that in their particular use can be replaced with
356 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
357 /// to PHI nodes and the like. They aren't entirely dead (there might be
358 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
359 /// want to swap this particular input for undef to simplify the use lists of
361 SmallVector<Use *, 8> DeadOperands;
363 /// \brief The underlying storage for auxiliary memcpy and memset info.
364 SmallDenseMap<MemTransferInst *, MemTransferOffsets, 4> MemTransferInstData;
366 /// \brief A side datastructure used when building up the partitions and uses.
368 /// This mapping is only really used during the initial building of the
369 /// partitioning so that we can retain information about PHI and select nodes
371 SmallDenseMap<Instruction *, std::pair<uint64_t, bool> > PHIOrSelectSizes;
373 /// \brief Auxiliary information for particular PHI or select operands.
374 SmallDenseMap<Use *, std::pair<unsigned, unsigned>, 4> PHIOrSelectOpMap;
376 /// \brief A utility routine called from the constructor.
378 /// This does what it says on the tin. It is the key of the alloca partition
379 /// splitting and merging. After it is called we have the desired disjoint
380 /// collection of partitions.
381 void splitAndMergePartitions();
385 template <typename DerivedT, typename RetT>
386 class AllocaPartitioning::BuilderBase
387 : public InstVisitor<DerivedT, RetT> {
389 BuilderBase(const TargetData &TD, AllocaInst &AI, AllocaPartitioning &P)
391 AllocSize(TD.getTypeAllocSize(AI.getAllocatedType())),
397 const TargetData &TD;
398 const uint64_t AllocSize;
399 AllocaPartitioning &P;
401 SmallPtrSet<Use *, 8> VisitedUses;
407 SmallVector<OffsetUse, 8> Queue;
409 // The active offset and use while visiting.
413 void enqueueUsers(Instruction &I, int64_t UserOffset) {
414 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
416 if (VisitedUses.insert(&UI.getUse())) {
417 OffsetUse OU = { &UI.getUse(), UserOffset };
423 bool computeConstantGEPOffset(GetElementPtrInst &GEPI, int64_t &GEPOffset) {
425 for (gep_type_iterator GTI = gep_type_begin(GEPI), GTE = gep_type_end(GEPI);
427 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
433 // Handle a struct index, which adds its field offset to the pointer.
434 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
435 unsigned ElementIdx = OpC->getZExtValue();
436 const StructLayout *SL = TD.getStructLayout(STy);
437 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
438 // Check that we can continue to model this GEP in a signed 64-bit offset.
439 if (ElementOffset > INT64_MAX ||
441 ((uint64_t)GEPOffset + ElementOffset) > INT64_MAX)) {
442 DEBUG(dbgs() << "WARNING: Encountered a cumulative offset exceeding "
443 << "what can be represented in an int64_t!\n"
444 << " alloca: " << P.AI << "\n");
448 GEPOffset = ElementOffset + (uint64_t)-GEPOffset;
450 GEPOffset += ElementOffset;
454 APInt Index = OpC->getValue().sextOrTrunc(TD.getPointerSizeInBits());
455 Index *= APInt(Index.getBitWidth(),
456 TD.getTypeAllocSize(GTI.getIndexedType()));
457 Index += APInt(Index.getBitWidth(), (uint64_t)GEPOffset,
459 // Check if the result can be stored in our int64_t offset.
460 if (!Index.isSignedIntN(sizeof(GEPOffset) * 8)) {
461 DEBUG(dbgs() << "WARNING: Encountered a cumulative offset exceeding "
462 << "what can be represented in an int64_t!\n"
463 << " alloca: " << P.AI << "\n");
467 GEPOffset = Index.getSExtValue();
472 Value *foldSelectInst(SelectInst &SI) {
473 // If the condition being selected on is a constant or the same value is
474 // being selected between, fold the select. Yes this does (rarely) happen
476 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
477 return SI.getOperand(1+CI->isZero());
478 if (SI.getOperand(1) == SI.getOperand(2)) {
479 assert(*U == SI.getOperand(1));
480 return SI.getOperand(1);
486 /// \brief Builder for the alloca partitioning.
488 /// This class builds an alloca partitioning by recursively visiting the uses
489 /// of an alloca and splitting the partitions for each load and store at each
491 class AllocaPartitioning::PartitionBuilder
492 : public BuilderBase<PartitionBuilder, bool> {
493 friend class InstVisitor<PartitionBuilder, bool>;
495 SmallDenseMap<Instruction *, unsigned> MemTransferPartitionMap;
498 PartitionBuilder(const TargetData &TD, AllocaInst &AI, AllocaPartitioning &P)
499 : BuilderBase<PartitionBuilder, bool>(TD, AI, P) {}
501 /// \brief Run the builder over the allocation.
503 // Note that we have to re-evaluate size on each trip through the loop as
504 // the queue grows at the tail.
505 for (unsigned Idx = 0; Idx < Queue.size(); ++Idx) {
507 Offset = Queue[Idx].Offset;
508 if (!visit(cast<Instruction>(U->getUser())))
515 bool markAsEscaping(Instruction &I) {
516 P.PointerEscapingInstr = &I;
520 void insertUse(Instruction &I, int64_t Offset, uint64_t Size,
521 bool IsSplittable = false) {
522 // Completely skip uses which have a zero size or don't overlap the
525 (Offset >= 0 && (uint64_t)Offset >= AllocSize) ||
526 (Offset < 0 && (uint64_t)-Offset >= Size)) {
527 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
528 << " which starts past the end of the " << AllocSize
530 << " alloca: " << P.AI << "\n"
531 << " use: " << I << "\n");
535 // Clamp the start to the beginning of the allocation.
537 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
538 << " to start at the beginning of the alloca:\n"
539 << " alloca: " << P.AI << "\n"
540 << " use: " << I << "\n");
541 Size -= (uint64_t)-Offset;
545 uint64_t BeginOffset = Offset, EndOffset = BeginOffset + Size;
547 // Clamp the end offset to the end of the allocation. Note that this is
548 // formulated to handle even the case where "BeginOffset + Size" overflows.
549 assert(AllocSize >= BeginOffset); // Established above.
550 if (Size > AllocSize - BeginOffset) {
551 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
552 << " to remain within the " << AllocSize << " byte alloca:\n"
553 << " alloca: " << P.AI << "\n"
554 << " use: " << I << "\n");
555 EndOffset = AllocSize;
558 // See if we can just add a user onto the last slot currently occupied.
559 if (!P.Partitions.empty() &&
560 P.Partitions.back().BeginOffset == BeginOffset &&
561 P.Partitions.back().EndOffset == EndOffset) {
562 P.Partitions.back().IsSplittable &= IsSplittable;
566 Partition New(BeginOffset, EndOffset, IsSplittable);
567 P.Partitions.push_back(New);
570 bool handleLoadOrStore(Type *Ty, Instruction &I, int64_t Offset) {
571 uint64_t Size = TD.getTypeStoreSize(Ty);
573 // If this memory access can be shown to *statically* extend outside the
574 // bounds of of the allocation, it's behavior is undefined, so simply
575 // ignore it. Note that this is more strict than the generic clamping
576 // behavior of insertUse. We also try to handle cases which might run the
578 // FIXME: We should instead consider the pointer to have escaped if this
579 // function is being instrumented for addressing bugs or race conditions.
580 if (Offset < 0 || (uint64_t)Offset >= AllocSize ||
581 Size > (AllocSize - (uint64_t)Offset)) {
582 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte "
583 << (isa<LoadInst>(I) ? "load" : "store") << " @" << Offset
584 << " which extends past the end of the " << AllocSize
586 << " alloca: " << P.AI << "\n"
587 << " use: " << I << "\n");
591 insertUse(I, Offset, Size);
595 bool visitBitCastInst(BitCastInst &BC) {
596 enqueueUsers(BC, Offset);
600 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
602 if (!computeConstantGEPOffset(GEPI, GEPOffset))
603 return markAsEscaping(GEPI);
605 enqueueUsers(GEPI, GEPOffset);
609 bool visitLoadInst(LoadInst &LI) {
610 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
611 "All simple FCA loads should have been pre-split");
612 return handleLoadOrStore(LI.getType(), LI, Offset);
615 bool visitStoreInst(StoreInst &SI) {
616 Value *ValOp = SI.getValueOperand();
618 return markAsEscaping(SI);
620 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
621 "All simple FCA stores should have been pre-split");
622 return handleLoadOrStore(ValOp->getType(), SI, Offset);
626 bool visitMemSetInst(MemSetInst &II) {
627 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
628 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
629 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
630 insertUse(II, Offset, Size, Length);
634 bool visitMemTransferInst(MemTransferInst &II) {
635 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
636 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
638 // Zero-length mem transfer intrinsics can be ignored entirely.
641 MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
643 // Only intrinsics with a constant length can be split.
644 Offsets.IsSplittable = Length;
646 if (*U != II.getRawDest()) {
647 assert(*U == II.getRawSource());
648 Offsets.SourceBegin = Offset;
649 Offsets.SourceEnd = Offset + Size;
651 Offsets.DestBegin = Offset;
652 Offsets.DestEnd = Offset + Size;
655 insertUse(II, Offset, Size, Offsets.IsSplittable);
656 unsigned NewIdx = P.Partitions.size() - 1;
658 SmallDenseMap<Instruction *, unsigned>::const_iterator PMI;
659 bool Inserted = false;
660 llvm::tie(PMI, Inserted)
661 = MemTransferPartitionMap.insert(std::make_pair(&II, NewIdx));
662 if (Offsets.IsSplittable &&
663 (!Inserted || II.getRawSource() == II.getRawDest())) {
664 // We've found a memory transfer intrinsic which refers to the alloca as
665 // both a source and dest. This is detected either by direct equality of
666 // the operand values, or when we visit the intrinsic twice due to two
667 // different chains of values leading to it. We refuse to split these to
668 // simplify splitting logic. If possible, SROA will still split them into
669 // separate allocas and then re-analyze.
670 Offsets.IsSplittable = false;
671 P.Partitions[PMI->second].IsSplittable = false;
672 P.Partitions[NewIdx].IsSplittable = false;
678 // Disable SRoA for any intrinsics except for lifetime invariants.
679 // FIXME: What about debug instrinsics? This matches old behavior, but
680 // doesn't make sense.
681 bool visitIntrinsicInst(IntrinsicInst &II) {
682 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
683 II.getIntrinsicID() == Intrinsic::lifetime_end) {
684 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
685 uint64_t Size = std::min(AllocSize - Offset, Length->getLimitedValue());
686 insertUse(II, Offset, Size, true);
690 return markAsEscaping(II);
693 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
694 // We consider any PHI or select that results in a direct load or store of
695 // the same offset to be a viable use for partitioning purposes. These uses
696 // are considered unsplittable and the size is the maximum loaded or stored
698 SmallPtrSet<Instruction *, 4> Visited;
699 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
700 Visited.insert(Root);
701 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
702 // If there are no loads or stores, the access is dead. We mark that as
703 // a size zero access.
706 Instruction *I, *UsedI;
707 llvm::tie(UsedI, I) = Uses.pop_back_val();
709 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
710 Size = std::max(Size, TD.getTypeStoreSize(LI->getType()));
713 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
714 Value *Op = SI->getOperand(0);
717 Size = std::max(Size, TD.getTypeStoreSize(Op->getType()));
721 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
722 if (!GEP->hasAllZeroIndices())
724 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
725 !isa<SelectInst>(I)) {
729 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
731 if (Visited.insert(cast<Instruction>(*UI)))
732 Uses.push_back(std::make_pair(I, cast<Instruction>(*UI)));
733 } while (!Uses.empty());
738 bool visitPHINode(PHINode &PN) {
739 // See if we already have computed info on this node.
740 std::pair<uint64_t, bool> &PHIInfo = P.PHIOrSelectSizes[&PN];
742 PHIInfo.second = true;
743 insertUse(PN, Offset, PHIInfo.first);
747 // Check for an unsafe use of the PHI node.
748 if (Instruction *EscapingI = hasUnsafePHIOrSelectUse(&PN, PHIInfo.first))
749 return markAsEscaping(*EscapingI);
751 insertUse(PN, Offset, PHIInfo.first);
755 bool visitSelectInst(SelectInst &SI) {
756 if (Value *Result = foldSelectInst(SI)) {
758 // If the result of the constant fold will be the pointer, recurse
759 // through the select as if we had RAUW'ed it.
760 enqueueUsers(SI, Offset);
765 // See if we already have computed info on this node.
766 std::pair<uint64_t, bool> &SelectInfo = P.PHIOrSelectSizes[&SI];
767 if (SelectInfo.first) {
768 SelectInfo.second = true;
769 insertUse(SI, Offset, SelectInfo.first);
773 // Check for an unsafe use of the PHI node.
774 if (Instruction *EscapingI = hasUnsafePHIOrSelectUse(&SI, SelectInfo.first))
775 return markAsEscaping(*EscapingI);
777 insertUse(SI, Offset, SelectInfo.first);
781 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
782 bool visitInstruction(Instruction &I) { return markAsEscaping(I); }
786 /// \brief Use adder for the alloca partitioning.
788 /// This class adds the uses of an alloca to all of the partitions which they
789 /// use. For splittable partitions, this can end up doing essentially a linear
790 /// walk of the partitions, but the number of steps remains bounded by the
791 /// total result instruction size:
792 /// - The number of partitions is a result of the number unsplittable
793 /// instructions using the alloca.
794 /// - The number of users of each partition is at worst the total number of
795 /// splittable instructions using the alloca.
796 /// Thus we will produce N * M instructions in the end, where N are the number
797 /// of unsplittable uses and M are the number of splittable. This visitor does
798 /// the exact same number of updates to the partitioning.
800 /// In the more common case, this visitor will leverage the fact that the
801 /// partition space is pre-sorted, and do a logarithmic search for the
802 /// partition needed, making the total visit a classical ((N + M) * log(N))
803 /// complexity operation.
804 class AllocaPartitioning::UseBuilder : public BuilderBase<UseBuilder> {
805 friend class InstVisitor<UseBuilder>;
807 /// \brief Set to de-duplicate dead instructions found in the use walk.
808 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
811 UseBuilder(const TargetData &TD, AllocaInst &AI, AllocaPartitioning &P)
812 : BuilderBase<UseBuilder>(TD, AI, P) {}
814 /// \brief Run the builder over the allocation.
816 // Note that we have to re-evaluate size on each trip through the loop as
817 // the queue grows at the tail.
818 for (unsigned Idx = 0; Idx < Queue.size(); ++Idx) {
820 Offset = Queue[Idx].Offset;
821 this->visit(cast<Instruction>(U->getUser()));
826 void markAsDead(Instruction &I) {
827 if (VisitedDeadInsts.insert(&I))
828 P.DeadUsers.push_back(&I);
831 void insertUse(Instruction &User, int64_t Offset, uint64_t Size) {
832 // If the use has a zero size or extends outside of the allocation, record
833 // it as a dead use for elimination later.
834 if (Size == 0 || (uint64_t)Offset >= AllocSize ||
835 (Offset < 0 && (uint64_t)-Offset >= Size))
836 return markAsDead(User);
838 // Clamp the start to the beginning of the allocation.
840 Size -= (uint64_t)-Offset;
844 uint64_t BeginOffset = Offset, EndOffset = BeginOffset + Size;
846 // Clamp the end offset to the end of the allocation. Note that this is
847 // formulated to handle even the case where "BeginOffset + Size" overflows.
848 assert(AllocSize >= BeginOffset); // Established above.
849 if (Size > AllocSize - BeginOffset)
850 EndOffset = AllocSize;
852 // NB: This only works if we have zero overlapping partitions.
853 iterator B = std::lower_bound(P.begin(), P.end(), BeginOffset);
854 if (B != P.begin() && llvm::prior(B)->EndOffset > BeginOffset)
856 for (iterator I = B, E = P.end(); I != E && I->BeginOffset < EndOffset;
858 PartitionUse NewPU(std::max(I->BeginOffset, BeginOffset),
859 std::min(I->EndOffset, EndOffset), U);
860 P.use_push_back(I, NewPU);
861 if (isa<PHINode>(U->getUser()) || isa<SelectInst>(U->getUser()))
862 P.PHIOrSelectOpMap[U]
863 = std::make_pair(I - P.begin(), P.Uses[I - P.begin()].size() - 1);
867 void handleLoadOrStore(Type *Ty, Instruction &I, int64_t Offset) {
868 uint64_t Size = TD.getTypeStoreSize(Ty);
870 // If this memory access can be shown to *statically* extend outside the
871 // bounds of of the allocation, it's behavior is undefined, so simply
872 // ignore it. Note that this is more strict than the generic clamping
873 // behavior of insertUse.
874 if (Offset < 0 || (uint64_t)Offset >= AllocSize ||
875 Size > (AllocSize - (uint64_t)Offset))
876 return markAsDead(I);
878 insertUse(I, Offset, Size);
881 void visitBitCastInst(BitCastInst &BC) {
883 return markAsDead(BC);
885 enqueueUsers(BC, Offset);
888 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
889 if (GEPI.use_empty())
890 return markAsDead(GEPI);
893 if (!computeConstantGEPOffset(GEPI, GEPOffset))
894 llvm_unreachable("Unable to compute constant offset for use");
896 enqueueUsers(GEPI, GEPOffset);
899 void visitLoadInst(LoadInst &LI) {
900 handleLoadOrStore(LI.getType(), LI, Offset);
903 void visitStoreInst(StoreInst &SI) {
904 handleLoadOrStore(SI.getOperand(0)->getType(), SI, Offset);
907 void visitMemSetInst(MemSetInst &II) {
908 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
909 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
910 insertUse(II, Offset, Size);
913 void visitMemTransferInst(MemTransferInst &II) {
914 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
915 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
916 insertUse(II, Offset, Size);
919 void visitIntrinsicInst(IntrinsicInst &II) {
920 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
921 II.getIntrinsicID() == Intrinsic::lifetime_end);
923 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
924 insertUse(II, Offset,
925 std::min(AllocSize - Offset, Length->getLimitedValue()));
928 void insertPHIOrSelect(Instruction &User, uint64_t Offset) {
929 uint64_t Size = P.PHIOrSelectSizes.lookup(&User).first;
931 // For PHI and select operands outside the alloca, we can't nuke the entire
932 // phi or select -- the other side might still be relevant, so we special
933 // case them here and use a separate structure to track the operands
934 // themselves which should be replaced with undef.
935 if (Offset >= AllocSize) {
936 P.DeadOperands.push_back(U);
940 insertUse(User, Offset, Size);
942 void visitPHINode(PHINode &PN) {
944 return markAsDead(PN);
946 insertPHIOrSelect(PN, Offset);
948 void visitSelectInst(SelectInst &SI) {
950 return markAsDead(SI);
952 if (Value *Result = foldSelectInst(SI)) {
954 // If the result of the constant fold will be the pointer, recurse
955 // through the select as if we had RAUW'ed it.
956 enqueueUsers(SI, Offset);
958 // Otherwise the operand to the select is dead, and we can replace it
960 P.DeadOperands.push_back(U);
965 insertPHIOrSelect(SI, Offset);
968 /// \brief Unreachable, we've already visited the alloca once.
969 void visitInstruction(Instruction &I) {
970 llvm_unreachable("Unhandled instruction in use builder.");
974 void AllocaPartitioning::splitAndMergePartitions() {
975 size_t NumDeadPartitions = 0;
977 // Track the range of splittable partitions that we pass when accumulating
978 // overlapping unsplittable partitions.
979 uint64_t SplitEndOffset = 0ull;
981 Partition New(0ull, 0ull, false);
983 for (unsigned i = 0, j = i, e = Partitions.size(); i != e; i = j) {
986 if (!Partitions[i].IsSplittable || New.BeginOffset == New.EndOffset) {
987 assert(New.BeginOffset == New.EndOffset);
990 assert(New.IsSplittable);
991 New.EndOffset = std::max(New.EndOffset, Partitions[i].EndOffset);
993 assert(New.BeginOffset != New.EndOffset);
995 // Scan the overlapping partitions.
996 while (j != e && New.EndOffset > Partitions[j].BeginOffset) {
997 // If the new partition we are forming is splittable, stop at the first
998 // unsplittable partition.
999 if (New.IsSplittable && !Partitions[j].IsSplittable)
1002 // Grow the new partition to include any equally splittable range. 'j' is
1003 // always equally splittable when New is splittable, but when New is not
1004 // splittable, we may subsume some (or part of some) splitable partition
1005 // without growing the new one.
1006 if (New.IsSplittable == Partitions[j].IsSplittable) {
1007 New.EndOffset = std::max(New.EndOffset, Partitions[j].EndOffset);
1009 assert(!New.IsSplittable);
1010 assert(Partitions[j].IsSplittable);
1011 SplitEndOffset = std::max(SplitEndOffset, Partitions[j].EndOffset);
1014 Partitions[j].BeginOffset = Partitions[j].EndOffset = UINT64_MAX;
1015 ++NumDeadPartitions;
1019 // If the new partition is splittable, chop off the end as soon as the
1020 // unsplittable subsequent partition starts and ensure we eventually cover
1021 // the splittable area.
1022 if (j != e && New.IsSplittable) {
1023 SplitEndOffset = std::max(SplitEndOffset, New.EndOffset);
1024 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
1027 // Add the new partition if it differs from the original one and is
1028 // non-empty. We can end up with an empty partition here if it was
1029 // splittable but there is an unsplittable one that starts at the same
1031 if (New != Partitions[i]) {
1032 if (New.BeginOffset != New.EndOffset)
1033 Partitions.push_back(New);
1034 // Mark the old one for removal.
1035 Partitions[i].BeginOffset = Partitions[i].EndOffset = UINT64_MAX;
1036 ++NumDeadPartitions;
1039 New.BeginOffset = New.EndOffset;
1040 if (!New.IsSplittable) {
1041 New.EndOffset = std::max(New.EndOffset, SplitEndOffset);
1042 if (j != e && !Partitions[j].IsSplittable)
1043 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
1044 New.IsSplittable = true;
1045 // If there is a trailing splittable partition which won't be fused into
1046 // the next splittable partition go ahead and add it onto the partitions
1048 if (New.BeginOffset < New.EndOffset &&
1049 (j == e || !Partitions[j].IsSplittable ||
1050 New.EndOffset < Partitions[j].BeginOffset)) {
1051 Partitions.push_back(New);
1052 New.BeginOffset = New.EndOffset = 0ull;
1057 // Re-sort the partitions now that they have been split and merged into
1058 // disjoint set of partitions. Also remove any of the dead partitions we've
1059 // replaced in the process.
1060 std::sort(Partitions.begin(), Partitions.end());
1061 if (NumDeadPartitions) {
1062 assert(Partitions.back().BeginOffset == UINT64_MAX);
1063 assert(Partitions.back().EndOffset == UINT64_MAX);
1064 assert((ptrdiff_t)NumDeadPartitions ==
1065 std::count(Partitions.begin(), Partitions.end(), Partitions.back()));
1067 Partitions.erase(Partitions.end() - NumDeadPartitions, Partitions.end());
1070 AllocaPartitioning::AllocaPartitioning(const TargetData &TD, AllocaInst &AI)
1075 PointerEscapingInstr(0) {
1076 PartitionBuilder PB(TD, AI, *this);
1080 if (Partitions.size() > 1) {
1081 // Sort the uses. This arranges for the offsets to be in ascending order,
1082 // and the sizes to be in descending order.
1083 std::sort(Partitions.begin(), Partitions.end());
1085 // Intersect splittability for all partitions with equal offsets and sizes.
1086 // Then remove all but the first so that we have a sequence of non-equal but
1087 // potentially overlapping partitions.
1088 for (iterator I = Partitions.begin(), J = I, E = Partitions.end(); I != E;
1091 while (J != E && *I == *J) {
1092 I->IsSplittable &= J->IsSplittable;
1096 Partitions.erase(std::unique(Partitions.begin(), Partitions.end()),
1099 // Split splittable and merge unsplittable partitions into a disjoint set
1100 // of partitions over the used space of the allocation.
1101 splitAndMergePartitions();
1104 // Now build up the user lists for each of these disjoint partitions by
1105 // re-walking the recursive users of the alloca.
1106 Uses.resize(Partitions.size());
1107 UseBuilder UB(TD, AI, *this);
1111 Type *AllocaPartitioning::getCommonType(iterator I) const {
1113 for (const_use_iterator UI = use_begin(I), UE = use_end(I); UI != UE; ++UI) {
1115 continue; // Skip dead uses.
1116 if (isa<IntrinsicInst>(*UI->U->getUser()))
1118 if (UI->BeginOffset != I->BeginOffset || UI->EndOffset != I->EndOffset)
1122 if (LoadInst *LI = dyn_cast<LoadInst>(UI->U->getUser())) {
1123 UserTy = LI->getType();
1124 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI->U->getUser())) {
1125 UserTy = SI->getValueOperand()->getType();
1128 if (Ty && Ty != UserTy)
1136 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1138 void AllocaPartitioning::print(raw_ostream &OS, const_iterator I,
1139 StringRef Indent) const {
1140 OS << Indent << "partition #" << (I - begin())
1141 << " [" << I->BeginOffset << "," << I->EndOffset << ")"
1142 << (I->IsSplittable ? " (splittable)" : "")
1143 << (Uses[I - begin()].empty() ? " (zero uses)" : "")
1147 void AllocaPartitioning::printUsers(raw_ostream &OS, const_iterator I,
1148 StringRef Indent) const {
1149 for (const_use_iterator UI = use_begin(I), UE = use_end(I);
1152 continue; // Skip dead uses.
1153 OS << Indent << " [" << UI->BeginOffset << "," << UI->EndOffset << ") "
1154 << "used by: " << *UI->U->getUser() << "\n";
1155 if (MemTransferInst *II = dyn_cast<MemTransferInst>(UI->U->getUser())) {
1156 const MemTransferOffsets &MTO = MemTransferInstData.lookup(II);
1158 if (!MTO.IsSplittable)
1159 IsDest = UI->BeginOffset == MTO.DestBegin;
1161 IsDest = MTO.DestBegin != 0u;
1162 OS << Indent << " (original " << (IsDest ? "dest" : "source") << ": "
1163 << "[" << (IsDest ? MTO.DestBegin : MTO.SourceBegin)
1164 << "," << (IsDest ? MTO.DestEnd : MTO.SourceEnd) << ")\n";
1169 void AllocaPartitioning::print(raw_ostream &OS) const {
1170 if (PointerEscapingInstr) {
1171 OS << "No partitioning for alloca: " << AI << "\n"
1172 << " A pointer to this alloca escaped by:\n"
1173 << " " << *PointerEscapingInstr << "\n";
1177 OS << "Partitioning of alloca: " << AI << "\n";
1179 for (const_iterator I = begin(), E = end(); I != E; ++I, ++Num) {
1185 void AllocaPartitioning::dump(const_iterator I) const { print(dbgs(), I); }
1186 void AllocaPartitioning::dump() const { print(dbgs()); }
1188 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1192 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
1194 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
1195 /// the loads and stores of an alloca instruction, as well as updating its
1196 /// debug information. This is used when a domtree is unavailable and thus
1197 /// mem2reg in its full form can't be used to handle promotion of allocas to
1199 class AllocaPromoter : public LoadAndStorePromoter {
1203 SmallVector<DbgDeclareInst *, 4> DDIs;
1204 SmallVector<DbgValueInst *, 4> DVIs;
1207 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
1208 AllocaInst &AI, DIBuilder &DIB)
1209 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
1211 void run(const SmallVectorImpl<Instruction*> &Insts) {
1212 // Remember which alloca we're promoting (for isInstInList).
1213 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
1214 for (Value::use_iterator UI = DebugNode->use_begin(),
1215 UE = DebugNode->use_end();
1217 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
1218 DDIs.push_back(DDI);
1219 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
1220 DVIs.push_back(DVI);
1223 LoadAndStorePromoter::run(Insts);
1224 AI.eraseFromParent();
1225 while (!DDIs.empty())
1226 DDIs.pop_back_val()->eraseFromParent();
1227 while (!DVIs.empty())
1228 DVIs.pop_back_val()->eraseFromParent();
1231 virtual bool isInstInList(Instruction *I,
1232 const SmallVectorImpl<Instruction*> &Insts) const {
1233 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1234 return LI->getOperand(0) == &AI;
1235 return cast<StoreInst>(I)->getPointerOperand() == &AI;
1238 virtual void updateDebugInfo(Instruction *Inst) const {
1239 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(),
1240 E = DDIs.end(); I != E; ++I) {
1241 DbgDeclareInst *DDI = *I;
1242 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
1243 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1244 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
1245 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1247 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(),
1248 E = DVIs.end(); I != E; ++I) {
1249 DbgValueInst *DVI = *I;
1251 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1252 // If an argument is zero extended then use argument directly. The ZExt
1253 // may be zapped by an optimization pass in future.
1254 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1255 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
1256 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1257 Arg = dyn_cast<Argument>(SExt->getOperand(0));
1259 Arg = SI->getOperand(0);
1260 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
1261 Arg = LI->getOperand(0);
1265 Instruction *DbgVal =
1266 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
1268 DbgVal->setDebugLoc(DVI->getDebugLoc());
1272 } // end anon namespace
1276 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
1278 /// This pass takes allocations which can be completely analyzed (that is, they
1279 /// don't escape) and tries to turn them into scalar SSA values. There are
1280 /// a few steps to this process.
1282 /// 1) It takes allocations of aggregates and analyzes the ways in which they
1283 /// are used to try to split them into smaller allocations, ideally of
1284 /// a single scalar data type. It will split up memcpy and memset accesses
1285 /// as necessary and try to isolate invidual scalar accesses.
1286 /// 2) It will transform accesses into forms which are suitable for SSA value
1287 /// promotion. This can be replacing a memset with a scalar store of an
1288 /// integer value, or it can involve speculating operations on a PHI or
1289 /// select to be a PHI or select of the results.
1290 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
1291 /// onto insert and extract operations on a vector value, and convert them to
1292 /// this form. By doing so, it will enable promotion of vector aggregates to
1293 /// SSA vector values.
1294 class SROA : public FunctionPass {
1295 const bool RequiresDomTree;
1298 const TargetData *TD;
1301 /// \brief Worklist of alloca instructions to simplify.
1303 /// Each alloca in the function is added to this. Each new alloca formed gets
1304 /// added to it as well to recursively simplify unless that alloca can be
1305 /// directly promoted. Finally, each time we rewrite a use of an alloca other
1306 /// the one being actively rewritten, we add it back onto the list if not
1307 /// already present to ensure it is re-visited.
1308 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
1310 /// \brief A collection of instructions to delete.
1311 /// We try to batch deletions to simplify code and make things a bit more
1313 SmallVector<Instruction *, 8> DeadInsts;
1315 /// \brief A set to prevent repeatedly marking an instruction split into many
1316 /// uses as dead. Only used to guard insertion into DeadInsts.
1317 SmallPtrSet<Instruction *, 4> DeadSplitInsts;
1319 /// \brief Post-promotion worklist.
1321 /// Sometimes we discover an alloca which has a high probability of becoming
1322 /// viable for SROA after a round of promotion takes place. In those cases,
1323 /// the alloca is enqueued here for re-processing.
1325 /// Note that we have to be very careful to clear allocas out of this list in
1326 /// the event they are deleted.
1327 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist;
1329 /// \brief A collection of alloca instructions we can directly promote.
1330 std::vector<AllocaInst *> PromotableAllocas;
1333 SROA(bool RequiresDomTree = true)
1334 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
1335 C(0), TD(0), DT(0) {
1336 initializeSROAPass(*PassRegistry::getPassRegistry());
1338 bool runOnFunction(Function &F);
1339 void getAnalysisUsage(AnalysisUsage &AU) const;
1341 const char *getPassName() const { return "SROA"; }
1345 friend class PHIOrSelectSpeculator;
1346 friend class AllocaPartitionRewriter;
1347 friend class AllocaPartitionVectorRewriter;
1349 bool rewriteAllocaPartition(AllocaInst &AI,
1350 AllocaPartitioning &P,
1351 AllocaPartitioning::iterator PI);
1352 bool splitAlloca(AllocaInst &AI, AllocaPartitioning &P);
1353 bool runOnAlloca(AllocaInst &AI);
1354 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
1355 bool promoteAllocas(Function &F);
1361 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
1362 return new SROA(RequiresDomTree);
1365 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
1367 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
1368 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
1372 /// \brief Visitor to speculate PHIs and Selects where possible.
1373 class PHIOrSelectSpeculator : public InstVisitor<PHIOrSelectSpeculator> {
1374 // Befriend the base class so it can delegate to private visit methods.
1375 friend class llvm::InstVisitor<PHIOrSelectSpeculator>;
1377 const TargetData &TD;
1378 AllocaPartitioning &P;
1382 PHIOrSelectSpeculator(const TargetData &TD, AllocaPartitioning &P, SROA &Pass)
1383 : TD(TD), P(P), Pass(Pass) {}
1385 /// \brief Visit the users of an alloca partition and rewrite them.
1386 void visitUsers(AllocaPartitioning::const_iterator PI) {
1387 // Note that we need to use an index here as the underlying vector of uses
1388 // may be grown during speculation. However, we never need to re-visit the
1389 // new uses, and so we can use the initial size bound.
1390 for (unsigned Idx = 0, Size = P.use_size(PI); Idx != Size; ++Idx) {
1391 const AllocaPartitioning::PartitionUse &PU = P.getUse(PI, Idx);
1393 continue; // Skip dead use.
1395 visit(cast<Instruction>(PU.U->getUser()));
1400 // By default, skip this instruction.
1401 void visitInstruction(Instruction &I) {}
1403 /// PHI instructions that use an alloca and are subsequently loaded can be
1404 /// rewritten to load both input pointers in the pred blocks and then PHI the
1405 /// results, allowing the load of the alloca to be promoted.
1407 /// %P2 = phi [i32* %Alloca, i32* %Other]
1408 /// %V = load i32* %P2
1410 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1412 /// %V2 = load i32* %Other
1414 /// %V = phi [i32 %V1, i32 %V2]
1416 /// We can do this to a select if its only uses are loads and if the operands
1417 /// to the select can be loaded unconditionally.
1419 /// FIXME: This should be hoisted into a generic utility, likely in
1420 /// Transforms/Util/Local.h
1421 bool isSafePHIToSpeculate(PHINode &PN, SmallVectorImpl<LoadInst *> &Loads) {
1422 // For now, we can only do this promotion if the load is in the same block
1423 // as the PHI, and if there are no stores between the phi and load.
1424 // TODO: Allow recursive phi users.
1425 // TODO: Allow stores.
1426 BasicBlock *BB = PN.getParent();
1427 unsigned MaxAlign = 0;
1428 for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end();
1430 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1431 if (LI == 0 || !LI->isSimple()) return false;
1433 // For now we only allow loads in the same block as the PHI. This is
1434 // a common case that happens when instcombine merges two loads through
1436 if (LI->getParent() != BB) return false;
1438 // Ensure that there are no instructions between the PHI and the load that
1440 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1441 if (BBI->mayWriteToMemory())
1444 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1445 Loads.push_back(LI);
1448 // We can only transform this if it is safe to push the loads into the
1449 // predecessor blocks. The only thing to watch out for is that we can't put
1450 // a possibly trapping load in the predecessor if it is a critical edge.
1451 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num;
1453 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1454 Value *InVal = PN.getIncomingValue(Idx);
1456 // If the value is produced by the terminator of the predecessor (an
1457 // invoke) or it has side-effects, there is no valid place to put a load
1458 // in the predecessor.
1459 if (TI == InVal || TI->mayHaveSideEffects())
1462 // If the predecessor has a single successor, then the edge isn't
1464 if (TI->getNumSuccessors() == 1)
1467 // If this pointer is always safe to load, or if we can prove that there
1468 // is already a load in the block, then we can move the load to the pred
1470 if (InVal->isDereferenceablePointer() ||
1471 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, &TD))
1480 void visitPHINode(PHINode &PN) {
1481 DEBUG(dbgs() << " original: " << PN << "\n");
1483 SmallVector<LoadInst *, 4> Loads;
1484 if (!isSafePHIToSpeculate(PN, Loads))
1487 assert(!Loads.empty());
1489 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1490 IRBuilder<> PHIBuilder(&PN);
1491 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1492 PN.getName() + ".sroa.speculated");
1494 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1495 // matter which one we get and if any differ, it doesn't matter.
1496 LoadInst *SomeLoad = cast<LoadInst>(Loads.back());
1497 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1498 unsigned Align = SomeLoad->getAlignment();
1500 // Rewrite all loads of the PN to use the new PHI.
1502 LoadInst *LI = Loads.pop_back_val();
1503 LI->replaceAllUsesWith(NewPN);
1504 Pass.DeadInsts.push_back(LI);
1505 } while (!Loads.empty());
1507 // Inject loads into all of the pred blocks.
1508 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1509 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1510 TerminatorInst *TI = Pred->getTerminator();
1511 Use *InUse = &PN.getOperandUse(PN.getOperandNumForIncomingValue(Idx));
1512 Value *InVal = PN.getIncomingValue(Idx);
1513 IRBuilder<> PredBuilder(TI);
1516 = PredBuilder.CreateLoad(InVal, (PN.getName() + ".sroa.speculate.load." +
1518 ++NumLoadsSpeculated;
1519 Load->setAlignment(Align);
1521 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1522 NewPN->addIncoming(Load, Pred);
1524 Instruction *Ptr = dyn_cast<Instruction>(InVal);
1526 // No uses to rewrite.
1529 // Try to lookup and rewrite any partition uses corresponding to this phi
1531 AllocaPartitioning::iterator PI
1532 = P.findPartitionForPHIOrSelectOperand(InUse);
1536 // Replace the Use in the PartitionUse for this operand with the Use
1538 AllocaPartitioning::use_iterator UI
1539 = P.findPartitionUseForPHIOrSelectOperand(InUse);
1540 assert(isa<PHINode>(*UI->U->getUser()));
1541 UI->U = &Load->getOperandUse(Load->getPointerOperandIndex());
1543 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1546 /// Select instructions that use an alloca and are subsequently loaded can be
1547 /// rewritten to load both input pointers and then select between the result,
1548 /// allowing the load of the alloca to be promoted.
1550 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1551 /// %V = load i32* %P2
1553 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1554 /// %V2 = load i32* %Other
1555 /// %V = select i1 %cond, i32 %V1, i32 %V2
1557 /// We can do this to a select if its only uses are loads and if the operand
1558 /// to the select can be loaded unconditionally.
1559 bool isSafeSelectToSpeculate(SelectInst &SI,
1560 SmallVectorImpl<LoadInst *> &Loads) {
1561 Value *TValue = SI.getTrueValue();
1562 Value *FValue = SI.getFalseValue();
1563 bool TDerefable = TValue->isDereferenceablePointer();
1564 bool FDerefable = FValue->isDereferenceablePointer();
1566 for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end();
1568 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1569 if (LI == 0 || !LI->isSimple()) return false;
1571 // Both operands to the select need to be dereferencable, either
1572 // absolutely (e.g. allocas) or at this point because we can see other
1574 if (!TDerefable && !isSafeToLoadUnconditionally(TValue, LI,
1575 LI->getAlignment(), &TD))
1577 if (!FDerefable && !isSafeToLoadUnconditionally(FValue, LI,
1578 LI->getAlignment(), &TD))
1580 Loads.push_back(LI);
1586 void visitSelectInst(SelectInst &SI) {
1587 DEBUG(dbgs() << " original: " << SI << "\n");
1588 IRBuilder<> IRB(&SI);
1590 // If the select isn't safe to speculate, just use simple logic to emit it.
1591 SmallVector<LoadInst *, 4> Loads;
1592 if (!isSafeSelectToSpeculate(SI, Loads))
1595 Use *Ops[2] = { &SI.getOperandUse(1), &SI.getOperandUse(2) };
1596 AllocaPartitioning::iterator PIs[2];
1597 AllocaPartitioning::PartitionUse PUs[2];
1598 for (unsigned i = 0, e = 2; i != e; ++i) {
1599 PIs[i] = P.findPartitionForPHIOrSelectOperand(Ops[i]);
1600 if (PIs[i] != P.end()) {
1601 // If the pointer is within the partitioning, remove the select from
1602 // its uses. We'll add in the new loads below.
1603 AllocaPartitioning::use_iterator UI
1604 = P.findPartitionUseForPHIOrSelectOperand(Ops[i]);
1606 // Clear out the use here so that the offsets into the use list remain
1607 // stable but this use is ignored when rewriting.
1612 Value *TV = SI.getTrueValue();
1613 Value *FV = SI.getFalseValue();
1614 // Replace the loads of the select with a select of two loads.
1615 while (!Loads.empty()) {
1616 LoadInst *LI = Loads.pop_back_val();
1618 IRB.SetInsertPoint(LI);
1620 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1622 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1623 NumLoadsSpeculated += 2;
1625 // Transfer alignment and TBAA info if present.
1626 TL->setAlignment(LI->getAlignment());
1627 FL->setAlignment(LI->getAlignment());
1628 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1629 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
1630 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
1633 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1634 LI->getName() + ".sroa.speculated");
1636 LoadInst *Loads[2] = { TL, FL };
1637 for (unsigned i = 0, e = 2; i != e; ++i) {
1638 if (PIs[i] != P.end()) {
1639 Use *LoadUse = &Loads[i]->getOperandUse(0);
1640 assert(PUs[i].U->get() == LoadUse->get());
1642 P.use_push_back(PIs[i], PUs[i]);
1646 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1647 LI->replaceAllUsesWith(V);
1648 Pass.DeadInsts.push_back(LI);
1654 /// \brief Accumulate the constant offsets in a GEP into a single APInt offset.
1656 /// If the provided GEP is all-constant, the total byte offset formed by the
1657 /// GEP is computed and Offset is set to it. If the GEP has any non-constant
1658 /// operands, the function returns false and the value of Offset is unmodified.
1659 static bool accumulateGEPOffsets(const TargetData &TD, GEPOperator &GEP,
1661 APInt GEPOffset(Offset.getBitWidth(), 0);
1662 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1663 GTI != GTE; ++GTI) {
1664 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
1667 if (OpC->isZero()) continue;
1669 // Handle a struct index, which adds its field offset to the pointer.
1670 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1671 unsigned ElementIdx = OpC->getZExtValue();
1672 const StructLayout *SL = TD.getStructLayout(STy);
1673 GEPOffset += APInt(Offset.getBitWidth(),
1674 SL->getElementOffset(ElementIdx));
1678 APInt TypeSize(Offset.getBitWidth(),
1679 TD.getTypeAllocSize(GTI.getIndexedType()));
1680 if (VectorType *VTy = dyn_cast<VectorType>(*GTI)) {
1681 assert((VTy->getScalarSizeInBits() % 8) == 0 &&
1682 "vector element size is not a multiple of 8, cannot GEP over it");
1683 TypeSize = VTy->getScalarSizeInBits() / 8;
1686 GEPOffset += OpC->getValue().sextOrTrunc(Offset.getBitWidth()) * TypeSize;
1692 /// \brief Build a GEP out of a base pointer and indices.
1694 /// This will return the BasePtr if that is valid, or build a new GEP
1695 /// instruction using the IRBuilder if GEP-ing is needed.
1696 static Value *buildGEP(IRBuilder<> &IRB, Value *BasePtr,
1697 SmallVectorImpl<Value *> &Indices,
1698 const Twine &Prefix) {
1699 if (Indices.empty())
1702 // A single zero index is a no-op, so check for this and avoid building a GEP
1704 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1707 return IRB.CreateInBoundsGEP(BasePtr, Indices, Prefix + ".idx");
1710 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1711 /// TargetTy without changing the offset of the pointer.
1713 /// This routine assumes we've already established a properly offset GEP with
1714 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1715 /// zero-indices down through type layers until we find one the same as
1716 /// TargetTy. If we can't find one with the same type, we at least try to use
1717 /// one with the same size. If none of that works, we just produce the GEP as
1718 /// indicated by Indices to have the correct offset.
1719 static Value *getNaturalGEPWithType(IRBuilder<> &IRB, const TargetData &TD,
1720 Value *BasePtr, Type *Ty, Type *TargetTy,
1721 SmallVectorImpl<Value *> &Indices,
1722 const Twine &Prefix) {
1724 return buildGEP(IRB, BasePtr, Indices, Prefix);
1726 // See if we can descend into a struct and locate a field with the correct
1728 unsigned NumLayers = 0;
1729 Type *ElementTy = Ty;
1731 if (ElementTy->isPointerTy())
1733 if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
1734 ElementTy = SeqTy->getElementType();
1735 Indices.push_back(IRB.getInt(APInt(TD.getPointerSizeInBits(), 0)));
1736 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1737 ElementTy = *STy->element_begin();
1738 Indices.push_back(IRB.getInt32(0));
1743 } while (ElementTy != TargetTy);
1744 if (ElementTy != TargetTy)
1745 Indices.erase(Indices.end() - NumLayers, Indices.end());
1747 return buildGEP(IRB, BasePtr, Indices, Prefix);
1750 /// \brief Recursively compute indices for a natural GEP.
1752 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1753 /// element types adding appropriate indices for the GEP.
1754 static Value *getNaturalGEPRecursively(IRBuilder<> &IRB, const TargetData &TD,
1755 Value *Ptr, Type *Ty, APInt &Offset,
1757 SmallVectorImpl<Value *> &Indices,
1758 const Twine &Prefix) {
1760 return getNaturalGEPWithType(IRB, TD, Ptr, Ty, TargetTy, Indices, Prefix);
1762 // We can't recurse through pointer types.
1763 if (Ty->isPointerTy())
1766 // We try to analyze GEPs over vectors here, but note that these GEPs are
1767 // extremely poorly defined currently. The long-term goal is to remove GEPing
1768 // over a vector from the IR completely.
1769 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1770 unsigned ElementSizeInBits = VecTy->getScalarSizeInBits();
1771 if (ElementSizeInBits % 8)
1772 return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
1773 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1774 APInt NumSkippedElements = Offset.udiv(ElementSize);
1775 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1777 Offset -= NumSkippedElements * ElementSize;
1778 Indices.push_back(IRB.getInt(NumSkippedElements));
1779 return getNaturalGEPRecursively(IRB, TD, Ptr, VecTy->getElementType(),
1780 Offset, TargetTy, Indices, Prefix);
1783 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1784 Type *ElementTy = ArrTy->getElementType();
1785 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1786 APInt NumSkippedElements = Offset.udiv(ElementSize);
1787 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1790 Offset -= NumSkippedElements * ElementSize;
1791 Indices.push_back(IRB.getInt(NumSkippedElements));
1792 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1796 StructType *STy = dyn_cast<StructType>(Ty);
1800 const StructLayout *SL = TD.getStructLayout(STy);
1801 uint64_t StructOffset = Offset.getZExtValue();
1802 if (StructOffset >= SL->getSizeInBytes())
1804 unsigned Index = SL->getElementContainingOffset(StructOffset);
1805 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1806 Type *ElementTy = STy->getElementType(Index);
1807 if (Offset.uge(TD.getTypeAllocSize(ElementTy)))
1808 return 0; // The offset points into alignment padding.
1810 Indices.push_back(IRB.getInt32(Index));
1811 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1815 /// \brief Get a natural GEP from a base pointer to a particular offset and
1816 /// resulting in a particular type.
1818 /// The goal is to produce a "natural" looking GEP that works with the existing
1819 /// composite types to arrive at the appropriate offset and element type for
1820 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1821 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1822 /// Indices, and setting Ty to the result subtype.
1824 /// If no natural GEP can be constructed, this function returns null.
1825 static Value *getNaturalGEPWithOffset(IRBuilder<> &IRB, const TargetData &TD,
1826 Value *Ptr, APInt Offset, Type *TargetTy,
1827 SmallVectorImpl<Value *> &Indices,
1828 const Twine &Prefix) {
1829 PointerType *Ty = cast<PointerType>(Ptr->getType());
1831 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1833 if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
1836 Type *ElementTy = Ty->getElementType();
1837 if (!ElementTy->isSized())
1838 return 0; // We can't GEP through an unsized element.
1839 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1840 if (ElementSize == 0)
1841 return 0; // Zero-length arrays can't help us build a natural GEP.
1842 APInt NumSkippedElements = Offset.udiv(ElementSize);
1844 Offset -= NumSkippedElements * ElementSize;
1845 Indices.push_back(IRB.getInt(NumSkippedElements));
1846 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1850 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1851 /// resulting pointer has PointerTy.
1853 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1854 /// and produces the pointer type desired. Where it cannot, it will try to use
1855 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1856 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1857 /// bitcast to the type.
1859 /// The strategy for finding the more natural GEPs is to peel off layers of the
1860 /// pointer, walking back through bit casts and GEPs, searching for a base
1861 /// pointer from which we can compute a natural GEP with the desired
1862 /// properities. The algorithm tries to fold as many constant indices into
1863 /// a single GEP as possible, thus making each GEP more independent of the
1864 /// surrounding code.
1865 static Value *getAdjustedPtr(IRBuilder<> &IRB, const TargetData &TD,
1866 Value *Ptr, APInt Offset, Type *PointerTy,
1867 const Twine &Prefix) {
1868 // Even though we don't look through PHI nodes, we could be called on an
1869 // instruction in an unreachable block, which may be on a cycle.
1870 SmallPtrSet<Value *, 4> Visited;
1871 Visited.insert(Ptr);
1872 SmallVector<Value *, 4> Indices;
1874 // We may end up computing an offset pointer that has the wrong type. If we
1875 // never are able to compute one directly that has the correct type, we'll
1876 // fall back to it, so keep it around here.
1877 Value *OffsetPtr = 0;
1879 // Remember any i8 pointer we come across to re-use if we need to do a raw
1882 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1884 Type *TargetTy = PointerTy->getPointerElementType();
1887 // First fold any existing GEPs into the offset.
1888 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1889 APInt GEPOffset(Offset.getBitWidth(), 0);
1890 if (!accumulateGEPOffsets(TD, *GEP, GEPOffset))
1892 Offset += GEPOffset;
1893 Ptr = GEP->getPointerOperand();
1894 if (!Visited.insert(Ptr))
1898 // See if we can perform a natural GEP here.
1900 if (Value *P = getNaturalGEPWithOffset(IRB, TD, Ptr, Offset, TargetTy,
1902 if (P->getType() == PointerTy) {
1903 // Zap any offset pointer that we ended up computing in previous rounds.
1904 if (OffsetPtr && OffsetPtr->use_empty())
1905 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1906 I->eraseFromParent();
1914 // Stash this pointer if we've found an i8*.
1915 if (Ptr->getType()->isIntegerTy(8)) {
1917 Int8PtrOffset = Offset;
1920 // Peel off a layer of the pointer and update the offset appropriately.
1921 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1922 Ptr = cast<Operator>(Ptr)->getOperand(0);
1923 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1924 if (GA->mayBeOverridden())
1926 Ptr = GA->getAliasee();
1930 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1931 } while (Visited.insert(Ptr));
1935 Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
1936 Prefix + ".raw_cast");
1937 Int8PtrOffset = Offset;
1940 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1941 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1942 Prefix + ".raw_idx");
1946 // On the off chance we were targeting i8*, guard the bitcast here.
1947 if (Ptr->getType() != PointerTy)
1948 Ptr = IRB.CreateBitCast(Ptr, PointerTy, Prefix + ".cast");
1953 /// \brief Test whether the given alloca partition can be promoted to a vector.
1955 /// This is a quick test to check whether we can rewrite a particular alloca
1956 /// partition (and its newly formed alloca) into a vector alloca with only
1957 /// whole-vector loads and stores such that it could be promoted to a vector
1958 /// SSA value. We only can ensure this for a limited set of operations, and we
1959 /// don't want to do the rewrites unless we are confident that the result will
1960 /// be promotable, so we have an early test here.
1961 static bool isVectorPromotionViable(const TargetData &TD,
1963 AllocaPartitioning &P,
1964 uint64_t PartitionBeginOffset,
1965 uint64_t PartitionEndOffset,
1966 AllocaPartitioning::const_use_iterator I,
1967 AllocaPartitioning::const_use_iterator E) {
1968 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
1972 uint64_t VecSize = TD.getTypeSizeInBits(Ty);
1973 uint64_t ElementSize = Ty->getScalarSizeInBits();
1975 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1976 // that aren't byte sized.
1977 if (ElementSize % 8)
1979 assert((VecSize % 8) == 0 && "vector size not a multiple of element size?");
1983 for (; I != E; ++I) {
1985 continue; // Skip dead use.
1987 uint64_t BeginOffset = I->BeginOffset - PartitionBeginOffset;
1988 uint64_t BeginIndex = BeginOffset / ElementSize;
1989 if (BeginIndex * ElementSize != BeginOffset ||
1990 BeginIndex >= Ty->getNumElements())
1992 uint64_t EndOffset = I->EndOffset - PartitionBeginOffset;
1993 uint64_t EndIndex = EndOffset / ElementSize;
1994 if (EndIndex * ElementSize != EndOffset ||
1995 EndIndex > Ty->getNumElements())
1998 // FIXME: We should build shuffle vector instructions to handle
1999 // non-element-sized accesses.
2000 if ((EndOffset - BeginOffset) != ElementSize &&
2001 (EndOffset - BeginOffset) != VecSize)
2004 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
2005 if (MI->isVolatile())
2007 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
2008 const AllocaPartitioning::MemTransferOffsets &MTO
2009 = P.getMemTransferOffsets(*MTI);
2010 if (!MTO.IsSplittable)
2013 } else if (I->U->get()->getType()->getPointerElementType()->isStructTy()) {
2014 // Disable vector promotion when there are loads or stores of an FCA.
2016 } else if (!isa<LoadInst>(I->U->getUser()) &&
2017 !isa<StoreInst>(I->U->getUser())) {
2024 /// \brief Test whether the given alloca partition can be promoted to an int.
2026 /// This is a quick test to check whether we can rewrite a particular alloca
2027 /// partition (and its newly formed alloca) into an integer alloca suitable for
2028 /// promotion to an SSA value. We only can ensure this for a limited set of
2029 /// operations, and we don't want to do the rewrites unless we are confident
2030 /// that the result will be promotable, so we have an early test here.
2031 static bool isIntegerPromotionViable(const TargetData &TD,
2033 uint64_t AllocBeginOffset,
2034 AllocaPartitioning &P,
2035 AllocaPartitioning::const_use_iterator I,
2036 AllocaPartitioning::const_use_iterator E) {
2037 IntegerType *Ty = dyn_cast<IntegerType>(AllocaTy);
2038 if (!Ty || 8*TD.getTypeStoreSize(Ty) != Ty->getBitWidth())
2041 // Check the uses to ensure the uses are (likely) promoteable integer uses.
2042 // Also ensure that the alloca has a covering load or store. We don't want
2043 // promote because of some other unsplittable entry (which we may make
2044 // splittable later) and lose the ability to promote each element access.
2045 bool WholeAllocaOp = false;
2046 for (; I != E; ++I) {
2048 continue; // Skip dead use.
2050 // We can't reasonably handle cases where the load or store extends past
2051 // the end of the aloca's type and into its padding.
2052 if ((I->EndOffset - AllocBeginOffset) > TD.getTypeStoreSize(Ty))
2055 if (LoadInst *LI = dyn_cast<LoadInst>(I->U->getUser())) {
2056 if (LI->isVolatile() || !LI->getType()->isIntegerTy())
2058 if (LI->getType() == Ty)
2059 WholeAllocaOp = true;
2060 } else if (StoreInst *SI = dyn_cast<StoreInst>(I->U->getUser())) {
2061 if (SI->isVolatile() || !SI->getValueOperand()->getType()->isIntegerTy())
2063 if (SI->getValueOperand()->getType() == Ty)
2064 WholeAllocaOp = true;
2065 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
2066 if (MI->isVolatile())
2068 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
2069 const AllocaPartitioning::MemTransferOffsets &MTO
2070 = P.getMemTransferOffsets(*MTI);
2071 if (!MTO.IsSplittable)
2078 return WholeAllocaOp;
2082 /// \brief Visitor to rewrite instructions using a partition of an alloca to
2083 /// use a new alloca.
2085 /// Also implements the rewriting to vector-based accesses when the partition
2086 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2088 class AllocaPartitionRewriter : public InstVisitor<AllocaPartitionRewriter,
2090 // Befriend the base class so it can delegate to private visit methods.
2091 friend class llvm::InstVisitor<AllocaPartitionRewriter, bool>;
2093 const TargetData &TD;
2094 AllocaPartitioning &P;
2096 AllocaInst &OldAI, &NewAI;
2097 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2099 // If we are rewriting an alloca partition which can be written as pure
2100 // vector operations, we stash extra information here. When VecTy is
2101 // non-null, we have some strict guarantees about the rewriten alloca:
2102 // - The new alloca is exactly the size of the vector type here.
2103 // - The accesses all either map to the entire vector or to a single
2105 // - The set of accessing instructions is only one of those handled above
2106 // in isVectorPromotionViable. Generally these are the same access kinds
2107 // which are promotable via mem2reg.
2110 uint64_t ElementSize;
2112 // This is a convenience and flag variable that will be null unless the new
2113 // alloca has a promotion-targeted integer type due to passing
2114 // isIntegerPromotionViable above. If it is non-null does, the desired
2115 // integer type will be stored here for easy access during rewriting.
2116 IntegerType *IntPromotionTy;
2118 // The offset of the partition user currently being rewritten.
2119 uint64_t BeginOffset, EndOffset;
2121 Instruction *OldPtr;
2123 // The name prefix to use when rewriting instructions for this alloca.
2124 std::string NamePrefix;
2127 AllocaPartitionRewriter(const TargetData &TD, AllocaPartitioning &P,
2128 AllocaPartitioning::iterator PI,
2129 SROA &Pass, AllocaInst &OldAI, AllocaInst &NewAI,
2130 uint64_t NewBeginOffset, uint64_t NewEndOffset)
2131 : TD(TD), P(P), Pass(Pass),
2132 OldAI(OldAI), NewAI(NewAI),
2133 NewAllocaBeginOffset(NewBeginOffset),
2134 NewAllocaEndOffset(NewEndOffset),
2135 VecTy(), ElementTy(), ElementSize(), IntPromotionTy(),
2136 BeginOffset(), EndOffset() {
2139 /// \brief Visit the users of the alloca partition and rewrite them.
2140 bool visitUsers(AllocaPartitioning::const_use_iterator I,
2141 AllocaPartitioning::const_use_iterator E) {
2142 if (isVectorPromotionViable(TD, NewAI.getAllocatedType(), P,
2143 NewAllocaBeginOffset, NewAllocaEndOffset,
2146 VecTy = cast<VectorType>(NewAI.getAllocatedType());
2147 ElementTy = VecTy->getElementType();
2148 assert((VecTy->getScalarSizeInBits() % 8) == 0 &&
2149 "Only multiple-of-8 sized vector elements are viable");
2150 ElementSize = VecTy->getScalarSizeInBits() / 8;
2151 } else if (isIntegerPromotionViable(TD, NewAI.getAllocatedType(),
2152 NewAllocaBeginOffset, P, I, E)) {
2153 IntPromotionTy = cast<IntegerType>(NewAI.getAllocatedType());
2155 bool CanSROA = true;
2156 for (; I != E; ++I) {
2158 continue; // Skip dead uses.
2159 BeginOffset = I->BeginOffset;
2160 EndOffset = I->EndOffset;
2162 OldPtr = cast<Instruction>(I->U->get());
2163 NamePrefix = (Twine(NewAI.getName()) + "." + Twine(BeginOffset)).str();
2164 CanSROA &= visit(cast<Instruction>(I->U->getUser()));
2176 // Every instruction which can end up as a user must have a rewrite rule.
2177 bool visitInstruction(Instruction &I) {
2178 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2179 llvm_unreachable("No rewrite rule for this instruction!");
2182 Twine getName(const Twine &Suffix) {
2183 return NamePrefix + Suffix;
2186 Value *getAdjustedAllocaPtr(IRBuilder<> &IRB, Type *PointerTy) {
2187 assert(BeginOffset >= NewAllocaBeginOffset);
2188 APInt Offset(TD.getPointerSizeInBits(), BeginOffset - NewAllocaBeginOffset);
2189 return getAdjustedPtr(IRB, TD, &NewAI, Offset, PointerTy, getName(""));
2192 /// \brief Compute suitable alignment to access an offset into the new alloca.
2193 unsigned getOffsetAlign(uint64_t Offset) {
2194 unsigned NewAIAlign = NewAI.getAlignment();
2196 NewAIAlign = TD.getABITypeAlignment(NewAI.getAllocatedType());
2197 return MinAlign(NewAIAlign, Offset);
2200 /// \brief Compute suitable alignment to access this partition of the new
2202 unsigned getPartitionAlign() {
2203 return getOffsetAlign(BeginOffset - NewAllocaBeginOffset);
2206 /// \brief Compute suitable alignment to access a type at an offset of the
2209 /// \returns zero if the type's ABI alignment is a suitable alignment,
2210 /// otherwise returns the maximal suitable alignment.
2211 unsigned getOffsetTypeAlign(Type *Ty, uint64_t Offset) {
2212 unsigned Align = getOffsetAlign(Offset);
2213 return Align == TD.getABITypeAlignment(Ty) ? 0 : Align;
2216 /// \brief Compute suitable alignment to access a type at the beginning of
2217 /// this partition of the new alloca.
2219 /// See \c getOffsetTypeAlign for details; this routine delegates to it.
2220 unsigned getPartitionTypeAlign(Type *Ty) {
2221 return getOffsetTypeAlign(Ty, BeginOffset - NewAllocaBeginOffset);
2224 ConstantInt *getIndex(IRBuilder<> &IRB, uint64_t Offset) {
2225 assert(VecTy && "Can only call getIndex when rewriting a vector");
2226 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2227 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2228 uint32_t Index = RelOffset / ElementSize;
2229 assert(Index * ElementSize == RelOffset);
2230 return IRB.getInt32(Index);
2233 Value *extractInteger(IRBuilder<> &IRB, IntegerType *TargetTy,
2235 assert(IntPromotionTy && "Alloca is not an integer we can extract from");
2236 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2238 assert(Offset >= NewAllocaBeginOffset && "Out of bounds offset");
2239 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2240 assert(TD.getTypeStoreSize(TargetTy) + RelOffset <=
2241 TD.getTypeStoreSize(IntPromotionTy) &&
2242 "Element load outside of alloca store");
2243 uint64_t ShAmt = 8*RelOffset;
2244 if (TD.isBigEndian())
2245 ShAmt = 8*(TD.getTypeStoreSize(IntPromotionTy) -
2246 TD.getTypeStoreSize(TargetTy) - RelOffset);
2248 V = IRB.CreateLShr(V, ShAmt, getName(".shift"));
2249 if (TargetTy != IntPromotionTy) {
2250 assert(TargetTy->getBitWidth() < IntPromotionTy->getBitWidth() &&
2251 "Cannot extract to a larger integer!");
2252 V = IRB.CreateTrunc(V, TargetTy, getName(".trunc"));
2257 StoreInst *insertInteger(IRBuilder<> &IRB, Value *V, uint64_t Offset) {
2258 IntegerType *Ty = cast<IntegerType>(V->getType());
2259 if (Ty == IntPromotionTy)
2260 return IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2262 assert(Ty->getBitWidth() < IntPromotionTy->getBitWidth() &&
2263 "Cannot insert a larger integer!");
2264 V = IRB.CreateZExt(V, IntPromotionTy, getName(".ext"));
2265 assert(Offset >= NewAllocaBeginOffset && "Out of bounds offset");
2266 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2267 assert(TD.getTypeStoreSize(Ty) + RelOffset <=
2268 TD.getTypeStoreSize(IntPromotionTy) &&
2269 "Element store outside of alloca store");
2270 uint64_t ShAmt = 8*RelOffset;
2271 if (TD.isBigEndian())
2272 ShAmt = 8*(TD.getTypeStoreSize(IntPromotionTy) - TD.getTypeStoreSize(Ty)
2275 V = IRB.CreateShl(V, ShAmt, getName(".shift"));
2277 APInt Mask = ~Ty->getMask().zext(IntPromotionTy->getBitWidth()).shl(ShAmt);
2278 Value *Old = IRB.CreateAnd(IRB.CreateAlignedLoad(&NewAI,
2279 NewAI.getAlignment(),
2280 getName(".oldload")),
2281 Mask, getName(".mask"));
2282 return IRB.CreateAlignedStore(IRB.CreateOr(Old, V, getName(".insert")),
2283 &NewAI, NewAI.getAlignment());
2286 void deleteIfTriviallyDead(Value *V) {
2287 Instruction *I = cast<Instruction>(V);
2288 if (isInstructionTriviallyDead(I))
2289 Pass.DeadInsts.push_back(I);
2292 Value *getValueCast(IRBuilder<> &IRB, Value *V, Type *Ty) {
2293 if (V->getType()->isIntegerTy() && Ty->isPointerTy())
2294 return IRB.CreateIntToPtr(V, Ty);
2295 if (V->getType()->isPointerTy() && Ty->isIntegerTy())
2296 return IRB.CreatePtrToInt(V, Ty);
2298 return IRB.CreateBitCast(V, Ty);
2301 bool rewriteVectorizedLoadInst(IRBuilder<> &IRB, LoadInst &LI, Value *OldOp) {
2303 if (LI.getType() == VecTy->getElementType() ||
2304 BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
2305 Result = IRB.CreateExtractElement(
2306 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), getName(".load")),
2307 getIndex(IRB, BeginOffset), getName(".extract"));
2309 Result = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2312 if (Result->getType() != LI.getType())
2313 Result = getValueCast(IRB, Result, LI.getType());
2314 LI.replaceAllUsesWith(Result);
2315 Pass.DeadInsts.push_back(&LI);
2317 DEBUG(dbgs() << " to: " << *Result << "\n");
2321 bool rewriteIntegerLoad(IRBuilder<> &IRB, LoadInst &LI) {
2322 assert(!LI.isVolatile());
2323 Value *Result = extractInteger(IRB, cast<IntegerType>(LI.getType()),
2325 LI.replaceAllUsesWith(Result);
2326 Pass.DeadInsts.push_back(&LI);
2327 DEBUG(dbgs() << " to: " << *Result << "\n");
2331 bool visitLoadInst(LoadInst &LI) {
2332 DEBUG(dbgs() << " original: " << LI << "\n");
2333 Value *OldOp = LI.getOperand(0);
2334 assert(OldOp == OldPtr);
2335 IRBuilder<> IRB(&LI);
2338 return rewriteVectorizedLoadInst(IRB, LI, OldOp);
2340 return rewriteIntegerLoad(IRB, LI);
2342 Value *NewPtr = getAdjustedAllocaPtr(IRB,
2343 LI.getPointerOperand()->getType());
2344 LI.setOperand(0, NewPtr);
2345 LI.setAlignment(getPartitionTypeAlign(LI.getType()));
2346 DEBUG(dbgs() << " to: " << LI << "\n");
2348 deleteIfTriviallyDead(OldOp);
2349 return NewPtr == &NewAI && !LI.isVolatile();
2352 bool rewriteVectorizedStoreInst(IRBuilder<> &IRB, StoreInst &SI,
2354 Value *V = SI.getValueOperand();
2355 if (V->getType() == ElementTy ||
2356 BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
2357 if (V->getType() != ElementTy)
2358 V = getValueCast(IRB, V, ElementTy);
2359 LoadInst *LI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2361 V = IRB.CreateInsertElement(LI, V, getIndex(IRB, BeginOffset),
2362 getName(".insert"));
2363 } else if (V->getType() != VecTy) {
2364 V = getValueCast(IRB, V, VecTy);
2366 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2367 Pass.DeadInsts.push_back(&SI);
2370 DEBUG(dbgs() << " to: " << *Store << "\n");
2374 bool rewriteIntegerStore(IRBuilder<> &IRB, StoreInst &SI) {
2375 assert(!SI.isVolatile());
2376 StoreInst *Store = insertInteger(IRB, SI.getValueOperand(), BeginOffset);
2377 Pass.DeadInsts.push_back(&SI);
2379 DEBUG(dbgs() << " to: " << *Store << "\n");
2383 bool visitStoreInst(StoreInst &SI) {
2384 DEBUG(dbgs() << " original: " << SI << "\n");
2385 Value *OldOp = SI.getOperand(1);
2386 assert(OldOp == OldPtr);
2387 IRBuilder<> IRB(&SI);
2390 return rewriteVectorizedStoreInst(IRB, SI, OldOp);
2392 return rewriteIntegerStore(IRB, SI);
2394 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2395 // alloca that should be re-examined after promoting this alloca.
2396 if (SI.getValueOperand()->getType()->isPointerTy())
2397 if (AllocaInst *AI = dyn_cast<AllocaInst>(SI.getValueOperand()
2398 ->stripInBoundsOffsets()))
2399 Pass.PostPromotionWorklist.insert(AI);
2401 Value *NewPtr = getAdjustedAllocaPtr(IRB,
2402 SI.getPointerOperand()->getType());
2403 SI.setOperand(1, NewPtr);
2404 SI.setAlignment(getPartitionTypeAlign(SI.getValueOperand()->getType()));
2405 DEBUG(dbgs() << " to: " << SI << "\n");
2407 deleteIfTriviallyDead(OldOp);
2408 return NewPtr == &NewAI && !SI.isVolatile();
2411 bool visitMemSetInst(MemSetInst &II) {
2412 DEBUG(dbgs() << " original: " << II << "\n");
2413 IRBuilder<> IRB(&II);
2414 assert(II.getRawDest() == OldPtr);
2416 // If the memset has a variable size, it cannot be split, just adjust the
2417 // pointer to the new alloca.
2418 if (!isa<Constant>(II.getLength())) {
2419 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2420 Type *CstTy = II.getAlignmentCst()->getType();
2421 II.setAlignment(ConstantInt::get(CstTy, getPartitionAlign()));
2423 deleteIfTriviallyDead(OldPtr);
2427 // Record this instruction for deletion.
2428 if (Pass.DeadSplitInsts.insert(&II))
2429 Pass.DeadInsts.push_back(&II);
2431 Type *AllocaTy = NewAI.getAllocatedType();
2432 Type *ScalarTy = AllocaTy->getScalarType();
2434 // If this doesn't map cleanly onto the alloca type, and that type isn't
2435 // a single value type, just emit a memset.
2436 if (!VecTy && (BeginOffset != NewAllocaBeginOffset ||
2437 EndOffset != NewAllocaEndOffset ||
2438 !AllocaTy->isSingleValueType() ||
2439 !TD.isLegalInteger(TD.getTypeSizeInBits(ScalarTy)))) {
2440 Type *SizeTy = II.getLength()->getType();
2441 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2443 = IRB.CreateMemSet(getAdjustedAllocaPtr(IRB,
2444 II.getRawDest()->getType()),
2445 II.getValue(), Size, getPartitionAlign(),
2448 DEBUG(dbgs() << " to: " << *New << "\n");
2452 // If we can represent this as a simple value, we have to build the actual
2453 // value to store, which requires expanding the byte present in memset to
2454 // a sensible representation for the alloca type. This is essentially
2455 // splatting the byte to a sufficiently wide integer, bitcasting to the
2456 // desired scalar type, and splatting it across any desired vector type.
2457 Value *V = II.getValue();
2458 IntegerType *VTy = cast<IntegerType>(V->getType());
2459 Type *IntTy = Type::getIntNTy(VTy->getContext(),
2460 TD.getTypeSizeInBits(ScalarTy));
2461 if (TD.getTypeSizeInBits(ScalarTy) > VTy->getBitWidth())
2462 V = IRB.CreateMul(IRB.CreateZExt(V, IntTy, getName(".zext")),
2463 ConstantExpr::getUDiv(
2464 Constant::getAllOnesValue(IntTy),
2465 ConstantExpr::getZExt(
2466 Constant::getAllOnesValue(V->getType()),
2468 getName(".isplat"));
2469 if (V->getType() != ScalarTy) {
2470 if (ScalarTy->isPointerTy())
2471 V = IRB.CreateIntToPtr(V, ScalarTy);
2472 else if (ScalarTy->isPrimitiveType() || ScalarTy->isVectorTy())
2473 V = IRB.CreateBitCast(V, ScalarTy);
2474 else if (ScalarTy->isIntegerTy())
2475 llvm_unreachable("Computed different integer types with equal widths");
2477 llvm_unreachable("Invalid scalar type");
2480 // If this is an element-wide memset of a vectorizable alloca, insert it.
2481 if (VecTy && (BeginOffset > NewAllocaBeginOffset ||
2482 EndOffset < NewAllocaEndOffset)) {
2483 StoreInst *Store = IRB.CreateAlignedStore(
2484 IRB.CreateInsertElement(IRB.CreateAlignedLoad(&NewAI,
2485 NewAI.getAlignment(),
2487 V, getIndex(IRB, BeginOffset),
2488 getName(".insert")),
2489 &NewAI, NewAI.getAlignment());
2491 DEBUG(dbgs() << " to: " << *Store << "\n");
2495 // Splat to a vector if needed.
2496 if (VectorType *VecTy = dyn_cast<VectorType>(AllocaTy)) {
2497 VectorType *SplatSourceTy = VectorType::get(V->getType(), 1);
2498 V = IRB.CreateShuffleVector(
2499 IRB.CreateInsertElement(UndefValue::get(SplatSourceTy), V,
2500 IRB.getInt32(0), getName(".vsplat.insert")),
2501 UndefValue::get(SplatSourceTy),
2502 ConstantVector::getSplat(VecTy->getNumElements(), IRB.getInt32(0)),
2503 getName(".vsplat.shuffle"));
2504 assert(V->getType() == VecTy);
2507 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2510 DEBUG(dbgs() << " to: " << *New << "\n");
2511 return !II.isVolatile();
2514 bool visitMemTransferInst(MemTransferInst &II) {
2515 // Rewriting of memory transfer instructions can be a bit tricky. We break
2516 // them into two categories: split intrinsics and unsplit intrinsics.
2518 DEBUG(dbgs() << " original: " << II << "\n");
2519 IRBuilder<> IRB(&II);
2521 assert(II.getRawSource() == OldPtr || II.getRawDest() == OldPtr);
2522 bool IsDest = II.getRawDest() == OldPtr;
2524 const AllocaPartitioning::MemTransferOffsets &MTO
2525 = P.getMemTransferOffsets(II);
2527 // Compute the relative offset within the transfer.
2528 unsigned IntPtrWidth = TD.getPointerSizeInBits();
2529 APInt RelOffset(IntPtrWidth, BeginOffset - (IsDest ? MTO.DestBegin
2530 : MTO.SourceBegin));
2532 unsigned Align = II.getAlignment();
2534 Align = MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(),
2535 MinAlign(II.getAlignment(), getPartitionAlign()));
2537 // For unsplit intrinsics, we simply modify the source and destination
2538 // pointers in place. This isn't just an optimization, it is a matter of
2539 // correctness. With unsplit intrinsics we may be dealing with transfers
2540 // within a single alloca before SROA ran, or with transfers that have
2541 // a variable length. We may also be dealing with memmove instead of
2542 // memcpy, and so simply updating the pointers is the necessary for us to
2543 // update both source and dest of a single call.
2544 if (!MTO.IsSplittable) {
2545 Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource();
2547 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2549 II.setSource(getAdjustedAllocaPtr(IRB, II.getRawSource()->getType()));
2551 Type *CstTy = II.getAlignmentCst()->getType();
2552 II.setAlignment(ConstantInt::get(CstTy, Align));
2554 DEBUG(dbgs() << " to: " << II << "\n");
2555 deleteIfTriviallyDead(OldOp);
2558 // For split transfer intrinsics we have an incredibly useful assurance:
2559 // the source and destination do not reside within the same alloca, and at
2560 // least one of them does not escape. This means that we can replace
2561 // memmove with memcpy, and we don't need to worry about all manner of
2562 // downsides to splitting and transforming the operations.
2564 // If this doesn't map cleanly onto the alloca type, and that type isn't
2565 // a single value type, just emit a memcpy.
2567 = !VecTy && (BeginOffset != NewAllocaBeginOffset ||
2568 EndOffset != NewAllocaEndOffset ||
2569 !NewAI.getAllocatedType()->isSingleValueType());
2571 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2572 // size hasn't been shrunk based on analysis of the viable range, this is
2574 if (EmitMemCpy && &OldAI == &NewAI) {
2575 uint64_t OrigBegin = IsDest ? MTO.DestBegin : MTO.SourceBegin;
2576 uint64_t OrigEnd = IsDest ? MTO.DestEnd : MTO.SourceEnd;
2577 // Ensure the start lines up.
2578 assert(BeginOffset == OrigBegin);
2581 // Rewrite the size as needed.
2582 if (EndOffset != OrigEnd)
2583 II.setLength(ConstantInt::get(II.getLength()->getType(),
2584 EndOffset - BeginOffset));
2587 // Record this instruction for deletion.
2588 if (Pass.DeadSplitInsts.insert(&II))
2589 Pass.DeadInsts.push_back(&II);
2591 bool IsVectorElement = VecTy && (BeginOffset > NewAllocaBeginOffset ||
2592 EndOffset < NewAllocaEndOffset);
2594 Type *OtherPtrTy = IsDest ? II.getRawSource()->getType()
2595 : II.getRawDest()->getType();
2597 OtherPtrTy = IsVectorElement ? VecTy->getElementType()->getPointerTo()
2600 // Compute the other pointer, folding as much as possible to produce
2601 // a single, simple GEP in most cases.
2602 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2603 OtherPtr = getAdjustedPtr(IRB, TD, OtherPtr, RelOffset, OtherPtrTy,
2604 getName("." + OtherPtr->getName()));
2606 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2607 // alloca that should be re-examined after rewriting this instruction.
2609 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets()))
2610 Pass.Worklist.insert(AI);
2614 = getAdjustedAllocaPtr(IRB, IsDest ? II.getRawDest()->getType()
2615 : II.getRawSource()->getType());
2616 Type *SizeTy = II.getLength()->getType();
2617 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2619 CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
2620 IsDest ? OtherPtr : OurPtr,
2621 Size, Align, II.isVolatile());
2623 DEBUG(dbgs() << " to: " << *New << "\n");
2627 // Note that we clamp the alignment to 1 here as a 0 alignment for a memcpy
2628 // is equivalent to 1, but that isn't true if we end up rewriting this as
2633 Value *SrcPtr = OtherPtr;
2634 Value *DstPtr = &NewAI;
2636 std::swap(SrcPtr, DstPtr);
2639 if (IsVectorElement && !IsDest) {
2640 // We have to extract rather than load.
2641 Src = IRB.CreateExtractElement(
2642 IRB.CreateAlignedLoad(SrcPtr, Align, getName(".copyload")),
2643 getIndex(IRB, BeginOffset),
2644 getName(".copyextract"));
2646 Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(),
2647 getName(".copyload"));
2650 if (IsVectorElement && IsDest) {
2651 // We have to insert into a loaded copy before storing.
2652 Src = IRB.CreateInsertElement(
2653 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), getName(".load")),
2654 Src, getIndex(IRB, BeginOffset),
2655 getName(".insert"));
2658 StoreInst *Store = cast<StoreInst>(
2659 IRB.CreateAlignedStore(Src, DstPtr, Align, II.isVolatile()));
2661 DEBUG(dbgs() << " to: " << *Store << "\n");
2662 return !II.isVolatile();
2665 bool visitIntrinsicInst(IntrinsicInst &II) {
2666 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2667 II.getIntrinsicID() == Intrinsic::lifetime_end);
2668 DEBUG(dbgs() << " original: " << II << "\n");
2669 IRBuilder<> IRB(&II);
2670 assert(II.getArgOperand(1) == OldPtr);
2672 // Record this instruction for deletion.
2673 if (Pass.DeadSplitInsts.insert(&II))
2674 Pass.DeadInsts.push_back(&II);
2677 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2678 EndOffset - BeginOffset);
2679 Value *Ptr = getAdjustedAllocaPtr(IRB, II.getArgOperand(1)->getType());
2681 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2682 New = IRB.CreateLifetimeStart(Ptr, Size);
2684 New = IRB.CreateLifetimeEnd(Ptr, Size);
2686 DEBUG(dbgs() << " to: " << *New << "\n");
2690 bool visitPHINode(PHINode &PN) {
2691 DEBUG(dbgs() << " original: " << PN << "\n");
2693 // We would like to compute a new pointer in only one place, but have it be
2694 // as local as possible to the PHI. To do that, we re-use the location of
2695 // the old pointer, which necessarily must be in the right position to
2696 // dominate the PHI.
2697 IRBuilder<> PtrBuilder(cast<Instruction>(OldPtr));
2699 Value *NewPtr = getAdjustedAllocaPtr(PtrBuilder, OldPtr->getType());
2700 // Replace the operands which were using the old pointer.
2701 User::op_iterator OI = PN.op_begin(), OE = PN.op_end();
2702 for (; OI != OE; ++OI)
2706 DEBUG(dbgs() << " to: " << PN << "\n");
2707 deleteIfTriviallyDead(OldPtr);
2711 bool visitSelectInst(SelectInst &SI) {
2712 DEBUG(dbgs() << " original: " << SI << "\n");
2713 IRBuilder<> IRB(&SI);
2715 // Find the operand we need to rewrite here.
2716 bool IsTrueVal = SI.getTrueValue() == OldPtr;
2718 assert(SI.getFalseValue() != OldPtr && "Pointer is both operands!");
2720 assert(SI.getFalseValue() == OldPtr && "Pointer isn't an operand!");
2722 Value *NewPtr = getAdjustedAllocaPtr(IRB, OldPtr->getType());
2723 SI.setOperand(IsTrueVal ? 1 : 2, NewPtr);
2724 DEBUG(dbgs() << " to: " << SI << "\n");
2725 deleteIfTriviallyDead(OldPtr);
2733 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2735 /// This pass aggressively rewrites all aggregate loads and stores on
2736 /// a particular pointer (or any pointer derived from it which we can identify)
2737 /// with scalar loads and stores.
2738 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2739 // Befriend the base class so it can delegate to private visit methods.
2740 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2742 const TargetData &TD;
2744 /// Queue of pointer uses to analyze and potentially rewrite.
2745 SmallVector<Use *, 8> Queue;
2747 /// Set to prevent us from cycling with phi nodes and loops.
2748 SmallPtrSet<User *, 8> Visited;
2750 /// The current pointer use being rewritten. This is used to dig up the used
2751 /// value (as opposed to the user).
2755 AggLoadStoreRewriter(const TargetData &TD) : TD(TD) {}
2757 /// Rewrite loads and stores through a pointer and all pointers derived from
2759 bool rewrite(Instruction &I) {
2760 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2762 bool Changed = false;
2763 while (!Queue.empty()) {
2764 U = Queue.pop_back_val();
2765 Changed |= visit(cast<Instruction>(U->getUser()));
2771 /// Enqueue all the users of the given instruction for further processing.
2772 /// This uses a set to de-duplicate users.
2773 void enqueueUsers(Instruction &I) {
2774 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
2776 if (Visited.insert(*UI))
2777 Queue.push_back(&UI.getUse());
2780 // Conservative default is to not rewrite anything.
2781 bool visitInstruction(Instruction &I) { return false; }
2783 /// \brief Generic recursive split emission class.
2784 template <typename Derived>
2787 /// The builder used to form new instructions.
2789 /// The indices which to be used with insert- or extractvalue to select the
2790 /// appropriate value within the aggregate.
2791 SmallVector<unsigned, 4> Indices;
2792 /// The indices to a GEP instruction which will move Ptr to the correct slot
2793 /// within the aggregate.
2794 SmallVector<Value *, 4> GEPIndices;
2795 /// The base pointer of the original op, used as a base for GEPing the
2796 /// split operations.
2799 /// Initialize the splitter with an insertion point, Ptr and start with a
2800 /// single zero GEP index.
2801 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
2802 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
2805 /// \brief Generic recursive split emission routine.
2807 /// This method recursively splits an aggregate op (load or store) into
2808 /// scalar or vector ops. It splits recursively until it hits a single value
2809 /// and emits that single value operation via the template argument.
2811 /// The logic of this routine relies on GEPs and insertvalue and
2812 /// extractvalue all operating with the same fundamental index list, merely
2813 /// formatted differently (GEPs need actual values).
2815 /// \param Ty The type being split recursively into smaller ops.
2816 /// \param Agg The aggregate value being built up or stored, depending on
2817 /// whether this is splitting a load or a store respectively.
2818 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
2819 if (Ty->isSingleValueType())
2820 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
2822 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2823 unsigned OldSize = Indices.size();
2825 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
2827 assert(Indices.size() == OldSize && "Did not return to the old size");
2828 Indices.push_back(Idx);
2829 GEPIndices.push_back(IRB.getInt32(Idx));
2830 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
2831 GEPIndices.pop_back();
2837 if (StructType *STy = dyn_cast<StructType>(Ty)) {
2838 unsigned OldSize = Indices.size();
2840 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
2842 assert(Indices.size() == OldSize && "Did not return to the old size");
2843 Indices.push_back(Idx);
2844 GEPIndices.push_back(IRB.getInt32(Idx));
2845 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
2846 GEPIndices.pop_back();
2852 llvm_unreachable("Only arrays and structs are aggregate loadable types");
2856 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
2857 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2858 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
2860 /// Emit a leaf load of a single value. This is called at the leaves of the
2861 /// recursive emission to actually load values.
2862 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2863 assert(Ty->isSingleValueType());
2864 // Load the single value and insert it using the indices.
2865 Value *Load = IRB.CreateLoad(IRB.CreateInBoundsGEP(Ptr, GEPIndices,
2868 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
2869 DEBUG(dbgs() << " to: " << *Load << "\n");
2873 bool visitLoadInst(LoadInst &LI) {
2874 assert(LI.getPointerOperand() == *U);
2875 if (!LI.isSimple() || LI.getType()->isSingleValueType())
2878 // We have an aggregate being loaded, split it apart.
2879 DEBUG(dbgs() << " original: " << LI << "\n");
2880 LoadOpSplitter Splitter(&LI, *U);
2881 Value *V = UndefValue::get(LI.getType());
2882 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
2883 LI.replaceAllUsesWith(V);
2884 LI.eraseFromParent();
2888 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
2889 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2890 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
2892 /// Emit a leaf store of a single value. This is called at the leaves of the
2893 /// recursive emission to actually produce stores.
2894 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2895 assert(Ty->isSingleValueType());
2896 // Extract the single value and store it using the indices.
2897 Value *Store = IRB.CreateStore(
2898 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
2899 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
2901 DEBUG(dbgs() << " to: " << *Store << "\n");
2905 bool visitStoreInst(StoreInst &SI) {
2906 if (!SI.isSimple() || SI.getPointerOperand() != *U)
2908 Value *V = SI.getValueOperand();
2909 if (V->getType()->isSingleValueType())
2912 // We have an aggregate being stored, split it apart.
2913 DEBUG(dbgs() << " original: " << SI << "\n");
2914 StoreOpSplitter Splitter(&SI, *U);
2915 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
2916 SI.eraseFromParent();
2920 bool visitBitCastInst(BitCastInst &BC) {
2925 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
2930 bool visitPHINode(PHINode &PN) {
2935 bool visitSelectInst(SelectInst &SI) {
2942 /// \brief Try to find a partition of the aggregate type passed in for a given
2943 /// offset and size.
2945 /// This recurses through the aggregate type and tries to compute a subtype
2946 /// based on the offset and size. When the offset and size span a sub-section
2947 /// of an array, it will even compute a new array type for that sub-section,
2948 /// and the same for structs.
2950 /// Note that this routine is very strict and tries to find a partition of the
2951 /// type which produces the *exact* right offset and size. It is not forgiving
2952 /// when the size or offset cause either end of type-based partition to be off.
2953 /// Also, this is a best-effort routine. It is reasonable to give up and not
2954 /// return a type if necessary.
2955 static Type *getTypePartition(const TargetData &TD, Type *Ty,
2956 uint64_t Offset, uint64_t Size) {
2957 if (Offset == 0 && TD.getTypeAllocSize(Ty) == Size)
2960 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
2961 // We can't partition pointers...
2962 if (SeqTy->isPointerTy())
2965 Type *ElementTy = SeqTy->getElementType();
2966 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
2967 uint64_t NumSkippedElements = Offset / ElementSize;
2968 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy))
2969 if (NumSkippedElements >= ArrTy->getNumElements())
2971 if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy))
2972 if (NumSkippedElements >= VecTy->getNumElements())
2974 Offset -= NumSkippedElements * ElementSize;
2976 // First check if we need to recurse.
2977 if (Offset > 0 || Size < ElementSize) {
2978 // Bail if the partition ends in a different array element.
2979 if ((Offset + Size) > ElementSize)
2981 // Recurse through the element type trying to peel off offset bytes.
2982 return getTypePartition(TD, ElementTy, Offset, Size);
2984 assert(Offset == 0);
2986 if (Size == ElementSize)
2988 assert(Size > ElementSize);
2989 uint64_t NumElements = Size / ElementSize;
2990 if (NumElements * ElementSize != Size)
2992 return ArrayType::get(ElementTy, NumElements);
2995 StructType *STy = dyn_cast<StructType>(Ty);
2999 const StructLayout *SL = TD.getStructLayout(STy);
3000 if (Offset >= SL->getSizeInBytes())
3002 uint64_t EndOffset = Offset + Size;
3003 if (EndOffset > SL->getSizeInBytes())
3006 unsigned Index = SL->getElementContainingOffset(Offset);
3007 Offset -= SL->getElementOffset(Index);
3009 Type *ElementTy = STy->getElementType(Index);
3010 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
3011 if (Offset >= ElementSize)
3012 return 0; // The offset points into alignment padding.
3014 // See if any partition must be contained by the element.
3015 if (Offset > 0 || Size < ElementSize) {
3016 if ((Offset + Size) > ElementSize)
3018 return getTypePartition(TD, ElementTy, Offset, Size);
3020 assert(Offset == 0);
3022 if (Size == ElementSize)
3025 StructType::element_iterator EI = STy->element_begin() + Index,
3026 EE = STy->element_end();
3027 if (EndOffset < SL->getSizeInBytes()) {
3028 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3029 if (Index == EndIndex)
3030 return 0; // Within a single element and its padding.
3032 // Don't try to form "natural" types if the elements don't line up with the
3034 // FIXME: We could potentially recurse down through the last element in the
3035 // sub-struct to find a natural end point.
3036 if (SL->getElementOffset(EndIndex) != EndOffset)
3039 assert(Index < EndIndex);
3040 EE = STy->element_begin() + EndIndex;
3043 // Try to build up a sub-structure.
3044 SmallVector<Type *, 4> ElementTys;
3046 ElementTys.push_back(*EI++);
3048 StructType *SubTy = StructType::get(STy->getContext(), ElementTys,
3050 const StructLayout *SubSL = TD.getStructLayout(SubTy);
3051 if (Size != SubSL->getSizeInBytes())
3052 return 0; // The sub-struct doesn't have quite the size needed.
3057 /// \brief Rewrite an alloca partition's users.
3059 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3060 /// to rewrite uses of an alloca partition to be conducive for SSA value
3061 /// promotion. If the partition needs a new, more refined alloca, this will
3062 /// build that new alloca, preserving as much type information as possible, and
3063 /// rewrite the uses of the old alloca to point at the new one and have the
3064 /// appropriate new offsets. It also evaluates how successful the rewrite was
3065 /// at enabling promotion and if it was successful queues the alloca to be
3067 bool SROA::rewriteAllocaPartition(AllocaInst &AI,
3068 AllocaPartitioning &P,
3069 AllocaPartitioning::iterator PI) {
3070 uint64_t AllocaSize = PI->EndOffset - PI->BeginOffset;
3071 bool IsLive = false;
3072 for (AllocaPartitioning::use_iterator UI = P.use_begin(PI),
3074 UI != UE && !IsLive; ++UI)
3078 return false; // No live uses left of this partition.
3080 DEBUG(dbgs() << "Speculating PHIs and selects in partition "
3081 << "[" << PI->BeginOffset << "," << PI->EndOffset << ")\n");
3083 PHIOrSelectSpeculator Speculator(*TD, P, *this);
3084 DEBUG(dbgs() << " speculating ");
3085 DEBUG(P.print(dbgs(), PI, ""));
3086 Speculator.visitUsers(PI);
3088 // Try to compute a friendly type for this partition of the alloca. This
3089 // won't always succeed, in which case we fall back to a legal integer type
3090 // or an i8 array of an appropriate size.
3092 if (Type *PartitionTy = P.getCommonType(PI))
3093 if (TD->getTypeAllocSize(PartitionTy) >= AllocaSize)
3094 AllocaTy = PartitionTy;
3096 if (Type *PartitionTy = getTypePartition(*TD, AI.getAllocatedType(),
3097 PI->BeginOffset, AllocaSize))
3098 AllocaTy = PartitionTy;
3100 (AllocaTy->isArrayTy() &&
3101 AllocaTy->getArrayElementType()->isIntegerTy())) &&
3102 TD->isLegalInteger(AllocaSize * 8))
3103 AllocaTy = Type::getIntNTy(*C, AllocaSize * 8);
3105 AllocaTy = ArrayType::get(Type::getInt8Ty(*C), AllocaSize);
3106 assert(TD->getTypeAllocSize(AllocaTy) >= AllocaSize);
3108 // Check for the case where we're going to rewrite to a new alloca of the
3109 // exact same type as the original, and with the same access offsets. In that
3110 // case, re-use the existing alloca, but still run through the rewriter to
3111 // performe phi and select speculation.
3113 if (AllocaTy == AI.getAllocatedType()) {
3114 assert(PI->BeginOffset == 0 &&
3115 "Non-zero begin offset but same alloca type");
3116 assert(PI == P.begin() && "Begin offset is zero on later partition");
3119 unsigned Alignment = AI.getAlignment();
3121 // The minimum alignment which users can rely on when the explicit
3122 // alignment is omitted or zero is that required by the ABI for this
3124 Alignment = TD->getABITypeAlignment(AI.getAllocatedType());
3126 Alignment = MinAlign(Alignment, PI->BeginOffset);
3127 // If we will get at least this much alignment from the type alone, leave
3128 // the alloca's alignment unconstrained.
3129 if (Alignment <= TD->getABITypeAlignment(AllocaTy))
3131 NewAI = new AllocaInst(AllocaTy, 0, Alignment,
3132 AI.getName() + ".sroa." + Twine(PI - P.begin()),
3137 DEBUG(dbgs() << "Rewriting alloca partition "
3138 << "[" << PI->BeginOffset << "," << PI->EndOffset << ") to: "
3141 // Track the high watermark of the post-promotion worklist. We will reset it
3142 // to this point if the alloca is not in fact scheduled for promotion.
3143 unsigned PPWOldSize = PostPromotionWorklist.size();
3145 AllocaPartitionRewriter Rewriter(*TD, P, PI, *this, AI, *NewAI,
3146 PI->BeginOffset, PI->EndOffset);
3147 DEBUG(dbgs() << " rewriting ");
3148 DEBUG(P.print(dbgs(), PI, ""));
3149 bool Promotable = Rewriter.visitUsers(P.use_begin(PI), P.use_end(PI));
3151 DEBUG(dbgs() << " and queuing for promotion\n");
3152 PromotableAllocas.push_back(NewAI);
3153 } else if (NewAI != &AI) {
3154 // If we can't promote the alloca, iterate on it to check for new
3155 // refinements exposed by splitting the current alloca. Don't iterate on an
3156 // alloca which didn't actually change and didn't get promoted.
3157 Worklist.insert(NewAI);
3160 // Drop any post-promotion work items if promotion didn't happen.
3162 while (PostPromotionWorklist.size() > PPWOldSize)
3163 PostPromotionWorklist.pop_back();
3168 /// \brief Walks the partitioning of an alloca rewriting uses of each partition.
3169 bool SROA::splitAlloca(AllocaInst &AI, AllocaPartitioning &P) {
3170 bool Changed = false;
3171 for (AllocaPartitioning::iterator PI = P.begin(), PE = P.end(); PI != PE;
3173 Changed |= rewriteAllocaPartition(AI, P, PI);
3178 /// \brief Analyze an alloca for SROA.
3180 /// This analyzes the alloca to ensure we can reason about it, builds
3181 /// a partitioning of the alloca, and then hands it off to be split and
3182 /// rewritten as needed.
3183 bool SROA::runOnAlloca(AllocaInst &AI) {
3184 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3185 ++NumAllocasAnalyzed;
3187 // Special case dead allocas, as they're trivial.
3188 if (AI.use_empty()) {
3189 AI.eraseFromParent();
3193 // Skip alloca forms that this analysis can't handle.
3194 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3195 TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
3198 bool Changed = false;
3200 // First, split any FCA loads and stores touching this alloca to promote
3201 // better splitting and promotion opportunities.
3202 AggLoadStoreRewriter AggRewriter(*TD);
3203 Changed |= AggRewriter.rewrite(AI);
3205 // Build the partition set using a recursive instruction-visiting builder.
3206 AllocaPartitioning P(*TD, AI);
3207 DEBUG(P.print(dbgs()));
3211 // No partitions to split. Leave the dead alloca for a later pass to clean up.
3212 if (P.begin() == P.end())
3215 // Delete all the dead users of this alloca before splitting and rewriting it.
3216 for (AllocaPartitioning::dead_user_iterator DI = P.dead_user_begin(),
3217 DE = P.dead_user_end();
3220 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3221 DeadInsts.push_back(*DI);
3223 for (AllocaPartitioning::dead_op_iterator DO = P.dead_op_begin(),
3224 DE = P.dead_op_end();
3227 // Clobber the use with an undef value.
3228 **DO = UndefValue::get(OldV->getType());
3229 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3230 if (isInstructionTriviallyDead(OldI)) {
3232 DeadInsts.push_back(OldI);
3236 return splitAlloca(AI, P) || Changed;
3239 /// \brief Delete the dead instructions accumulated in this run.
3241 /// Recursively deletes the dead instructions we've accumulated. This is done
3242 /// at the very end to maximize locality of the recursive delete and to
3243 /// minimize the problems of invalidated instruction pointers as such pointers
3244 /// are used heavily in the intermediate stages of the algorithm.
3246 /// We also record the alloca instructions deleted here so that they aren't
3247 /// subsequently handed to mem2reg to promote.
3248 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3249 DeadSplitInsts.clear();
3250 while (!DeadInsts.empty()) {
3251 Instruction *I = DeadInsts.pop_back_val();
3252 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3254 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
3255 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
3256 // Zero out the operand and see if it becomes trivially dead.
3258 if (isInstructionTriviallyDead(U))
3259 DeadInsts.push_back(U);
3262 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3263 DeletedAllocas.insert(AI);
3266 I->eraseFromParent();
3270 /// \brief Promote the allocas, using the best available technique.
3272 /// This attempts to promote whatever allocas have been identified as viable in
3273 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3274 /// If there is a domtree available, we attempt to promote using the full power
3275 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3276 /// based on the SSAUpdater utilities. This function returns whether any
3277 /// promotion occured.
3278 bool SROA::promoteAllocas(Function &F) {
3279 if (PromotableAllocas.empty())
3282 NumPromoted += PromotableAllocas.size();
3284 if (DT && !ForceSSAUpdater) {
3285 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3286 PromoteMemToReg(PromotableAllocas, *DT);
3287 PromotableAllocas.clear();
3291 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3293 DIBuilder DIB(*F.getParent());
3294 SmallVector<Instruction*, 64> Insts;
3296 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3297 AllocaInst *AI = PromotableAllocas[Idx];
3298 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
3300 Instruction *I = cast<Instruction>(*UI++);
3301 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3302 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3303 // leading to them) here. Eventually it should use them to optimize the
3304 // scalar values produced.
3305 if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
3306 assert(onlyUsedByLifetimeMarkers(I) &&
3307 "Found a bitcast used outside of a lifetime marker.");
3308 while (!I->use_empty())
3309 cast<Instruction>(*I->use_begin())->eraseFromParent();
3310 I->eraseFromParent();
3313 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3314 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3315 II->getIntrinsicID() == Intrinsic::lifetime_end);
3316 II->eraseFromParent();
3322 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3326 PromotableAllocas.clear();
3331 /// \brief A predicate to test whether an alloca belongs to a set.
3332 class IsAllocaInSet {
3333 typedef SmallPtrSet<AllocaInst *, 4> SetType;
3337 typedef AllocaInst *argument_type;
3339 IsAllocaInSet(const SetType &Set) : Set(Set) {}
3340 bool operator()(AllocaInst *AI) const { return Set.count(AI); }
3344 bool SROA::runOnFunction(Function &F) {
3345 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3346 C = &F.getContext();
3347 TD = getAnalysisIfAvailable<TargetData>();
3349 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3352 DT = getAnalysisIfAvailable<DominatorTree>();
3354 BasicBlock &EntryBB = F.getEntryBlock();
3355 for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end());
3357 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3358 Worklist.insert(AI);
3360 bool Changed = false;
3361 // A set of deleted alloca instruction pointers which should be removed from
3362 // the list of promotable allocas.
3363 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3366 while (!Worklist.empty()) {
3367 Changed |= runOnAlloca(*Worklist.pop_back_val());
3368 deleteDeadInstructions(DeletedAllocas);
3370 // Remove the deleted allocas from various lists so that we don't try to
3371 // continue processing them.
3372 if (!DeletedAllocas.empty()) {
3373 Worklist.remove_if(IsAllocaInSet(DeletedAllocas));
3374 PostPromotionWorklist.remove_if(IsAllocaInSet(DeletedAllocas));
3375 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3376 PromotableAllocas.end(),
3377 IsAllocaInSet(DeletedAllocas)),
3378 PromotableAllocas.end());
3379 DeletedAllocas.clear();
3383 Changed |= promoteAllocas(F);
3385 Worklist = PostPromotionWorklist;
3386 PostPromotionWorklist.clear();
3387 } while (!Worklist.empty());
3392 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3393 if (RequiresDomTree)
3394 AU.addRequired<DominatorTree>();
3395 AU.setPreservesCFG();