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 /// \brief Test whether a partition has been marked as dead.
141 bool isDead() const {
142 if (BeginOffset == UINT64_MAX) {
143 assert(EndOffset == UINT64_MAX);
149 /// \brief Kill a partition.
150 /// This is accomplished by setting both its beginning and end offset to
151 /// the maximum possible value.
153 assert(!isDead() && "He's Dead, Jim!");
154 BeginOffset = EndOffset = UINT64_MAX;
157 Partition() : ByteRange(), IsSplittable() {}
158 Partition(uint64_t BeginOffset, uint64_t EndOffset, bool IsSplittable)
159 : ByteRange(BeginOffset, EndOffset), IsSplittable(IsSplittable) {}
162 /// \brief A particular use of a partition of the alloca.
164 /// This structure is used to associate uses of a partition with it. They
165 /// mark the range of bytes which are referenced by a particular instruction,
166 /// and includes a handle to the user itself and the pointer value in use.
167 /// The bounds of these uses are determined by intersecting the bounds of the
168 /// memory use itself with a particular partition. As a consequence there is
169 /// intentionally overlap between various uses of the same partition.
170 struct PartitionUse : public ByteRange {
171 /// \brief The use in question. Provides access to both user and used value.
173 /// Note that this may be null if the partition use is *dead*, that is, it
174 /// should be ignored.
177 PartitionUse() : ByteRange(), U() {}
178 PartitionUse(uint64_t BeginOffset, uint64_t EndOffset, Use *U)
179 : ByteRange(BeginOffset, EndOffset), U(U) {}
182 /// \brief Construct a partitioning of a particular alloca.
184 /// Construction does most of the work for partitioning the alloca. This
185 /// performs the necessary walks of users and builds a partitioning from it.
186 AllocaPartitioning(const TargetData &TD, AllocaInst &AI);
188 /// \brief Test whether a pointer to the allocation escapes our analysis.
190 /// If this is true, the partitioning is never fully built and should be
192 bool isEscaped() const { return PointerEscapingInstr; }
194 /// \brief Support for iterating over the partitions.
196 typedef SmallVectorImpl<Partition>::iterator iterator;
197 iterator begin() { return Partitions.begin(); }
198 iterator end() { return Partitions.end(); }
200 typedef SmallVectorImpl<Partition>::const_iterator const_iterator;
201 const_iterator begin() const { return Partitions.begin(); }
202 const_iterator end() const { return Partitions.end(); }
205 /// \brief Support for iterating over and manipulating a particular
206 /// partition's uses.
208 /// The iteration support provided for uses is more limited, but also
209 /// includes some manipulation routines to support rewriting the uses of
210 /// partitions during SROA.
212 typedef SmallVectorImpl<PartitionUse>::iterator use_iterator;
213 use_iterator use_begin(unsigned Idx) { return Uses[Idx].begin(); }
214 use_iterator use_begin(const_iterator I) { return Uses[I - begin()].begin(); }
215 use_iterator use_end(unsigned Idx) { return Uses[Idx].end(); }
216 use_iterator use_end(const_iterator I) { return Uses[I - begin()].end(); }
218 typedef SmallVectorImpl<PartitionUse>::const_iterator const_use_iterator;
219 const_use_iterator use_begin(unsigned Idx) const { return Uses[Idx].begin(); }
220 const_use_iterator use_begin(const_iterator I) const {
221 return Uses[I - begin()].begin();
223 const_use_iterator use_end(unsigned Idx) const { return Uses[Idx].end(); }
224 const_use_iterator use_end(const_iterator I) const {
225 return Uses[I - begin()].end();
228 unsigned use_size(unsigned Idx) const { return Uses[Idx].size(); }
229 unsigned use_size(const_iterator I) const { return Uses[I - begin()].size(); }
230 const PartitionUse &getUse(unsigned PIdx, unsigned UIdx) const {
231 return Uses[PIdx][UIdx];
233 const PartitionUse &getUse(const_iterator I, unsigned UIdx) const {
234 return Uses[I - begin()][UIdx];
237 void use_push_back(unsigned Idx, const PartitionUse &PU) {
238 Uses[Idx].push_back(PU);
240 void use_push_back(const_iterator I, const PartitionUse &PU) {
241 Uses[I - begin()].push_back(PU);
245 /// \brief Allow iterating the dead users for this alloca.
247 /// These are instructions which will never actually use the alloca as they
248 /// are outside the allocated range. They are safe to replace with undef and
251 typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator;
252 dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); }
253 dead_user_iterator dead_user_end() const { return DeadUsers.end(); }
256 /// \brief Allow iterating the dead expressions referring to this alloca.
258 /// These are operands which have cannot actually be used to refer to the
259 /// alloca as they are outside its range and the user doesn't correct for
260 /// that. These mostly consist of PHI node inputs and the like which we just
261 /// need to replace with undef.
263 typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator;
264 dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); }
265 dead_op_iterator dead_op_end() const { return DeadOperands.end(); }
268 /// \brief MemTransferInst auxiliary data.
269 /// This struct provides some auxiliary data about memory transfer
270 /// intrinsics such as memcpy and memmove. These intrinsics can use two
271 /// different ranges within the same alloca, and provide other challenges to
272 /// correctly represent. We stash extra data to help us untangle this
273 /// after the partitioning is complete.
274 struct MemTransferOffsets {
275 /// The destination begin and end offsets when the destination is within
276 /// this alloca. If the end offset is zero the destination is not within
278 uint64_t DestBegin, DestEnd;
280 /// The source begin and end offsets when the source is within this alloca.
281 /// If the end offset is zero, the source is not within this alloca.
282 uint64_t SourceBegin, SourceEnd;
284 /// Flag for whether an alloca is splittable.
287 MemTransferOffsets getMemTransferOffsets(MemTransferInst &II) const {
288 return MemTransferInstData.lookup(&II);
291 /// \brief Map from a PHI or select operand back to a partition.
293 /// When manipulating PHI nodes or selects, they can use more than one
294 /// partition of an alloca. We store a special mapping to allow finding the
295 /// partition referenced by each of these operands, if any.
296 iterator findPartitionForPHIOrSelectOperand(Use *U) {
297 SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt
298 = PHIOrSelectOpMap.find(U);
299 if (MapIt == PHIOrSelectOpMap.end())
302 return begin() + MapIt->second.first;
305 /// \brief Map from a PHI or select operand back to the specific use of
308 /// Similar to mapping these operands back to the partitions, this maps
309 /// directly to the use structure of that partition.
310 use_iterator findPartitionUseForPHIOrSelectOperand(Use *U) {
311 SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt
312 = PHIOrSelectOpMap.find(U);
313 assert(MapIt != PHIOrSelectOpMap.end());
314 return Uses[MapIt->second.first].begin() + MapIt->second.second;
317 /// \brief Compute a common type among the uses of a particular partition.
319 /// This routines walks all of the uses of a particular partition and tries
320 /// to find a common type between them. Untyped operations such as memset and
321 /// memcpy are ignored.
322 Type *getCommonType(iterator I) const;
324 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
325 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
326 void printUsers(raw_ostream &OS, const_iterator I,
327 StringRef Indent = " ") const;
328 void print(raw_ostream &OS) const;
329 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump(const_iterator I) const;
330 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump() const;
334 template <typename DerivedT, typename RetT = void> class BuilderBase;
335 class PartitionBuilder;
336 friend class AllocaPartitioning::PartitionBuilder;
338 friend class AllocaPartitioning::UseBuilder;
341 /// \brief Handle to alloca instruction to simplify method interfaces.
345 /// \brief The instruction responsible for this alloca having no partitioning.
347 /// When an instruction (potentially) escapes the pointer to the alloca, we
348 /// store a pointer to that here and abort trying to partition the alloca.
349 /// This will be null if the alloca is partitioned successfully.
350 Instruction *PointerEscapingInstr;
352 /// \brief The partitions of the alloca.
354 /// We store a vector of the partitions over the alloca here. This vector is
355 /// sorted by increasing begin offset, and then by decreasing end offset. See
356 /// the Partition inner class for more details. Initially (during
357 /// construction) there are overlaps, but we form a disjoint sequence of
358 /// partitions while finishing construction and a fully constructed object is
359 /// expected to always have this as a disjoint space.
360 SmallVector<Partition, 8> Partitions;
362 /// \brief The uses of the partitions.
364 /// This is essentially a mapping from each partition to a list of uses of
365 /// that partition. The mapping is done with a Uses vector that has the exact
366 /// same number of entries as the partition vector. Each entry is itself
367 /// a vector of the uses.
368 SmallVector<SmallVector<PartitionUse, 2>, 8> Uses;
370 /// \brief Instructions which will become dead if we rewrite the alloca.
372 /// Note that these are not separated by partition. This is because we expect
373 /// a partitioned alloca to be completely rewritten or not rewritten at all.
374 /// If rewritten, all these instructions can simply be removed and replaced
375 /// with undef as they come from outside of the allocated space.
376 SmallVector<Instruction *, 8> DeadUsers;
378 /// \brief Operands which will become dead if we rewrite the alloca.
380 /// These are operands that in their particular use can be replaced with
381 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
382 /// to PHI nodes and the like. They aren't entirely dead (there might be
383 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
384 /// want to swap this particular input for undef to simplify the use lists of
386 SmallVector<Use *, 8> DeadOperands;
388 /// \brief The underlying storage for auxiliary memcpy and memset info.
389 SmallDenseMap<MemTransferInst *, MemTransferOffsets, 4> MemTransferInstData;
391 /// \brief A side datastructure used when building up the partitions and uses.
393 /// This mapping is only really used during the initial building of the
394 /// partitioning so that we can retain information about PHI and select nodes
396 SmallDenseMap<Instruction *, std::pair<uint64_t, bool> > PHIOrSelectSizes;
398 /// \brief Auxiliary information for particular PHI or select operands.
399 SmallDenseMap<Use *, std::pair<unsigned, unsigned>, 4> PHIOrSelectOpMap;
401 /// \brief A utility routine called from the constructor.
403 /// This does what it says on the tin. It is the key of the alloca partition
404 /// splitting and merging. After it is called we have the desired disjoint
405 /// collection of partitions.
406 void splitAndMergePartitions();
410 template <typename DerivedT, typename RetT>
411 class AllocaPartitioning::BuilderBase
412 : public InstVisitor<DerivedT, RetT> {
414 BuilderBase(const TargetData &TD, AllocaInst &AI, AllocaPartitioning &P)
416 AllocSize(TD.getTypeAllocSize(AI.getAllocatedType())),
422 const TargetData &TD;
423 const uint64_t AllocSize;
424 AllocaPartitioning &P;
426 SmallPtrSet<Use *, 8> VisitedUses;
432 SmallVector<OffsetUse, 8> Queue;
434 // The active offset and use while visiting.
438 void enqueueUsers(Instruction &I, int64_t UserOffset) {
439 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
441 if (VisitedUses.insert(&UI.getUse())) {
442 OffsetUse OU = { &UI.getUse(), UserOffset };
448 bool computeConstantGEPOffset(GetElementPtrInst &GEPI, int64_t &GEPOffset) {
450 for (gep_type_iterator GTI = gep_type_begin(GEPI), GTE = gep_type_end(GEPI);
452 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
458 // Handle a struct index, which adds its field offset to the pointer.
459 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
460 unsigned ElementIdx = OpC->getZExtValue();
461 const StructLayout *SL = TD.getStructLayout(STy);
462 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
463 // Check that we can continue to model this GEP in a signed 64-bit offset.
464 if (ElementOffset > INT64_MAX ||
466 ((uint64_t)GEPOffset + ElementOffset) > INT64_MAX)) {
467 DEBUG(dbgs() << "WARNING: Encountered a cumulative offset exceeding "
468 << "what can be represented in an int64_t!\n"
469 << " alloca: " << P.AI << "\n");
473 GEPOffset = ElementOffset + (uint64_t)-GEPOffset;
475 GEPOffset += ElementOffset;
479 APInt Index = OpC->getValue().sextOrTrunc(TD.getPointerSizeInBits());
480 Index *= APInt(Index.getBitWidth(),
481 TD.getTypeAllocSize(GTI.getIndexedType()));
482 Index += APInt(Index.getBitWidth(), (uint64_t)GEPOffset,
484 // Check if the result can be stored in our int64_t offset.
485 if (!Index.isSignedIntN(sizeof(GEPOffset) * 8)) {
486 DEBUG(dbgs() << "WARNING: Encountered a cumulative offset exceeding "
487 << "what can be represented in an int64_t!\n"
488 << " alloca: " << P.AI << "\n");
492 GEPOffset = Index.getSExtValue();
497 Value *foldSelectInst(SelectInst &SI) {
498 // If the condition being selected on is a constant or the same value is
499 // being selected between, fold the select. Yes this does (rarely) happen
501 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
502 return SI.getOperand(1+CI->isZero());
503 if (SI.getOperand(1) == SI.getOperand(2)) {
504 assert(*U == SI.getOperand(1));
505 return SI.getOperand(1);
511 /// \brief Builder for the alloca partitioning.
513 /// This class builds an alloca partitioning by recursively visiting the uses
514 /// of an alloca and splitting the partitions for each load and store at each
516 class AllocaPartitioning::PartitionBuilder
517 : public BuilderBase<PartitionBuilder, bool> {
518 friend class InstVisitor<PartitionBuilder, bool>;
520 SmallDenseMap<Instruction *, unsigned> MemTransferPartitionMap;
523 PartitionBuilder(const TargetData &TD, AllocaInst &AI, AllocaPartitioning &P)
524 : BuilderBase<PartitionBuilder, bool>(TD, AI, P) {}
526 /// \brief Run the builder over the allocation.
528 // Note that we have to re-evaluate size on each trip through the loop as
529 // the queue grows at the tail.
530 for (unsigned Idx = 0; Idx < Queue.size(); ++Idx) {
532 Offset = Queue[Idx].Offset;
533 if (!visit(cast<Instruction>(U->getUser())))
540 bool markAsEscaping(Instruction &I) {
541 P.PointerEscapingInstr = &I;
545 void insertUse(Instruction &I, int64_t Offset, uint64_t Size,
546 bool IsSplittable = false) {
547 // Completely skip uses which have a zero size or don't overlap the
550 (Offset >= 0 && (uint64_t)Offset >= AllocSize) ||
551 (Offset < 0 && (uint64_t)-Offset >= Size)) {
552 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
553 << " which starts past the end of the " << AllocSize
555 << " alloca: " << P.AI << "\n"
556 << " use: " << I << "\n");
560 // Clamp the start to the beginning of the allocation.
562 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
563 << " to start at the beginning of the alloca:\n"
564 << " alloca: " << P.AI << "\n"
565 << " use: " << I << "\n");
566 Size -= (uint64_t)-Offset;
570 uint64_t BeginOffset = Offset, EndOffset = BeginOffset + Size;
572 // Clamp the end offset to the end of the allocation. Note that this is
573 // formulated to handle even the case where "BeginOffset + Size" overflows.
574 assert(AllocSize >= BeginOffset); // Established above.
575 if (Size > AllocSize - BeginOffset) {
576 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
577 << " to remain within the " << AllocSize << " byte alloca:\n"
578 << " alloca: " << P.AI << "\n"
579 << " use: " << I << "\n");
580 EndOffset = AllocSize;
583 Partition New(BeginOffset, EndOffset, IsSplittable);
584 P.Partitions.push_back(New);
587 bool handleLoadOrStore(Type *Ty, Instruction &I, int64_t Offset) {
588 uint64_t Size = TD.getTypeStoreSize(Ty);
590 // If this memory access can be shown to *statically* extend outside the
591 // bounds of of the allocation, it's behavior is undefined, so simply
592 // ignore it. Note that this is more strict than the generic clamping
593 // behavior of insertUse. We also try to handle cases which might run the
595 // FIXME: We should instead consider the pointer to have escaped if this
596 // function is being instrumented for addressing bugs or race conditions.
597 if (Offset < 0 || (uint64_t)Offset >= AllocSize ||
598 Size > (AllocSize - (uint64_t)Offset)) {
599 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte "
600 << (isa<LoadInst>(I) ? "load" : "store") << " @" << Offset
601 << " which extends past the end of the " << AllocSize
603 << " alloca: " << P.AI << "\n"
604 << " use: " << I << "\n");
608 insertUse(I, Offset, Size);
612 bool visitBitCastInst(BitCastInst &BC) {
613 enqueueUsers(BC, Offset);
617 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
619 if (!computeConstantGEPOffset(GEPI, GEPOffset))
620 return markAsEscaping(GEPI);
622 enqueueUsers(GEPI, GEPOffset);
626 bool visitLoadInst(LoadInst &LI) {
627 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
628 "All simple FCA loads should have been pre-split");
629 return handleLoadOrStore(LI.getType(), LI, Offset);
632 bool visitStoreInst(StoreInst &SI) {
633 Value *ValOp = SI.getValueOperand();
635 return markAsEscaping(SI);
637 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
638 "All simple FCA stores should have been pre-split");
639 return handleLoadOrStore(ValOp->getType(), SI, Offset);
643 bool visitMemSetInst(MemSetInst &II) {
644 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
645 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
646 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
647 insertUse(II, Offset, Size, Length);
651 bool visitMemTransferInst(MemTransferInst &II) {
652 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
653 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
655 // Zero-length mem transfer intrinsics can be ignored entirely.
658 MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
660 // Only intrinsics with a constant length can be split.
661 Offsets.IsSplittable = Length;
663 if (*U == II.getRawDest()) {
664 Offsets.DestBegin = Offset;
665 Offsets.DestEnd = Offset + Size;
667 if (*U == II.getRawSource()) {
668 Offsets.SourceBegin = Offset;
669 Offsets.SourceEnd = Offset + Size;
672 // If we have set up end offsets for both the source and the destination,
673 // we have found both sides of this transfer pointing at the same alloca.
674 bool SeenBothEnds = Offsets.SourceEnd && Offsets.DestEnd;
675 if (SeenBothEnds && II.getRawDest() != II.getRawSource()) {
676 unsigned PrevIdx = MemTransferPartitionMap[&II];
678 // Check if the begin offsets match and this is a non-volatile transfer.
679 // In that case, we can completely elide the transfer.
680 if (!II.isVolatile() && Offsets.SourceBegin == Offsets.DestBegin) {
681 P.Partitions[PrevIdx].kill();
685 // Otherwise we have an offset transfer within the same alloca. We can't
687 P.Partitions[PrevIdx].IsSplittable = Offsets.IsSplittable = false;
688 } else if (SeenBothEnds) {
689 // Handle the case where this exact use provides both ends of the
691 assert(II.getRawDest() == II.getRawSource());
693 // For non-volatile transfers this is a no-op.
694 if (!II.isVolatile())
697 // Otherwise just suppress splitting.
698 Offsets.IsSplittable = false;
702 // Insert the use now that we've fixed up the splittable nature.
703 insertUse(II, Offset, Size, Offsets.IsSplittable);
705 // Setup the mapping from intrinsic to partition of we've not seen both
706 // ends of this transfer.
708 unsigned NewIdx = P.Partitions.size() - 1;
710 = MemTransferPartitionMap.insert(std::make_pair(&II, NewIdx)).second;
712 "Already have intrinsic in map but haven't seen both ends");
718 // Disable SRoA for any intrinsics except for lifetime invariants.
719 // FIXME: What about debug instrinsics? This matches old behavior, but
720 // doesn't make sense.
721 bool visitIntrinsicInst(IntrinsicInst &II) {
722 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
723 II.getIntrinsicID() == Intrinsic::lifetime_end) {
724 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
725 uint64_t Size = std::min(AllocSize - Offset, Length->getLimitedValue());
726 insertUse(II, Offset, Size, true);
730 return markAsEscaping(II);
733 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
734 // We consider any PHI or select that results in a direct load or store of
735 // the same offset to be a viable use for partitioning purposes. These uses
736 // are considered unsplittable and the size is the maximum loaded or stored
738 SmallPtrSet<Instruction *, 4> Visited;
739 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
740 Visited.insert(Root);
741 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
742 // If there are no loads or stores, the access is dead. We mark that as
743 // a size zero access.
746 Instruction *I, *UsedI;
747 llvm::tie(UsedI, I) = Uses.pop_back_val();
749 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
750 Size = std::max(Size, TD.getTypeStoreSize(LI->getType()));
753 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
754 Value *Op = SI->getOperand(0);
757 Size = std::max(Size, TD.getTypeStoreSize(Op->getType()));
761 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
762 if (!GEP->hasAllZeroIndices())
764 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
765 !isa<SelectInst>(I)) {
769 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
771 if (Visited.insert(cast<Instruction>(*UI)))
772 Uses.push_back(std::make_pair(I, cast<Instruction>(*UI)));
773 } while (!Uses.empty());
778 bool visitPHINode(PHINode &PN) {
779 // See if we already have computed info on this node.
780 std::pair<uint64_t, bool> &PHIInfo = P.PHIOrSelectSizes[&PN];
782 PHIInfo.second = true;
783 insertUse(PN, Offset, PHIInfo.first);
787 // Check for an unsafe use of the PHI node.
788 if (Instruction *EscapingI = hasUnsafePHIOrSelectUse(&PN, PHIInfo.first))
789 return markAsEscaping(*EscapingI);
791 insertUse(PN, Offset, PHIInfo.first);
795 bool visitSelectInst(SelectInst &SI) {
796 if (Value *Result = foldSelectInst(SI)) {
798 // If the result of the constant fold will be the pointer, recurse
799 // through the select as if we had RAUW'ed it.
800 enqueueUsers(SI, Offset);
805 // See if we already have computed info on this node.
806 std::pair<uint64_t, bool> &SelectInfo = P.PHIOrSelectSizes[&SI];
807 if (SelectInfo.first) {
808 SelectInfo.second = true;
809 insertUse(SI, Offset, SelectInfo.first);
813 // Check for an unsafe use of the PHI node.
814 if (Instruction *EscapingI = hasUnsafePHIOrSelectUse(&SI, SelectInfo.first))
815 return markAsEscaping(*EscapingI);
817 insertUse(SI, Offset, SelectInfo.first);
821 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
822 bool visitInstruction(Instruction &I) { return markAsEscaping(I); }
826 /// \brief Use adder for the alloca partitioning.
828 /// This class adds the uses of an alloca to all of the partitions which they
829 /// use. For splittable partitions, this can end up doing essentially a linear
830 /// walk of the partitions, but the number of steps remains bounded by the
831 /// total result instruction size:
832 /// - The number of partitions is a result of the number unsplittable
833 /// instructions using the alloca.
834 /// - The number of users of each partition is at worst the total number of
835 /// splittable instructions using the alloca.
836 /// Thus we will produce N * M instructions in the end, where N are the number
837 /// of unsplittable uses and M are the number of splittable. This visitor does
838 /// the exact same number of updates to the partitioning.
840 /// In the more common case, this visitor will leverage the fact that the
841 /// partition space is pre-sorted, and do a logarithmic search for the
842 /// partition needed, making the total visit a classical ((N + M) * log(N))
843 /// complexity operation.
844 class AllocaPartitioning::UseBuilder : public BuilderBase<UseBuilder> {
845 friend class InstVisitor<UseBuilder>;
847 /// \brief Set to de-duplicate dead instructions found in the use walk.
848 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
851 UseBuilder(const TargetData &TD, AllocaInst &AI, AllocaPartitioning &P)
852 : BuilderBase<UseBuilder>(TD, AI, P) {}
854 /// \brief Run the builder over the allocation.
856 // Note that we have to re-evaluate size on each trip through the loop as
857 // the queue grows at the tail.
858 for (unsigned Idx = 0; Idx < Queue.size(); ++Idx) {
860 Offset = Queue[Idx].Offset;
861 this->visit(cast<Instruction>(U->getUser()));
866 void markAsDead(Instruction &I) {
867 if (VisitedDeadInsts.insert(&I))
868 P.DeadUsers.push_back(&I);
871 void insertUse(Instruction &User, int64_t Offset, uint64_t Size) {
872 // If the use has a zero size or extends outside of the allocation, record
873 // it as a dead use for elimination later.
874 if (Size == 0 || (uint64_t)Offset >= AllocSize ||
875 (Offset < 0 && (uint64_t)-Offset >= Size))
876 return markAsDead(User);
878 // Clamp the start to the beginning of the allocation.
880 Size -= (uint64_t)-Offset;
884 uint64_t BeginOffset = Offset, EndOffset = BeginOffset + Size;
886 // Clamp the end offset to the end of the allocation. Note that this is
887 // formulated to handle even the case where "BeginOffset + Size" overflows.
888 assert(AllocSize >= BeginOffset); // Established above.
889 if (Size > AllocSize - BeginOffset)
890 EndOffset = AllocSize;
892 // NB: This only works if we have zero overlapping partitions.
893 iterator B = std::lower_bound(P.begin(), P.end(), BeginOffset);
894 if (B != P.begin() && llvm::prior(B)->EndOffset > BeginOffset)
896 for (iterator I = B, E = P.end(); I != E && I->BeginOffset < EndOffset;
898 PartitionUse NewPU(std::max(I->BeginOffset, BeginOffset),
899 std::min(I->EndOffset, EndOffset), U);
900 P.use_push_back(I, NewPU);
901 if (isa<PHINode>(U->getUser()) || isa<SelectInst>(U->getUser()))
902 P.PHIOrSelectOpMap[U]
903 = std::make_pair(I - P.begin(), P.Uses[I - P.begin()].size() - 1);
907 void handleLoadOrStore(Type *Ty, Instruction &I, int64_t Offset) {
908 uint64_t Size = TD.getTypeStoreSize(Ty);
910 // If this memory access can be shown to *statically* extend outside the
911 // bounds of of the allocation, it's behavior is undefined, so simply
912 // ignore it. Note that this is more strict than the generic clamping
913 // behavior of insertUse.
914 if (Offset < 0 || (uint64_t)Offset >= AllocSize ||
915 Size > (AllocSize - (uint64_t)Offset))
916 return markAsDead(I);
918 insertUse(I, Offset, Size);
921 void visitBitCastInst(BitCastInst &BC) {
923 return markAsDead(BC);
925 enqueueUsers(BC, Offset);
928 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
929 if (GEPI.use_empty())
930 return markAsDead(GEPI);
933 if (!computeConstantGEPOffset(GEPI, GEPOffset))
934 llvm_unreachable("Unable to compute constant offset for use");
936 enqueueUsers(GEPI, GEPOffset);
939 void visitLoadInst(LoadInst &LI) {
940 handleLoadOrStore(LI.getType(), LI, Offset);
943 void visitStoreInst(StoreInst &SI) {
944 handleLoadOrStore(SI.getOperand(0)->getType(), SI, Offset);
947 void visitMemSetInst(MemSetInst &II) {
948 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
949 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
950 insertUse(II, Offset, Size);
953 void visitMemTransferInst(MemTransferInst &II) {
954 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
955 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
957 return markAsDead(II);
959 MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
960 if (!II.isVolatile() && Offsets.DestEnd && Offsets.SourceEnd &&
961 Offsets.DestBegin == Offsets.SourceBegin)
962 return markAsDead(II); // Skip identity transfers without side-effects.
964 insertUse(II, Offset, Size);
967 void visitIntrinsicInst(IntrinsicInst &II) {
968 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
969 II.getIntrinsicID() == Intrinsic::lifetime_end);
971 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
972 insertUse(II, Offset,
973 std::min(AllocSize - Offset, Length->getLimitedValue()));
976 void insertPHIOrSelect(Instruction &User, uint64_t Offset) {
977 uint64_t Size = P.PHIOrSelectSizes.lookup(&User).first;
979 // For PHI and select operands outside the alloca, we can't nuke the entire
980 // phi or select -- the other side might still be relevant, so we special
981 // case them here and use a separate structure to track the operands
982 // themselves which should be replaced with undef.
983 if (Offset >= AllocSize) {
984 P.DeadOperands.push_back(U);
988 insertUse(User, Offset, Size);
990 void visitPHINode(PHINode &PN) {
992 return markAsDead(PN);
994 insertPHIOrSelect(PN, Offset);
996 void visitSelectInst(SelectInst &SI) {
998 return markAsDead(SI);
1000 if (Value *Result = foldSelectInst(SI)) {
1002 // If the result of the constant fold will be the pointer, recurse
1003 // through the select as if we had RAUW'ed it.
1004 enqueueUsers(SI, Offset);
1006 // Otherwise the operand to the select is dead, and we can replace it
1008 P.DeadOperands.push_back(U);
1013 insertPHIOrSelect(SI, Offset);
1016 /// \brief Unreachable, we've already visited the alloca once.
1017 void visitInstruction(Instruction &I) {
1018 llvm_unreachable("Unhandled instruction in use builder.");
1022 void AllocaPartitioning::splitAndMergePartitions() {
1023 size_t NumDeadPartitions = 0;
1025 // Track the range of splittable partitions that we pass when accumulating
1026 // overlapping unsplittable partitions.
1027 uint64_t SplitEndOffset = 0ull;
1029 Partition New(0ull, 0ull, false);
1031 for (unsigned i = 0, j = i, e = Partitions.size(); i != e; i = j) {
1034 if (!Partitions[i].IsSplittable || New.BeginOffset == New.EndOffset) {
1035 assert(New.BeginOffset == New.EndOffset);
1036 New = Partitions[i];
1038 assert(New.IsSplittable);
1039 New.EndOffset = std::max(New.EndOffset, Partitions[i].EndOffset);
1041 assert(New.BeginOffset != New.EndOffset);
1043 // Scan the overlapping partitions.
1044 while (j != e && New.EndOffset > Partitions[j].BeginOffset) {
1045 // If the new partition we are forming is splittable, stop at the first
1046 // unsplittable partition.
1047 if (New.IsSplittable && !Partitions[j].IsSplittable)
1050 // Grow the new partition to include any equally splittable range. 'j' is
1051 // always equally splittable when New is splittable, but when New is not
1052 // splittable, we may subsume some (or part of some) splitable partition
1053 // without growing the new one.
1054 if (New.IsSplittable == Partitions[j].IsSplittable) {
1055 New.EndOffset = std::max(New.EndOffset, Partitions[j].EndOffset);
1057 assert(!New.IsSplittable);
1058 assert(Partitions[j].IsSplittable);
1059 SplitEndOffset = std::max(SplitEndOffset, Partitions[j].EndOffset);
1062 Partitions[j].kill();
1063 ++NumDeadPartitions;
1067 // If the new partition is splittable, chop off the end as soon as the
1068 // unsplittable subsequent partition starts and ensure we eventually cover
1069 // the splittable area.
1070 if (j != e && New.IsSplittable) {
1071 SplitEndOffset = std::max(SplitEndOffset, New.EndOffset);
1072 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
1075 // Add the new partition if it differs from the original one and is
1076 // non-empty. We can end up with an empty partition here if it was
1077 // splittable but there is an unsplittable one that starts at the same
1079 if (New != Partitions[i]) {
1080 if (New.BeginOffset != New.EndOffset)
1081 Partitions.push_back(New);
1082 // Mark the old one for removal.
1083 Partitions[i].kill();
1084 ++NumDeadPartitions;
1087 New.BeginOffset = New.EndOffset;
1088 if (!New.IsSplittable) {
1089 New.EndOffset = std::max(New.EndOffset, SplitEndOffset);
1090 if (j != e && !Partitions[j].IsSplittable)
1091 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
1092 New.IsSplittable = true;
1093 // If there is a trailing splittable partition which won't be fused into
1094 // the next splittable partition go ahead and add it onto the partitions
1096 if (New.BeginOffset < New.EndOffset &&
1097 (j == e || !Partitions[j].IsSplittable ||
1098 New.EndOffset < Partitions[j].BeginOffset)) {
1099 Partitions.push_back(New);
1100 New.BeginOffset = New.EndOffset = 0ull;
1105 // Re-sort the partitions now that they have been split and merged into
1106 // disjoint set of partitions. Also remove any of the dead partitions we've
1107 // replaced in the process.
1108 std::sort(Partitions.begin(), Partitions.end());
1109 if (NumDeadPartitions) {
1110 assert(Partitions.back().isDead());
1111 assert((ptrdiff_t)NumDeadPartitions ==
1112 std::count(Partitions.begin(), Partitions.end(), Partitions.back()));
1114 Partitions.erase(Partitions.end() - NumDeadPartitions, Partitions.end());
1117 AllocaPartitioning::AllocaPartitioning(const TargetData &TD, AllocaInst &AI)
1122 PointerEscapingInstr(0) {
1123 PartitionBuilder PB(TD, AI, *this);
1127 // Sort the uses. This arranges for the offsets to be in ascending order,
1128 // and the sizes to be in descending order.
1129 std::sort(Partitions.begin(), Partitions.end());
1131 // Remove any partitions from the back which are marked as dead.
1132 while (!Partitions.empty() && Partitions.back().isDead())
1133 Partitions.pop_back();
1135 if (Partitions.size() > 1) {
1136 // Intersect splittability for all partitions with equal offsets and sizes.
1137 // Then remove all but the first so that we have a sequence of non-equal but
1138 // potentially overlapping partitions.
1139 for (iterator I = Partitions.begin(), J = I, E = Partitions.end(); I != E;
1142 while (J != E && *I == *J) {
1143 I->IsSplittable &= J->IsSplittable;
1147 Partitions.erase(std::unique(Partitions.begin(), Partitions.end()),
1150 // Split splittable and merge unsplittable partitions into a disjoint set
1151 // of partitions over the used space of the allocation.
1152 splitAndMergePartitions();
1155 // Now build up the user lists for each of these disjoint partitions by
1156 // re-walking the recursive users of the alloca.
1157 Uses.resize(Partitions.size());
1158 UseBuilder UB(TD, AI, *this);
1162 Type *AllocaPartitioning::getCommonType(iterator I) const {
1164 for (const_use_iterator UI = use_begin(I), UE = use_end(I); UI != UE; ++UI) {
1166 continue; // Skip dead uses.
1167 if (isa<IntrinsicInst>(*UI->U->getUser()))
1169 if (UI->BeginOffset != I->BeginOffset || UI->EndOffset != I->EndOffset)
1173 if (LoadInst *LI = dyn_cast<LoadInst>(UI->U->getUser())) {
1174 UserTy = LI->getType();
1175 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI->U->getUser())) {
1176 UserTy = SI->getValueOperand()->getType();
1179 if (Ty && Ty != UserTy)
1187 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1189 void AllocaPartitioning::print(raw_ostream &OS, const_iterator I,
1190 StringRef Indent) const {
1191 OS << Indent << "partition #" << (I - begin())
1192 << " [" << I->BeginOffset << "," << I->EndOffset << ")"
1193 << (I->IsSplittable ? " (splittable)" : "")
1194 << (Uses[I - begin()].empty() ? " (zero uses)" : "")
1198 void AllocaPartitioning::printUsers(raw_ostream &OS, const_iterator I,
1199 StringRef Indent) const {
1200 for (const_use_iterator UI = use_begin(I), UE = use_end(I);
1203 continue; // Skip dead uses.
1204 OS << Indent << " [" << UI->BeginOffset << "," << UI->EndOffset << ") "
1205 << "used by: " << *UI->U->getUser() << "\n";
1206 if (MemTransferInst *II = dyn_cast<MemTransferInst>(UI->U->getUser())) {
1207 const MemTransferOffsets &MTO = MemTransferInstData.lookup(II);
1209 if (!MTO.IsSplittable)
1210 IsDest = UI->BeginOffset == MTO.DestBegin;
1212 IsDest = MTO.DestBegin != 0u;
1213 OS << Indent << " (original " << (IsDest ? "dest" : "source") << ": "
1214 << "[" << (IsDest ? MTO.DestBegin : MTO.SourceBegin)
1215 << "," << (IsDest ? MTO.DestEnd : MTO.SourceEnd) << ")\n";
1220 void AllocaPartitioning::print(raw_ostream &OS) const {
1221 if (PointerEscapingInstr) {
1222 OS << "No partitioning for alloca: " << AI << "\n"
1223 << " A pointer to this alloca escaped by:\n"
1224 << " " << *PointerEscapingInstr << "\n";
1228 OS << "Partitioning of alloca: " << AI << "\n";
1230 for (const_iterator I = begin(), E = end(); I != E; ++I, ++Num) {
1236 void AllocaPartitioning::dump(const_iterator I) const { print(dbgs(), I); }
1237 void AllocaPartitioning::dump() const { print(dbgs()); }
1239 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1243 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
1245 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
1246 /// the loads and stores of an alloca instruction, as well as updating its
1247 /// debug information. This is used when a domtree is unavailable and thus
1248 /// mem2reg in its full form can't be used to handle promotion of allocas to
1250 class AllocaPromoter : public LoadAndStorePromoter {
1254 SmallVector<DbgDeclareInst *, 4> DDIs;
1255 SmallVector<DbgValueInst *, 4> DVIs;
1258 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
1259 AllocaInst &AI, DIBuilder &DIB)
1260 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
1262 void run(const SmallVectorImpl<Instruction*> &Insts) {
1263 // Remember which alloca we're promoting (for isInstInList).
1264 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
1265 for (Value::use_iterator UI = DebugNode->use_begin(),
1266 UE = DebugNode->use_end();
1268 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
1269 DDIs.push_back(DDI);
1270 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
1271 DVIs.push_back(DVI);
1274 LoadAndStorePromoter::run(Insts);
1275 AI.eraseFromParent();
1276 while (!DDIs.empty())
1277 DDIs.pop_back_val()->eraseFromParent();
1278 while (!DVIs.empty())
1279 DVIs.pop_back_val()->eraseFromParent();
1282 virtual bool isInstInList(Instruction *I,
1283 const SmallVectorImpl<Instruction*> &Insts) const {
1284 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1285 return LI->getOperand(0) == &AI;
1286 return cast<StoreInst>(I)->getPointerOperand() == &AI;
1289 virtual void updateDebugInfo(Instruction *Inst) const {
1290 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(),
1291 E = DDIs.end(); I != E; ++I) {
1292 DbgDeclareInst *DDI = *I;
1293 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
1294 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1295 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
1296 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1298 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(),
1299 E = DVIs.end(); I != E; ++I) {
1300 DbgValueInst *DVI = *I;
1302 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1303 // If an argument is zero extended then use argument directly. The ZExt
1304 // may be zapped by an optimization pass in future.
1305 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1306 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
1307 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1308 Arg = dyn_cast<Argument>(SExt->getOperand(0));
1310 Arg = SI->getOperand(0);
1311 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
1312 Arg = LI->getOperand(0);
1316 Instruction *DbgVal =
1317 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
1319 DbgVal->setDebugLoc(DVI->getDebugLoc());
1323 } // end anon namespace
1327 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
1329 /// This pass takes allocations which can be completely analyzed (that is, they
1330 /// don't escape) and tries to turn them into scalar SSA values. There are
1331 /// a few steps to this process.
1333 /// 1) It takes allocations of aggregates and analyzes the ways in which they
1334 /// are used to try to split them into smaller allocations, ideally of
1335 /// a single scalar data type. It will split up memcpy and memset accesses
1336 /// as necessary and try to isolate invidual scalar accesses.
1337 /// 2) It will transform accesses into forms which are suitable for SSA value
1338 /// promotion. This can be replacing a memset with a scalar store of an
1339 /// integer value, or it can involve speculating operations on a PHI or
1340 /// select to be a PHI or select of the results.
1341 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
1342 /// onto insert and extract operations on a vector value, and convert them to
1343 /// this form. By doing so, it will enable promotion of vector aggregates to
1344 /// SSA vector values.
1345 class SROA : public FunctionPass {
1346 const bool RequiresDomTree;
1349 const TargetData *TD;
1352 /// \brief Worklist of alloca instructions to simplify.
1354 /// Each alloca in the function is added to this. Each new alloca formed gets
1355 /// added to it as well to recursively simplify unless that alloca can be
1356 /// directly promoted. Finally, each time we rewrite a use of an alloca other
1357 /// the one being actively rewritten, we add it back onto the list if not
1358 /// already present to ensure it is re-visited.
1359 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
1361 /// \brief A collection of instructions to delete.
1362 /// We try to batch deletions to simplify code and make things a bit more
1364 SmallVector<Instruction *, 8> DeadInsts;
1366 /// \brief A set to prevent repeatedly marking an instruction split into many
1367 /// uses as dead. Only used to guard insertion into DeadInsts.
1368 SmallPtrSet<Instruction *, 4> DeadSplitInsts;
1370 /// \brief Post-promotion worklist.
1372 /// Sometimes we discover an alloca which has a high probability of becoming
1373 /// viable for SROA after a round of promotion takes place. In those cases,
1374 /// the alloca is enqueued here for re-processing.
1376 /// Note that we have to be very careful to clear allocas out of this list in
1377 /// the event they are deleted.
1378 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist;
1380 /// \brief A collection of alloca instructions we can directly promote.
1381 std::vector<AllocaInst *> PromotableAllocas;
1384 SROA(bool RequiresDomTree = true)
1385 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
1386 C(0), TD(0), DT(0) {
1387 initializeSROAPass(*PassRegistry::getPassRegistry());
1389 bool runOnFunction(Function &F);
1390 void getAnalysisUsage(AnalysisUsage &AU) const;
1392 const char *getPassName() const { return "SROA"; }
1396 friend class PHIOrSelectSpeculator;
1397 friend class AllocaPartitionRewriter;
1398 friend class AllocaPartitionVectorRewriter;
1400 bool rewriteAllocaPartition(AllocaInst &AI,
1401 AllocaPartitioning &P,
1402 AllocaPartitioning::iterator PI);
1403 bool splitAlloca(AllocaInst &AI, AllocaPartitioning &P);
1404 bool runOnAlloca(AllocaInst &AI);
1405 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
1406 bool promoteAllocas(Function &F);
1412 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
1413 return new SROA(RequiresDomTree);
1416 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
1418 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
1419 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
1423 /// \brief Visitor to speculate PHIs and Selects where possible.
1424 class PHIOrSelectSpeculator : public InstVisitor<PHIOrSelectSpeculator> {
1425 // Befriend the base class so it can delegate to private visit methods.
1426 friend class llvm::InstVisitor<PHIOrSelectSpeculator>;
1428 const TargetData &TD;
1429 AllocaPartitioning &P;
1433 PHIOrSelectSpeculator(const TargetData &TD, AllocaPartitioning &P, SROA &Pass)
1434 : TD(TD), P(P), Pass(Pass) {}
1436 /// \brief Visit the users of an alloca partition and rewrite them.
1437 void visitUsers(AllocaPartitioning::const_iterator PI) {
1438 // Note that we need to use an index here as the underlying vector of uses
1439 // may be grown during speculation. However, we never need to re-visit the
1440 // new uses, and so we can use the initial size bound.
1441 for (unsigned Idx = 0, Size = P.use_size(PI); Idx != Size; ++Idx) {
1442 const AllocaPartitioning::PartitionUse &PU = P.getUse(PI, Idx);
1444 continue; // Skip dead use.
1446 visit(cast<Instruction>(PU.U->getUser()));
1451 // By default, skip this instruction.
1452 void visitInstruction(Instruction &I) {}
1454 /// PHI instructions that use an alloca and are subsequently loaded can be
1455 /// rewritten to load both input pointers in the pred blocks and then PHI the
1456 /// results, allowing the load of the alloca to be promoted.
1458 /// %P2 = phi [i32* %Alloca, i32* %Other]
1459 /// %V = load i32* %P2
1461 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1463 /// %V2 = load i32* %Other
1465 /// %V = phi [i32 %V1, i32 %V2]
1467 /// We can do this to a select if its only uses are loads and if the operands
1468 /// to the select can be loaded unconditionally.
1470 /// FIXME: This should be hoisted into a generic utility, likely in
1471 /// Transforms/Util/Local.h
1472 bool isSafePHIToSpeculate(PHINode &PN, SmallVectorImpl<LoadInst *> &Loads) {
1473 // For now, we can only do this promotion if the load is in the same block
1474 // as the PHI, and if there are no stores between the phi and load.
1475 // TODO: Allow recursive phi users.
1476 // TODO: Allow stores.
1477 BasicBlock *BB = PN.getParent();
1478 unsigned MaxAlign = 0;
1479 for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end();
1481 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1482 if (LI == 0 || !LI->isSimple()) return false;
1484 // For now we only allow loads in the same block as the PHI. This is
1485 // a common case that happens when instcombine merges two loads through
1487 if (LI->getParent() != BB) return false;
1489 // Ensure that there are no instructions between the PHI and the load that
1491 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1492 if (BBI->mayWriteToMemory())
1495 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1496 Loads.push_back(LI);
1499 // We can only transform this if it is safe to push the loads into the
1500 // predecessor blocks. The only thing to watch out for is that we can't put
1501 // a possibly trapping load in the predecessor if it is a critical edge.
1502 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num;
1504 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1505 Value *InVal = PN.getIncomingValue(Idx);
1507 // If the value is produced by the terminator of the predecessor (an
1508 // invoke) or it has side-effects, there is no valid place to put a load
1509 // in the predecessor.
1510 if (TI == InVal || TI->mayHaveSideEffects())
1513 // If the predecessor has a single successor, then the edge isn't
1515 if (TI->getNumSuccessors() == 1)
1518 // If this pointer is always safe to load, or if we can prove that there
1519 // is already a load in the block, then we can move the load to the pred
1521 if (InVal->isDereferenceablePointer() ||
1522 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, &TD))
1531 void visitPHINode(PHINode &PN) {
1532 DEBUG(dbgs() << " original: " << PN << "\n");
1534 SmallVector<LoadInst *, 4> Loads;
1535 if (!isSafePHIToSpeculate(PN, Loads))
1538 assert(!Loads.empty());
1540 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1541 IRBuilder<> PHIBuilder(&PN);
1542 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1543 PN.getName() + ".sroa.speculated");
1545 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1546 // matter which one we get and if any differ, it doesn't matter.
1547 LoadInst *SomeLoad = cast<LoadInst>(Loads.back());
1548 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1549 unsigned Align = SomeLoad->getAlignment();
1551 // Rewrite all loads of the PN to use the new PHI.
1553 LoadInst *LI = Loads.pop_back_val();
1554 LI->replaceAllUsesWith(NewPN);
1555 Pass.DeadInsts.push_back(LI);
1556 } while (!Loads.empty());
1558 // Inject loads into all of the pred blocks.
1559 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1560 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1561 TerminatorInst *TI = Pred->getTerminator();
1562 Use *InUse = &PN.getOperandUse(PN.getOperandNumForIncomingValue(Idx));
1563 Value *InVal = PN.getIncomingValue(Idx);
1564 IRBuilder<> PredBuilder(TI);
1567 = PredBuilder.CreateLoad(InVal, (PN.getName() + ".sroa.speculate.load." +
1569 ++NumLoadsSpeculated;
1570 Load->setAlignment(Align);
1572 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1573 NewPN->addIncoming(Load, Pred);
1575 Instruction *Ptr = dyn_cast<Instruction>(InVal);
1577 // No uses to rewrite.
1580 // Try to lookup and rewrite any partition uses corresponding to this phi
1582 AllocaPartitioning::iterator PI
1583 = P.findPartitionForPHIOrSelectOperand(InUse);
1587 // Replace the Use in the PartitionUse for this operand with the Use
1589 AllocaPartitioning::use_iterator UI
1590 = P.findPartitionUseForPHIOrSelectOperand(InUse);
1591 assert(isa<PHINode>(*UI->U->getUser()));
1592 UI->U = &Load->getOperandUse(Load->getPointerOperandIndex());
1594 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1597 /// Select instructions that use an alloca and are subsequently loaded can be
1598 /// rewritten to load both input pointers and then select between the result,
1599 /// allowing the load of the alloca to be promoted.
1601 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1602 /// %V = load i32* %P2
1604 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1605 /// %V2 = load i32* %Other
1606 /// %V = select i1 %cond, i32 %V1, i32 %V2
1608 /// We can do this to a select if its only uses are loads and if the operand
1609 /// to the select can be loaded unconditionally.
1610 bool isSafeSelectToSpeculate(SelectInst &SI,
1611 SmallVectorImpl<LoadInst *> &Loads) {
1612 Value *TValue = SI.getTrueValue();
1613 Value *FValue = SI.getFalseValue();
1614 bool TDerefable = TValue->isDereferenceablePointer();
1615 bool FDerefable = FValue->isDereferenceablePointer();
1617 for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end();
1619 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1620 if (LI == 0 || !LI->isSimple()) return false;
1622 // Both operands to the select need to be dereferencable, either
1623 // absolutely (e.g. allocas) or at this point because we can see other
1625 if (!TDerefable && !isSafeToLoadUnconditionally(TValue, LI,
1626 LI->getAlignment(), &TD))
1628 if (!FDerefable && !isSafeToLoadUnconditionally(FValue, LI,
1629 LI->getAlignment(), &TD))
1631 Loads.push_back(LI);
1637 void visitSelectInst(SelectInst &SI) {
1638 DEBUG(dbgs() << " original: " << SI << "\n");
1639 IRBuilder<> IRB(&SI);
1641 // If the select isn't safe to speculate, just use simple logic to emit it.
1642 SmallVector<LoadInst *, 4> Loads;
1643 if (!isSafeSelectToSpeculate(SI, Loads))
1646 Use *Ops[2] = { &SI.getOperandUse(1), &SI.getOperandUse(2) };
1647 AllocaPartitioning::iterator PIs[2];
1648 AllocaPartitioning::PartitionUse PUs[2];
1649 for (unsigned i = 0, e = 2; i != e; ++i) {
1650 PIs[i] = P.findPartitionForPHIOrSelectOperand(Ops[i]);
1651 if (PIs[i] != P.end()) {
1652 // If the pointer is within the partitioning, remove the select from
1653 // its uses. We'll add in the new loads below.
1654 AllocaPartitioning::use_iterator UI
1655 = P.findPartitionUseForPHIOrSelectOperand(Ops[i]);
1657 // Clear out the use here so that the offsets into the use list remain
1658 // stable but this use is ignored when rewriting.
1663 Value *TV = SI.getTrueValue();
1664 Value *FV = SI.getFalseValue();
1665 // Replace the loads of the select with a select of two loads.
1666 while (!Loads.empty()) {
1667 LoadInst *LI = Loads.pop_back_val();
1669 IRB.SetInsertPoint(LI);
1671 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1673 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1674 NumLoadsSpeculated += 2;
1676 // Transfer alignment and TBAA info if present.
1677 TL->setAlignment(LI->getAlignment());
1678 FL->setAlignment(LI->getAlignment());
1679 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1680 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
1681 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
1684 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1685 LI->getName() + ".sroa.speculated");
1687 LoadInst *Loads[2] = { TL, FL };
1688 for (unsigned i = 0, e = 2; i != e; ++i) {
1689 if (PIs[i] != P.end()) {
1690 Use *LoadUse = &Loads[i]->getOperandUse(0);
1691 assert(PUs[i].U->get() == LoadUse->get());
1693 P.use_push_back(PIs[i], PUs[i]);
1697 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1698 LI->replaceAllUsesWith(V);
1699 Pass.DeadInsts.push_back(LI);
1705 /// \brief Accumulate the constant offsets in a GEP into a single APInt offset.
1707 /// If the provided GEP is all-constant, the total byte offset formed by the
1708 /// GEP is computed and Offset is set to it. If the GEP has any non-constant
1709 /// operands, the function returns false and the value of Offset is unmodified.
1710 static bool accumulateGEPOffsets(const TargetData &TD, GEPOperator &GEP,
1712 APInt GEPOffset(Offset.getBitWidth(), 0);
1713 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1714 GTI != GTE; ++GTI) {
1715 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
1718 if (OpC->isZero()) continue;
1720 // Handle a struct index, which adds its field offset to the pointer.
1721 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1722 unsigned ElementIdx = OpC->getZExtValue();
1723 const StructLayout *SL = TD.getStructLayout(STy);
1724 GEPOffset += APInt(Offset.getBitWidth(),
1725 SL->getElementOffset(ElementIdx));
1729 APInt TypeSize(Offset.getBitWidth(),
1730 TD.getTypeAllocSize(GTI.getIndexedType()));
1731 if (VectorType *VTy = dyn_cast<VectorType>(*GTI)) {
1732 assert((VTy->getScalarSizeInBits() % 8) == 0 &&
1733 "vector element size is not a multiple of 8, cannot GEP over it");
1734 TypeSize = VTy->getScalarSizeInBits() / 8;
1737 GEPOffset += OpC->getValue().sextOrTrunc(Offset.getBitWidth()) * TypeSize;
1743 /// \brief Build a GEP out of a base pointer and indices.
1745 /// This will return the BasePtr if that is valid, or build a new GEP
1746 /// instruction using the IRBuilder if GEP-ing is needed.
1747 static Value *buildGEP(IRBuilder<> &IRB, Value *BasePtr,
1748 SmallVectorImpl<Value *> &Indices,
1749 const Twine &Prefix) {
1750 if (Indices.empty())
1753 // A single zero index is a no-op, so check for this and avoid building a GEP
1755 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1758 return IRB.CreateInBoundsGEP(BasePtr, Indices, Prefix + ".idx");
1761 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1762 /// TargetTy without changing the offset of the pointer.
1764 /// This routine assumes we've already established a properly offset GEP with
1765 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1766 /// zero-indices down through type layers until we find one the same as
1767 /// TargetTy. If we can't find one with the same type, we at least try to use
1768 /// one with the same size. If none of that works, we just produce the GEP as
1769 /// indicated by Indices to have the correct offset.
1770 static Value *getNaturalGEPWithType(IRBuilder<> &IRB, const TargetData &TD,
1771 Value *BasePtr, Type *Ty, Type *TargetTy,
1772 SmallVectorImpl<Value *> &Indices,
1773 const Twine &Prefix) {
1775 return buildGEP(IRB, BasePtr, Indices, Prefix);
1777 // See if we can descend into a struct and locate a field with the correct
1779 unsigned NumLayers = 0;
1780 Type *ElementTy = Ty;
1782 if (ElementTy->isPointerTy())
1784 if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
1785 ElementTy = SeqTy->getElementType();
1786 Indices.push_back(IRB.getInt(APInt(TD.getPointerSizeInBits(), 0)));
1787 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1788 ElementTy = *STy->element_begin();
1789 Indices.push_back(IRB.getInt32(0));
1794 } while (ElementTy != TargetTy);
1795 if (ElementTy != TargetTy)
1796 Indices.erase(Indices.end() - NumLayers, Indices.end());
1798 return buildGEP(IRB, BasePtr, Indices, Prefix);
1801 /// \brief Recursively compute indices for a natural GEP.
1803 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1804 /// element types adding appropriate indices for the GEP.
1805 static Value *getNaturalGEPRecursively(IRBuilder<> &IRB, const TargetData &TD,
1806 Value *Ptr, Type *Ty, APInt &Offset,
1808 SmallVectorImpl<Value *> &Indices,
1809 const Twine &Prefix) {
1811 return getNaturalGEPWithType(IRB, TD, Ptr, Ty, TargetTy, Indices, Prefix);
1813 // We can't recurse through pointer types.
1814 if (Ty->isPointerTy())
1817 // We try to analyze GEPs over vectors here, but note that these GEPs are
1818 // extremely poorly defined currently. The long-term goal is to remove GEPing
1819 // over a vector from the IR completely.
1820 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1821 unsigned ElementSizeInBits = VecTy->getScalarSizeInBits();
1822 if (ElementSizeInBits % 8)
1823 return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
1824 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1825 APInt NumSkippedElements = Offset.udiv(ElementSize);
1826 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1828 Offset -= NumSkippedElements * ElementSize;
1829 Indices.push_back(IRB.getInt(NumSkippedElements));
1830 return getNaturalGEPRecursively(IRB, TD, Ptr, VecTy->getElementType(),
1831 Offset, TargetTy, Indices, Prefix);
1834 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1835 Type *ElementTy = ArrTy->getElementType();
1836 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1837 APInt NumSkippedElements = Offset.udiv(ElementSize);
1838 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1841 Offset -= NumSkippedElements * ElementSize;
1842 Indices.push_back(IRB.getInt(NumSkippedElements));
1843 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1847 StructType *STy = dyn_cast<StructType>(Ty);
1851 const StructLayout *SL = TD.getStructLayout(STy);
1852 uint64_t StructOffset = Offset.getZExtValue();
1853 if (StructOffset >= SL->getSizeInBytes())
1855 unsigned Index = SL->getElementContainingOffset(StructOffset);
1856 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1857 Type *ElementTy = STy->getElementType(Index);
1858 if (Offset.uge(TD.getTypeAllocSize(ElementTy)))
1859 return 0; // The offset points into alignment padding.
1861 Indices.push_back(IRB.getInt32(Index));
1862 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1866 /// \brief Get a natural GEP from a base pointer to a particular offset and
1867 /// resulting in a particular type.
1869 /// The goal is to produce a "natural" looking GEP that works with the existing
1870 /// composite types to arrive at the appropriate offset and element type for
1871 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1872 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1873 /// Indices, and setting Ty to the result subtype.
1875 /// If no natural GEP can be constructed, this function returns null.
1876 static Value *getNaturalGEPWithOffset(IRBuilder<> &IRB, const TargetData &TD,
1877 Value *Ptr, APInt Offset, Type *TargetTy,
1878 SmallVectorImpl<Value *> &Indices,
1879 const Twine &Prefix) {
1880 PointerType *Ty = cast<PointerType>(Ptr->getType());
1882 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1884 if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
1887 Type *ElementTy = Ty->getElementType();
1888 if (!ElementTy->isSized())
1889 return 0; // We can't GEP through an unsized element.
1890 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1891 if (ElementSize == 0)
1892 return 0; // Zero-length arrays can't help us build a natural GEP.
1893 APInt NumSkippedElements = Offset.udiv(ElementSize);
1895 Offset -= NumSkippedElements * ElementSize;
1896 Indices.push_back(IRB.getInt(NumSkippedElements));
1897 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1901 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1902 /// resulting pointer has PointerTy.
1904 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1905 /// and produces the pointer type desired. Where it cannot, it will try to use
1906 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1907 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1908 /// bitcast to the type.
1910 /// The strategy for finding the more natural GEPs is to peel off layers of the
1911 /// pointer, walking back through bit casts and GEPs, searching for a base
1912 /// pointer from which we can compute a natural GEP with the desired
1913 /// properities. The algorithm tries to fold as many constant indices into
1914 /// a single GEP as possible, thus making each GEP more independent of the
1915 /// surrounding code.
1916 static Value *getAdjustedPtr(IRBuilder<> &IRB, const TargetData &TD,
1917 Value *Ptr, APInt Offset, Type *PointerTy,
1918 const Twine &Prefix) {
1919 // Even though we don't look through PHI nodes, we could be called on an
1920 // instruction in an unreachable block, which may be on a cycle.
1921 SmallPtrSet<Value *, 4> Visited;
1922 Visited.insert(Ptr);
1923 SmallVector<Value *, 4> Indices;
1925 // We may end up computing an offset pointer that has the wrong type. If we
1926 // never are able to compute one directly that has the correct type, we'll
1927 // fall back to it, so keep it around here.
1928 Value *OffsetPtr = 0;
1930 // Remember any i8 pointer we come across to re-use if we need to do a raw
1933 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1935 Type *TargetTy = PointerTy->getPointerElementType();
1938 // First fold any existing GEPs into the offset.
1939 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1940 APInt GEPOffset(Offset.getBitWidth(), 0);
1941 if (!accumulateGEPOffsets(TD, *GEP, GEPOffset))
1943 Offset += GEPOffset;
1944 Ptr = GEP->getPointerOperand();
1945 if (!Visited.insert(Ptr))
1949 // See if we can perform a natural GEP here.
1951 if (Value *P = getNaturalGEPWithOffset(IRB, TD, Ptr, Offset, TargetTy,
1953 if (P->getType() == PointerTy) {
1954 // Zap any offset pointer that we ended up computing in previous rounds.
1955 if (OffsetPtr && OffsetPtr->use_empty())
1956 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1957 I->eraseFromParent();
1965 // Stash this pointer if we've found an i8*.
1966 if (Ptr->getType()->isIntegerTy(8)) {
1968 Int8PtrOffset = Offset;
1971 // Peel off a layer of the pointer and update the offset appropriately.
1972 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1973 Ptr = cast<Operator>(Ptr)->getOperand(0);
1974 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1975 if (GA->mayBeOverridden())
1977 Ptr = GA->getAliasee();
1981 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1982 } while (Visited.insert(Ptr));
1986 Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
1987 Prefix + ".raw_cast");
1988 Int8PtrOffset = Offset;
1991 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1992 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1993 Prefix + ".raw_idx");
1997 // On the off chance we were targeting i8*, guard the bitcast here.
1998 if (Ptr->getType() != PointerTy)
1999 Ptr = IRB.CreateBitCast(Ptr, PointerTy, Prefix + ".cast");
2004 /// \brief Test whether the given alloca partition can be promoted to a vector.
2006 /// This is a quick test to check whether we can rewrite a particular alloca
2007 /// partition (and its newly formed alloca) into a vector alloca with only
2008 /// whole-vector loads and stores such that it could be promoted to a vector
2009 /// SSA value. We only can ensure this for a limited set of operations, and we
2010 /// don't want to do the rewrites unless we are confident that the result will
2011 /// be promotable, so we have an early test here.
2012 static bool isVectorPromotionViable(const TargetData &TD,
2014 AllocaPartitioning &P,
2015 uint64_t PartitionBeginOffset,
2016 uint64_t PartitionEndOffset,
2017 AllocaPartitioning::const_use_iterator I,
2018 AllocaPartitioning::const_use_iterator E) {
2019 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
2023 uint64_t VecSize = TD.getTypeSizeInBits(Ty);
2024 uint64_t ElementSize = Ty->getScalarSizeInBits();
2026 // While the definition of LLVM vectors is bitpacked, we don't support sizes
2027 // that aren't byte sized.
2028 if (ElementSize % 8)
2030 assert((VecSize % 8) == 0 && "vector size not a multiple of element size?");
2034 for (; I != E; ++I) {
2036 continue; // Skip dead use.
2038 uint64_t BeginOffset = I->BeginOffset - PartitionBeginOffset;
2039 uint64_t BeginIndex = BeginOffset / ElementSize;
2040 if (BeginIndex * ElementSize != BeginOffset ||
2041 BeginIndex >= Ty->getNumElements())
2043 uint64_t EndOffset = I->EndOffset - PartitionBeginOffset;
2044 uint64_t EndIndex = EndOffset / ElementSize;
2045 if (EndIndex * ElementSize != EndOffset ||
2046 EndIndex > Ty->getNumElements())
2049 // FIXME: We should build shuffle vector instructions to handle
2050 // non-element-sized accesses.
2051 if ((EndOffset - BeginOffset) != ElementSize &&
2052 (EndOffset - BeginOffset) != VecSize)
2055 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
2056 if (MI->isVolatile())
2058 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
2059 const AllocaPartitioning::MemTransferOffsets &MTO
2060 = P.getMemTransferOffsets(*MTI);
2061 if (!MTO.IsSplittable)
2064 } else if (I->U->get()->getType()->getPointerElementType()->isStructTy()) {
2065 // Disable vector promotion when there are loads or stores of an FCA.
2067 } else if (!isa<LoadInst>(I->U->getUser()) &&
2068 !isa<StoreInst>(I->U->getUser())) {
2075 /// \brief Test whether the given alloca partition can be promoted to an int.
2077 /// This is a quick test to check whether we can rewrite a particular alloca
2078 /// partition (and its newly formed alloca) into an integer alloca suitable for
2079 /// promotion to an SSA value. We only can ensure this for a limited set of
2080 /// operations, and we don't want to do the rewrites unless we are confident
2081 /// that the result will be promotable, so we have an early test here.
2082 static bool isIntegerPromotionViable(const TargetData &TD,
2084 uint64_t AllocBeginOffset,
2085 AllocaPartitioning &P,
2086 AllocaPartitioning::const_use_iterator I,
2087 AllocaPartitioning::const_use_iterator E) {
2088 IntegerType *Ty = dyn_cast<IntegerType>(AllocaTy);
2089 if (!Ty || 8*TD.getTypeStoreSize(Ty) != Ty->getBitWidth())
2092 // Check the uses to ensure the uses are (likely) promoteable integer uses.
2093 // Also ensure that the alloca has a covering load or store. We don't want
2094 // promote because of some other unsplittable entry (which we may make
2095 // splittable later) and lose the ability to promote each element access.
2096 bool WholeAllocaOp = false;
2097 for (; I != E; ++I) {
2099 continue; // Skip dead use.
2101 // We can't reasonably handle cases where the load or store extends past
2102 // the end of the aloca's type and into its padding.
2103 if ((I->EndOffset - AllocBeginOffset) > TD.getTypeStoreSize(Ty))
2106 if (LoadInst *LI = dyn_cast<LoadInst>(I->U->getUser())) {
2107 if (LI->isVolatile() || !LI->getType()->isIntegerTy())
2109 if (LI->getType() == Ty)
2110 WholeAllocaOp = true;
2111 } else if (StoreInst *SI = dyn_cast<StoreInst>(I->U->getUser())) {
2112 if (SI->isVolatile() || !SI->getValueOperand()->getType()->isIntegerTy())
2114 if (SI->getValueOperand()->getType() == Ty)
2115 WholeAllocaOp = true;
2116 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
2117 if (MI->isVolatile())
2119 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
2120 const AllocaPartitioning::MemTransferOffsets &MTO
2121 = P.getMemTransferOffsets(*MTI);
2122 if (!MTO.IsSplittable)
2129 return WholeAllocaOp;
2133 /// \brief Visitor to rewrite instructions using a partition of an alloca to
2134 /// use a new alloca.
2136 /// Also implements the rewriting to vector-based accesses when the partition
2137 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2139 class AllocaPartitionRewriter : public InstVisitor<AllocaPartitionRewriter,
2141 // Befriend the base class so it can delegate to private visit methods.
2142 friend class llvm::InstVisitor<AllocaPartitionRewriter, bool>;
2144 const TargetData &TD;
2145 AllocaPartitioning &P;
2147 AllocaInst &OldAI, &NewAI;
2148 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2150 // If we are rewriting an alloca partition which can be written as pure
2151 // vector operations, we stash extra information here. When VecTy is
2152 // non-null, we have some strict guarantees about the rewriten alloca:
2153 // - The new alloca is exactly the size of the vector type here.
2154 // - The accesses all either map to the entire vector or to a single
2156 // - The set of accessing instructions is only one of those handled above
2157 // in isVectorPromotionViable. Generally these are the same access kinds
2158 // which are promotable via mem2reg.
2161 uint64_t ElementSize;
2163 // This is a convenience and flag variable that will be null unless the new
2164 // alloca has a promotion-targeted integer type due to passing
2165 // isIntegerPromotionViable above. If it is non-null does, the desired
2166 // integer type will be stored here for easy access during rewriting.
2167 IntegerType *IntPromotionTy;
2169 // The offset of the partition user currently being rewritten.
2170 uint64_t BeginOffset, EndOffset;
2172 Instruction *OldPtr;
2174 // The name prefix to use when rewriting instructions for this alloca.
2175 std::string NamePrefix;
2178 AllocaPartitionRewriter(const TargetData &TD, AllocaPartitioning &P,
2179 AllocaPartitioning::iterator PI,
2180 SROA &Pass, AllocaInst &OldAI, AllocaInst &NewAI,
2181 uint64_t NewBeginOffset, uint64_t NewEndOffset)
2182 : TD(TD), P(P), Pass(Pass),
2183 OldAI(OldAI), NewAI(NewAI),
2184 NewAllocaBeginOffset(NewBeginOffset),
2185 NewAllocaEndOffset(NewEndOffset),
2186 VecTy(), ElementTy(), ElementSize(), IntPromotionTy(),
2187 BeginOffset(), EndOffset() {
2190 /// \brief Visit the users of the alloca partition and rewrite them.
2191 bool visitUsers(AllocaPartitioning::const_use_iterator I,
2192 AllocaPartitioning::const_use_iterator E) {
2193 if (isVectorPromotionViable(TD, NewAI.getAllocatedType(), P,
2194 NewAllocaBeginOffset, NewAllocaEndOffset,
2197 VecTy = cast<VectorType>(NewAI.getAllocatedType());
2198 ElementTy = VecTy->getElementType();
2199 assert((VecTy->getScalarSizeInBits() % 8) == 0 &&
2200 "Only multiple-of-8 sized vector elements are viable");
2201 ElementSize = VecTy->getScalarSizeInBits() / 8;
2202 } else if (isIntegerPromotionViable(TD, NewAI.getAllocatedType(),
2203 NewAllocaBeginOffset, P, I, E)) {
2204 IntPromotionTy = cast<IntegerType>(NewAI.getAllocatedType());
2206 bool CanSROA = true;
2207 for (; I != E; ++I) {
2209 continue; // Skip dead uses.
2210 BeginOffset = I->BeginOffset;
2211 EndOffset = I->EndOffset;
2213 OldPtr = cast<Instruction>(I->U->get());
2214 NamePrefix = (Twine(NewAI.getName()) + "." + Twine(BeginOffset)).str();
2215 CanSROA &= visit(cast<Instruction>(I->U->getUser()));
2227 // Every instruction which can end up as a user must have a rewrite rule.
2228 bool visitInstruction(Instruction &I) {
2229 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2230 llvm_unreachable("No rewrite rule for this instruction!");
2233 Twine getName(const Twine &Suffix) {
2234 return NamePrefix + Suffix;
2237 Value *getAdjustedAllocaPtr(IRBuilder<> &IRB, Type *PointerTy) {
2238 assert(BeginOffset >= NewAllocaBeginOffset);
2239 APInt Offset(TD.getPointerSizeInBits(), BeginOffset - NewAllocaBeginOffset);
2240 return getAdjustedPtr(IRB, TD, &NewAI, Offset, PointerTy, getName(""));
2243 /// \brief Compute suitable alignment to access an offset into the new alloca.
2244 unsigned getOffsetAlign(uint64_t Offset) {
2245 unsigned NewAIAlign = NewAI.getAlignment();
2247 NewAIAlign = TD.getABITypeAlignment(NewAI.getAllocatedType());
2248 return MinAlign(NewAIAlign, Offset);
2251 /// \brief Compute suitable alignment to access this partition of the new
2253 unsigned getPartitionAlign() {
2254 return getOffsetAlign(BeginOffset - NewAllocaBeginOffset);
2257 /// \brief Compute suitable alignment to access a type at an offset of the
2260 /// \returns zero if the type's ABI alignment is a suitable alignment,
2261 /// otherwise returns the maximal suitable alignment.
2262 unsigned getOffsetTypeAlign(Type *Ty, uint64_t Offset) {
2263 unsigned Align = getOffsetAlign(Offset);
2264 return Align == TD.getABITypeAlignment(Ty) ? 0 : Align;
2267 /// \brief Compute suitable alignment to access a type at the beginning of
2268 /// this partition of the new alloca.
2270 /// See \c getOffsetTypeAlign for details; this routine delegates to it.
2271 unsigned getPartitionTypeAlign(Type *Ty) {
2272 return getOffsetTypeAlign(Ty, BeginOffset - NewAllocaBeginOffset);
2275 ConstantInt *getIndex(IRBuilder<> &IRB, uint64_t Offset) {
2276 assert(VecTy && "Can only call getIndex when rewriting a vector");
2277 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2278 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2279 uint32_t Index = RelOffset / ElementSize;
2280 assert(Index * ElementSize == RelOffset);
2281 return IRB.getInt32(Index);
2284 Value *extractInteger(IRBuilder<> &IRB, IntegerType *TargetTy,
2286 assert(IntPromotionTy && "Alloca is not an integer we can extract from");
2287 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2289 assert(Offset >= NewAllocaBeginOffset && "Out of bounds offset");
2290 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2291 assert(TD.getTypeStoreSize(TargetTy) + RelOffset <=
2292 TD.getTypeStoreSize(IntPromotionTy) &&
2293 "Element load outside of alloca store");
2294 uint64_t ShAmt = 8*RelOffset;
2295 if (TD.isBigEndian())
2296 ShAmt = 8*(TD.getTypeStoreSize(IntPromotionTy) -
2297 TD.getTypeStoreSize(TargetTy) - RelOffset);
2299 V = IRB.CreateLShr(V, ShAmt, getName(".shift"));
2300 if (TargetTy != IntPromotionTy) {
2301 assert(TargetTy->getBitWidth() < IntPromotionTy->getBitWidth() &&
2302 "Cannot extract to a larger integer!");
2303 V = IRB.CreateTrunc(V, TargetTy, getName(".trunc"));
2308 StoreInst *insertInteger(IRBuilder<> &IRB, Value *V, uint64_t Offset) {
2309 IntegerType *Ty = cast<IntegerType>(V->getType());
2310 if (Ty == IntPromotionTy)
2311 return IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2313 assert(Ty->getBitWidth() < IntPromotionTy->getBitWidth() &&
2314 "Cannot insert a larger integer!");
2315 V = IRB.CreateZExt(V, IntPromotionTy, getName(".ext"));
2316 assert(Offset >= NewAllocaBeginOffset && "Out of bounds offset");
2317 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2318 assert(TD.getTypeStoreSize(Ty) + RelOffset <=
2319 TD.getTypeStoreSize(IntPromotionTy) &&
2320 "Element store outside of alloca store");
2321 uint64_t ShAmt = 8*RelOffset;
2322 if (TD.isBigEndian())
2323 ShAmt = 8*(TD.getTypeStoreSize(IntPromotionTy) - TD.getTypeStoreSize(Ty)
2326 V = IRB.CreateShl(V, ShAmt, getName(".shift"));
2328 APInt Mask = ~Ty->getMask().zext(IntPromotionTy->getBitWidth()).shl(ShAmt);
2329 Value *Old = IRB.CreateAnd(IRB.CreateAlignedLoad(&NewAI,
2330 NewAI.getAlignment(),
2331 getName(".oldload")),
2332 Mask, getName(".mask"));
2333 return IRB.CreateAlignedStore(IRB.CreateOr(Old, V, getName(".insert")),
2334 &NewAI, NewAI.getAlignment());
2337 void deleteIfTriviallyDead(Value *V) {
2338 Instruction *I = cast<Instruction>(V);
2339 if (isInstructionTriviallyDead(I))
2340 Pass.DeadInsts.push_back(I);
2343 Value *getValueCast(IRBuilder<> &IRB, Value *V, Type *Ty) {
2344 if (V->getType()->isIntegerTy() && Ty->isPointerTy())
2345 return IRB.CreateIntToPtr(V, Ty);
2346 if (V->getType()->isPointerTy() && Ty->isIntegerTy())
2347 return IRB.CreatePtrToInt(V, Ty);
2349 return IRB.CreateBitCast(V, Ty);
2352 bool rewriteVectorizedLoadInst(IRBuilder<> &IRB, LoadInst &LI, Value *OldOp) {
2354 if (LI.getType() == VecTy->getElementType() ||
2355 BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
2356 Result = IRB.CreateExtractElement(
2357 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), getName(".load")),
2358 getIndex(IRB, BeginOffset), getName(".extract"));
2360 Result = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2363 if (Result->getType() != LI.getType())
2364 Result = getValueCast(IRB, Result, LI.getType());
2365 LI.replaceAllUsesWith(Result);
2366 Pass.DeadInsts.push_back(&LI);
2368 DEBUG(dbgs() << " to: " << *Result << "\n");
2372 bool rewriteIntegerLoad(IRBuilder<> &IRB, LoadInst &LI) {
2373 assert(!LI.isVolatile());
2374 Value *Result = extractInteger(IRB, cast<IntegerType>(LI.getType()),
2376 LI.replaceAllUsesWith(Result);
2377 Pass.DeadInsts.push_back(&LI);
2378 DEBUG(dbgs() << " to: " << *Result << "\n");
2382 bool visitLoadInst(LoadInst &LI) {
2383 DEBUG(dbgs() << " original: " << LI << "\n");
2384 Value *OldOp = LI.getOperand(0);
2385 assert(OldOp == OldPtr);
2386 IRBuilder<> IRB(&LI);
2389 return rewriteVectorizedLoadInst(IRB, LI, OldOp);
2391 return rewriteIntegerLoad(IRB, LI);
2393 Value *NewPtr = getAdjustedAllocaPtr(IRB,
2394 LI.getPointerOperand()->getType());
2395 LI.setOperand(0, NewPtr);
2396 LI.setAlignment(getPartitionTypeAlign(LI.getType()));
2397 DEBUG(dbgs() << " to: " << LI << "\n");
2399 deleteIfTriviallyDead(OldOp);
2400 return NewPtr == &NewAI && !LI.isVolatile();
2403 bool rewriteVectorizedStoreInst(IRBuilder<> &IRB, StoreInst &SI,
2405 Value *V = SI.getValueOperand();
2406 if (V->getType() == ElementTy ||
2407 BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
2408 if (V->getType() != ElementTy)
2409 V = getValueCast(IRB, V, ElementTy);
2410 LoadInst *LI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2412 V = IRB.CreateInsertElement(LI, V, getIndex(IRB, BeginOffset),
2413 getName(".insert"));
2414 } else if (V->getType() != VecTy) {
2415 V = getValueCast(IRB, V, VecTy);
2417 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2418 Pass.DeadInsts.push_back(&SI);
2421 DEBUG(dbgs() << " to: " << *Store << "\n");
2425 bool rewriteIntegerStore(IRBuilder<> &IRB, StoreInst &SI) {
2426 assert(!SI.isVolatile());
2427 StoreInst *Store = insertInteger(IRB, SI.getValueOperand(), BeginOffset);
2428 Pass.DeadInsts.push_back(&SI);
2430 DEBUG(dbgs() << " to: " << *Store << "\n");
2434 bool visitStoreInst(StoreInst &SI) {
2435 DEBUG(dbgs() << " original: " << SI << "\n");
2436 Value *OldOp = SI.getOperand(1);
2437 assert(OldOp == OldPtr);
2438 IRBuilder<> IRB(&SI);
2441 return rewriteVectorizedStoreInst(IRB, SI, OldOp);
2443 return rewriteIntegerStore(IRB, SI);
2445 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2446 // alloca that should be re-examined after promoting this alloca.
2447 if (SI.getValueOperand()->getType()->isPointerTy())
2448 if (AllocaInst *AI = dyn_cast<AllocaInst>(SI.getValueOperand()
2449 ->stripInBoundsOffsets()))
2450 Pass.PostPromotionWorklist.insert(AI);
2452 Value *NewPtr = getAdjustedAllocaPtr(IRB,
2453 SI.getPointerOperand()->getType());
2454 SI.setOperand(1, NewPtr);
2455 SI.setAlignment(getPartitionTypeAlign(SI.getValueOperand()->getType()));
2456 DEBUG(dbgs() << " to: " << SI << "\n");
2458 deleteIfTriviallyDead(OldOp);
2459 return NewPtr == &NewAI && !SI.isVolatile();
2462 bool visitMemSetInst(MemSetInst &II) {
2463 DEBUG(dbgs() << " original: " << II << "\n");
2464 IRBuilder<> IRB(&II);
2465 assert(II.getRawDest() == OldPtr);
2467 // If the memset has a variable size, it cannot be split, just adjust the
2468 // pointer to the new alloca.
2469 if (!isa<Constant>(II.getLength())) {
2470 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2471 Type *CstTy = II.getAlignmentCst()->getType();
2472 II.setAlignment(ConstantInt::get(CstTy, getPartitionAlign()));
2474 deleteIfTriviallyDead(OldPtr);
2478 // Record this instruction for deletion.
2479 if (Pass.DeadSplitInsts.insert(&II))
2480 Pass.DeadInsts.push_back(&II);
2482 Type *AllocaTy = NewAI.getAllocatedType();
2483 Type *ScalarTy = AllocaTy->getScalarType();
2485 // If this doesn't map cleanly onto the alloca type, and that type isn't
2486 // a single value type, just emit a memset.
2487 if (!VecTy && (BeginOffset != NewAllocaBeginOffset ||
2488 EndOffset != NewAllocaEndOffset ||
2489 !AllocaTy->isSingleValueType() ||
2490 !TD.isLegalInteger(TD.getTypeSizeInBits(ScalarTy)))) {
2491 Type *SizeTy = II.getLength()->getType();
2492 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2494 = IRB.CreateMemSet(getAdjustedAllocaPtr(IRB,
2495 II.getRawDest()->getType()),
2496 II.getValue(), Size, getPartitionAlign(),
2499 DEBUG(dbgs() << " to: " << *New << "\n");
2503 // If we can represent this as a simple value, we have to build the actual
2504 // value to store, which requires expanding the byte present in memset to
2505 // a sensible representation for the alloca type. This is essentially
2506 // splatting the byte to a sufficiently wide integer, bitcasting to the
2507 // desired scalar type, and splatting it across any desired vector type.
2508 Value *V = II.getValue();
2509 IntegerType *VTy = cast<IntegerType>(V->getType());
2510 Type *IntTy = Type::getIntNTy(VTy->getContext(),
2511 TD.getTypeSizeInBits(ScalarTy));
2512 if (TD.getTypeSizeInBits(ScalarTy) > VTy->getBitWidth())
2513 V = IRB.CreateMul(IRB.CreateZExt(V, IntTy, getName(".zext")),
2514 ConstantExpr::getUDiv(
2515 Constant::getAllOnesValue(IntTy),
2516 ConstantExpr::getZExt(
2517 Constant::getAllOnesValue(V->getType()),
2519 getName(".isplat"));
2520 if (V->getType() != ScalarTy) {
2521 if (ScalarTy->isPointerTy())
2522 V = IRB.CreateIntToPtr(V, ScalarTy);
2523 else if (ScalarTy->isPrimitiveType() || ScalarTy->isVectorTy())
2524 V = IRB.CreateBitCast(V, ScalarTy);
2525 else if (ScalarTy->isIntegerTy())
2526 llvm_unreachable("Computed different integer types with equal widths");
2528 llvm_unreachable("Invalid scalar type");
2531 // If this is an element-wide memset of a vectorizable alloca, insert it.
2532 if (VecTy && (BeginOffset > NewAllocaBeginOffset ||
2533 EndOffset < NewAllocaEndOffset)) {
2534 StoreInst *Store = IRB.CreateAlignedStore(
2535 IRB.CreateInsertElement(IRB.CreateAlignedLoad(&NewAI,
2536 NewAI.getAlignment(),
2538 V, getIndex(IRB, BeginOffset),
2539 getName(".insert")),
2540 &NewAI, NewAI.getAlignment());
2542 DEBUG(dbgs() << " to: " << *Store << "\n");
2546 // Splat to a vector if needed.
2547 if (VectorType *VecTy = dyn_cast<VectorType>(AllocaTy)) {
2548 VectorType *SplatSourceTy = VectorType::get(V->getType(), 1);
2549 V = IRB.CreateShuffleVector(
2550 IRB.CreateInsertElement(UndefValue::get(SplatSourceTy), V,
2551 IRB.getInt32(0), getName(".vsplat.insert")),
2552 UndefValue::get(SplatSourceTy),
2553 ConstantVector::getSplat(VecTy->getNumElements(), IRB.getInt32(0)),
2554 getName(".vsplat.shuffle"));
2555 assert(V->getType() == VecTy);
2558 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2561 DEBUG(dbgs() << " to: " << *New << "\n");
2562 return !II.isVolatile();
2565 bool visitMemTransferInst(MemTransferInst &II) {
2566 // Rewriting of memory transfer instructions can be a bit tricky. We break
2567 // them into two categories: split intrinsics and unsplit intrinsics.
2569 DEBUG(dbgs() << " original: " << II << "\n");
2570 IRBuilder<> IRB(&II);
2572 assert(II.getRawSource() == OldPtr || II.getRawDest() == OldPtr);
2573 bool IsDest = II.getRawDest() == OldPtr;
2575 const AllocaPartitioning::MemTransferOffsets &MTO
2576 = P.getMemTransferOffsets(II);
2578 // Compute the relative offset within the transfer.
2579 unsigned IntPtrWidth = TD.getPointerSizeInBits();
2580 APInt RelOffset(IntPtrWidth, BeginOffset - (IsDest ? MTO.DestBegin
2581 : MTO.SourceBegin));
2583 unsigned Align = II.getAlignment();
2585 Align = MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(),
2586 MinAlign(II.getAlignment(), getPartitionAlign()));
2588 // For unsplit intrinsics, we simply modify the source and destination
2589 // pointers in place. This isn't just an optimization, it is a matter of
2590 // correctness. With unsplit intrinsics we may be dealing with transfers
2591 // within a single alloca before SROA ran, or with transfers that have
2592 // a variable length. We may also be dealing with memmove instead of
2593 // memcpy, and so simply updating the pointers is the necessary for us to
2594 // update both source and dest of a single call.
2595 if (!MTO.IsSplittable) {
2596 Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource();
2598 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2600 II.setSource(getAdjustedAllocaPtr(IRB, II.getRawSource()->getType()));
2602 Type *CstTy = II.getAlignmentCst()->getType();
2603 II.setAlignment(ConstantInt::get(CstTy, Align));
2605 DEBUG(dbgs() << " to: " << II << "\n");
2606 deleteIfTriviallyDead(OldOp);
2609 // For split transfer intrinsics we have an incredibly useful assurance:
2610 // the source and destination do not reside within the same alloca, and at
2611 // least one of them does not escape. This means that we can replace
2612 // memmove with memcpy, and we don't need to worry about all manner of
2613 // downsides to splitting and transforming the operations.
2615 // If this doesn't map cleanly onto the alloca type, and that type isn't
2616 // a single value type, just emit a memcpy.
2618 = !VecTy && (BeginOffset != NewAllocaBeginOffset ||
2619 EndOffset != NewAllocaEndOffset ||
2620 !NewAI.getAllocatedType()->isSingleValueType());
2622 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2623 // size hasn't been shrunk based on analysis of the viable range, this is
2625 if (EmitMemCpy && &OldAI == &NewAI) {
2626 uint64_t OrigBegin = IsDest ? MTO.DestBegin : MTO.SourceBegin;
2627 uint64_t OrigEnd = IsDest ? MTO.DestEnd : MTO.SourceEnd;
2628 // Ensure the start lines up.
2629 assert(BeginOffset == OrigBegin);
2632 // Rewrite the size as needed.
2633 if (EndOffset != OrigEnd)
2634 II.setLength(ConstantInt::get(II.getLength()->getType(),
2635 EndOffset - BeginOffset));
2638 // Record this instruction for deletion.
2639 if (Pass.DeadSplitInsts.insert(&II))
2640 Pass.DeadInsts.push_back(&II);
2642 bool IsVectorElement = VecTy && (BeginOffset > NewAllocaBeginOffset ||
2643 EndOffset < NewAllocaEndOffset);
2645 Type *OtherPtrTy = IsDest ? II.getRawSource()->getType()
2646 : II.getRawDest()->getType();
2648 OtherPtrTy = IsVectorElement ? VecTy->getElementType()->getPointerTo()
2651 // Compute the other pointer, folding as much as possible to produce
2652 // a single, simple GEP in most cases.
2653 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2654 OtherPtr = getAdjustedPtr(IRB, TD, OtherPtr, RelOffset, OtherPtrTy,
2655 getName("." + OtherPtr->getName()));
2657 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2658 // alloca that should be re-examined after rewriting this instruction.
2660 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets()))
2661 Pass.Worklist.insert(AI);
2665 = getAdjustedAllocaPtr(IRB, IsDest ? II.getRawDest()->getType()
2666 : II.getRawSource()->getType());
2667 Type *SizeTy = II.getLength()->getType();
2668 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2670 CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
2671 IsDest ? OtherPtr : OurPtr,
2672 Size, Align, II.isVolatile());
2674 DEBUG(dbgs() << " to: " << *New << "\n");
2678 // Note that we clamp the alignment to 1 here as a 0 alignment for a memcpy
2679 // is equivalent to 1, but that isn't true if we end up rewriting this as
2684 Value *SrcPtr = OtherPtr;
2685 Value *DstPtr = &NewAI;
2687 std::swap(SrcPtr, DstPtr);
2690 if (IsVectorElement && !IsDest) {
2691 // We have to extract rather than load.
2692 Src = IRB.CreateExtractElement(
2693 IRB.CreateAlignedLoad(SrcPtr, Align, getName(".copyload")),
2694 getIndex(IRB, BeginOffset),
2695 getName(".copyextract"));
2697 Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(),
2698 getName(".copyload"));
2701 if (IsVectorElement && IsDest) {
2702 // We have to insert into a loaded copy before storing.
2703 Src = IRB.CreateInsertElement(
2704 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), getName(".load")),
2705 Src, getIndex(IRB, BeginOffset),
2706 getName(".insert"));
2709 StoreInst *Store = cast<StoreInst>(
2710 IRB.CreateAlignedStore(Src, DstPtr, Align, II.isVolatile()));
2712 DEBUG(dbgs() << " to: " << *Store << "\n");
2713 return !II.isVolatile();
2716 bool visitIntrinsicInst(IntrinsicInst &II) {
2717 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2718 II.getIntrinsicID() == Intrinsic::lifetime_end);
2719 DEBUG(dbgs() << " original: " << II << "\n");
2720 IRBuilder<> IRB(&II);
2721 assert(II.getArgOperand(1) == OldPtr);
2723 // Record this instruction for deletion.
2724 if (Pass.DeadSplitInsts.insert(&II))
2725 Pass.DeadInsts.push_back(&II);
2728 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2729 EndOffset - BeginOffset);
2730 Value *Ptr = getAdjustedAllocaPtr(IRB, II.getArgOperand(1)->getType());
2732 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2733 New = IRB.CreateLifetimeStart(Ptr, Size);
2735 New = IRB.CreateLifetimeEnd(Ptr, Size);
2737 DEBUG(dbgs() << " to: " << *New << "\n");
2741 bool visitPHINode(PHINode &PN) {
2742 DEBUG(dbgs() << " original: " << PN << "\n");
2744 // We would like to compute a new pointer in only one place, but have it be
2745 // as local as possible to the PHI. To do that, we re-use the location of
2746 // the old pointer, which necessarily must be in the right position to
2747 // dominate the PHI.
2748 IRBuilder<> PtrBuilder(cast<Instruction>(OldPtr));
2750 Value *NewPtr = getAdjustedAllocaPtr(PtrBuilder, OldPtr->getType());
2751 // Replace the operands which were using the old pointer.
2752 User::op_iterator OI = PN.op_begin(), OE = PN.op_end();
2753 for (; OI != OE; ++OI)
2757 DEBUG(dbgs() << " to: " << PN << "\n");
2758 deleteIfTriviallyDead(OldPtr);
2762 bool visitSelectInst(SelectInst &SI) {
2763 DEBUG(dbgs() << " original: " << SI << "\n");
2764 IRBuilder<> IRB(&SI);
2766 // Find the operand we need to rewrite here.
2767 bool IsTrueVal = SI.getTrueValue() == OldPtr;
2769 assert(SI.getFalseValue() != OldPtr && "Pointer is both operands!");
2771 assert(SI.getFalseValue() == OldPtr && "Pointer isn't an operand!");
2773 Value *NewPtr = getAdjustedAllocaPtr(IRB, OldPtr->getType());
2774 SI.setOperand(IsTrueVal ? 1 : 2, NewPtr);
2775 DEBUG(dbgs() << " to: " << SI << "\n");
2776 deleteIfTriviallyDead(OldPtr);
2784 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2786 /// This pass aggressively rewrites all aggregate loads and stores on
2787 /// a particular pointer (or any pointer derived from it which we can identify)
2788 /// with scalar loads and stores.
2789 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2790 // Befriend the base class so it can delegate to private visit methods.
2791 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2793 const TargetData &TD;
2795 /// Queue of pointer uses to analyze and potentially rewrite.
2796 SmallVector<Use *, 8> Queue;
2798 /// Set to prevent us from cycling with phi nodes and loops.
2799 SmallPtrSet<User *, 8> Visited;
2801 /// The current pointer use being rewritten. This is used to dig up the used
2802 /// value (as opposed to the user).
2806 AggLoadStoreRewriter(const TargetData &TD) : TD(TD) {}
2808 /// Rewrite loads and stores through a pointer and all pointers derived from
2810 bool rewrite(Instruction &I) {
2811 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2813 bool Changed = false;
2814 while (!Queue.empty()) {
2815 U = Queue.pop_back_val();
2816 Changed |= visit(cast<Instruction>(U->getUser()));
2822 /// Enqueue all the users of the given instruction for further processing.
2823 /// This uses a set to de-duplicate users.
2824 void enqueueUsers(Instruction &I) {
2825 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
2827 if (Visited.insert(*UI))
2828 Queue.push_back(&UI.getUse());
2831 // Conservative default is to not rewrite anything.
2832 bool visitInstruction(Instruction &I) { return false; }
2834 /// \brief Generic recursive split emission class.
2835 template <typename Derived>
2838 /// The builder used to form new instructions.
2840 /// The indices which to be used with insert- or extractvalue to select the
2841 /// appropriate value within the aggregate.
2842 SmallVector<unsigned, 4> Indices;
2843 /// The indices to a GEP instruction which will move Ptr to the correct slot
2844 /// within the aggregate.
2845 SmallVector<Value *, 4> GEPIndices;
2846 /// The base pointer of the original op, used as a base for GEPing the
2847 /// split operations.
2850 /// Initialize the splitter with an insertion point, Ptr and start with a
2851 /// single zero GEP index.
2852 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
2853 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
2856 /// \brief Generic recursive split emission routine.
2858 /// This method recursively splits an aggregate op (load or store) into
2859 /// scalar or vector ops. It splits recursively until it hits a single value
2860 /// and emits that single value operation via the template argument.
2862 /// The logic of this routine relies on GEPs and insertvalue and
2863 /// extractvalue all operating with the same fundamental index list, merely
2864 /// formatted differently (GEPs need actual values).
2866 /// \param Ty The type being split recursively into smaller ops.
2867 /// \param Agg The aggregate value being built up or stored, depending on
2868 /// whether this is splitting a load or a store respectively.
2869 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
2870 if (Ty->isSingleValueType())
2871 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
2873 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2874 unsigned OldSize = Indices.size();
2876 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
2878 assert(Indices.size() == OldSize && "Did not return to the old size");
2879 Indices.push_back(Idx);
2880 GEPIndices.push_back(IRB.getInt32(Idx));
2881 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
2882 GEPIndices.pop_back();
2888 if (StructType *STy = dyn_cast<StructType>(Ty)) {
2889 unsigned OldSize = Indices.size();
2891 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
2893 assert(Indices.size() == OldSize && "Did not return to the old size");
2894 Indices.push_back(Idx);
2895 GEPIndices.push_back(IRB.getInt32(Idx));
2896 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
2897 GEPIndices.pop_back();
2903 llvm_unreachable("Only arrays and structs are aggregate loadable types");
2907 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
2908 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2909 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
2911 /// Emit a leaf load of a single value. This is called at the leaves of the
2912 /// recursive emission to actually load values.
2913 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2914 assert(Ty->isSingleValueType());
2915 // Load the single value and insert it using the indices.
2916 Value *Load = IRB.CreateLoad(IRB.CreateInBoundsGEP(Ptr, GEPIndices,
2919 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
2920 DEBUG(dbgs() << " to: " << *Load << "\n");
2924 bool visitLoadInst(LoadInst &LI) {
2925 assert(LI.getPointerOperand() == *U);
2926 if (!LI.isSimple() || LI.getType()->isSingleValueType())
2929 // We have an aggregate being loaded, split it apart.
2930 DEBUG(dbgs() << " original: " << LI << "\n");
2931 LoadOpSplitter Splitter(&LI, *U);
2932 Value *V = UndefValue::get(LI.getType());
2933 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
2934 LI.replaceAllUsesWith(V);
2935 LI.eraseFromParent();
2939 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
2940 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2941 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
2943 /// Emit a leaf store of a single value. This is called at the leaves of the
2944 /// recursive emission to actually produce stores.
2945 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2946 assert(Ty->isSingleValueType());
2947 // Extract the single value and store it using the indices.
2948 Value *Store = IRB.CreateStore(
2949 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
2950 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
2952 DEBUG(dbgs() << " to: " << *Store << "\n");
2956 bool visitStoreInst(StoreInst &SI) {
2957 if (!SI.isSimple() || SI.getPointerOperand() != *U)
2959 Value *V = SI.getValueOperand();
2960 if (V->getType()->isSingleValueType())
2963 // We have an aggregate being stored, split it apart.
2964 DEBUG(dbgs() << " original: " << SI << "\n");
2965 StoreOpSplitter Splitter(&SI, *U);
2966 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
2967 SI.eraseFromParent();
2971 bool visitBitCastInst(BitCastInst &BC) {
2976 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
2981 bool visitPHINode(PHINode &PN) {
2986 bool visitSelectInst(SelectInst &SI) {
2993 /// \brief Try to find a partition of the aggregate type passed in for a given
2994 /// offset and size.
2996 /// This recurses through the aggregate type and tries to compute a subtype
2997 /// based on the offset and size. When the offset and size span a sub-section
2998 /// of an array, it will even compute a new array type for that sub-section,
2999 /// and the same for structs.
3001 /// Note that this routine is very strict and tries to find a partition of the
3002 /// type which produces the *exact* right offset and size. It is not forgiving
3003 /// when the size or offset cause either end of type-based partition to be off.
3004 /// Also, this is a best-effort routine. It is reasonable to give up and not
3005 /// return a type if necessary.
3006 static Type *getTypePartition(const TargetData &TD, Type *Ty,
3007 uint64_t Offset, uint64_t Size) {
3008 if (Offset == 0 && TD.getTypeAllocSize(Ty) == Size)
3011 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3012 // We can't partition pointers...
3013 if (SeqTy->isPointerTy())
3016 Type *ElementTy = SeqTy->getElementType();
3017 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
3018 uint64_t NumSkippedElements = Offset / ElementSize;
3019 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy))
3020 if (NumSkippedElements >= ArrTy->getNumElements())
3022 if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy))
3023 if (NumSkippedElements >= VecTy->getNumElements())
3025 Offset -= NumSkippedElements * ElementSize;
3027 // First check if we need to recurse.
3028 if (Offset > 0 || Size < ElementSize) {
3029 // Bail if the partition ends in a different array element.
3030 if ((Offset + Size) > ElementSize)
3032 // Recurse through the element type trying to peel off offset bytes.
3033 return getTypePartition(TD, ElementTy, Offset, Size);
3035 assert(Offset == 0);
3037 if (Size == ElementSize)
3039 assert(Size > ElementSize);
3040 uint64_t NumElements = Size / ElementSize;
3041 if (NumElements * ElementSize != Size)
3043 return ArrayType::get(ElementTy, NumElements);
3046 StructType *STy = dyn_cast<StructType>(Ty);
3050 const StructLayout *SL = TD.getStructLayout(STy);
3051 if (Offset >= SL->getSizeInBytes())
3053 uint64_t EndOffset = Offset + Size;
3054 if (EndOffset > SL->getSizeInBytes())
3057 unsigned Index = SL->getElementContainingOffset(Offset);
3058 Offset -= SL->getElementOffset(Index);
3060 Type *ElementTy = STy->getElementType(Index);
3061 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
3062 if (Offset >= ElementSize)
3063 return 0; // The offset points into alignment padding.
3065 // See if any partition must be contained by the element.
3066 if (Offset > 0 || Size < ElementSize) {
3067 if ((Offset + Size) > ElementSize)
3069 return getTypePartition(TD, ElementTy, Offset, Size);
3071 assert(Offset == 0);
3073 if (Size == ElementSize)
3076 StructType::element_iterator EI = STy->element_begin() + Index,
3077 EE = STy->element_end();
3078 if (EndOffset < SL->getSizeInBytes()) {
3079 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3080 if (Index == EndIndex)
3081 return 0; // Within a single element and its padding.
3083 // Don't try to form "natural" types if the elements don't line up with the
3085 // FIXME: We could potentially recurse down through the last element in the
3086 // sub-struct to find a natural end point.
3087 if (SL->getElementOffset(EndIndex) != EndOffset)
3090 assert(Index < EndIndex);
3091 EE = STy->element_begin() + EndIndex;
3094 // Try to build up a sub-structure.
3095 SmallVector<Type *, 4> ElementTys;
3097 ElementTys.push_back(*EI++);
3099 StructType *SubTy = StructType::get(STy->getContext(), ElementTys,
3101 const StructLayout *SubSL = TD.getStructLayout(SubTy);
3102 if (Size != SubSL->getSizeInBytes())
3103 return 0; // The sub-struct doesn't have quite the size needed.
3108 /// \brief Rewrite an alloca partition's users.
3110 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3111 /// to rewrite uses of an alloca partition to be conducive for SSA value
3112 /// promotion. If the partition needs a new, more refined alloca, this will
3113 /// build that new alloca, preserving as much type information as possible, and
3114 /// rewrite the uses of the old alloca to point at the new one and have the
3115 /// appropriate new offsets. It also evaluates how successful the rewrite was
3116 /// at enabling promotion and if it was successful queues the alloca to be
3118 bool SROA::rewriteAllocaPartition(AllocaInst &AI,
3119 AllocaPartitioning &P,
3120 AllocaPartitioning::iterator PI) {
3121 uint64_t AllocaSize = PI->EndOffset - PI->BeginOffset;
3122 bool IsLive = false;
3123 for (AllocaPartitioning::use_iterator UI = P.use_begin(PI),
3125 UI != UE && !IsLive; ++UI)
3129 return false; // No live uses left of this partition.
3131 DEBUG(dbgs() << "Speculating PHIs and selects in partition "
3132 << "[" << PI->BeginOffset << "," << PI->EndOffset << ")\n");
3134 PHIOrSelectSpeculator Speculator(*TD, P, *this);
3135 DEBUG(dbgs() << " speculating ");
3136 DEBUG(P.print(dbgs(), PI, ""));
3137 Speculator.visitUsers(PI);
3139 // Try to compute a friendly type for this partition of the alloca. This
3140 // won't always succeed, in which case we fall back to a legal integer type
3141 // or an i8 array of an appropriate size.
3143 if (Type *PartitionTy = P.getCommonType(PI))
3144 if (TD->getTypeAllocSize(PartitionTy) >= AllocaSize)
3145 AllocaTy = PartitionTy;
3147 if (Type *PartitionTy = getTypePartition(*TD, AI.getAllocatedType(),
3148 PI->BeginOffset, AllocaSize))
3149 AllocaTy = PartitionTy;
3151 (AllocaTy->isArrayTy() &&
3152 AllocaTy->getArrayElementType()->isIntegerTy())) &&
3153 TD->isLegalInteger(AllocaSize * 8))
3154 AllocaTy = Type::getIntNTy(*C, AllocaSize * 8);
3156 AllocaTy = ArrayType::get(Type::getInt8Ty(*C), AllocaSize);
3157 assert(TD->getTypeAllocSize(AllocaTy) >= AllocaSize);
3159 // Check for the case where we're going to rewrite to a new alloca of the
3160 // exact same type as the original, and with the same access offsets. In that
3161 // case, re-use the existing alloca, but still run through the rewriter to
3162 // performe phi and select speculation.
3164 if (AllocaTy == AI.getAllocatedType()) {
3165 assert(PI->BeginOffset == 0 &&
3166 "Non-zero begin offset but same alloca type");
3167 assert(PI == P.begin() && "Begin offset is zero on later partition");
3170 unsigned Alignment = AI.getAlignment();
3172 // The minimum alignment which users can rely on when the explicit
3173 // alignment is omitted or zero is that required by the ABI for this
3175 Alignment = TD->getABITypeAlignment(AI.getAllocatedType());
3177 Alignment = MinAlign(Alignment, PI->BeginOffset);
3178 // If we will get at least this much alignment from the type alone, leave
3179 // the alloca's alignment unconstrained.
3180 if (Alignment <= TD->getABITypeAlignment(AllocaTy))
3182 NewAI = new AllocaInst(AllocaTy, 0, Alignment,
3183 AI.getName() + ".sroa." + Twine(PI - P.begin()),
3188 DEBUG(dbgs() << "Rewriting alloca partition "
3189 << "[" << PI->BeginOffset << "," << PI->EndOffset << ") to: "
3192 // Track the high watermark of the post-promotion worklist. We will reset it
3193 // to this point if the alloca is not in fact scheduled for promotion.
3194 unsigned PPWOldSize = PostPromotionWorklist.size();
3196 AllocaPartitionRewriter Rewriter(*TD, P, PI, *this, AI, *NewAI,
3197 PI->BeginOffset, PI->EndOffset);
3198 DEBUG(dbgs() << " rewriting ");
3199 DEBUG(P.print(dbgs(), PI, ""));
3200 bool Promotable = Rewriter.visitUsers(P.use_begin(PI), P.use_end(PI));
3202 DEBUG(dbgs() << " and queuing for promotion\n");
3203 PromotableAllocas.push_back(NewAI);
3204 } else if (NewAI != &AI) {
3205 // If we can't promote the alloca, iterate on it to check for new
3206 // refinements exposed by splitting the current alloca. Don't iterate on an
3207 // alloca which didn't actually change and didn't get promoted.
3208 Worklist.insert(NewAI);
3211 // Drop any post-promotion work items if promotion didn't happen.
3213 while (PostPromotionWorklist.size() > PPWOldSize)
3214 PostPromotionWorklist.pop_back();
3219 /// \brief Walks the partitioning of an alloca rewriting uses of each partition.
3220 bool SROA::splitAlloca(AllocaInst &AI, AllocaPartitioning &P) {
3221 bool Changed = false;
3222 for (AllocaPartitioning::iterator PI = P.begin(), PE = P.end(); PI != PE;
3224 Changed |= rewriteAllocaPartition(AI, P, PI);
3229 /// \brief Analyze an alloca for SROA.
3231 /// This analyzes the alloca to ensure we can reason about it, builds
3232 /// a partitioning of the alloca, and then hands it off to be split and
3233 /// rewritten as needed.
3234 bool SROA::runOnAlloca(AllocaInst &AI) {
3235 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3236 ++NumAllocasAnalyzed;
3238 // Special case dead allocas, as they're trivial.
3239 if (AI.use_empty()) {
3240 AI.eraseFromParent();
3244 // Skip alloca forms that this analysis can't handle.
3245 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3246 TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
3249 bool Changed = false;
3251 // First, split any FCA loads and stores touching this alloca to promote
3252 // better splitting and promotion opportunities.
3253 AggLoadStoreRewriter AggRewriter(*TD);
3254 Changed |= AggRewriter.rewrite(AI);
3256 // Build the partition set using a recursive instruction-visiting builder.
3257 AllocaPartitioning P(*TD, AI);
3258 DEBUG(P.print(dbgs()));
3262 // Delete all the dead users of this alloca before splitting and rewriting it.
3263 for (AllocaPartitioning::dead_user_iterator DI = P.dead_user_begin(),
3264 DE = P.dead_user_end();
3267 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3268 DeadInsts.push_back(*DI);
3270 for (AllocaPartitioning::dead_op_iterator DO = P.dead_op_begin(),
3271 DE = P.dead_op_end();
3274 // Clobber the use with an undef value.
3275 **DO = UndefValue::get(OldV->getType());
3276 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3277 if (isInstructionTriviallyDead(OldI)) {
3279 DeadInsts.push_back(OldI);
3283 // No partitions to split. Leave the dead alloca for a later pass to clean up.
3284 if (P.begin() == P.end())
3287 return splitAlloca(AI, P) || Changed;
3290 /// \brief Delete the dead instructions accumulated in this run.
3292 /// Recursively deletes the dead instructions we've accumulated. This is done
3293 /// at the very end to maximize locality of the recursive delete and to
3294 /// minimize the problems of invalidated instruction pointers as such pointers
3295 /// are used heavily in the intermediate stages of the algorithm.
3297 /// We also record the alloca instructions deleted here so that they aren't
3298 /// subsequently handed to mem2reg to promote.
3299 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3300 DeadSplitInsts.clear();
3301 while (!DeadInsts.empty()) {
3302 Instruction *I = DeadInsts.pop_back_val();
3303 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3305 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
3306 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
3307 // Zero out the operand and see if it becomes trivially dead.
3309 if (isInstructionTriviallyDead(U))
3310 DeadInsts.push_back(U);
3313 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3314 DeletedAllocas.insert(AI);
3317 I->eraseFromParent();
3321 /// \brief Promote the allocas, using the best available technique.
3323 /// This attempts to promote whatever allocas have been identified as viable in
3324 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3325 /// If there is a domtree available, we attempt to promote using the full power
3326 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3327 /// based on the SSAUpdater utilities. This function returns whether any
3328 /// promotion occured.
3329 bool SROA::promoteAllocas(Function &F) {
3330 if (PromotableAllocas.empty())
3333 NumPromoted += PromotableAllocas.size();
3335 if (DT && !ForceSSAUpdater) {
3336 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3337 PromoteMemToReg(PromotableAllocas, *DT);
3338 PromotableAllocas.clear();
3342 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3344 DIBuilder DIB(*F.getParent());
3345 SmallVector<Instruction*, 64> Insts;
3347 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3348 AllocaInst *AI = PromotableAllocas[Idx];
3349 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
3351 Instruction *I = cast<Instruction>(*UI++);
3352 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3353 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3354 // leading to them) here. Eventually it should use them to optimize the
3355 // scalar values produced.
3356 if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
3357 assert(onlyUsedByLifetimeMarkers(I) &&
3358 "Found a bitcast used outside of a lifetime marker.");
3359 while (!I->use_empty())
3360 cast<Instruction>(*I->use_begin())->eraseFromParent();
3361 I->eraseFromParent();
3364 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3365 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3366 II->getIntrinsicID() == Intrinsic::lifetime_end);
3367 II->eraseFromParent();
3373 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3377 PromotableAllocas.clear();
3382 /// \brief A predicate to test whether an alloca belongs to a set.
3383 class IsAllocaInSet {
3384 typedef SmallPtrSet<AllocaInst *, 4> SetType;
3388 typedef AllocaInst *argument_type;
3390 IsAllocaInSet(const SetType &Set) : Set(Set) {}
3391 bool operator()(AllocaInst *AI) const { return Set.count(AI); }
3395 bool SROA::runOnFunction(Function &F) {
3396 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3397 C = &F.getContext();
3398 TD = getAnalysisIfAvailable<TargetData>();
3400 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3403 DT = getAnalysisIfAvailable<DominatorTree>();
3405 BasicBlock &EntryBB = F.getEntryBlock();
3406 for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end());
3408 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3409 Worklist.insert(AI);
3411 bool Changed = false;
3412 // A set of deleted alloca instruction pointers which should be removed from
3413 // the list of promotable allocas.
3414 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3417 while (!Worklist.empty()) {
3418 Changed |= runOnAlloca(*Worklist.pop_back_val());
3419 deleteDeadInstructions(DeletedAllocas);
3421 // Remove the deleted allocas from various lists so that we don't try to
3422 // continue processing them.
3423 if (!DeletedAllocas.empty()) {
3424 Worklist.remove_if(IsAllocaInSet(DeletedAllocas));
3425 PostPromotionWorklist.remove_if(IsAllocaInSet(DeletedAllocas));
3426 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3427 PromotableAllocas.end(),
3428 IsAllocaInSet(DeletedAllocas)),
3429 PromotableAllocas.end());
3430 DeletedAllocas.clear();
3434 Changed |= promoteAllocas(F);
3436 Worklist = PostPromotionWorklist;
3437 PostPromotionWorklist.clear();
3438 } while (!Worklist.empty());
3443 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3444 if (RequiresDomTree)
3445 AU.addRequired<DominatorTree>();
3446 AU.setPreservesCFG();