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/DataLayout.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 DataLayout &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 DataLayout &TD, AllocaInst &AI, AllocaPartitioning &P)
416 AllocSize(TD.getTypeAllocSize(AI.getAllocatedType())),
422 const DataLayout &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 DataLayout &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");
719 // Disable SRoA for any intrinsics except for lifetime invariants.
720 // FIXME: What about debug instrinsics? This matches old behavior, but
721 // doesn't make sense.
722 bool visitIntrinsicInst(IntrinsicInst &II) {
723 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
724 II.getIntrinsicID() == Intrinsic::lifetime_end) {
725 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
726 uint64_t Size = std::min(AllocSize - Offset, Length->getLimitedValue());
727 insertUse(II, Offset, Size, true);
731 return markAsEscaping(II);
734 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
735 // We consider any PHI or select that results in a direct load or store of
736 // the same offset to be a viable use for partitioning purposes. These uses
737 // are considered unsplittable and the size is the maximum loaded or stored
739 SmallPtrSet<Instruction *, 4> Visited;
740 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
741 Visited.insert(Root);
742 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
743 // If there are no loads or stores, the access is dead. We mark that as
744 // a size zero access.
747 Instruction *I, *UsedI;
748 llvm::tie(UsedI, I) = Uses.pop_back_val();
750 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
751 Size = std::max(Size, TD.getTypeStoreSize(LI->getType()));
754 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
755 Value *Op = SI->getOperand(0);
758 Size = std::max(Size, TD.getTypeStoreSize(Op->getType()));
762 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
763 if (!GEP->hasAllZeroIndices())
765 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
766 !isa<SelectInst>(I)) {
770 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
772 if (Visited.insert(cast<Instruction>(*UI)))
773 Uses.push_back(std::make_pair(I, cast<Instruction>(*UI)));
774 } while (!Uses.empty());
779 bool visitPHINode(PHINode &PN) {
780 // See if we already have computed info on this node.
781 std::pair<uint64_t, bool> &PHIInfo = P.PHIOrSelectSizes[&PN];
783 PHIInfo.second = true;
784 insertUse(PN, Offset, PHIInfo.first);
788 // Check for an unsafe use of the PHI node.
789 if (Instruction *EscapingI = hasUnsafePHIOrSelectUse(&PN, PHIInfo.first))
790 return markAsEscaping(*EscapingI);
792 insertUse(PN, Offset, PHIInfo.first);
796 bool visitSelectInst(SelectInst &SI) {
797 if (Value *Result = foldSelectInst(SI)) {
799 // If the result of the constant fold will be the pointer, recurse
800 // through the select as if we had RAUW'ed it.
801 enqueueUsers(SI, Offset);
806 // See if we already have computed info on this node.
807 std::pair<uint64_t, bool> &SelectInfo = P.PHIOrSelectSizes[&SI];
808 if (SelectInfo.first) {
809 SelectInfo.second = true;
810 insertUse(SI, Offset, SelectInfo.first);
814 // Check for an unsafe use of the PHI node.
815 if (Instruction *EscapingI = hasUnsafePHIOrSelectUse(&SI, SelectInfo.first))
816 return markAsEscaping(*EscapingI);
818 insertUse(SI, Offset, SelectInfo.first);
822 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
823 bool visitInstruction(Instruction &I) { return markAsEscaping(I); }
827 /// \brief Use adder for the alloca partitioning.
829 /// This class adds the uses of an alloca to all of the partitions which they
830 /// use. For splittable partitions, this can end up doing essentially a linear
831 /// walk of the partitions, but the number of steps remains bounded by the
832 /// total result instruction size:
833 /// - The number of partitions is a result of the number unsplittable
834 /// instructions using the alloca.
835 /// - The number of users of each partition is at worst the total number of
836 /// splittable instructions using the alloca.
837 /// Thus we will produce N * M instructions in the end, where N are the number
838 /// of unsplittable uses and M are the number of splittable. This visitor does
839 /// the exact same number of updates to the partitioning.
841 /// In the more common case, this visitor will leverage the fact that the
842 /// partition space is pre-sorted, and do a logarithmic search for the
843 /// partition needed, making the total visit a classical ((N + M) * log(N))
844 /// complexity operation.
845 class AllocaPartitioning::UseBuilder : public BuilderBase<UseBuilder> {
846 friend class InstVisitor<UseBuilder>;
848 /// \brief Set to de-duplicate dead instructions found in the use walk.
849 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
852 UseBuilder(const DataLayout &TD, AllocaInst &AI, AllocaPartitioning &P)
853 : BuilderBase<UseBuilder>(TD, AI, P) {}
855 /// \brief Run the builder over the allocation.
857 // Note that we have to re-evaluate size on each trip through the loop as
858 // the queue grows at the tail.
859 for (unsigned Idx = 0; Idx < Queue.size(); ++Idx) {
861 Offset = Queue[Idx].Offset;
862 this->visit(cast<Instruction>(U->getUser()));
867 void markAsDead(Instruction &I) {
868 if (VisitedDeadInsts.insert(&I))
869 P.DeadUsers.push_back(&I);
872 void insertUse(Instruction &User, int64_t Offset, uint64_t Size) {
873 // If the use has a zero size or extends outside of the allocation, record
874 // it as a dead use for elimination later.
875 if (Size == 0 || (uint64_t)Offset >= AllocSize ||
876 (Offset < 0 && (uint64_t)-Offset >= Size))
877 return markAsDead(User);
879 // Clamp the start to the beginning of the allocation.
881 Size -= (uint64_t)-Offset;
885 uint64_t BeginOffset = Offset, EndOffset = BeginOffset + Size;
887 // Clamp the end offset to the end of the allocation. Note that this is
888 // formulated to handle even the case where "BeginOffset + Size" overflows.
889 assert(AllocSize >= BeginOffset); // Established above.
890 if (Size > AllocSize - BeginOffset)
891 EndOffset = AllocSize;
893 // NB: This only works if we have zero overlapping partitions.
894 iterator B = std::lower_bound(P.begin(), P.end(), BeginOffset);
895 if (B != P.begin() && llvm::prior(B)->EndOffset > BeginOffset)
897 for (iterator I = B, E = P.end(); I != E && I->BeginOffset < EndOffset;
899 PartitionUse NewPU(std::max(I->BeginOffset, BeginOffset),
900 std::min(I->EndOffset, EndOffset), U);
901 P.use_push_back(I, NewPU);
902 if (isa<PHINode>(U->getUser()) || isa<SelectInst>(U->getUser()))
903 P.PHIOrSelectOpMap[U]
904 = std::make_pair(I - P.begin(), P.Uses[I - P.begin()].size() - 1);
908 void handleLoadOrStore(Type *Ty, Instruction &I, int64_t Offset) {
909 uint64_t Size = TD.getTypeStoreSize(Ty);
911 // If this memory access can be shown to *statically* extend outside the
912 // bounds of of the allocation, it's behavior is undefined, so simply
913 // ignore it. Note that this is more strict than the generic clamping
914 // behavior of insertUse.
915 if (Offset < 0 || (uint64_t)Offset >= AllocSize ||
916 Size > (AllocSize - (uint64_t)Offset))
917 return markAsDead(I);
919 insertUse(I, Offset, Size);
922 void visitBitCastInst(BitCastInst &BC) {
924 return markAsDead(BC);
926 enqueueUsers(BC, Offset);
929 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
930 if (GEPI.use_empty())
931 return markAsDead(GEPI);
934 if (!computeConstantGEPOffset(GEPI, GEPOffset))
935 llvm_unreachable("Unable to compute constant offset for use");
937 enqueueUsers(GEPI, GEPOffset);
940 void visitLoadInst(LoadInst &LI) {
941 handleLoadOrStore(LI.getType(), LI, Offset);
944 void visitStoreInst(StoreInst &SI) {
945 handleLoadOrStore(SI.getOperand(0)->getType(), SI, Offset);
948 void visitMemSetInst(MemSetInst &II) {
949 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
950 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
951 insertUse(II, Offset, Size);
954 void visitMemTransferInst(MemTransferInst &II) {
955 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
956 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
958 return markAsDead(II);
960 MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
961 if (!II.isVolatile() && Offsets.DestEnd && Offsets.SourceEnd &&
962 Offsets.DestBegin == Offsets.SourceBegin)
963 return markAsDead(II); // Skip identity transfers without side-effects.
965 insertUse(II, Offset, Size);
968 void visitIntrinsicInst(IntrinsicInst &II) {
969 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
970 II.getIntrinsicID() == Intrinsic::lifetime_end);
972 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
973 insertUse(II, Offset,
974 std::min(AllocSize - Offset, Length->getLimitedValue()));
977 void insertPHIOrSelect(Instruction &User, uint64_t Offset) {
978 uint64_t Size = P.PHIOrSelectSizes.lookup(&User).first;
980 // For PHI and select operands outside the alloca, we can't nuke the entire
981 // phi or select -- the other side might still be relevant, so we special
982 // case them here and use a separate structure to track the operands
983 // themselves which should be replaced with undef.
984 if (Offset >= AllocSize) {
985 P.DeadOperands.push_back(U);
989 insertUse(User, Offset, Size);
991 void visitPHINode(PHINode &PN) {
993 return markAsDead(PN);
995 insertPHIOrSelect(PN, Offset);
997 void visitSelectInst(SelectInst &SI) {
999 return markAsDead(SI);
1001 if (Value *Result = foldSelectInst(SI)) {
1003 // If the result of the constant fold will be the pointer, recurse
1004 // through the select as if we had RAUW'ed it.
1005 enqueueUsers(SI, Offset);
1007 // Otherwise the operand to the select is dead, and we can replace it
1009 P.DeadOperands.push_back(U);
1014 insertPHIOrSelect(SI, Offset);
1017 /// \brief Unreachable, we've already visited the alloca once.
1018 void visitInstruction(Instruction &I) {
1019 llvm_unreachable("Unhandled instruction in use builder.");
1023 void AllocaPartitioning::splitAndMergePartitions() {
1024 size_t NumDeadPartitions = 0;
1026 // Track the range of splittable partitions that we pass when accumulating
1027 // overlapping unsplittable partitions.
1028 uint64_t SplitEndOffset = 0ull;
1030 Partition New(0ull, 0ull, false);
1032 for (unsigned i = 0, j = i, e = Partitions.size(); i != e; i = j) {
1035 if (!Partitions[i].IsSplittable || New.BeginOffset == New.EndOffset) {
1036 assert(New.BeginOffset == New.EndOffset);
1037 New = Partitions[i];
1039 assert(New.IsSplittable);
1040 New.EndOffset = std::max(New.EndOffset, Partitions[i].EndOffset);
1042 assert(New.BeginOffset != New.EndOffset);
1044 // Scan the overlapping partitions.
1045 while (j != e && New.EndOffset > Partitions[j].BeginOffset) {
1046 // If the new partition we are forming is splittable, stop at the first
1047 // unsplittable partition.
1048 if (New.IsSplittable && !Partitions[j].IsSplittable)
1051 // Grow the new partition to include any equally splittable range. 'j' is
1052 // always equally splittable when New is splittable, but when New is not
1053 // splittable, we may subsume some (or part of some) splitable partition
1054 // without growing the new one.
1055 if (New.IsSplittable == Partitions[j].IsSplittable) {
1056 New.EndOffset = std::max(New.EndOffset, Partitions[j].EndOffset);
1058 assert(!New.IsSplittable);
1059 assert(Partitions[j].IsSplittable);
1060 SplitEndOffset = std::max(SplitEndOffset, Partitions[j].EndOffset);
1063 Partitions[j].kill();
1064 ++NumDeadPartitions;
1068 // If the new partition is splittable, chop off the end as soon as the
1069 // unsplittable subsequent partition starts and ensure we eventually cover
1070 // the splittable area.
1071 if (j != e && New.IsSplittable) {
1072 SplitEndOffset = std::max(SplitEndOffset, New.EndOffset);
1073 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
1076 // Add the new partition if it differs from the original one and is
1077 // non-empty. We can end up with an empty partition here if it was
1078 // splittable but there is an unsplittable one that starts at the same
1080 if (New != Partitions[i]) {
1081 if (New.BeginOffset != New.EndOffset)
1082 Partitions.push_back(New);
1083 // Mark the old one for removal.
1084 Partitions[i].kill();
1085 ++NumDeadPartitions;
1088 New.BeginOffset = New.EndOffset;
1089 if (!New.IsSplittable) {
1090 New.EndOffset = std::max(New.EndOffset, SplitEndOffset);
1091 if (j != e && !Partitions[j].IsSplittable)
1092 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
1093 New.IsSplittable = true;
1094 // If there is a trailing splittable partition which won't be fused into
1095 // the next splittable partition go ahead and add it onto the partitions
1097 if (New.BeginOffset < New.EndOffset &&
1098 (j == e || !Partitions[j].IsSplittable ||
1099 New.EndOffset < Partitions[j].BeginOffset)) {
1100 Partitions.push_back(New);
1101 New.BeginOffset = New.EndOffset = 0ull;
1106 // Re-sort the partitions now that they have been split and merged into
1107 // disjoint set of partitions. Also remove any of the dead partitions we've
1108 // replaced in the process.
1109 std::sort(Partitions.begin(), Partitions.end());
1110 if (NumDeadPartitions) {
1111 assert(Partitions.back().isDead());
1112 assert((ptrdiff_t)NumDeadPartitions ==
1113 std::count(Partitions.begin(), Partitions.end(), Partitions.back()));
1115 Partitions.erase(Partitions.end() - NumDeadPartitions, Partitions.end());
1118 AllocaPartitioning::AllocaPartitioning(const DataLayout &TD, AllocaInst &AI)
1123 PointerEscapingInstr(0) {
1124 PartitionBuilder PB(TD, AI, *this);
1128 // Sort the uses. This arranges for the offsets to be in ascending order,
1129 // and the sizes to be in descending order.
1130 std::sort(Partitions.begin(), Partitions.end());
1132 // Remove any partitions from the back which are marked as dead.
1133 while (!Partitions.empty() && Partitions.back().isDead())
1134 Partitions.pop_back();
1136 if (Partitions.size() > 1) {
1137 // Intersect splittability for all partitions with equal offsets and sizes.
1138 // Then remove all but the first so that we have a sequence of non-equal but
1139 // potentially overlapping partitions.
1140 for (iterator I = Partitions.begin(), J = I, E = Partitions.end(); I != E;
1143 while (J != E && *I == *J) {
1144 I->IsSplittable &= J->IsSplittable;
1148 Partitions.erase(std::unique(Partitions.begin(), Partitions.end()),
1151 // Split splittable and merge unsplittable partitions into a disjoint set
1152 // of partitions over the used space of the allocation.
1153 splitAndMergePartitions();
1156 // Now build up the user lists for each of these disjoint partitions by
1157 // re-walking the recursive users of the alloca.
1158 Uses.resize(Partitions.size());
1159 UseBuilder UB(TD, AI, *this);
1163 Type *AllocaPartitioning::getCommonType(iterator I) const {
1165 for (const_use_iterator UI = use_begin(I), UE = use_end(I); UI != UE; ++UI) {
1167 continue; // Skip dead uses.
1168 if (isa<IntrinsicInst>(*UI->U->getUser()))
1170 if (UI->BeginOffset != I->BeginOffset || UI->EndOffset != I->EndOffset)
1174 if (LoadInst *LI = dyn_cast<LoadInst>(UI->U->getUser())) {
1175 UserTy = LI->getType();
1176 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI->U->getUser())) {
1177 UserTy = SI->getValueOperand()->getType();
1180 if (Ty && Ty != UserTy)
1188 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1190 void AllocaPartitioning::print(raw_ostream &OS, const_iterator I,
1191 StringRef Indent) const {
1192 OS << Indent << "partition #" << (I - begin())
1193 << " [" << I->BeginOffset << "," << I->EndOffset << ")"
1194 << (I->IsSplittable ? " (splittable)" : "")
1195 << (Uses[I - begin()].empty() ? " (zero uses)" : "")
1199 void AllocaPartitioning::printUsers(raw_ostream &OS, const_iterator I,
1200 StringRef Indent) const {
1201 for (const_use_iterator UI = use_begin(I), UE = use_end(I);
1204 continue; // Skip dead uses.
1205 OS << Indent << " [" << UI->BeginOffset << "," << UI->EndOffset << ") "
1206 << "used by: " << *UI->U->getUser() << "\n";
1207 if (MemTransferInst *II = dyn_cast<MemTransferInst>(UI->U->getUser())) {
1208 const MemTransferOffsets &MTO = MemTransferInstData.lookup(II);
1210 if (!MTO.IsSplittable)
1211 IsDest = UI->BeginOffset == MTO.DestBegin;
1213 IsDest = MTO.DestBegin != 0u;
1214 OS << Indent << " (original " << (IsDest ? "dest" : "source") << ": "
1215 << "[" << (IsDest ? MTO.DestBegin : MTO.SourceBegin)
1216 << "," << (IsDest ? MTO.DestEnd : MTO.SourceEnd) << ")\n";
1221 void AllocaPartitioning::print(raw_ostream &OS) const {
1222 if (PointerEscapingInstr) {
1223 OS << "No partitioning for alloca: " << AI << "\n"
1224 << " A pointer to this alloca escaped by:\n"
1225 << " " << *PointerEscapingInstr << "\n";
1229 OS << "Partitioning of alloca: " << AI << "\n";
1231 for (const_iterator I = begin(), E = end(); I != E; ++I, ++Num) {
1237 void AllocaPartitioning::dump(const_iterator I) const { print(dbgs(), I); }
1238 void AllocaPartitioning::dump() const { print(dbgs()); }
1240 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1244 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
1246 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
1247 /// the loads and stores of an alloca instruction, as well as updating its
1248 /// debug information. This is used when a domtree is unavailable and thus
1249 /// mem2reg in its full form can't be used to handle promotion of allocas to
1251 class AllocaPromoter : public LoadAndStorePromoter {
1255 SmallVector<DbgDeclareInst *, 4> DDIs;
1256 SmallVector<DbgValueInst *, 4> DVIs;
1259 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
1260 AllocaInst &AI, DIBuilder &DIB)
1261 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
1263 void run(const SmallVectorImpl<Instruction*> &Insts) {
1264 // Remember which alloca we're promoting (for isInstInList).
1265 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
1266 for (Value::use_iterator UI = DebugNode->use_begin(),
1267 UE = DebugNode->use_end();
1269 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
1270 DDIs.push_back(DDI);
1271 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
1272 DVIs.push_back(DVI);
1275 LoadAndStorePromoter::run(Insts);
1276 AI.eraseFromParent();
1277 while (!DDIs.empty())
1278 DDIs.pop_back_val()->eraseFromParent();
1279 while (!DVIs.empty())
1280 DVIs.pop_back_val()->eraseFromParent();
1283 virtual bool isInstInList(Instruction *I,
1284 const SmallVectorImpl<Instruction*> &Insts) const {
1285 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1286 return LI->getOperand(0) == &AI;
1287 return cast<StoreInst>(I)->getPointerOperand() == &AI;
1290 virtual void updateDebugInfo(Instruction *Inst) const {
1291 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(),
1292 E = DDIs.end(); I != E; ++I) {
1293 DbgDeclareInst *DDI = *I;
1294 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
1295 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1296 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
1297 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1299 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(),
1300 E = DVIs.end(); I != E; ++I) {
1301 DbgValueInst *DVI = *I;
1303 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1304 // If an argument is zero extended then use argument directly. The ZExt
1305 // may be zapped by an optimization pass in future.
1306 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1307 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
1308 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1309 Arg = dyn_cast<Argument>(SExt->getOperand(0));
1311 Arg = SI->getOperand(0);
1312 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
1313 Arg = LI->getOperand(0);
1317 Instruction *DbgVal =
1318 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
1320 DbgVal->setDebugLoc(DVI->getDebugLoc());
1324 } // end anon namespace
1328 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
1330 /// This pass takes allocations which can be completely analyzed (that is, they
1331 /// don't escape) and tries to turn them into scalar SSA values. There are
1332 /// a few steps to this process.
1334 /// 1) It takes allocations of aggregates and analyzes the ways in which they
1335 /// are used to try to split them into smaller allocations, ideally of
1336 /// a single scalar data type. It will split up memcpy and memset accesses
1337 /// as necessary and try to isolate invidual scalar accesses.
1338 /// 2) It will transform accesses into forms which are suitable for SSA value
1339 /// promotion. This can be replacing a memset with a scalar store of an
1340 /// integer value, or it can involve speculating operations on a PHI or
1341 /// select to be a PHI or select of the results.
1342 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
1343 /// onto insert and extract operations on a vector value, and convert them to
1344 /// this form. By doing so, it will enable promotion of vector aggregates to
1345 /// SSA vector values.
1346 class SROA : public FunctionPass {
1347 const bool RequiresDomTree;
1350 const DataLayout *TD;
1353 /// \brief Worklist of alloca instructions to simplify.
1355 /// Each alloca in the function is added to this. Each new alloca formed gets
1356 /// added to it as well to recursively simplify unless that alloca can be
1357 /// directly promoted. Finally, each time we rewrite a use of an alloca other
1358 /// the one being actively rewritten, we add it back onto the list if not
1359 /// already present to ensure it is re-visited.
1360 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
1362 /// \brief A collection of instructions to delete.
1363 /// We try to batch deletions to simplify code and make things a bit more
1365 SmallVector<Instruction *, 8> DeadInsts;
1367 /// \brief A set to prevent repeatedly marking an instruction split into many
1368 /// uses as dead. Only used to guard insertion into DeadInsts.
1369 SmallPtrSet<Instruction *, 4> DeadSplitInsts;
1371 /// \brief Post-promotion worklist.
1373 /// Sometimes we discover an alloca which has a high probability of becoming
1374 /// viable for SROA after a round of promotion takes place. In those cases,
1375 /// the alloca is enqueued here for re-processing.
1377 /// Note that we have to be very careful to clear allocas out of this list in
1378 /// the event they are deleted.
1379 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist;
1381 /// \brief A collection of alloca instructions we can directly promote.
1382 std::vector<AllocaInst *> PromotableAllocas;
1385 SROA(bool RequiresDomTree = true)
1386 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
1387 C(0), TD(0), DT(0) {
1388 initializeSROAPass(*PassRegistry::getPassRegistry());
1390 bool runOnFunction(Function &F);
1391 void getAnalysisUsage(AnalysisUsage &AU) const;
1393 const char *getPassName() const { return "SROA"; }
1397 friend class PHIOrSelectSpeculator;
1398 friend class AllocaPartitionRewriter;
1399 friend class AllocaPartitionVectorRewriter;
1401 bool rewriteAllocaPartition(AllocaInst &AI,
1402 AllocaPartitioning &P,
1403 AllocaPartitioning::iterator PI);
1404 bool splitAlloca(AllocaInst &AI, AllocaPartitioning &P);
1405 bool runOnAlloca(AllocaInst &AI);
1406 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
1407 bool promoteAllocas(Function &F);
1413 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
1414 return new SROA(RequiresDomTree);
1417 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
1419 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
1420 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
1424 /// \brief Visitor to speculate PHIs and Selects where possible.
1425 class PHIOrSelectSpeculator : public InstVisitor<PHIOrSelectSpeculator> {
1426 // Befriend the base class so it can delegate to private visit methods.
1427 friend class llvm::InstVisitor<PHIOrSelectSpeculator>;
1429 const DataLayout &TD;
1430 AllocaPartitioning &P;
1434 PHIOrSelectSpeculator(const DataLayout &TD, AllocaPartitioning &P, SROA &Pass)
1435 : TD(TD), P(P), Pass(Pass) {}
1437 /// \brief Visit the users of an alloca partition and rewrite them.
1438 void visitUsers(AllocaPartitioning::const_iterator PI) {
1439 // Note that we need to use an index here as the underlying vector of uses
1440 // may be grown during speculation. However, we never need to re-visit the
1441 // new uses, and so we can use the initial size bound.
1442 for (unsigned Idx = 0, Size = P.use_size(PI); Idx != Size; ++Idx) {
1443 const AllocaPartitioning::PartitionUse &PU = P.getUse(PI, Idx);
1445 continue; // Skip dead use.
1447 visit(cast<Instruction>(PU.U->getUser()));
1452 // By default, skip this instruction.
1453 void visitInstruction(Instruction &I) {}
1455 /// PHI instructions that use an alloca and are subsequently loaded can be
1456 /// rewritten to load both input pointers in the pred blocks and then PHI the
1457 /// results, allowing the load of the alloca to be promoted.
1459 /// %P2 = phi [i32* %Alloca, i32* %Other]
1460 /// %V = load i32* %P2
1462 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1464 /// %V2 = load i32* %Other
1466 /// %V = phi [i32 %V1, i32 %V2]
1468 /// We can do this to a select if its only uses are loads and if the operands
1469 /// to the select can be loaded unconditionally.
1471 /// FIXME: This should be hoisted into a generic utility, likely in
1472 /// Transforms/Util/Local.h
1473 bool isSafePHIToSpeculate(PHINode &PN, SmallVectorImpl<LoadInst *> &Loads) {
1474 // For now, we can only do this promotion if the load is in the same block
1475 // as the PHI, and if there are no stores between the phi and load.
1476 // TODO: Allow recursive phi users.
1477 // TODO: Allow stores.
1478 BasicBlock *BB = PN.getParent();
1479 unsigned MaxAlign = 0;
1480 for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end();
1482 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1483 if (LI == 0 || !LI->isSimple()) return false;
1485 // For now we only allow loads in the same block as the PHI. This is
1486 // a common case that happens when instcombine merges two loads through
1488 if (LI->getParent() != BB) return false;
1490 // Ensure that there are no instructions between the PHI and the load that
1492 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1493 if (BBI->mayWriteToMemory())
1496 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1497 Loads.push_back(LI);
1500 // We can only transform this if it is safe to push the loads into the
1501 // predecessor blocks. The only thing to watch out for is that we can't put
1502 // a possibly trapping load in the predecessor if it is a critical edge.
1503 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num;
1505 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1506 Value *InVal = PN.getIncomingValue(Idx);
1508 // If the value is produced by the terminator of the predecessor (an
1509 // invoke) or it has side-effects, there is no valid place to put a load
1510 // in the predecessor.
1511 if (TI == InVal || TI->mayHaveSideEffects())
1514 // If the predecessor has a single successor, then the edge isn't
1516 if (TI->getNumSuccessors() == 1)
1519 // If this pointer is always safe to load, or if we can prove that there
1520 // is already a load in the block, then we can move the load to the pred
1522 if (InVal->isDereferenceablePointer() ||
1523 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, &TD))
1532 void visitPHINode(PHINode &PN) {
1533 DEBUG(dbgs() << " original: " << PN << "\n");
1535 SmallVector<LoadInst *, 4> Loads;
1536 if (!isSafePHIToSpeculate(PN, Loads))
1539 assert(!Loads.empty());
1541 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1542 IRBuilder<> PHIBuilder(&PN);
1543 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1544 PN.getName() + ".sroa.speculated");
1546 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1547 // matter which one we get and if any differ, it doesn't matter.
1548 LoadInst *SomeLoad = cast<LoadInst>(Loads.back());
1549 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1550 unsigned Align = SomeLoad->getAlignment();
1552 // Rewrite all loads of the PN to use the new PHI.
1554 LoadInst *LI = Loads.pop_back_val();
1555 LI->replaceAllUsesWith(NewPN);
1556 Pass.DeadInsts.push_back(LI);
1557 } while (!Loads.empty());
1559 // Inject loads into all of the pred blocks.
1560 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1561 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1562 TerminatorInst *TI = Pred->getTerminator();
1563 Use *InUse = &PN.getOperandUse(PN.getOperandNumForIncomingValue(Idx));
1564 Value *InVal = PN.getIncomingValue(Idx);
1565 IRBuilder<> PredBuilder(TI);
1568 = PredBuilder.CreateLoad(InVal, (PN.getName() + ".sroa.speculate.load." +
1570 ++NumLoadsSpeculated;
1571 Load->setAlignment(Align);
1573 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1574 NewPN->addIncoming(Load, Pred);
1576 Instruction *Ptr = dyn_cast<Instruction>(InVal);
1578 // No uses to rewrite.
1581 // Try to lookup and rewrite any partition uses corresponding to this phi
1583 AllocaPartitioning::iterator PI
1584 = P.findPartitionForPHIOrSelectOperand(InUse);
1588 // Replace the Use in the PartitionUse for this operand with the Use
1590 AllocaPartitioning::use_iterator UI
1591 = P.findPartitionUseForPHIOrSelectOperand(InUse);
1592 assert(isa<PHINode>(*UI->U->getUser()));
1593 UI->U = &Load->getOperandUse(Load->getPointerOperandIndex());
1595 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1598 /// Select instructions that use an alloca and are subsequently loaded can be
1599 /// rewritten to load both input pointers and then select between the result,
1600 /// allowing the load of the alloca to be promoted.
1602 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1603 /// %V = load i32* %P2
1605 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1606 /// %V2 = load i32* %Other
1607 /// %V = select i1 %cond, i32 %V1, i32 %V2
1609 /// We can do this to a select if its only uses are loads and if the operand
1610 /// to the select can be loaded unconditionally.
1611 bool isSafeSelectToSpeculate(SelectInst &SI,
1612 SmallVectorImpl<LoadInst *> &Loads) {
1613 Value *TValue = SI.getTrueValue();
1614 Value *FValue = SI.getFalseValue();
1615 bool TDerefable = TValue->isDereferenceablePointer();
1616 bool FDerefable = FValue->isDereferenceablePointer();
1618 for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end();
1620 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1621 if (LI == 0 || !LI->isSimple()) return false;
1623 // Both operands to the select need to be dereferencable, either
1624 // absolutely (e.g. allocas) or at this point because we can see other
1626 if (!TDerefable && !isSafeToLoadUnconditionally(TValue, LI,
1627 LI->getAlignment(), &TD))
1629 if (!FDerefable && !isSafeToLoadUnconditionally(FValue, LI,
1630 LI->getAlignment(), &TD))
1632 Loads.push_back(LI);
1638 void visitSelectInst(SelectInst &SI) {
1639 DEBUG(dbgs() << " original: " << SI << "\n");
1640 IRBuilder<> IRB(&SI);
1642 // If the select isn't safe to speculate, just use simple logic to emit it.
1643 SmallVector<LoadInst *, 4> Loads;
1644 if (!isSafeSelectToSpeculate(SI, Loads))
1647 Use *Ops[2] = { &SI.getOperandUse(1), &SI.getOperandUse(2) };
1648 AllocaPartitioning::iterator PIs[2];
1649 AllocaPartitioning::PartitionUse PUs[2];
1650 for (unsigned i = 0, e = 2; i != e; ++i) {
1651 PIs[i] = P.findPartitionForPHIOrSelectOperand(Ops[i]);
1652 if (PIs[i] != P.end()) {
1653 // If the pointer is within the partitioning, remove the select from
1654 // its uses. We'll add in the new loads below.
1655 AllocaPartitioning::use_iterator UI
1656 = P.findPartitionUseForPHIOrSelectOperand(Ops[i]);
1658 // Clear out the use here so that the offsets into the use list remain
1659 // stable but this use is ignored when rewriting.
1664 Value *TV = SI.getTrueValue();
1665 Value *FV = SI.getFalseValue();
1666 // Replace the loads of the select with a select of two loads.
1667 while (!Loads.empty()) {
1668 LoadInst *LI = Loads.pop_back_val();
1670 IRB.SetInsertPoint(LI);
1672 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1674 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1675 NumLoadsSpeculated += 2;
1677 // Transfer alignment and TBAA info if present.
1678 TL->setAlignment(LI->getAlignment());
1679 FL->setAlignment(LI->getAlignment());
1680 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1681 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
1682 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
1685 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1686 LI->getName() + ".sroa.speculated");
1688 LoadInst *Loads[2] = { TL, FL };
1689 for (unsigned i = 0, e = 2; i != e; ++i) {
1690 if (PIs[i] != P.end()) {
1691 Use *LoadUse = &Loads[i]->getOperandUse(0);
1692 assert(PUs[i].U->get() == LoadUse->get());
1694 P.use_push_back(PIs[i], PUs[i]);
1698 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1699 LI->replaceAllUsesWith(V);
1700 Pass.DeadInsts.push_back(LI);
1706 /// \brief Accumulate the constant offsets in a GEP into a single APInt offset.
1708 /// If the provided GEP is all-constant, the total byte offset formed by the
1709 /// GEP is computed and Offset is set to it. If the GEP has any non-constant
1710 /// operands, the function returns false and the value of Offset is unmodified.
1711 static bool accumulateGEPOffsets(const DataLayout &TD, GEPOperator &GEP,
1713 APInt GEPOffset(Offset.getBitWidth(), 0);
1714 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1715 GTI != GTE; ++GTI) {
1716 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
1719 if (OpC->isZero()) continue;
1721 // Handle a struct index, which adds its field offset to the pointer.
1722 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1723 unsigned ElementIdx = OpC->getZExtValue();
1724 const StructLayout *SL = TD.getStructLayout(STy);
1725 GEPOffset += APInt(Offset.getBitWidth(),
1726 SL->getElementOffset(ElementIdx));
1730 APInt TypeSize(Offset.getBitWidth(),
1731 TD.getTypeAllocSize(GTI.getIndexedType()));
1732 if (VectorType *VTy = dyn_cast<VectorType>(*GTI)) {
1733 assert((VTy->getScalarSizeInBits() % 8) == 0 &&
1734 "vector element size is not a multiple of 8, cannot GEP over it");
1735 TypeSize = VTy->getScalarSizeInBits() / 8;
1738 GEPOffset += OpC->getValue().sextOrTrunc(Offset.getBitWidth()) * TypeSize;
1744 /// \brief Build a GEP out of a base pointer and indices.
1746 /// This will return the BasePtr if that is valid, or build a new GEP
1747 /// instruction using the IRBuilder if GEP-ing is needed.
1748 static Value *buildGEP(IRBuilder<> &IRB, Value *BasePtr,
1749 SmallVectorImpl<Value *> &Indices,
1750 const Twine &Prefix) {
1751 if (Indices.empty())
1754 // A single zero index is a no-op, so check for this and avoid building a GEP
1756 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1759 return IRB.CreateInBoundsGEP(BasePtr, Indices, Prefix + ".idx");
1762 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1763 /// TargetTy without changing the offset of the pointer.
1765 /// This routine assumes we've already established a properly offset GEP with
1766 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1767 /// zero-indices down through type layers until we find one the same as
1768 /// TargetTy. If we can't find one with the same type, we at least try to use
1769 /// one with the same size. If none of that works, we just produce the GEP as
1770 /// indicated by Indices to have the correct offset.
1771 static Value *getNaturalGEPWithType(IRBuilder<> &IRB, const DataLayout &TD,
1772 Value *BasePtr, Type *Ty, Type *TargetTy,
1773 SmallVectorImpl<Value *> &Indices,
1774 const Twine &Prefix) {
1776 return buildGEP(IRB, BasePtr, Indices, Prefix);
1778 // See if we can descend into a struct and locate a field with the correct
1780 unsigned NumLayers = 0;
1781 Type *ElementTy = Ty;
1783 if (ElementTy->isPointerTy())
1785 if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
1786 ElementTy = SeqTy->getElementType();
1787 Indices.push_back(IRB.getInt(APInt(TD.getPointerSizeInBits(), 0)));
1788 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1789 if (STy->element_begin() == STy->element_end())
1790 break; // Nothing left to descend into.
1791 ElementTy = *STy->element_begin();
1792 Indices.push_back(IRB.getInt32(0));
1797 } while (ElementTy != TargetTy);
1798 if (ElementTy != TargetTy)
1799 Indices.erase(Indices.end() - NumLayers, Indices.end());
1801 return buildGEP(IRB, BasePtr, Indices, Prefix);
1804 /// \brief Recursively compute indices for a natural GEP.
1806 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1807 /// element types adding appropriate indices for the GEP.
1808 static Value *getNaturalGEPRecursively(IRBuilder<> &IRB, const DataLayout &TD,
1809 Value *Ptr, Type *Ty, APInt &Offset,
1811 SmallVectorImpl<Value *> &Indices,
1812 const Twine &Prefix) {
1814 return getNaturalGEPWithType(IRB, TD, Ptr, Ty, TargetTy, Indices, Prefix);
1816 // We can't recurse through pointer types.
1817 if (Ty->isPointerTy())
1820 // We try to analyze GEPs over vectors here, but note that these GEPs are
1821 // extremely poorly defined currently. The long-term goal is to remove GEPing
1822 // over a vector from the IR completely.
1823 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1824 unsigned ElementSizeInBits = VecTy->getScalarSizeInBits();
1825 if (ElementSizeInBits % 8)
1826 return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
1827 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1828 APInt NumSkippedElements = Offset.udiv(ElementSize);
1829 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1831 Offset -= NumSkippedElements * ElementSize;
1832 Indices.push_back(IRB.getInt(NumSkippedElements));
1833 return getNaturalGEPRecursively(IRB, TD, Ptr, VecTy->getElementType(),
1834 Offset, TargetTy, Indices, Prefix);
1837 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1838 Type *ElementTy = ArrTy->getElementType();
1839 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1840 APInt NumSkippedElements = Offset.udiv(ElementSize);
1841 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1844 Offset -= NumSkippedElements * ElementSize;
1845 Indices.push_back(IRB.getInt(NumSkippedElements));
1846 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1850 StructType *STy = dyn_cast<StructType>(Ty);
1854 const StructLayout *SL = TD.getStructLayout(STy);
1855 uint64_t StructOffset = Offset.getZExtValue();
1856 if (StructOffset >= SL->getSizeInBytes())
1858 unsigned Index = SL->getElementContainingOffset(StructOffset);
1859 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1860 Type *ElementTy = STy->getElementType(Index);
1861 if (Offset.uge(TD.getTypeAllocSize(ElementTy)))
1862 return 0; // The offset points into alignment padding.
1864 Indices.push_back(IRB.getInt32(Index));
1865 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1869 /// \brief Get a natural GEP from a base pointer to a particular offset and
1870 /// resulting in a particular type.
1872 /// The goal is to produce a "natural" looking GEP that works with the existing
1873 /// composite types to arrive at the appropriate offset and element type for
1874 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1875 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1876 /// Indices, and setting Ty to the result subtype.
1878 /// If no natural GEP can be constructed, this function returns null.
1879 static Value *getNaturalGEPWithOffset(IRBuilder<> &IRB, const DataLayout &TD,
1880 Value *Ptr, APInt Offset, Type *TargetTy,
1881 SmallVectorImpl<Value *> &Indices,
1882 const Twine &Prefix) {
1883 PointerType *Ty = cast<PointerType>(Ptr->getType());
1885 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1887 if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
1890 Type *ElementTy = Ty->getElementType();
1891 if (!ElementTy->isSized())
1892 return 0; // We can't GEP through an unsized element.
1893 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1894 if (ElementSize == 0)
1895 return 0; // Zero-length arrays can't help us build a natural GEP.
1896 APInt NumSkippedElements = Offset.udiv(ElementSize);
1898 Offset -= NumSkippedElements * ElementSize;
1899 Indices.push_back(IRB.getInt(NumSkippedElements));
1900 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1904 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1905 /// resulting pointer has PointerTy.
1907 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1908 /// and produces the pointer type desired. Where it cannot, it will try to use
1909 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1910 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1911 /// bitcast to the type.
1913 /// The strategy for finding the more natural GEPs is to peel off layers of the
1914 /// pointer, walking back through bit casts and GEPs, searching for a base
1915 /// pointer from which we can compute a natural GEP with the desired
1916 /// properities. The algorithm tries to fold as many constant indices into
1917 /// a single GEP as possible, thus making each GEP more independent of the
1918 /// surrounding code.
1919 static Value *getAdjustedPtr(IRBuilder<> &IRB, const DataLayout &TD,
1920 Value *Ptr, APInt Offset, Type *PointerTy,
1921 const Twine &Prefix) {
1922 // Even though we don't look through PHI nodes, we could be called on an
1923 // instruction in an unreachable block, which may be on a cycle.
1924 SmallPtrSet<Value *, 4> Visited;
1925 Visited.insert(Ptr);
1926 SmallVector<Value *, 4> Indices;
1928 // We may end up computing an offset pointer that has the wrong type. If we
1929 // never are able to compute one directly that has the correct type, we'll
1930 // fall back to it, so keep it around here.
1931 Value *OffsetPtr = 0;
1933 // Remember any i8 pointer we come across to re-use if we need to do a raw
1936 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1938 Type *TargetTy = PointerTy->getPointerElementType();
1941 // First fold any existing GEPs into the offset.
1942 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1943 APInt GEPOffset(Offset.getBitWidth(), 0);
1944 if (!accumulateGEPOffsets(TD, *GEP, GEPOffset))
1946 Offset += GEPOffset;
1947 Ptr = GEP->getPointerOperand();
1948 if (!Visited.insert(Ptr))
1952 // See if we can perform a natural GEP here.
1954 if (Value *P = getNaturalGEPWithOffset(IRB, TD, Ptr, Offset, TargetTy,
1956 if (P->getType() == PointerTy) {
1957 // Zap any offset pointer that we ended up computing in previous rounds.
1958 if (OffsetPtr && OffsetPtr->use_empty())
1959 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1960 I->eraseFromParent();
1968 // Stash this pointer if we've found an i8*.
1969 if (Ptr->getType()->isIntegerTy(8)) {
1971 Int8PtrOffset = Offset;
1974 // Peel off a layer of the pointer and update the offset appropriately.
1975 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1976 Ptr = cast<Operator>(Ptr)->getOperand(0);
1977 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1978 if (GA->mayBeOverridden())
1980 Ptr = GA->getAliasee();
1984 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1985 } while (Visited.insert(Ptr));
1989 Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
1990 Prefix + ".raw_cast");
1991 Int8PtrOffset = Offset;
1994 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1995 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1996 Prefix + ".raw_idx");
2000 // On the off chance we were targeting i8*, guard the bitcast here.
2001 if (Ptr->getType() != PointerTy)
2002 Ptr = IRB.CreateBitCast(Ptr, PointerTy, Prefix + ".cast");
2007 /// \brief Test whether we can convert a value from the old to the new type.
2009 /// This predicate should be used to guard calls to convertValue in order to
2010 /// ensure that we only try to convert viable values. The strategy is that we
2011 /// will peel off single element struct and array wrappings to get to an
2012 /// underlying value, and convert that value.
2013 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
2016 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
2018 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
2021 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
2022 if (NewTy->isPointerTy() && OldTy->isPointerTy())
2024 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
2032 /// \brief Generic routine to convert an SSA value to a value of a different
2035 /// This will try various different casting techniques, such as bitcasts,
2036 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
2037 /// two types for viability with this routine.
2038 static Value *convertValue(const DataLayout &DL, IRBuilder<> &IRB, Value *V,
2040 assert(canConvertValue(DL, V->getType(), Ty) &&
2041 "Value not convertable to type");
2042 if (V->getType() == Ty)
2044 if (V->getType()->isIntegerTy() && Ty->isPointerTy())
2045 return IRB.CreateIntToPtr(V, Ty);
2046 if (V->getType()->isPointerTy() && Ty->isIntegerTy())
2047 return IRB.CreatePtrToInt(V, Ty);
2049 return IRB.CreateBitCast(V, Ty);
2052 /// \brief Test whether the given alloca partition can be promoted to a vector.
2054 /// This is a quick test to check whether we can rewrite a particular alloca
2055 /// partition (and its newly formed alloca) into a vector alloca with only
2056 /// whole-vector loads and stores such that it could be promoted to a vector
2057 /// SSA value. We only can ensure this for a limited set of operations, and we
2058 /// don't want to do the rewrites unless we are confident that the result will
2059 /// be promotable, so we have an early test here.
2060 static bool isVectorPromotionViable(const DataLayout &TD,
2062 AllocaPartitioning &P,
2063 uint64_t PartitionBeginOffset,
2064 uint64_t PartitionEndOffset,
2065 AllocaPartitioning::const_use_iterator I,
2066 AllocaPartitioning::const_use_iterator E) {
2067 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
2071 uint64_t VecSize = TD.getTypeSizeInBits(Ty);
2072 uint64_t ElementSize = Ty->getScalarSizeInBits();
2074 // While the definition of LLVM vectors is bitpacked, we don't support sizes
2075 // that aren't byte sized.
2076 if (ElementSize % 8)
2078 assert((VecSize % 8) == 0 && "vector size not a multiple of element size?");
2082 for (; I != E; ++I) {
2084 continue; // Skip dead use.
2086 uint64_t BeginOffset = I->BeginOffset - PartitionBeginOffset;
2087 uint64_t BeginIndex = BeginOffset / ElementSize;
2088 if (BeginIndex * ElementSize != BeginOffset ||
2089 BeginIndex >= Ty->getNumElements())
2091 uint64_t EndOffset = I->EndOffset - PartitionBeginOffset;
2092 uint64_t EndIndex = EndOffset / ElementSize;
2093 if (EndIndex * ElementSize != EndOffset ||
2094 EndIndex > Ty->getNumElements())
2097 // FIXME: We should build shuffle vector instructions to handle
2098 // non-element-sized accesses.
2099 if ((EndOffset - BeginOffset) != ElementSize &&
2100 (EndOffset - BeginOffset) != VecSize)
2103 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
2104 if (MI->isVolatile())
2106 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
2107 const AllocaPartitioning::MemTransferOffsets &MTO
2108 = P.getMemTransferOffsets(*MTI);
2109 if (!MTO.IsSplittable)
2112 } else if (I->U->get()->getType()->getPointerElementType()->isStructTy()) {
2113 // Disable vector promotion when there are loads or stores of an FCA.
2115 } else if (!isa<LoadInst>(I->U->getUser()) &&
2116 !isa<StoreInst>(I->U->getUser())) {
2123 /// \brief Test whether the given alloca partition's integer operations can be
2124 /// widened to promotable ones.
2126 /// This is a quick test to check whether we can rewrite the integer loads and
2127 /// stores to a particular alloca into wider loads and stores and be able to
2128 /// promote the resulting alloca.
2129 static bool isIntegerWideningViable(const DataLayout &TD,
2131 uint64_t AllocBeginOffset,
2132 AllocaPartitioning &P,
2133 AllocaPartitioning::const_use_iterator I,
2134 AllocaPartitioning::const_use_iterator E) {
2135 uint64_t SizeInBits = TD.getTypeSizeInBits(AllocaTy);
2137 // Don't try to handle allocas with bit-padding.
2138 if (SizeInBits != TD.getTypeStoreSizeInBits(AllocaTy))
2141 uint64_t Size = TD.getTypeStoreSize(AllocaTy);
2143 // Check the uses to ensure the uses are (likely) promoteable integer uses.
2144 // Also ensure that the alloca has a covering load or store. We don't want
2145 // to widen the integer operotains only to fail to promote due to some other
2146 // unsplittable entry (which we may make splittable later).
2147 bool WholeAllocaOp = false;
2148 for (; I != E; ++I) {
2150 continue; // Skip dead use.
2152 uint64_t RelBegin = I->BeginOffset - AllocBeginOffset;
2153 uint64_t RelEnd = I->EndOffset - AllocBeginOffset;
2155 // We can't reasonably handle cases where the load or store extends past
2156 // the end of the aloca's type and into its padding.
2160 if (LoadInst *LI = dyn_cast<LoadInst>(I->U->getUser())) {
2161 if (LI->isVolatile())
2163 if (RelBegin == 0 && RelEnd == Size)
2164 WholeAllocaOp = true;
2165 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
2166 if (ITy->getBitWidth() < TD.getTypeStoreSize(ITy))
2170 // Non-integer loads need to be convertible from the alloca type so that
2171 // they are promotable.
2172 if (RelBegin != 0 || RelEnd != Size ||
2173 !canConvertValue(TD, AllocaTy, LI->getType()))
2175 } else if (StoreInst *SI = dyn_cast<StoreInst>(I->U->getUser())) {
2176 Type *ValueTy = SI->getValueOperand()->getType();
2177 if (SI->isVolatile())
2179 if (RelBegin == 0 && RelEnd == Size)
2180 WholeAllocaOp = true;
2181 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
2182 if (ITy->getBitWidth() < TD.getTypeStoreSize(ITy))
2186 // Non-integer stores need to be convertible to the alloca type so that
2187 // they are promotable.
2188 if (RelBegin != 0 || RelEnd != Size ||
2189 !canConvertValue(TD, ValueTy, AllocaTy))
2191 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
2192 if (MI->isVolatile())
2194 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
2195 const AllocaPartitioning::MemTransferOffsets &MTO
2196 = P.getMemTransferOffsets(*MTI);
2197 if (!MTO.IsSplittable)
2200 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->U->getUser())) {
2201 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2202 II->getIntrinsicID() != Intrinsic::lifetime_end)
2208 return WholeAllocaOp;
2212 /// \brief Visitor to rewrite instructions using a partition of an alloca to
2213 /// use a new alloca.
2215 /// Also implements the rewriting to vector-based accesses when the partition
2216 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2218 class AllocaPartitionRewriter : public InstVisitor<AllocaPartitionRewriter,
2220 // Befriend the base class so it can delegate to private visit methods.
2221 friend class llvm::InstVisitor<AllocaPartitionRewriter, bool>;
2223 const DataLayout &TD;
2224 AllocaPartitioning &P;
2226 AllocaInst &OldAI, &NewAI;
2227 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2230 // If we are rewriting an alloca partition which can be written as pure
2231 // vector operations, we stash extra information here. When VecTy is
2232 // non-null, we have some strict guarantees about the rewriten alloca:
2233 // - The new alloca is exactly the size of the vector type here.
2234 // - The accesses all either map to the entire vector or to a single
2236 // - The set of accessing instructions is only one of those handled above
2237 // in isVectorPromotionViable. Generally these are the same access kinds
2238 // which are promotable via mem2reg.
2241 uint64_t ElementSize;
2243 // This is a convenience and flag variable that will be null unless the new
2244 // alloca's integer operations should be widened to this integer type due to
2245 // passing isIntegerWideningViable above. If it is non-null, the desired
2246 // integer type will be stored here for easy access during rewriting.
2249 // The offset of the partition user currently being rewritten.
2250 uint64_t BeginOffset, EndOffset;
2252 Instruction *OldPtr;
2254 // The name prefix to use when rewriting instructions for this alloca.
2255 std::string NamePrefix;
2258 AllocaPartitionRewriter(const DataLayout &TD, AllocaPartitioning &P,
2259 AllocaPartitioning::iterator PI,
2260 SROA &Pass, AllocaInst &OldAI, AllocaInst &NewAI,
2261 uint64_t NewBeginOffset, uint64_t NewEndOffset)
2262 : TD(TD), P(P), Pass(Pass),
2263 OldAI(OldAI), NewAI(NewAI),
2264 NewAllocaBeginOffset(NewBeginOffset),
2265 NewAllocaEndOffset(NewEndOffset),
2266 NewAllocaTy(NewAI.getAllocatedType()),
2267 VecTy(), ElementTy(), ElementSize(), IntTy(),
2268 BeginOffset(), EndOffset() {
2271 /// \brief Visit the users of the alloca partition and rewrite them.
2272 bool visitUsers(AllocaPartitioning::const_use_iterator I,
2273 AllocaPartitioning::const_use_iterator E) {
2274 if (isVectorPromotionViable(TD, NewAI.getAllocatedType(), P,
2275 NewAllocaBeginOffset, NewAllocaEndOffset,
2278 VecTy = cast<VectorType>(NewAI.getAllocatedType());
2279 ElementTy = VecTy->getElementType();
2280 assert((VecTy->getScalarSizeInBits() % 8) == 0 &&
2281 "Only multiple-of-8 sized vector elements are viable");
2282 ElementSize = VecTy->getScalarSizeInBits() / 8;
2283 } else if (isIntegerWideningViable(TD, NewAI.getAllocatedType(),
2284 NewAllocaBeginOffset, P, I, E)) {
2285 IntTy = Type::getIntNTy(NewAI.getContext(),
2286 TD.getTypeSizeInBits(NewAI.getAllocatedType()));
2288 bool CanSROA = true;
2289 for (; I != E; ++I) {
2291 continue; // Skip dead uses.
2292 BeginOffset = I->BeginOffset;
2293 EndOffset = I->EndOffset;
2295 OldPtr = cast<Instruction>(I->U->get());
2296 NamePrefix = (Twine(NewAI.getName()) + "." + Twine(BeginOffset)).str();
2297 CanSROA &= visit(cast<Instruction>(I->U->getUser()));
2313 // Every instruction which can end up as a user must have a rewrite rule.
2314 bool visitInstruction(Instruction &I) {
2315 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2316 llvm_unreachable("No rewrite rule for this instruction!");
2319 Twine getName(const Twine &Suffix) {
2320 return NamePrefix + Suffix;
2323 Value *getAdjustedAllocaPtr(IRBuilder<> &IRB, Type *PointerTy) {
2324 assert(BeginOffset >= NewAllocaBeginOffset);
2325 APInt Offset(TD.getPointerSizeInBits(), BeginOffset - NewAllocaBeginOffset);
2326 return getAdjustedPtr(IRB, TD, &NewAI, Offset, PointerTy, getName(""));
2329 /// \brief Compute suitable alignment to access an offset into the new alloca.
2330 unsigned getOffsetAlign(uint64_t Offset) {
2331 unsigned NewAIAlign = NewAI.getAlignment();
2333 NewAIAlign = TD.getABITypeAlignment(NewAI.getAllocatedType());
2334 return MinAlign(NewAIAlign, Offset);
2337 /// \brief Compute suitable alignment to access this partition of the new
2339 unsigned getPartitionAlign() {
2340 return getOffsetAlign(BeginOffset - NewAllocaBeginOffset);
2343 /// \brief Compute suitable alignment to access a type at an offset of the
2346 /// \returns zero if the type's ABI alignment is a suitable alignment,
2347 /// otherwise returns the maximal suitable alignment.
2348 unsigned getOffsetTypeAlign(Type *Ty, uint64_t Offset) {
2349 unsigned Align = getOffsetAlign(Offset);
2350 return Align == TD.getABITypeAlignment(Ty) ? 0 : Align;
2353 /// \brief Compute suitable alignment to access a type at the beginning of
2354 /// this partition of the new alloca.
2356 /// See \c getOffsetTypeAlign for details; this routine delegates to it.
2357 unsigned getPartitionTypeAlign(Type *Ty) {
2358 return getOffsetTypeAlign(Ty, BeginOffset - NewAllocaBeginOffset);
2361 ConstantInt *getIndex(IRBuilder<> &IRB, uint64_t Offset) {
2362 assert(VecTy && "Can only call getIndex when rewriting a vector");
2363 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2364 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2365 uint32_t Index = RelOffset / ElementSize;
2366 assert(Index * ElementSize == RelOffset);
2367 return IRB.getInt32(Index);
2370 Value *extractInteger(IRBuilder<> &IRB, IntegerType *TargetTy,
2372 assert(IntTy && "We cannot extract an integer from the alloca");
2373 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2375 V = convertValue(TD, IRB, V, IntTy);
2376 assert(Offset >= NewAllocaBeginOffset && "Out of bounds offset");
2377 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2378 assert(TD.getTypeStoreSize(TargetTy) + RelOffset <=
2379 TD.getTypeStoreSize(IntTy) &&
2380 "Element load outside of alloca store");
2381 uint64_t ShAmt = 8*RelOffset;
2382 if (TD.isBigEndian())
2383 ShAmt = 8*(TD.getTypeStoreSize(IntTy) -
2384 TD.getTypeStoreSize(TargetTy) - RelOffset);
2386 V = IRB.CreateLShr(V, ShAmt, getName(".shift"));
2387 assert(TargetTy->getBitWidth() <= IntTy->getBitWidth() &&
2388 "Cannot extract to a larger integer!");
2389 if (TargetTy != IntTy)
2390 V = IRB.CreateTrunc(V, TargetTy, getName(".trunc"));
2394 StoreInst *insertInteger(IRBuilder<> &IRB, Value *V, uint64_t Offset) {
2395 IntegerType *Ty = cast<IntegerType>(V->getType());
2396 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2397 "Cannot insert a larger integer!");
2399 V = IRB.CreateZExt(V, IntTy, getName(".ext"));
2400 assert(Offset >= NewAllocaBeginOffset && "Out of bounds offset");
2401 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2402 assert(TD.getTypeStoreSize(Ty) + RelOffset <=
2403 TD.getTypeStoreSize(IntTy) &&
2404 "Element store outside of alloca store");
2405 uint64_t ShAmt = 8*RelOffset;
2406 if (TD.isBigEndian())
2407 ShAmt = 8*(TD.getTypeStoreSize(IntTy) - TD.getTypeStoreSize(Ty)
2410 V = IRB.CreateShl(V, ShAmt, getName(".shift"));
2412 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2413 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2414 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2415 getName(".oldload"));
2416 Old = convertValue(TD, IRB, Old, IntTy);
2417 Old = IRB.CreateAnd(Old, Mask, getName(".mask"));
2418 V = IRB.CreateOr(Old, V, getName(".insert"));
2420 V = convertValue(TD, IRB, V, NewAllocaTy);
2421 return IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2424 void deleteIfTriviallyDead(Value *V) {
2425 Instruction *I = cast<Instruction>(V);
2426 if (isInstructionTriviallyDead(I))
2427 Pass.DeadInsts.push_back(I);
2430 bool rewriteVectorizedLoadInst(IRBuilder<> &IRB, LoadInst &LI, Value *OldOp) {
2432 if (LI.getType() == VecTy->getElementType() ||
2433 BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
2434 Result = IRB.CreateExtractElement(
2435 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), getName(".load")),
2436 getIndex(IRB, BeginOffset), getName(".extract"));
2438 Result = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2441 if (Result->getType() != LI.getType())
2442 Result = convertValue(TD, IRB, Result, LI.getType());
2443 LI.replaceAllUsesWith(Result);
2444 Pass.DeadInsts.push_back(&LI);
2446 DEBUG(dbgs() << " to: " << *Result << "\n");
2450 bool rewriteIntegerLoad(IRBuilder<> &IRB, LoadInst &LI) {
2451 assert(!LI.isVolatile());
2452 Value *Result = extractInteger(IRB, cast<IntegerType>(LI.getType()),
2454 LI.replaceAllUsesWith(Result);
2455 Pass.DeadInsts.push_back(&LI);
2456 DEBUG(dbgs() << " to: " << *Result << "\n");
2460 bool visitLoadInst(LoadInst &LI) {
2461 DEBUG(dbgs() << " original: " << LI << "\n");
2462 Value *OldOp = LI.getOperand(0);
2463 assert(OldOp == OldPtr);
2464 IRBuilder<> IRB(&LI);
2467 return rewriteVectorizedLoadInst(IRB, LI, OldOp);
2468 if (IntTy && LI.getType()->isIntegerTy())
2469 return rewriteIntegerLoad(IRB, LI);
2471 if (BeginOffset == NewAllocaBeginOffset &&
2472 canConvertValue(TD, NewAllocaTy, LI.getType())) {
2473 Value *NewLI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2474 LI.isVolatile(), getName(".load"));
2475 Value *NewV = convertValue(TD, IRB, NewLI, LI.getType());
2476 LI.replaceAllUsesWith(NewV);
2477 Pass.DeadInsts.push_back(&LI);
2479 DEBUG(dbgs() << " to: " << *NewLI << "\n");
2480 return !LI.isVolatile();
2483 assert(!IntTy && "Invalid load found with int-op widening enabled");
2485 Value *NewPtr = getAdjustedAllocaPtr(IRB,
2486 LI.getPointerOperand()->getType());
2487 LI.setOperand(0, NewPtr);
2488 LI.setAlignment(getPartitionTypeAlign(LI.getType()));
2489 DEBUG(dbgs() << " to: " << LI << "\n");
2491 deleteIfTriviallyDead(OldOp);
2492 return NewPtr == &NewAI && !LI.isVolatile();
2495 bool rewriteVectorizedStoreInst(IRBuilder<> &IRB, StoreInst &SI,
2497 Value *V = SI.getValueOperand();
2498 if (V->getType() == ElementTy ||
2499 BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
2500 if (V->getType() != ElementTy)
2501 V = convertValue(TD, IRB, V, ElementTy);
2502 LoadInst *LI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2504 V = IRB.CreateInsertElement(LI, V, getIndex(IRB, BeginOffset),
2505 getName(".insert"));
2506 } else if (V->getType() != VecTy) {
2507 V = convertValue(TD, IRB, V, VecTy);
2509 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2510 Pass.DeadInsts.push_back(&SI);
2513 DEBUG(dbgs() << " to: " << *Store << "\n");
2517 bool rewriteIntegerStore(IRBuilder<> &IRB, StoreInst &SI) {
2518 assert(!SI.isVolatile());
2519 StoreInst *Store = insertInteger(IRB, SI.getValueOperand(), BeginOffset);
2520 Pass.DeadInsts.push_back(&SI);
2522 DEBUG(dbgs() << " to: " << *Store << "\n");
2526 bool visitStoreInst(StoreInst &SI) {
2527 DEBUG(dbgs() << " original: " << SI << "\n");
2528 Value *OldOp = SI.getOperand(1);
2529 assert(OldOp == OldPtr);
2530 IRBuilder<> IRB(&SI);
2533 return rewriteVectorizedStoreInst(IRB, SI, OldOp);
2534 Type *ValueTy = SI.getValueOperand()->getType();
2535 if (IntTy && ValueTy->isIntegerTy())
2536 return rewriteIntegerStore(IRB, SI);
2538 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2539 // alloca that should be re-examined after promoting this alloca.
2540 if (ValueTy->isPointerTy())
2541 if (AllocaInst *AI = dyn_cast<AllocaInst>(SI.getValueOperand()
2542 ->stripInBoundsOffsets()))
2543 Pass.PostPromotionWorklist.insert(AI);
2545 if (BeginOffset == NewAllocaBeginOffset &&
2546 canConvertValue(TD, ValueTy, NewAllocaTy)) {
2547 Value *NewV = convertValue(TD, IRB, SI.getValueOperand(), NewAllocaTy);
2548 StoreInst *NewSI = IRB.CreateAlignedStore(NewV, &NewAI, NewAI.getAlignment(),
2551 Pass.DeadInsts.push_back(&SI);
2553 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2554 return !SI.isVolatile();
2557 assert(!IntTy && "Invalid store found with int-op widening enabled");
2559 Value *NewPtr = getAdjustedAllocaPtr(IRB,
2560 SI.getPointerOperand()->getType());
2561 SI.setOperand(1, NewPtr);
2562 SI.setAlignment(getPartitionTypeAlign(SI.getValueOperand()->getType()));
2563 DEBUG(dbgs() << " to: " << SI << "\n");
2565 deleteIfTriviallyDead(OldOp);
2566 return NewPtr == &NewAI && !SI.isVolatile();
2569 bool visitMemSetInst(MemSetInst &II) {
2570 DEBUG(dbgs() << " original: " << II << "\n");
2571 IRBuilder<> IRB(&II);
2572 assert(II.getRawDest() == OldPtr);
2574 // If the memset has a variable size, it cannot be split, just adjust the
2575 // pointer to the new alloca.
2576 if (!isa<Constant>(II.getLength())) {
2577 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2578 Type *CstTy = II.getAlignmentCst()->getType();
2579 II.setAlignment(ConstantInt::get(CstTy, getPartitionAlign()));
2581 deleteIfTriviallyDead(OldPtr);
2585 // Record this instruction for deletion.
2586 if (Pass.DeadSplitInsts.insert(&II))
2587 Pass.DeadInsts.push_back(&II);
2589 Type *AllocaTy = NewAI.getAllocatedType();
2590 Type *ScalarTy = AllocaTy->getScalarType();
2592 // If this doesn't map cleanly onto the alloca type, and that type isn't
2593 // a single value type, just emit a memset.
2594 if (!VecTy && !IntTy &&
2595 (BeginOffset != NewAllocaBeginOffset ||
2596 EndOffset != NewAllocaEndOffset ||
2597 !AllocaTy->isSingleValueType() ||
2598 !TD.isLegalInteger(TD.getTypeSizeInBits(ScalarTy)))) {
2599 Type *SizeTy = II.getLength()->getType();
2600 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2602 = IRB.CreateMemSet(getAdjustedAllocaPtr(IRB,
2603 II.getRawDest()->getType()),
2604 II.getValue(), Size, getPartitionAlign(),
2607 DEBUG(dbgs() << " to: " << *New << "\n");
2611 // If we can represent this as a simple value, we have to build the actual
2612 // value to store, which requires expanding the byte present in memset to
2613 // a sensible representation for the alloca type. This is essentially
2614 // splatting the byte to a sufficiently wide integer, bitcasting to the
2615 // desired scalar type, and splatting it across any desired vector type.
2616 uint64_t Size = EndOffset - BeginOffset;
2617 Value *V = II.getValue();
2618 IntegerType *VTy = cast<IntegerType>(V->getType());
2619 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size*8);
2620 if (Size*8 > VTy->getBitWidth())
2621 V = IRB.CreateMul(IRB.CreateZExt(V, SplatIntTy, getName(".zext")),
2622 ConstantExpr::getUDiv(
2623 Constant::getAllOnesValue(SplatIntTy),
2624 ConstantExpr::getZExt(
2625 Constant::getAllOnesValue(V->getType()),
2627 getName(".isplat"));
2629 // If this is an element-wide memset of a vectorizable alloca, insert it.
2630 if (VecTy && (BeginOffset > NewAllocaBeginOffset ||
2631 EndOffset < NewAllocaEndOffset)) {
2632 if (V->getType() != ScalarTy)
2633 V = convertValue(TD, IRB, V, ScalarTy);
2634 StoreInst *Store = IRB.CreateAlignedStore(
2635 IRB.CreateInsertElement(IRB.CreateAlignedLoad(&NewAI,
2636 NewAI.getAlignment(),
2638 V, getIndex(IRB, BeginOffset),
2639 getName(".insert")),
2640 &NewAI, NewAI.getAlignment());
2642 DEBUG(dbgs() << " to: " << *Store << "\n");
2646 // If this is a memset on an alloca where we can widen stores, insert the
2648 if (IntTy && (BeginOffset > NewAllocaBeginOffset ||
2649 EndOffset < NewAllocaEndOffset)) {
2650 assert(!II.isVolatile());
2651 StoreInst *Store = insertInteger(IRB, V, BeginOffset);
2653 DEBUG(dbgs() << " to: " << *Store << "\n");
2657 if (V->getType() != AllocaTy)
2658 V = convertValue(TD, IRB, V, AllocaTy);
2660 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2663 DEBUG(dbgs() << " to: " << *New << "\n");
2664 return !II.isVolatile();
2667 bool visitMemTransferInst(MemTransferInst &II) {
2668 // Rewriting of memory transfer instructions can be a bit tricky. We break
2669 // them into two categories: split intrinsics and unsplit intrinsics.
2671 DEBUG(dbgs() << " original: " << II << "\n");
2672 IRBuilder<> IRB(&II);
2674 assert(II.getRawSource() == OldPtr || II.getRawDest() == OldPtr);
2675 bool IsDest = II.getRawDest() == OldPtr;
2677 const AllocaPartitioning::MemTransferOffsets &MTO
2678 = P.getMemTransferOffsets(II);
2680 // Compute the relative offset within the transfer.
2681 unsigned IntPtrWidth = TD.getPointerSizeInBits();
2682 APInt RelOffset(IntPtrWidth, BeginOffset - (IsDest ? MTO.DestBegin
2683 : MTO.SourceBegin));
2685 unsigned Align = II.getAlignment();
2687 Align = MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(),
2688 MinAlign(II.getAlignment(), getPartitionAlign()));
2690 // For unsplit intrinsics, we simply modify the source and destination
2691 // pointers in place. This isn't just an optimization, it is a matter of
2692 // correctness. With unsplit intrinsics we may be dealing with transfers
2693 // within a single alloca before SROA ran, or with transfers that have
2694 // a variable length. We may also be dealing with memmove instead of
2695 // memcpy, and so simply updating the pointers is the necessary for us to
2696 // update both source and dest of a single call.
2697 if (!MTO.IsSplittable) {
2698 Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource();
2700 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2702 II.setSource(getAdjustedAllocaPtr(IRB, II.getRawSource()->getType()));
2704 Type *CstTy = II.getAlignmentCst()->getType();
2705 II.setAlignment(ConstantInt::get(CstTy, Align));
2707 DEBUG(dbgs() << " to: " << II << "\n");
2708 deleteIfTriviallyDead(OldOp);
2711 // For split transfer intrinsics we have an incredibly useful assurance:
2712 // the source and destination do not reside within the same alloca, and at
2713 // least one of them does not escape. This means that we can replace
2714 // memmove with memcpy, and we don't need to worry about all manner of
2715 // downsides to splitting and transforming the operations.
2717 // If this doesn't map cleanly onto the alloca type, and that type isn't
2718 // a single value type, just emit a memcpy.
2720 = !VecTy && !IntTy && (BeginOffset != NewAllocaBeginOffset ||
2721 EndOffset != NewAllocaEndOffset ||
2722 !NewAI.getAllocatedType()->isSingleValueType());
2724 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2725 // size hasn't been shrunk based on analysis of the viable range, this is
2727 if (EmitMemCpy && &OldAI == &NewAI) {
2728 uint64_t OrigBegin = IsDest ? MTO.DestBegin : MTO.SourceBegin;
2729 uint64_t OrigEnd = IsDest ? MTO.DestEnd : MTO.SourceEnd;
2730 // Ensure the start lines up.
2731 assert(BeginOffset == OrigBegin);
2734 // Rewrite the size as needed.
2735 if (EndOffset != OrigEnd)
2736 II.setLength(ConstantInt::get(II.getLength()->getType(),
2737 EndOffset - BeginOffset));
2740 // Record this instruction for deletion.
2741 if (Pass.DeadSplitInsts.insert(&II))
2742 Pass.DeadInsts.push_back(&II);
2744 bool IsWholeAlloca = BeginOffset == NewAllocaBeginOffset &&
2745 EndOffset == NewAllocaEndOffset;
2746 bool IsVectorElement = VecTy && !IsWholeAlloca;
2747 uint64_t Size = EndOffset - BeginOffset;
2748 IntegerType *SubIntTy
2749 = IntTy ? Type::getIntNTy(IntTy->getContext(), Size*8) : 0;
2751 Type *OtherPtrTy = IsDest ? II.getRawSource()->getType()
2752 : II.getRawDest()->getType();
2754 if (IsVectorElement)
2755 OtherPtrTy = VecTy->getElementType()->getPointerTo();
2756 else if (IntTy && !IsWholeAlloca)
2757 OtherPtrTy = SubIntTy->getPointerTo();
2759 OtherPtrTy = NewAI.getType();
2762 // Compute the other pointer, folding as much as possible to produce
2763 // a single, simple GEP in most cases.
2764 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2765 OtherPtr = getAdjustedPtr(IRB, TD, OtherPtr, RelOffset, OtherPtrTy,
2766 getName("." + OtherPtr->getName()));
2768 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2769 // alloca that should be re-examined after rewriting this instruction.
2771 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets()))
2772 Pass.Worklist.insert(AI);
2776 = getAdjustedAllocaPtr(IRB, IsDest ? II.getRawDest()->getType()
2777 : II.getRawSource()->getType());
2778 Type *SizeTy = II.getLength()->getType();
2779 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2781 CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
2782 IsDest ? OtherPtr : OurPtr,
2783 Size, Align, II.isVolatile());
2785 DEBUG(dbgs() << " to: " << *New << "\n");
2789 // Note that we clamp the alignment to 1 here as a 0 alignment for a memcpy
2790 // is equivalent to 1, but that isn't true if we end up rewriting this as
2795 Value *SrcPtr = OtherPtr;
2796 Value *DstPtr = &NewAI;
2798 std::swap(SrcPtr, DstPtr);
2801 if (IsVectorElement && !IsDest) {
2802 // We have to extract rather than load.
2803 Src = IRB.CreateExtractElement(
2804 IRB.CreateAlignedLoad(SrcPtr, Align, getName(".copyload")),
2805 getIndex(IRB, BeginOffset),
2806 getName(".copyextract"));
2807 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2808 Src = extractInteger(IRB, SubIntTy, BeginOffset);
2810 Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(),
2811 getName(".copyload"));
2814 if (IntTy && !IsWholeAlloca && IsDest) {
2815 StoreInst *Store = insertInteger(IRB, Src, BeginOffset);
2817 DEBUG(dbgs() << " to: " << *Store << "\n");
2821 if (IsVectorElement && IsDest) {
2822 // We have to insert into a loaded copy before storing.
2823 Src = IRB.CreateInsertElement(
2824 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), getName(".load")),
2825 Src, getIndex(IRB, BeginOffset),
2826 getName(".insert"));
2829 StoreInst *Store = cast<StoreInst>(
2830 IRB.CreateAlignedStore(Src, DstPtr, Align, II.isVolatile()));
2832 DEBUG(dbgs() << " to: " << *Store << "\n");
2833 return !II.isVolatile();
2836 bool visitIntrinsicInst(IntrinsicInst &II) {
2837 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2838 II.getIntrinsicID() == Intrinsic::lifetime_end);
2839 DEBUG(dbgs() << " original: " << II << "\n");
2840 IRBuilder<> IRB(&II);
2841 assert(II.getArgOperand(1) == OldPtr);
2843 // Record this instruction for deletion.
2844 if (Pass.DeadSplitInsts.insert(&II))
2845 Pass.DeadInsts.push_back(&II);
2848 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2849 EndOffset - BeginOffset);
2850 Value *Ptr = getAdjustedAllocaPtr(IRB, II.getArgOperand(1)->getType());
2852 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2853 New = IRB.CreateLifetimeStart(Ptr, Size);
2855 New = IRB.CreateLifetimeEnd(Ptr, Size);
2857 DEBUG(dbgs() << " to: " << *New << "\n");
2861 bool visitPHINode(PHINode &PN) {
2862 DEBUG(dbgs() << " original: " << PN << "\n");
2864 // We would like to compute a new pointer in only one place, but have it be
2865 // as local as possible to the PHI. To do that, we re-use the location of
2866 // the old pointer, which necessarily must be in the right position to
2867 // dominate the PHI.
2868 IRBuilder<> PtrBuilder(cast<Instruction>(OldPtr));
2870 Value *NewPtr = getAdjustedAllocaPtr(PtrBuilder, OldPtr->getType());
2871 // Replace the operands which were using the old pointer.
2872 User::op_iterator OI = PN.op_begin(), OE = PN.op_end();
2873 for (; OI != OE; ++OI)
2877 DEBUG(dbgs() << " to: " << PN << "\n");
2878 deleteIfTriviallyDead(OldPtr);
2882 bool visitSelectInst(SelectInst &SI) {
2883 DEBUG(dbgs() << " original: " << SI << "\n");
2884 IRBuilder<> IRB(&SI);
2886 // Find the operand we need to rewrite here.
2887 bool IsTrueVal = SI.getTrueValue() == OldPtr;
2889 assert(SI.getFalseValue() != OldPtr && "Pointer is both operands!");
2891 assert(SI.getFalseValue() == OldPtr && "Pointer isn't an operand!");
2893 Value *NewPtr = getAdjustedAllocaPtr(IRB, OldPtr->getType());
2894 SI.setOperand(IsTrueVal ? 1 : 2, NewPtr);
2895 DEBUG(dbgs() << " to: " << SI << "\n");
2896 deleteIfTriviallyDead(OldPtr);
2904 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2906 /// This pass aggressively rewrites all aggregate loads and stores on
2907 /// a particular pointer (or any pointer derived from it which we can identify)
2908 /// with scalar loads and stores.
2909 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2910 // Befriend the base class so it can delegate to private visit methods.
2911 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2913 const DataLayout &TD;
2915 /// Queue of pointer uses to analyze and potentially rewrite.
2916 SmallVector<Use *, 8> Queue;
2918 /// Set to prevent us from cycling with phi nodes and loops.
2919 SmallPtrSet<User *, 8> Visited;
2921 /// The current pointer use being rewritten. This is used to dig up the used
2922 /// value (as opposed to the user).
2926 AggLoadStoreRewriter(const DataLayout &TD) : TD(TD) {}
2928 /// Rewrite loads and stores through a pointer and all pointers derived from
2930 bool rewrite(Instruction &I) {
2931 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2933 bool Changed = false;
2934 while (!Queue.empty()) {
2935 U = Queue.pop_back_val();
2936 Changed |= visit(cast<Instruction>(U->getUser()));
2942 /// Enqueue all the users of the given instruction for further processing.
2943 /// This uses a set to de-duplicate users.
2944 void enqueueUsers(Instruction &I) {
2945 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
2947 if (Visited.insert(*UI))
2948 Queue.push_back(&UI.getUse());
2951 // Conservative default is to not rewrite anything.
2952 bool visitInstruction(Instruction &I) { return false; }
2954 /// \brief Generic recursive split emission class.
2955 template <typename Derived>
2958 /// The builder used to form new instructions.
2960 /// The indices which to be used with insert- or extractvalue to select the
2961 /// appropriate value within the aggregate.
2962 SmallVector<unsigned, 4> Indices;
2963 /// The indices to a GEP instruction which will move Ptr to the correct slot
2964 /// within the aggregate.
2965 SmallVector<Value *, 4> GEPIndices;
2966 /// The base pointer of the original op, used as a base for GEPing the
2967 /// split operations.
2970 /// Initialize the splitter with an insertion point, Ptr and start with a
2971 /// single zero GEP index.
2972 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
2973 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
2976 /// \brief Generic recursive split emission routine.
2978 /// This method recursively splits an aggregate op (load or store) into
2979 /// scalar or vector ops. It splits recursively until it hits a single value
2980 /// and emits that single value operation via the template argument.
2982 /// The logic of this routine relies on GEPs and insertvalue and
2983 /// extractvalue all operating with the same fundamental index list, merely
2984 /// formatted differently (GEPs need actual values).
2986 /// \param Ty The type being split recursively into smaller ops.
2987 /// \param Agg The aggregate value being built up or stored, depending on
2988 /// whether this is splitting a load or a store respectively.
2989 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
2990 if (Ty->isSingleValueType())
2991 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
2993 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2994 unsigned OldSize = Indices.size();
2996 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
2998 assert(Indices.size() == OldSize && "Did not return to the old size");
2999 Indices.push_back(Idx);
3000 GEPIndices.push_back(IRB.getInt32(Idx));
3001 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3002 GEPIndices.pop_back();
3008 if (StructType *STy = dyn_cast<StructType>(Ty)) {
3009 unsigned OldSize = Indices.size();
3011 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3013 assert(Indices.size() == OldSize && "Did not return to the old size");
3014 Indices.push_back(Idx);
3015 GEPIndices.push_back(IRB.getInt32(Idx));
3016 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3017 GEPIndices.pop_back();
3023 llvm_unreachable("Only arrays and structs are aggregate loadable types");
3027 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3028 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3029 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
3031 /// Emit a leaf load of a single value. This is called at the leaves of the
3032 /// recursive emission to actually load values.
3033 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3034 assert(Ty->isSingleValueType());
3035 // Load the single value and insert it using the indices.
3036 Value *Load = IRB.CreateLoad(IRB.CreateInBoundsGEP(Ptr, GEPIndices,
3039 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3040 DEBUG(dbgs() << " to: " << *Load << "\n");
3044 bool visitLoadInst(LoadInst &LI) {
3045 assert(LI.getPointerOperand() == *U);
3046 if (!LI.isSimple() || LI.getType()->isSingleValueType())
3049 // We have an aggregate being loaded, split it apart.
3050 DEBUG(dbgs() << " original: " << LI << "\n");
3051 LoadOpSplitter Splitter(&LI, *U);
3052 Value *V = UndefValue::get(LI.getType());
3053 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3054 LI.replaceAllUsesWith(V);
3055 LI.eraseFromParent();
3059 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3060 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3061 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
3063 /// Emit a leaf store of a single value. This is called at the leaves of the
3064 /// recursive emission to actually produce stores.
3065 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3066 assert(Ty->isSingleValueType());
3067 // Extract the single value and store it using the indices.
3068 Value *Store = IRB.CreateStore(
3069 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
3070 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
3072 DEBUG(dbgs() << " to: " << *Store << "\n");
3076 bool visitStoreInst(StoreInst &SI) {
3077 if (!SI.isSimple() || SI.getPointerOperand() != *U)
3079 Value *V = SI.getValueOperand();
3080 if (V->getType()->isSingleValueType())
3083 // We have an aggregate being stored, split it apart.
3084 DEBUG(dbgs() << " original: " << SI << "\n");
3085 StoreOpSplitter Splitter(&SI, *U);
3086 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3087 SI.eraseFromParent();
3091 bool visitBitCastInst(BitCastInst &BC) {
3096 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3101 bool visitPHINode(PHINode &PN) {
3106 bool visitSelectInst(SelectInst &SI) {
3113 /// \brief Strip aggregate type wrapping.
3115 /// This removes no-op aggregate types wrapping an underlying type. It will
3116 /// strip as many layers of types as it can without changing either the type
3117 /// size or the allocated size.
3118 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
3119 if (Ty->isSingleValueType())
3122 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3123 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
3126 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3127 InnerTy = ArrTy->getElementType();
3128 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3129 const StructLayout *SL = DL.getStructLayout(STy);
3130 unsigned Index = SL->getElementContainingOffset(0);
3131 InnerTy = STy->getElementType(Index);
3136 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3137 TypeSize > DL.getTypeSizeInBits(InnerTy))
3140 return stripAggregateTypeWrapping(DL, InnerTy);
3143 /// \brief Try to find a partition of the aggregate type passed in for a given
3144 /// offset and size.
3146 /// This recurses through the aggregate type and tries to compute a subtype
3147 /// based on the offset and size. When the offset and size span a sub-section
3148 /// of an array, it will even compute a new array type for that sub-section,
3149 /// and the same for structs.
3151 /// Note that this routine is very strict and tries to find a partition of the
3152 /// type which produces the *exact* right offset and size. It is not forgiving
3153 /// when the size or offset cause either end of type-based partition to be off.
3154 /// Also, this is a best-effort routine. It is reasonable to give up and not
3155 /// return a type if necessary.
3156 static Type *getTypePartition(const DataLayout &TD, Type *Ty,
3157 uint64_t Offset, uint64_t Size) {
3158 if (Offset == 0 && TD.getTypeAllocSize(Ty) == Size)
3159 return stripAggregateTypeWrapping(TD, Ty);
3161 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3162 // We can't partition pointers...
3163 if (SeqTy->isPointerTy())
3166 Type *ElementTy = SeqTy->getElementType();
3167 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
3168 uint64_t NumSkippedElements = Offset / ElementSize;
3169 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy))
3170 if (NumSkippedElements >= ArrTy->getNumElements())
3172 if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy))
3173 if (NumSkippedElements >= VecTy->getNumElements())
3175 Offset -= NumSkippedElements * ElementSize;
3177 // First check if we need to recurse.
3178 if (Offset > 0 || Size < ElementSize) {
3179 // Bail if the partition ends in a different array element.
3180 if ((Offset + Size) > ElementSize)
3182 // Recurse through the element type trying to peel off offset bytes.
3183 return getTypePartition(TD, ElementTy, Offset, Size);
3185 assert(Offset == 0);
3187 if (Size == ElementSize)
3188 return stripAggregateTypeWrapping(TD, ElementTy);
3189 assert(Size > ElementSize);
3190 uint64_t NumElements = Size / ElementSize;
3191 if (NumElements * ElementSize != Size)
3193 return ArrayType::get(ElementTy, NumElements);
3196 StructType *STy = dyn_cast<StructType>(Ty);
3200 const StructLayout *SL = TD.getStructLayout(STy);
3201 if (Offset >= SL->getSizeInBytes())
3203 uint64_t EndOffset = Offset + Size;
3204 if (EndOffset > SL->getSizeInBytes())
3207 unsigned Index = SL->getElementContainingOffset(Offset);
3208 Offset -= SL->getElementOffset(Index);
3210 Type *ElementTy = STy->getElementType(Index);
3211 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
3212 if (Offset >= ElementSize)
3213 return 0; // The offset points into alignment padding.
3215 // See if any partition must be contained by the element.
3216 if (Offset > 0 || Size < ElementSize) {
3217 if ((Offset + Size) > ElementSize)
3219 return getTypePartition(TD, ElementTy, Offset, Size);
3221 assert(Offset == 0);
3223 if (Size == ElementSize)
3224 return stripAggregateTypeWrapping(TD, ElementTy);
3226 StructType::element_iterator EI = STy->element_begin() + Index,
3227 EE = STy->element_end();
3228 if (EndOffset < SL->getSizeInBytes()) {
3229 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3230 if (Index == EndIndex)
3231 return 0; // Within a single element and its padding.
3233 // Don't try to form "natural" types if the elements don't line up with the
3235 // FIXME: We could potentially recurse down through the last element in the
3236 // sub-struct to find a natural end point.
3237 if (SL->getElementOffset(EndIndex) != EndOffset)
3240 assert(Index < EndIndex);
3241 EE = STy->element_begin() + EndIndex;
3244 // Try to build up a sub-structure.
3245 SmallVector<Type *, 4> ElementTys;
3247 ElementTys.push_back(*EI++);
3249 StructType *SubTy = StructType::get(STy->getContext(), ElementTys,
3251 const StructLayout *SubSL = TD.getStructLayout(SubTy);
3252 if (Size != SubSL->getSizeInBytes())
3253 return 0; // The sub-struct doesn't have quite the size needed.
3258 /// \brief Rewrite an alloca partition's users.
3260 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3261 /// to rewrite uses of an alloca partition to be conducive for SSA value
3262 /// promotion. If the partition needs a new, more refined alloca, this will
3263 /// build that new alloca, preserving as much type information as possible, and
3264 /// rewrite the uses of the old alloca to point at the new one and have the
3265 /// appropriate new offsets. It also evaluates how successful the rewrite was
3266 /// at enabling promotion and if it was successful queues the alloca to be
3268 bool SROA::rewriteAllocaPartition(AllocaInst &AI,
3269 AllocaPartitioning &P,
3270 AllocaPartitioning::iterator PI) {
3271 uint64_t AllocaSize = PI->EndOffset - PI->BeginOffset;
3272 bool IsLive = false;
3273 for (AllocaPartitioning::use_iterator UI = P.use_begin(PI),
3275 UI != UE && !IsLive; ++UI)
3279 return false; // No live uses left of this partition.
3281 DEBUG(dbgs() << "Speculating PHIs and selects in partition "
3282 << "[" << PI->BeginOffset << "," << PI->EndOffset << ")\n");
3284 PHIOrSelectSpeculator Speculator(*TD, P, *this);
3285 DEBUG(dbgs() << " speculating ");
3286 DEBUG(P.print(dbgs(), PI, ""));
3287 Speculator.visitUsers(PI);
3289 // Try to compute a friendly type for this partition of the alloca. This
3290 // won't always succeed, in which case we fall back to a legal integer type
3291 // or an i8 array of an appropriate size.
3293 if (Type *PartitionTy = P.getCommonType(PI))
3294 if (TD->getTypeAllocSize(PartitionTy) >= AllocaSize)
3295 AllocaTy = PartitionTy;
3297 if (Type *PartitionTy = getTypePartition(*TD, AI.getAllocatedType(),
3298 PI->BeginOffset, AllocaSize))
3299 AllocaTy = PartitionTy;
3301 (AllocaTy->isArrayTy() &&
3302 AllocaTy->getArrayElementType()->isIntegerTy())) &&
3303 TD->isLegalInteger(AllocaSize * 8))
3304 AllocaTy = Type::getIntNTy(*C, AllocaSize * 8);
3306 AllocaTy = ArrayType::get(Type::getInt8Ty(*C), AllocaSize);
3307 assert(TD->getTypeAllocSize(AllocaTy) >= AllocaSize);
3309 // Check for the case where we're going to rewrite to a new alloca of the
3310 // exact same type as the original, and with the same access offsets. In that
3311 // case, re-use the existing alloca, but still run through the rewriter to
3312 // performe phi and select speculation.
3314 if (AllocaTy == AI.getAllocatedType()) {
3315 assert(PI->BeginOffset == 0 &&
3316 "Non-zero begin offset but same alloca type");
3317 assert(PI == P.begin() && "Begin offset is zero on later partition");
3320 unsigned Alignment = AI.getAlignment();
3322 // The minimum alignment which users can rely on when the explicit
3323 // alignment is omitted or zero is that required by the ABI for this
3325 Alignment = TD->getABITypeAlignment(AI.getAllocatedType());
3327 Alignment = MinAlign(Alignment, PI->BeginOffset);
3328 // If we will get at least this much alignment from the type alone, leave
3329 // the alloca's alignment unconstrained.
3330 if (Alignment <= TD->getABITypeAlignment(AllocaTy))
3332 NewAI = new AllocaInst(AllocaTy, 0, Alignment,
3333 AI.getName() + ".sroa." + Twine(PI - P.begin()),
3338 DEBUG(dbgs() << "Rewriting alloca partition "
3339 << "[" << PI->BeginOffset << "," << PI->EndOffset << ") to: "
3342 // Track the high watermark of the post-promotion worklist. We will reset it
3343 // to this point if the alloca is not in fact scheduled for promotion.
3344 unsigned PPWOldSize = PostPromotionWorklist.size();
3346 AllocaPartitionRewriter Rewriter(*TD, P, PI, *this, AI, *NewAI,
3347 PI->BeginOffset, PI->EndOffset);
3348 DEBUG(dbgs() << " rewriting ");
3349 DEBUG(P.print(dbgs(), PI, ""));
3350 bool Promotable = Rewriter.visitUsers(P.use_begin(PI), P.use_end(PI));
3352 DEBUG(dbgs() << " and queuing for promotion\n");
3353 PromotableAllocas.push_back(NewAI);
3354 } else if (NewAI != &AI) {
3355 // If we can't promote the alloca, iterate on it to check for new
3356 // refinements exposed by splitting the current alloca. Don't iterate on an
3357 // alloca which didn't actually change and didn't get promoted.
3358 Worklist.insert(NewAI);
3361 // Drop any post-promotion work items if promotion didn't happen.
3363 while (PostPromotionWorklist.size() > PPWOldSize)
3364 PostPromotionWorklist.pop_back();
3369 /// \brief Walks the partitioning of an alloca rewriting uses of each partition.
3370 bool SROA::splitAlloca(AllocaInst &AI, AllocaPartitioning &P) {
3371 bool Changed = false;
3372 for (AllocaPartitioning::iterator PI = P.begin(), PE = P.end(); PI != PE;
3374 Changed |= rewriteAllocaPartition(AI, P, PI);
3379 /// \brief Analyze an alloca for SROA.
3381 /// This analyzes the alloca to ensure we can reason about it, builds
3382 /// a partitioning of the alloca, and then hands it off to be split and
3383 /// rewritten as needed.
3384 bool SROA::runOnAlloca(AllocaInst &AI) {
3385 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3386 ++NumAllocasAnalyzed;
3388 // Special case dead allocas, as they're trivial.
3389 if (AI.use_empty()) {
3390 AI.eraseFromParent();
3394 // Skip alloca forms that this analysis can't handle.
3395 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3396 TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
3399 bool Changed = false;
3401 // First, split any FCA loads and stores touching this alloca to promote
3402 // better splitting and promotion opportunities.
3403 AggLoadStoreRewriter AggRewriter(*TD);
3404 Changed |= AggRewriter.rewrite(AI);
3406 // Build the partition set using a recursive instruction-visiting builder.
3407 AllocaPartitioning P(*TD, AI);
3408 DEBUG(P.print(dbgs()));
3412 // Delete all the dead users of this alloca before splitting and rewriting it.
3413 for (AllocaPartitioning::dead_user_iterator DI = P.dead_user_begin(),
3414 DE = P.dead_user_end();
3417 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3418 DeadInsts.push_back(*DI);
3420 for (AllocaPartitioning::dead_op_iterator DO = P.dead_op_begin(),
3421 DE = P.dead_op_end();
3424 // Clobber the use with an undef value.
3425 **DO = UndefValue::get(OldV->getType());
3426 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3427 if (isInstructionTriviallyDead(OldI)) {
3429 DeadInsts.push_back(OldI);
3433 // No partitions to split. Leave the dead alloca for a later pass to clean up.
3434 if (P.begin() == P.end())
3437 return splitAlloca(AI, P) || Changed;
3440 /// \brief Delete the dead instructions accumulated in this run.
3442 /// Recursively deletes the dead instructions we've accumulated. This is done
3443 /// at the very end to maximize locality of the recursive delete and to
3444 /// minimize the problems of invalidated instruction pointers as such pointers
3445 /// are used heavily in the intermediate stages of the algorithm.
3447 /// We also record the alloca instructions deleted here so that they aren't
3448 /// subsequently handed to mem2reg to promote.
3449 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3450 DeadSplitInsts.clear();
3451 while (!DeadInsts.empty()) {
3452 Instruction *I = DeadInsts.pop_back_val();
3453 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3455 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
3456 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
3457 // Zero out the operand and see if it becomes trivially dead.
3459 if (isInstructionTriviallyDead(U))
3460 DeadInsts.push_back(U);
3463 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3464 DeletedAllocas.insert(AI);
3467 I->eraseFromParent();
3471 /// \brief Promote the allocas, using the best available technique.
3473 /// This attempts to promote whatever allocas have been identified as viable in
3474 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3475 /// If there is a domtree available, we attempt to promote using the full power
3476 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3477 /// based on the SSAUpdater utilities. This function returns whether any
3478 /// promotion occured.
3479 bool SROA::promoteAllocas(Function &F) {
3480 if (PromotableAllocas.empty())
3483 NumPromoted += PromotableAllocas.size();
3485 if (DT && !ForceSSAUpdater) {
3486 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3487 PromoteMemToReg(PromotableAllocas, *DT);
3488 PromotableAllocas.clear();
3492 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3494 DIBuilder DIB(*F.getParent());
3495 SmallVector<Instruction*, 64> Insts;
3497 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3498 AllocaInst *AI = PromotableAllocas[Idx];
3499 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
3501 Instruction *I = cast<Instruction>(*UI++);
3502 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3503 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3504 // leading to them) here. Eventually it should use them to optimize the
3505 // scalar values produced.
3506 if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
3507 assert(onlyUsedByLifetimeMarkers(I) &&
3508 "Found a bitcast used outside of a lifetime marker.");
3509 while (!I->use_empty())
3510 cast<Instruction>(*I->use_begin())->eraseFromParent();
3511 I->eraseFromParent();
3514 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3515 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3516 II->getIntrinsicID() == Intrinsic::lifetime_end);
3517 II->eraseFromParent();
3523 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3527 PromotableAllocas.clear();
3532 /// \brief A predicate to test whether an alloca belongs to a set.
3533 class IsAllocaInSet {
3534 typedef SmallPtrSet<AllocaInst *, 4> SetType;
3538 typedef AllocaInst *argument_type;
3540 IsAllocaInSet(const SetType &Set) : Set(Set) {}
3541 bool operator()(AllocaInst *AI) const { return Set.count(AI); }
3545 bool SROA::runOnFunction(Function &F) {
3546 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3547 C = &F.getContext();
3548 TD = getAnalysisIfAvailable<DataLayout>();
3550 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3553 DT = getAnalysisIfAvailable<DominatorTree>();
3555 BasicBlock &EntryBB = F.getEntryBlock();
3556 for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end());
3558 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3559 Worklist.insert(AI);
3561 bool Changed = false;
3562 // A set of deleted alloca instruction pointers which should be removed from
3563 // the list of promotable allocas.
3564 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3567 while (!Worklist.empty()) {
3568 Changed |= runOnAlloca(*Worklist.pop_back_val());
3569 deleteDeadInstructions(DeletedAllocas);
3571 // Remove the deleted allocas from various lists so that we don't try to
3572 // continue processing them.
3573 if (!DeletedAllocas.empty()) {
3574 Worklist.remove_if(IsAllocaInSet(DeletedAllocas));
3575 PostPromotionWorklist.remove_if(IsAllocaInSet(DeletedAllocas));
3576 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3577 PromotableAllocas.end(),
3578 IsAllocaInSet(DeletedAllocas)),
3579 PromotableAllocas.end());
3580 DeletedAllocas.clear();
3584 Changed |= promoteAllocas(F);
3586 Worklist = PostPromotionWorklist;
3587 PostPromotionWorklist.clear();
3588 } while (!Worklist.empty());
3593 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3594 if (RequiresDomTree)
3595 AU.addRequired<DominatorTree>();
3596 AU.setPreservesCFG();