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/GlobalVariable.h"
34 #include "llvm/IRBuilder.h"
35 #include "llvm/Instructions.h"
36 #include "llvm/IntrinsicInst.h"
37 #include "llvm/LLVMContext.h"
38 #include "llvm/Module.h"
39 #include "llvm/Operator.h"
40 #include "llvm/Pass.h"
41 #include "llvm/ADT/SetVector.h"
42 #include "llvm/ADT/SmallVector.h"
43 #include "llvm/ADT/Statistic.h"
44 #include "llvm/ADT/STLExtras.h"
45 #include "llvm/ADT/TinyPtrVector.h"
46 #include "llvm/Analysis/Dominators.h"
47 #include "llvm/Analysis/Loads.h"
48 #include "llvm/Analysis/ValueTracking.h"
49 #include "llvm/Support/CallSite.h"
50 #include "llvm/Support/CommandLine.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/ErrorHandling.h"
53 #include "llvm/Support/GetElementPtrTypeIterator.h"
54 #include "llvm/Support/InstVisitor.h"
55 #include "llvm/Support/MathExtras.h"
56 #include "llvm/Support/ValueHandle.h"
57 #include "llvm/Support/raw_ostream.h"
58 #include "llvm/Target/TargetData.h"
59 #include "llvm/Transforms/Utils/Local.h"
60 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
61 #include "llvm/Transforms/Utils/SSAUpdater.h"
64 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
65 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
66 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
67 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
68 STATISTIC(NumDeleted, "Number of instructions deleted");
69 STATISTIC(NumVectorized, "Number of vectorized aggregates");
71 /// Hidden option to force the pass to not use DomTree and mem2reg, instead
72 /// forming SSA values through the SSAUpdater infrastructure.
74 ForceSSAUpdater("force-ssa-updater", cl::init(false), cl::Hidden);
77 /// \brief Alloca partitioning representation.
79 /// This class represents a partitioning of an alloca into slices, and
80 /// information about the nature of uses of each slice of the alloca. The goal
81 /// is that this information is sufficient to decide if and how to split the
82 /// alloca apart and replace slices with scalars. It is also intended that this
83 /// structure can capture the relevant information needed both to decide about
84 /// and to enact these transformations.
85 class AllocaPartitioning {
87 /// \brief A common base class for representing a half-open byte range.
89 /// \brief The beginning offset of the range.
92 /// \brief The ending offset, not included in the range.
95 ByteRange() : BeginOffset(), EndOffset() {}
96 ByteRange(uint64_t BeginOffset, uint64_t EndOffset)
97 : BeginOffset(BeginOffset), EndOffset(EndOffset) {}
99 /// \brief Support for ordering ranges.
101 /// This provides an ordering over ranges such that start offsets are
102 /// always increasing, and within equal start offsets, the end offsets are
103 /// decreasing. Thus the spanning range comes first in a cluster with the
104 /// same start position.
105 bool operator<(const ByteRange &RHS) const {
106 if (BeginOffset < RHS.BeginOffset) return true;
107 if (BeginOffset > RHS.BeginOffset) return false;
108 if (EndOffset > RHS.EndOffset) return true;
112 /// \brief Support comparison with a single offset to allow binary searches.
113 friend bool operator<(const ByteRange &LHS, uint64_t RHSOffset) {
114 return LHS.BeginOffset < RHSOffset;
117 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
118 const ByteRange &RHS) {
119 return LHSOffset < RHS.BeginOffset;
122 bool operator==(const ByteRange &RHS) const {
123 return BeginOffset == RHS.BeginOffset && EndOffset == RHS.EndOffset;
125 bool operator!=(const ByteRange &RHS) const { return !operator==(RHS); }
128 /// \brief A partition of an alloca.
130 /// This structure represents a contiguous partition of the alloca. These are
131 /// formed by examining the uses of the alloca. During formation, they may
132 /// overlap but once an AllocaPartitioning is built, the Partitions within it
133 /// are all disjoint.
134 struct Partition : public ByteRange {
135 /// \brief Whether this partition is splittable into smaller partitions.
137 /// We flag partitions as splittable when they are formed entirely due to
138 /// accesses by trivially splittable operations such as memset and memcpy.
140 /// FIXME: At some point we should consider loads and stores of FCAs to be
141 /// splittable and eagerly split them into scalar values.
144 Partition() : ByteRange(), IsSplittable() {}
145 Partition(uint64_t BeginOffset, uint64_t EndOffset, bool IsSplittable)
146 : ByteRange(BeginOffset, EndOffset), IsSplittable(IsSplittable) {}
149 /// \brief A particular use of a partition of the alloca.
151 /// This structure is used to associate uses of a partition with it. They
152 /// mark the range of bytes which are referenced by a particular instruction,
153 /// and includes a handle to the user itself and the pointer value in use.
154 /// The bounds of these uses are determined by intersecting the bounds of the
155 /// memory use itself with a particular partition. As a consequence there is
156 /// intentionally overlap between various uses of the same partition.
157 struct PartitionUse : public ByteRange {
158 /// \brief The use in question. Provides access to both user and used value.
161 PartitionUse() : ByteRange(), U() {}
162 PartitionUse(uint64_t BeginOffset, uint64_t EndOffset, Use *U)
163 : ByteRange(BeginOffset, EndOffset), U(U) {}
166 /// \brief Construct a partitioning of a particular alloca.
168 /// Construction does most of the work for partitioning the alloca. This
169 /// performs the necessary walks of users and builds a partitioning from it.
170 AllocaPartitioning(const TargetData &TD, AllocaInst &AI);
172 /// \brief Test whether a pointer to the allocation escapes our analysis.
174 /// If this is true, the partitioning is never fully built and should be
176 bool isEscaped() const { return PointerEscapingInstr; }
178 /// \brief Support for iterating over the partitions.
180 typedef SmallVectorImpl<Partition>::iterator iterator;
181 iterator begin() { return Partitions.begin(); }
182 iterator end() { return Partitions.end(); }
184 typedef SmallVectorImpl<Partition>::const_iterator const_iterator;
185 const_iterator begin() const { return Partitions.begin(); }
186 const_iterator end() const { return Partitions.end(); }
189 /// \brief Support for iterating over and manipulating a particular
190 /// partition's uses.
192 /// The iteration support provided for uses is more limited, but also
193 /// includes some manipulation routines to support rewriting the uses of
194 /// partitions during SROA.
196 typedef SmallVectorImpl<PartitionUse>::iterator use_iterator;
197 use_iterator use_begin(unsigned Idx) { return Uses[Idx].begin(); }
198 use_iterator use_begin(const_iterator I) { return Uses[I - begin()].begin(); }
199 use_iterator use_end(unsigned Idx) { return Uses[Idx].end(); }
200 use_iterator use_end(const_iterator I) { return Uses[I - begin()].end(); }
201 void use_push_back(unsigned Idx, const PartitionUse &PU) {
202 Uses[Idx].push_back(PU);
204 void use_push_back(const_iterator I, const PartitionUse &PU) {
205 Uses[I - begin()].push_back(PU);
207 void use_erase(unsigned Idx, use_iterator UI) { Uses[Idx].erase(UI); }
208 void use_erase(const_iterator I, use_iterator UI) {
209 Uses[I - begin()].erase(UI);
212 typedef SmallVectorImpl<PartitionUse>::const_iterator const_use_iterator;
213 const_use_iterator use_begin(unsigned Idx) const { return Uses[Idx].begin(); }
214 const_use_iterator use_begin(const_iterator I) const {
215 return Uses[I - begin()].begin();
217 const_use_iterator use_end(unsigned Idx) const { return Uses[Idx].end(); }
218 const_use_iterator use_end(const_iterator I) const {
219 return Uses[I - begin()].end();
223 /// \brief Allow iterating the dead users for this alloca.
225 /// These are instructions which will never actually use the alloca as they
226 /// are outside the allocated range. They are safe to replace with undef and
229 typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator;
230 dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); }
231 dead_user_iterator dead_user_end() const { return DeadUsers.end(); }
234 /// \brief Allow iterating the dead expressions referring to this alloca.
236 /// These are operands which have cannot actually be used to refer to the
237 /// alloca as they are outside its range and the user doesn't correct for
238 /// that. These mostly consist of PHI node inputs and the like which we just
239 /// need to replace with undef.
241 typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator;
242 dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); }
243 dead_op_iterator dead_op_end() const { return DeadOperands.end(); }
246 /// \brief MemTransferInst auxiliary data.
247 /// This struct provides some auxiliary data about memory transfer
248 /// intrinsics such as memcpy and memmove. These intrinsics can use two
249 /// different ranges within the same alloca, and provide other challenges to
250 /// correctly represent. We stash extra data to help us untangle this
251 /// after the partitioning is complete.
252 struct MemTransferOffsets {
253 uint64_t DestBegin, DestEnd;
254 uint64_t SourceBegin, SourceEnd;
257 MemTransferOffsets getMemTransferOffsets(MemTransferInst &II) const {
258 return MemTransferInstData.lookup(&II);
261 /// \brief Map from a PHI or select operand back to a partition.
263 /// When manipulating PHI nodes or selects, they can use more than one
264 /// partition of an alloca. We store a special mapping to allow finding the
265 /// partition referenced by each of these operands, if any.
266 iterator findPartitionForPHIOrSelectOperand(Use *U) {
267 SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt
268 = PHIOrSelectOpMap.find(U);
269 if (MapIt == PHIOrSelectOpMap.end())
272 return begin() + MapIt->second.first;
275 /// \brief Map from a PHI or select operand back to the specific use of
278 /// Similar to mapping these operands back to the partitions, this maps
279 /// directly to the use structure of that partition.
280 use_iterator findPartitionUseForPHIOrSelectOperand(Use *U) {
281 SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt
282 = PHIOrSelectOpMap.find(U);
283 assert(MapIt != PHIOrSelectOpMap.end());
284 return Uses[MapIt->second.first].begin() + MapIt->second.second;
287 /// \brief Compute a common type among the uses of a particular partition.
289 /// This routines walks all of the uses of a particular partition and tries
290 /// to find a common type between them. Untyped operations such as memset and
291 /// memcpy are ignored.
292 Type *getCommonType(iterator I) const;
294 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
295 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
296 void printUsers(raw_ostream &OS, const_iterator I,
297 StringRef Indent = " ") const;
298 void print(raw_ostream &OS) const;
299 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump(const_iterator I) const;
300 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump() const;
304 template <typename DerivedT, typename RetT = void> class BuilderBase;
305 class PartitionBuilder;
306 friend class AllocaPartitioning::PartitionBuilder;
308 friend class AllocaPartitioning::UseBuilder;
311 /// \brief Handle to alloca instruction to simplify method interfaces.
315 /// \brief The instruction responsible for this alloca having no partitioning.
317 /// When an instruction (potentially) escapes the pointer to the alloca, we
318 /// store a pointer to that here and abort trying to partition the alloca.
319 /// This will be null if the alloca is partitioned successfully.
320 Instruction *PointerEscapingInstr;
322 /// \brief The partitions of the alloca.
324 /// We store a vector of the partitions over the alloca here. This vector is
325 /// sorted by increasing begin offset, and then by decreasing end offset. See
326 /// the Partition inner class for more details. Initially (during
327 /// construction) there are overlaps, but we form a disjoint sequence of
328 /// partitions while finishing construction and a fully constructed object is
329 /// expected to always have this as a disjoint space.
330 SmallVector<Partition, 8> Partitions;
332 /// \brief The uses of the partitions.
334 /// This is essentially a mapping from each partition to a list of uses of
335 /// that partition. The mapping is done with a Uses vector that has the exact
336 /// same number of entries as the partition vector. Each entry is itself
337 /// a vector of the uses.
338 SmallVector<SmallVector<PartitionUse, 2>, 8> Uses;
340 /// \brief Instructions which will become dead if we rewrite the alloca.
342 /// Note that these are not separated by partition. This is because we expect
343 /// a partitioned alloca to be completely rewritten or not rewritten at all.
344 /// If rewritten, all these instructions can simply be removed and replaced
345 /// with undef as they come from outside of the allocated space.
346 SmallVector<Instruction *, 8> DeadUsers;
348 /// \brief Operands which will become dead if we rewrite the alloca.
350 /// These are operands that in their particular use can be replaced with
351 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
352 /// to PHI nodes and the like. They aren't entirely dead (there might be
353 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
354 /// want to swap this particular input for undef to simplify the use lists of
356 SmallVector<Use *, 8> DeadOperands;
358 /// \brief The underlying storage for auxiliary memcpy and memset info.
359 SmallDenseMap<MemTransferInst *, MemTransferOffsets, 4> MemTransferInstData;
361 /// \brief A side datastructure used when building up the partitions and uses.
363 /// This mapping is only really used during the initial building of the
364 /// partitioning so that we can retain information about PHI and select nodes
366 SmallDenseMap<Instruction *, std::pair<uint64_t, bool> > PHIOrSelectSizes;
368 /// \brief Auxiliary information for particular PHI or select operands.
369 SmallDenseMap<Use *, std::pair<unsigned, unsigned>, 4> PHIOrSelectOpMap;
371 /// \brief A utility routine called from the constructor.
373 /// This does what it says on the tin. It is the key of the alloca partition
374 /// splitting and merging. After it is called we have the desired disjoint
375 /// collection of partitions.
376 void splitAndMergePartitions();
380 template <typename DerivedT, typename RetT>
381 class AllocaPartitioning::BuilderBase
382 : public InstVisitor<DerivedT, RetT> {
384 BuilderBase(const TargetData &TD, AllocaInst &AI, AllocaPartitioning &P)
386 AllocSize(TD.getTypeAllocSize(AI.getAllocatedType())),
392 const TargetData &TD;
393 const uint64_t AllocSize;
394 AllocaPartitioning &P;
396 SmallPtrSet<Use *, 8> VisitedUses;
402 SmallVector<OffsetUse, 8> Queue;
404 // The active offset and use while visiting.
408 void enqueueUsers(Instruction &I, int64_t UserOffset) {
409 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
411 if (VisitedUses.insert(&UI.getUse())) {
412 OffsetUse OU = { &UI.getUse(), UserOffset };
418 bool computeConstantGEPOffset(GetElementPtrInst &GEPI, int64_t &GEPOffset) {
420 for (gep_type_iterator GTI = gep_type_begin(GEPI), GTE = gep_type_end(GEPI);
422 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
428 // Handle a struct index, which adds its field offset to the pointer.
429 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
430 unsigned ElementIdx = OpC->getZExtValue();
431 const StructLayout *SL = TD.getStructLayout(STy);
432 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
433 // Check that we can continue to model this GEP in a signed 64-bit offset.
434 if (ElementOffset > INT64_MAX ||
436 ((uint64_t)GEPOffset + ElementOffset) > INT64_MAX)) {
437 DEBUG(dbgs() << "WARNING: Encountered a cumulative offset exceeding "
438 << "what can be represented in an int64_t!\n"
439 << " alloca: " << P.AI << "\n");
443 GEPOffset = ElementOffset + (uint64_t)-GEPOffset;
445 GEPOffset += ElementOffset;
449 APInt Index = OpC->getValue().sextOrTrunc(TD.getPointerSizeInBits());
450 Index *= APInt(Index.getBitWidth(),
451 TD.getTypeAllocSize(GTI.getIndexedType()));
452 Index += APInt(Index.getBitWidth(), (uint64_t)GEPOffset,
454 // Check if the result can be stored in our int64_t offset.
455 if (!Index.isSignedIntN(sizeof(GEPOffset) * 8)) {
456 DEBUG(dbgs() << "WARNING: Encountered a cumulative offset exceeding "
457 << "what can be represented in an int64_t!\n"
458 << " alloca: " << P.AI << "\n");
462 GEPOffset = Index.getSExtValue();
467 Value *foldSelectInst(SelectInst &SI) {
468 // If the condition being selected on is a constant or the same value is
469 // being selected between, fold the select. Yes this does (rarely) happen
471 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
472 return SI.getOperand(1+CI->isZero());
473 if (SI.getOperand(1) == SI.getOperand(2)) {
474 assert(*U == SI.getOperand(1));
475 return SI.getOperand(1);
481 /// \brief Builder for the alloca partitioning.
483 /// This class builds an alloca partitioning by recursively visiting the uses
484 /// of an alloca and splitting the partitions for each load and store at each
486 class AllocaPartitioning::PartitionBuilder
487 : public BuilderBase<PartitionBuilder, bool> {
488 friend class InstVisitor<PartitionBuilder, bool>;
490 SmallDenseMap<Instruction *, unsigned> MemTransferPartitionMap;
493 PartitionBuilder(const TargetData &TD, AllocaInst &AI, AllocaPartitioning &P)
494 : BuilderBase<PartitionBuilder, bool>(TD, AI, P) {}
496 /// \brief Run the builder over the allocation.
498 // Note that we have to re-evaluate size on each trip through the loop as
499 // the queue grows at the tail.
500 for (unsigned Idx = 0; Idx < Queue.size(); ++Idx) {
502 Offset = Queue[Idx].Offset;
503 if (!visit(cast<Instruction>(U->getUser())))
510 bool markAsEscaping(Instruction &I) {
511 P.PointerEscapingInstr = &I;
515 void insertUse(Instruction &I, int64_t Offset, uint64_t Size,
516 bool IsSplittable = false) {
517 // Completely skip uses which have a zero size or don't overlap the
520 (Offset >= 0 && (uint64_t)Offset >= AllocSize) ||
521 (Offset < 0 && (uint64_t)-Offset >= Size)) {
522 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
523 << " which starts past the end of the " << AllocSize
525 << " alloca: " << P.AI << "\n"
526 << " use: " << I << "\n");
530 // Clamp the start to the beginning of the allocation.
532 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
533 << " to start at the beginning of the alloca:\n"
534 << " alloca: " << P.AI << "\n"
535 << " use: " << I << "\n");
536 Size -= (uint64_t)-Offset;
540 uint64_t BeginOffset = Offset, EndOffset = BeginOffset + Size;
542 // Clamp the end offset to the end of the allocation. Note that this is
543 // formulated to handle even the case where "BeginOffset + Size" overflows.
544 assert(AllocSize >= BeginOffset); // Established above.
545 if (Size > AllocSize - BeginOffset) {
546 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
547 << " to remain within the " << AllocSize << " byte alloca:\n"
548 << " alloca: " << P.AI << "\n"
549 << " use: " << I << "\n");
550 EndOffset = AllocSize;
553 // See if we can just add a user onto the last slot currently occupied.
554 if (!P.Partitions.empty() &&
555 P.Partitions.back().BeginOffset == BeginOffset &&
556 P.Partitions.back().EndOffset == EndOffset) {
557 P.Partitions.back().IsSplittable &= IsSplittable;
561 Partition New(BeginOffset, EndOffset, IsSplittable);
562 P.Partitions.push_back(New);
565 bool handleLoadOrStore(Type *Ty, Instruction &I, int64_t Offset) {
566 uint64_t Size = TD.getTypeStoreSize(Ty);
568 // If this memory access can be shown to *statically* extend outside the
569 // bounds of of the allocation, it's behavior is undefined, so simply
570 // ignore it. Note that this is more strict than the generic clamping
571 // behavior of insertUse. We also try to handle cases which might run the
573 // FIXME: We should instead consider the pointer to have escaped if this
574 // function is being instrumented for addressing bugs or race conditions.
575 if (Offset < 0 || (uint64_t)Offset >= AllocSize ||
576 Size > (AllocSize - (uint64_t)Offset)) {
577 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte "
578 << (isa<LoadInst>(I) ? "load" : "store") << " @" << Offset
579 << " which extends past the end of the " << AllocSize
581 << " alloca: " << P.AI << "\n"
582 << " use: " << I << "\n");
586 insertUse(I, Offset, Size);
590 bool visitBitCastInst(BitCastInst &BC) {
591 enqueueUsers(BC, Offset);
595 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
597 if (!computeConstantGEPOffset(GEPI, GEPOffset))
598 return markAsEscaping(GEPI);
600 enqueueUsers(GEPI, GEPOffset);
604 bool visitLoadInst(LoadInst &LI) {
605 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
606 "All simple FCA loads should have been pre-split");
607 return handleLoadOrStore(LI.getType(), LI, Offset);
610 bool visitStoreInst(StoreInst &SI) {
611 Value *ValOp = SI.getValueOperand();
613 return markAsEscaping(SI);
615 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
616 "All simple FCA stores should have been pre-split");
617 return handleLoadOrStore(ValOp->getType(), SI, Offset);
621 bool visitMemSetInst(MemSetInst &II) {
622 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
623 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
624 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
625 insertUse(II, Offset, Size, Length);
629 bool visitMemTransferInst(MemTransferInst &II) {
630 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
631 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
633 // Zero-length mem transfer intrinsics can be ignored entirely.
636 MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
638 // Only intrinsics with a constant length can be split.
639 Offsets.IsSplittable = Length;
641 if (*U != II.getRawDest()) {
642 assert(*U == II.getRawSource());
643 Offsets.SourceBegin = Offset;
644 Offsets.SourceEnd = Offset + Size;
646 Offsets.DestBegin = Offset;
647 Offsets.DestEnd = Offset + Size;
650 insertUse(II, Offset, Size, Offsets.IsSplittable);
651 unsigned NewIdx = P.Partitions.size() - 1;
653 SmallDenseMap<Instruction *, unsigned>::const_iterator PMI;
654 bool Inserted = false;
655 llvm::tie(PMI, Inserted)
656 = MemTransferPartitionMap.insert(std::make_pair(&II, NewIdx));
657 if (Offsets.IsSplittable &&
658 (!Inserted || II.getRawSource() == II.getRawDest())) {
659 // We've found a memory transfer intrinsic which refers to the alloca as
660 // both a source and dest. This is detected either by direct equality of
661 // the operand values, or when we visit the intrinsic twice due to two
662 // different chains of values leading to it. We refuse to split these to
663 // simplify splitting logic. If possible, SROA will still split them into
664 // separate allocas and then re-analyze.
665 Offsets.IsSplittable = false;
666 P.Partitions[PMI->second].IsSplittable = false;
667 P.Partitions[NewIdx].IsSplittable = false;
673 // Disable SRoA for any intrinsics except for lifetime invariants.
674 // FIXME: What about debug instrinsics? This matches old behavior, but
675 // doesn't make sense.
676 bool visitIntrinsicInst(IntrinsicInst &II) {
677 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
678 II.getIntrinsicID() == Intrinsic::lifetime_end) {
679 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
680 uint64_t Size = std::min(AllocSize - Offset, Length->getLimitedValue());
681 insertUse(II, Offset, Size, true);
685 return markAsEscaping(II);
688 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
689 // We consider any PHI or select that results in a direct load or store of
690 // the same offset to be a viable use for partitioning purposes. These uses
691 // are considered unsplittable and the size is the maximum loaded or stored
693 SmallPtrSet<Instruction *, 4> Visited;
694 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
695 Visited.insert(Root);
696 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
697 // If there are no loads or stores, the access is dead. We mark that as
698 // a size zero access.
701 Instruction *I, *UsedI;
702 llvm::tie(UsedI, I) = Uses.pop_back_val();
704 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
705 Size = std::max(Size, TD.getTypeStoreSize(LI->getType()));
708 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
709 Value *Op = SI->getOperand(0);
712 Size = std::max(Size, TD.getTypeStoreSize(Op->getType()));
716 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
717 if (!GEP->hasAllZeroIndices())
719 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
720 !isa<SelectInst>(I)) {
724 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
726 if (Visited.insert(cast<Instruction>(*UI)))
727 Uses.push_back(std::make_pair(I, cast<Instruction>(*UI)));
728 } while (!Uses.empty());
733 bool visitPHINode(PHINode &PN) {
734 // See if we already have computed info on this node.
735 std::pair<uint64_t, bool> &PHIInfo = P.PHIOrSelectSizes[&PN];
737 PHIInfo.second = true;
738 insertUse(PN, Offset, PHIInfo.first);
742 // Check for an unsafe use of the PHI node.
743 if (Instruction *EscapingI = hasUnsafePHIOrSelectUse(&PN, PHIInfo.first))
744 return markAsEscaping(*EscapingI);
746 insertUse(PN, Offset, PHIInfo.first);
750 bool visitSelectInst(SelectInst &SI) {
751 if (Value *Result = foldSelectInst(SI)) {
753 // If the result of the constant fold will be the pointer, recurse
754 // through the select as if we had RAUW'ed it.
755 enqueueUsers(SI, Offset);
760 // See if we already have computed info on this node.
761 std::pair<uint64_t, bool> &SelectInfo = P.PHIOrSelectSizes[&SI];
762 if (SelectInfo.first) {
763 SelectInfo.second = true;
764 insertUse(SI, Offset, SelectInfo.first);
768 // Check for an unsafe use of the PHI node.
769 if (Instruction *EscapingI = hasUnsafePHIOrSelectUse(&SI, SelectInfo.first))
770 return markAsEscaping(*EscapingI);
772 insertUse(SI, Offset, SelectInfo.first);
776 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
777 bool visitInstruction(Instruction &I) { return markAsEscaping(I); }
781 /// \brief Use adder for the alloca partitioning.
783 /// This class adds the uses of an alloca to all of the partitions which they
784 /// use. For splittable partitions, this can end up doing essentially a linear
785 /// walk of the partitions, but the number of steps remains bounded by the
786 /// total result instruction size:
787 /// - The number of partitions is a result of the number unsplittable
788 /// instructions using the alloca.
789 /// - The number of users of each partition is at worst the total number of
790 /// splittable instructions using the alloca.
791 /// Thus we will produce N * M instructions in the end, where N are the number
792 /// of unsplittable uses and M are the number of splittable. This visitor does
793 /// the exact same number of updates to the partitioning.
795 /// In the more common case, this visitor will leverage the fact that the
796 /// partition space is pre-sorted, and do a logarithmic search for the
797 /// partition needed, making the total visit a classical ((N + M) * log(N))
798 /// complexity operation.
799 class AllocaPartitioning::UseBuilder : public BuilderBase<UseBuilder> {
800 friend class InstVisitor<UseBuilder>;
802 /// \brief Set to de-duplicate dead instructions found in the use walk.
803 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
806 UseBuilder(const TargetData &TD, AllocaInst &AI, AllocaPartitioning &P)
807 : BuilderBase<UseBuilder>(TD, AI, P) {}
809 /// \brief Run the builder over the allocation.
811 // Note that we have to re-evaluate size on each trip through the loop as
812 // the queue grows at the tail.
813 for (unsigned Idx = 0; Idx < Queue.size(); ++Idx) {
815 Offset = Queue[Idx].Offset;
816 this->visit(cast<Instruction>(U->getUser()));
821 void markAsDead(Instruction &I) {
822 if (VisitedDeadInsts.insert(&I))
823 P.DeadUsers.push_back(&I);
826 void insertUse(Instruction &User, int64_t Offset, uint64_t Size) {
827 // If the use has a zero size or extends outside of the allocation, record
828 // it as a dead use for elimination later.
829 if (Size == 0 || (uint64_t)Offset >= AllocSize ||
830 (Offset < 0 && (uint64_t)-Offset >= Size))
831 return markAsDead(User);
833 // Clamp the start to the beginning of the allocation.
835 Size -= (uint64_t)-Offset;
839 uint64_t BeginOffset = Offset, EndOffset = BeginOffset + Size;
841 // Clamp the end offset to the end of the allocation. Note that this is
842 // formulated to handle even the case where "BeginOffset + Size" overflows.
843 assert(AllocSize >= BeginOffset); // Established above.
844 if (Size > AllocSize - BeginOffset)
845 EndOffset = AllocSize;
847 // NB: This only works if we have zero overlapping partitions.
848 iterator B = std::lower_bound(P.begin(), P.end(), BeginOffset);
849 if (B != P.begin() && llvm::prior(B)->EndOffset > BeginOffset)
851 for (iterator I = B, E = P.end(); I != E && I->BeginOffset < EndOffset;
853 PartitionUse NewPU(std::max(I->BeginOffset, BeginOffset),
854 std::min(I->EndOffset, EndOffset), U);
855 P.use_push_back(I, NewPU);
856 if (isa<PHINode>(U->getUser()) || isa<SelectInst>(U->getUser()))
857 P.PHIOrSelectOpMap[U]
858 = std::make_pair(I - P.begin(), P.Uses[I - P.begin()].size() - 1);
862 void handleLoadOrStore(Type *Ty, Instruction &I, int64_t Offset) {
863 uint64_t Size = TD.getTypeStoreSize(Ty);
865 // If this memory access can be shown to *statically* extend outside the
866 // bounds of of the allocation, it's behavior is undefined, so simply
867 // ignore it. Note that this is more strict than the generic clamping
868 // behavior of insertUse.
869 if (Offset < 0 || (uint64_t)Offset >= AllocSize ||
870 Size > (AllocSize - (uint64_t)Offset))
871 return markAsDead(I);
873 insertUse(I, Offset, Size);
876 void visitBitCastInst(BitCastInst &BC) {
878 return markAsDead(BC);
880 enqueueUsers(BC, Offset);
883 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
884 if (GEPI.use_empty())
885 return markAsDead(GEPI);
888 if (!computeConstantGEPOffset(GEPI, GEPOffset))
889 llvm_unreachable("Unable to compute constant offset for use");
891 enqueueUsers(GEPI, GEPOffset);
894 void visitLoadInst(LoadInst &LI) {
895 handleLoadOrStore(LI.getType(), LI, Offset);
898 void visitStoreInst(StoreInst &SI) {
899 handleLoadOrStore(SI.getOperand(0)->getType(), SI, Offset);
902 void visitMemSetInst(MemSetInst &II) {
903 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
904 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
905 insertUse(II, Offset, Size);
908 void visitMemTransferInst(MemTransferInst &II) {
909 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
910 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
911 insertUse(II, Offset, Size);
914 void visitIntrinsicInst(IntrinsicInst &II) {
915 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
916 II.getIntrinsicID() == Intrinsic::lifetime_end);
918 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
919 insertUse(II, Offset,
920 std::min(AllocSize - Offset, Length->getLimitedValue()));
923 void insertPHIOrSelect(Instruction &User, uint64_t Offset) {
924 uint64_t Size = P.PHIOrSelectSizes.lookup(&User).first;
926 // For PHI and select operands outside the alloca, we can't nuke the entire
927 // phi or select -- the other side might still be relevant, so we special
928 // case them here and use a separate structure to track the operands
929 // themselves which should be replaced with undef.
930 if (Offset >= AllocSize) {
931 P.DeadOperands.push_back(U);
935 insertUse(User, Offset, Size);
937 void visitPHINode(PHINode &PN) {
939 return markAsDead(PN);
941 insertPHIOrSelect(PN, Offset);
943 void visitSelectInst(SelectInst &SI) {
945 return markAsDead(SI);
947 if (Value *Result = foldSelectInst(SI)) {
949 // If the result of the constant fold will be the pointer, recurse
950 // through the select as if we had RAUW'ed it.
951 enqueueUsers(SI, Offset);
953 // Otherwise the operand to the select is dead, and we can replace it
955 P.DeadOperands.push_back(U);
960 insertPHIOrSelect(SI, Offset);
963 /// \brief Unreachable, we've already visited the alloca once.
964 void visitInstruction(Instruction &I) {
965 llvm_unreachable("Unhandled instruction in use builder.");
969 void AllocaPartitioning::splitAndMergePartitions() {
970 size_t NumDeadPartitions = 0;
972 // Track the range of splittable partitions that we pass when accumulating
973 // overlapping unsplittable partitions.
974 uint64_t SplitEndOffset = 0ull;
976 Partition New(0ull, 0ull, false);
978 for (unsigned i = 0, j = i, e = Partitions.size(); i != e; i = j) {
981 if (!Partitions[i].IsSplittable || New.BeginOffset == New.EndOffset) {
982 assert(New.BeginOffset == New.EndOffset);
985 assert(New.IsSplittable);
986 New.EndOffset = std::max(New.EndOffset, Partitions[i].EndOffset);
988 assert(New.BeginOffset != New.EndOffset);
990 // Scan the overlapping partitions.
991 while (j != e && New.EndOffset > Partitions[j].BeginOffset) {
992 // If the new partition we are forming is splittable, stop at the first
993 // unsplittable partition.
994 if (New.IsSplittable && !Partitions[j].IsSplittable)
997 // Grow the new partition to include any equally splittable range. 'j' is
998 // always equally splittable when New is splittable, but when New is not
999 // splittable, we may subsume some (or part of some) splitable partition
1000 // without growing the new one.
1001 if (New.IsSplittable == Partitions[j].IsSplittable) {
1002 New.EndOffset = std::max(New.EndOffset, Partitions[j].EndOffset);
1004 assert(!New.IsSplittable);
1005 assert(Partitions[j].IsSplittable);
1006 SplitEndOffset = std::max(SplitEndOffset, Partitions[j].EndOffset);
1009 Partitions[j].BeginOffset = Partitions[j].EndOffset = UINT64_MAX;
1010 ++NumDeadPartitions;
1014 // If the new partition is splittable, chop off the end as soon as the
1015 // unsplittable subsequent partition starts and ensure we eventually cover
1016 // the splittable area.
1017 if (j != e && New.IsSplittable) {
1018 SplitEndOffset = std::max(SplitEndOffset, New.EndOffset);
1019 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
1022 // Add the new partition if it differs from the original one and is
1023 // non-empty. We can end up with an empty partition here if it was
1024 // splittable but there is an unsplittable one that starts at the same
1026 if (New != Partitions[i]) {
1027 if (New.BeginOffset != New.EndOffset)
1028 Partitions.push_back(New);
1029 // Mark the old one for removal.
1030 Partitions[i].BeginOffset = Partitions[i].EndOffset = UINT64_MAX;
1031 ++NumDeadPartitions;
1034 New.BeginOffset = New.EndOffset;
1035 if (!New.IsSplittable) {
1036 New.EndOffset = std::max(New.EndOffset, SplitEndOffset);
1037 if (j != e && !Partitions[j].IsSplittable)
1038 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
1039 New.IsSplittable = true;
1040 // If there is a trailing splittable partition which won't be fused into
1041 // the next splittable partition go ahead and add it onto the partitions
1043 if (New.BeginOffset < New.EndOffset &&
1044 (j == e || !Partitions[j].IsSplittable ||
1045 New.EndOffset < Partitions[j].BeginOffset)) {
1046 Partitions.push_back(New);
1047 New.BeginOffset = New.EndOffset = 0ull;
1052 // Re-sort the partitions now that they have been split and merged into
1053 // disjoint set of partitions. Also remove any of the dead partitions we've
1054 // replaced in the process.
1055 std::sort(Partitions.begin(), Partitions.end());
1056 if (NumDeadPartitions) {
1057 assert(Partitions.back().BeginOffset == UINT64_MAX);
1058 assert(Partitions.back().EndOffset == UINT64_MAX);
1059 assert((ptrdiff_t)NumDeadPartitions ==
1060 std::count(Partitions.begin(), Partitions.end(), Partitions.back()));
1062 Partitions.erase(Partitions.end() - NumDeadPartitions, Partitions.end());
1065 AllocaPartitioning::AllocaPartitioning(const TargetData &TD, AllocaInst &AI)
1070 PointerEscapingInstr(0) {
1071 PartitionBuilder PB(TD, AI, *this);
1075 if (Partitions.size() > 1) {
1076 // Sort the uses. This arranges for the offsets to be in ascending order,
1077 // and the sizes to be in descending order.
1078 std::sort(Partitions.begin(), Partitions.end());
1080 // Intersect splittability for all partitions with equal offsets and sizes.
1081 // Then remove all but the first so that we have a sequence of non-equal but
1082 // potentially overlapping partitions.
1083 for (iterator I = Partitions.begin(), J = I, E = Partitions.end(); I != E;
1086 while (J != E && *I == *J) {
1087 I->IsSplittable &= J->IsSplittable;
1091 Partitions.erase(std::unique(Partitions.begin(), Partitions.end()),
1094 // Split splittable and merge unsplittable partitions into a disjoint set
1095 // of partitions over the used space of the allocation.
1096 splitAndMergePartitions();
1099 // Now build up the user lists for each of these disjoint partitions by
1100 // re-walking the recursive users of the alloca.
1101 Uses.resize(Partitions.size());
1102 UseBuilder UB(TD, AI, *this);
1106 Type *AllocaPartitioning::getCommonType(iterator I) const {
1108 for (const_use_iterator UI = use_begin(I), UE = use_end(I); UI != UE; ++UI) {
1109 if (isa<IntrinsicInst>(*UI->U->getUser()))
1111 if (UI->BeginOffset != I->BeginOffset || UI->EndOffset != I->EndOffset)
1115 if (LoadInst *LI = dyn_cast<LoadInst>(UI->U->getUser())) {
1116 UserTy = LI->getType();
1117 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI->U->getUser())) {
1118 UserTy = SI->getValueOperand()->getType();
1121 if (Ty && Ty != UserTy)
1129 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1131 void AllocaPartitioning::print(raw_ostream &OS, const_iterator I,
1132 StringRef Indent) const {
1133 OS << Indent << "partition #" << (I - begin())
1134 << " [" << I->BeginOffset << "," << I->EndOffset << ")"
1135 << (I->IsSplittable ? " (splittable)" : "")
1136 << (Uses[I - begin()].empty() ? " (zero uses)" : "")
1140 void AllocaPartitioning::printUsers(raw_ostream &OS, const_iterator I,
1141 StringRef Indent) const {
1142 for (const_use_iterator UI = use_begin(I), UE = use_end(I);
1144 OS << Indent << " [" << UI->BeginOffset << "," << UI->EndOffset << ") "
1145 << "used by: " << *UI->U->getUser() << "\n";
1146 if (MemTransferInst *II = dyn_cast<MemTransferInst>(UI->U->getUser())) {
1147 const MemTransferOffsets &MTO = MemTransferInstData.lookup(II);
1149 if (!MTO.IsSplittable)
1150 IsDest = UI->BeginOffset == MTO.DestBegin;
1152 IsDest = MTO.DestBegin != 0u;
1153 OS << Indent << " (original " << (IsDest ? "dest" : "source") << ": "
1154 << "[" << (IsDest ? MTO.DestBegin : MTO.SourceBegin)
1155 << "," << (IsDest ? MTO.DestEnd : MTO.SourceEnd) << ")\n";
1160 void AllocaPartitioning::print(raw_ostream &OS) const {
1161 if (PointerEscapingInstr) {
1162 OS << "No partitioning for alloca: " << AI << "\n"
1163 << " A pointer to this alloca escaped by:\n"
1164 << " " << *PointerEscapingInstr << "\n";
1168 OS << "Partitioning of alloca: " << AI << "\n";
1170 for (const_iterator I = begin(), E = end(); I != E; ++I, ++Num) {
1176 void AllocaPartitioning::dump(const_iterator I) const { print(dbgs(), I); }
1177 void AllocaPartitioning::dump() const { print(dbgs()); }
1179 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1183 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
1185 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
1186 /// the loads and stores of an alloca instruction, as well as updating its
1187 /// debug information. This is used when a domtree is unavailable and thus
1188 /// mem2reg in its full form can't be used to handle promotion of allocas to
1190 class AllocaPromoter : public LoadAndStorePromoter {
1194 SmallVector<DbgDeclareInst *, 4> DDIs;
1195 SmallVector<DbgValueInst *, 4> DVIs;
1198 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
1199 AllocaInst &AI, DIBuilder &DIB)
1200 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
1202 void run(const SmallVectorImpl<Instruction*> &Insts) {
1203 // Remember which alloca we're promoting (for isInstInList).
1204 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
1205 for (Value::use_iterator UI = DebugNode->use_begin(),
1206 UE = DebugNode->use_end();
1208 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
1209 DDIs.push_back(DDI);
1210 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
1211 DVIs.push_back(DVI);
1214 LoadAndStorePromoter::run(Insts);
1215 AI.eraseFromParent();
1216 while (!DDIs.empty())
1217 DDIs.pop_back_val()->eraseFromParent();
1218 while (!DVIs.empty())
1219 DVIs.pop_back_val()->eraseFromParent();
1222 virtual bool isInstInList(Instruction *I,
1223 const SmallVectorImpl<Instruction*> &Insts) const {
1224 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1225 return LI->getOperand(0) == &AI;
1226 return cast<StoreInst>(I)->getPointerOperand() == &AI;
1229 virtual void updateDebugInfo(Instruction *Inst) const {
1230 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(),
1231 E = DDIs.end(); I != E; ++I) {
1232 DbgDeclareInst *DDI = *I;
1233 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
1234 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1235 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
1236 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1238 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(),
1239 E = DVIs.end(); I != E; ++I) {
1240 DbgValueInst *DVI = *I;
1242 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1243 // If an argument is zero extended then use argument directly. The ZExt
1244 // may be zapped by an optimization pass in future.
1245 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1246 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
1247 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1248 Arg = dyn_cast<Argument>(SExt->getOperand(0));
1250 Arg = SI->getOperand(0);
1251 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
1252 Arg = LI->getOperand(0);
1256 Instruction *DbgVal =
1257 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
1259 DbgVal->setDebugLoc(DVI->getDebugLoc());
1263 } // end anon namespace
1267 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
1269 /// This pass takes allocations which can be completely analyzed (that is, they
1270 /// don't escape) and tries to turn them into scalar SSA values. There are
1271 /// a few steps to this process.
1273 /// 1) It takes allocations of aggregates and analyzes the ways in which they
1274 /// are used to try to split them into smaller allocations, ideally of
1275 /// a single scalar data type. It will split up memcpy and memset accesses
1276 /// as necessary and try to isolate invidual scalar accesses.
1277 /// 2) It will transform accesses into forms which are suitable for SSA value
1278 /// promotion. This can be replacing a memset with a scalar store of an
1279 /// integer value, or it can involve speculating operations on a PHI or
1280 /// select to be a PHI or select of the results.
1281 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
1282 /// onto insert and extract operations on a vector value, and convert them to
1283 /// this form. By doing so, it will enable promotion of vector aggregates to
1284 /// SSA vector values.
1285 class SROA : public FunctionPass {
1286 const bool RequiresDomTree;
1289 const TargetData *TD;
1292 /// \brief Worklist of alloca instructions to simplify.
1294 /// Each alloca in the function is added to this. Each new alloca formed gets
1295 /// added to it as well to recursively simplify unless that alloca can be
1296 /// directly promoted. Finally, each time we rewrite a use of an alloca other
1297 /// the one being actively rewritten, we add it back onto the list if not
1298 /// already present to ensure it is re-visited.
1299 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
1301 /// \brief A collection of instructions to delete.
1302 /// We try to batch deletions to simplify code and make things a bit more
1304 SmallVector<Instruction *, 8> DeadInsts;
1306 /// \brief A set to prevent repeatedly marking an instruction split into many
1307 /// uses as dead. Only used to guard insertion into DeadInsts.
1308 SmallPtrSet<Instruction *, 4> DeadSplitInsts;
1310 /// \brief A collection of alloca instructions we can directly promote.
1311 std::vector<AllocaInst *> PromotableAllocas;
1314 SROA(bool RequiresDomTree = true)
1315 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
1316 C(0), TD(0), DT(0) {
1317 initializeSROAPass(*PassRegistry::getPassRegistry());
1319 bool runOnFunction(Function &F);
1320 void getAnalysisUsage(AnalysisUsage &AU) const;
1322 const char *getPassName() const { return "SROA"; }
1326 friend class PHIOrSelectSpeculator;
1327 friend class AllocaPartitionRewriter;
1328 friend class AllocaPartitionVectorRewriter;
1330 bool rewriteAllocaPartition(AllocaInst &AI,
1331 AllocaPartitioning &P,
1332 AllocaPartitioning::iterator PI);
1333 bool splitAlloca(AllocaInst &AI, AllocaPartitioning &P);
1334 bool runOnAlloca(AllocaInst &AI);
1335 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
1336 bool promoteAllocas(Function &F);
1342 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
1343 return new SROA(RequiresDomTree);
1346 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
1348 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
1349 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
1352 /// \brief Accumulate the constant offsets in a GEP into a single APInt offset.
1354 /// If the provided GEP is all-constant, the total byte offset formed by the
1355 /// GEP is computed and Offset is set to it. If the GEP has any non-constant
1356 /// operands, the function returns false and the value of Offset is unmodified.
1357 static bool accumulateGEPOffsets(const TargetData &TD, GEPOperator &GEP,
1359 APInt GEPOffset(Offset.getBitWidth(), 0);
1360 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1361 GTI != GTE; ++GTI) {
1362 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
1365 if (OpC->isZero()) continue;
1367 // Handle a struct index, which adds its field offset to the pointer.
1368 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1369 unsigned ElementIdx = OpC->getZExtValue();
1370 const StructLayout *SL = TD.getStructLayout(STy);
1371 GEPOffset += APInt(Offset.getBitWidth(),
1372 SL->getElementOffset(ElementIdx));
1376 APInt TypeSize(Offset.getBitWidth(),
1377 TD.getTypeAllocSize(GTI.getIndexedType()));
1378 if (VectorType *VTy = dyn_cast<VectorType>(*GTI)) {
1379 assert((VTy->getScalarSizeInBits() % 8) == 0 &&
1380 "vector element size is not a multiple of 8, cannot GEP over it");
1381 TypeSize = VTy->getScalarSizeInBits() / 8;
1384 GEPOffset += OpC->getValue().sextOrTrunc(Offset.getBitWidth()) * TypeSize;
1390 /// \brief Build a GEP out of a base pointer and indices.
1392 /// This will return the BasePtr if that is valid, or build a new GEP
1393 /// instruction using the IRBuilder if GEP-ing is needed.
1394 static Value *buildGEP(IRBuilder<> &IRB, Value *BasePtr,
1395 SmallVectorImpl<Value *> &Indices,
1396 const Twine &Prefix) {
1397 if (Indices.empty())
1400 // A single zero index is a no-op, so check for this and avoid building a GEP
1402 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1405 return IRB.CreateInBoundsGEP(BasePtr, Indices, Prefix + ".idx");
1408 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1409 /// TargetTy without changing the offset of the pointer.
1411 /// This routine assumes we've already established a properly offset GEP with
1412 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1413 /// zero-indices down through type layers until we find one the same as
1414 /// TargetTy. If we can't find one with the same type, we at least try to use
1415 /// one with the same size. If none of that works, we just produce the GEP as
1416 /// indicated by Indices to have the correct offset.
1417 static Value *getNaturalGEPWithType(IRBuilder<> &IRB, const TargetData &TD,
1418 Value *BasePtr, Type *Ty, Type *TargetTy,
1419 SmallVectorImpl<Value *> &Indices,
1420 const Twine &Prefix) {
1422 return buildGEP(IRB, BasePtr, Indices, Prefix);
1424 // See if we can descend into a struct and locate a field with the correct
1426 unsigned NumLayers = 0;
1427 Type *ElementTy = Ty;
1429 if (ElementTy->isPointerTy())
1431 if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
1432 ElementTy = SeqTy->getElementType();
1433 Indices.push_back(IRB.getInt(APInt(TD.getPointerSizeInBits(), 0)));
1434 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1435 ElementTy = *STy->element_begin();
1436 Indices.push_back(IRB.getInt32(0));
1441 } while (ElementTy != TargetTy);
1442 if (ElementTy != TargetTy)
1443 Indices.erase(Indices.end() - NumLayers, Indices.end());
1445 return buildGEP(IRB, BasePtr, Indices, Prefix);
1448 /// \brief Recursively compute indices for a natural GEP.
1450 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1451 /// element types adding appropriate indices for the GEP.
1452 static Value *getNaturalGEPRecursively(IRBuilder<> &IRB, const TargetData &TD,
1453 Value *Ptr, Type *Ty, APInt &Offset,
1455 SmallVectorImpl<Value *> &Indices,
1456 const Twine &Prefix) {
1458 return getNaturalGEPWithType(IRB, TD, Ptr, Ty, TargetTy, Indices, Prefix);
1460 // We can't recurse through pointer types.
1461 if (Ty->isPointerTy())
1464 // We try to analyze GEPs over vectors here, but note that these GEPs are
1465 // extremely poorly defined currently. The long-term goal is to remove GEPing
1466 // over a vector from the IR completely.
1467 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1468 unsigned ElementSizeInBits = VecTy->getScalarSizeInBits();
1469 if (ElementSizeInBits % 8)
1470 return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
1471 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1472 APInt NumSkippedElements = Offset.udiv(ElementSize);
1473 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1475 Offset -= NumSkippedElements * ElementSize;
1476 Indices.push_back(IRB.getInt(NumSkippedElements));
1477 return getNaturalGEPRecursively(IRB, TD, Ptr, VecTy->getElementType(),
1478 Offset, TargetTy, Indices, Prefix);
1481 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1482 Type *ElementTy = ArrTy->getElementType();
1483 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1484 APInt NumSkippedElements = Offset.udiv(ElementSize);
1485 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1488 Offset -= NumSkippedElements * ElementSize;
1489 Indices.push_back(IRB.getInt(NumSkippedElements));
1490 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1494 StructType *STy = dyn_cast<StructType>(Ty);
1498 const StructLayout *SL = TD.getStructLayout(STy);
1499 uint64_t StructOffset = Offset.getZExtValue();
1500 if (StructOffset >= SL->getSizeInBytes())
1502 unsigned Index = SL->getElementContainingOffset(StructOffset);
1503 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1504 Type *ElementTy = STy->getElementType(Index);
1505 if (Offset.uge(TD.getTypeAllocSize(ElementTy)))
1506 return 0; // The offset points into alignment padding.
1508 Indices.push_back(IRB.getInt32(Index));
1509 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1513 /// \brief Get a natural GEP from a base pointer to a particular offset and
1514 /// resulting in a particular type.
1516 /// The goal is to produce a "natural" looking GEP that works with the existing
1517 /// composite types to arrive at the appropriate offset and element type for
1518 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1519 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1520 /// Indices, and setting Ty to the result subtype.
1522 /// If no natural GEP can be constructed, this function returns null.
1523 static Value *getNaturalGEPWithOffset(IRBuilder<> &IRB, const TargetData &TD,
1524 Value *Ptr, APInt Offset, Type *TargetTy,
1525 SmallVectorImpl<Value *> &Indices,
1526 const Twine &Prefix) {
1527 PointerType *Ty = cast<PointerType>(Ptr->getType());
1529 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1531 if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
1534 Type *ElementTy = Ty->getElementType();
1535 if (!ElementTy->isSized())
1536 return 0; // We can't GEP through an unsized element.
1537 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1538 if (ElementSize == 0)
1539 return 0; // Zero-length arrays can't help us build a natural GEP.
1540 APInt NumSkippedElements = Offset.udiv(ElementSize);
1542 Offset -= NumSkippedElements * ElementSize;
1543 Indices.push_back(IRB.getInt(NumSkippedElements));
1544 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1548 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1549 /// resulting pointer has PointerTy.
1551 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1552 /// and produces the pointer type desired. Where it cannot, it will try to use
1553 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1554 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1555 /// bitcast to the type.
1557 /// The strategy for finding the more natural GEPs is to peel off layers of the
1558 /// pointer, walking back through bit casts and GEPs, searching for a base
1559 /// pointer from which we can compute a natural GEP with the desired
1560 /// properities. The algorithm tries to fold as many constant indices into
1561 /// a single GEP as possible, thus making each GEP more independent of the
1562 /// surrounding code.
1563 static Value *getAdjustedPtr(IRBuilder<> &IRB, const TargetData &TD,
1564 Value *Ptr, APInt Offset, Type *PointerTy,
1565 const Twine &Prefix) {
1566 // Even though we don't look through PHI nodes, we could be called on an
1567 // instruction in an unreachable block, which may be on a cycle.
1568 SmallPtrSet<Value *, 4> Visited;
1569 Visited.insert(Ptr);
1570 SmallVector<Value *, 4> Indices;
1572 // We may end up computing an offset pointer that has the wrong type. If we
1573 // never are able to compute one directly that has the correct type, we'll
1574 // fall back to it, so keep it around here.
1575 Value *OffsetPtr = 0;
1577 // Remember any i8 pointer we come across to re-use if we need to do a raw
1580 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1582 Type *TargetTy = PointerTy->getPointerElementType();
1585 // First fold any existing GEPs into the offset.
1586 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1587 APInt GEPOffset(Offset.getBitWidth(), 0);
1588 if (!accumulateGEPOffsets(TD, *GEP, GEPOffset))
1590 Offset += GEPOffset;
1591 Ptr = GEP->getPointerOperand();
1592 if (!Visited.insert(Ptr))
1596 // See if we can perform a natural GEP here.
1598 if (Value *P = getNaturalGEPWithOffset(IRB, TD, Ptr, Offset, TargetTy,
1600 if (P->getType() == PointerTy) {
1601 // Zap any offset pointer that we ended up computing in previous rounds.
1602 if (OffsetPtr && OffsetPtr->use_empty())
1603 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1604 I->eraseFromParent();
1612 // Stash this pointer if we've found an i8*.
1613 if (Ptr->getType()->isIntegerTy(8)) {
1615 Int8PtrOffset = Offset;
1618 // Peel off a layer of the pointer and update the offset appropriately.
1619 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1620 Ptr = cast<Operator>(Ptr)->getOperand(0);
1621 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1622 if (GA->mayBeOverridden())
1624 Ptr = GA->getAliasee();
1628 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1629 } while (Visited.insert(Ptr));
1633 Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
1634 Prefix + ".raw_cast");
1635 Int8PtrOffset = Offset;
1638 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1639 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1640 Prefix + ".raw_idx");
1644 // On the off chance we were targeting i8*, guard the bitcast here.
1645 if (Ptr->getType() != PointerTy)
1646 Ptr = IRB.CreateBitCast(Ptr, PointerTy, Prefix + ".cast");
1651 /// \brief Test whether the given alloca partition can be promoted to a vector.
1653 /// This is a quick test to check whether we can rewrite a particular alloca
1654 /// partition (and its newly formed alloca) into a vector alloca with only
1655 /// whole-vector loads and stores such that it could be promoted to a vector
1656 /// SSA value. We only can ensure this for a limited set of operations, and we
1657 /// don't want to do the rewrites unless we are confident that the result will
1658 /// be promotable, so we have an early test here.
1659 static bool isVectorPromotionViable(const TargetData &TD,
1661 AllocaPartitioning &P,
1662 uint64_t PartitionBeginOffset,
1663 uint64_t PartitionEndOffset,
1664 AllocaPartitioning::const_use_iterator I,
1665 AllocaPartitioning::const_use_iterator E) {
1666 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
1670 uint64_t VecSize = TD.getTypeSizeInBits(Ty);
1671 uint64_t ElementSize = Ty->getScalarSizeInBits();
1673 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1674 // that aren't byte sized.
1675 if (ElementSize % 8)
1677 assert((VecSize % 8) == 0 && "vector size not a multiple of element size?");
1681 for (; I != E; ++I) {
1682 uint64_t BeginOffset = I->BeginOffset - PartitionBeginOffset;
1683 uint64_t BeginIndex = BeginOffset / ElementSize;
1684 if (BeginIndex * ElementSize != BeginOffset ||
1685 BeginIndex >= Ty->getNumElements())
1687 uint64_t EndOffset = I->EndOffset - PartitionBeginOffset;
1688 uint64_t EndIndex = EndOffset / ElementSize;
1689 if (EndIndex * ElementSize != EndOffset ||
1690 EndIndex > Ty->getNumElements())
1693 // FIXME: We should build shuffle vector instructions to handle
1694 // non-element-sized accesses.
1695 if ((EndOffset - BeginOffset) != ElementSize &&
1696 (EndOffset - BeginOffset) != VecSize)
1699 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
1700 if (MI->isVolatile())
1702 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
1703 const AllocaPartitioning::MemTransferOffsets &MTO
1704 = P.getMemTransferOffsets(*MTI);
1705 if (!MTO.IsSplittable)
1708 } else if (I->U->get()->getType()->getPointerElementType()->isStructTy()) {
1709 // Disable vector promotion when there are loads or stores of an FCA.
1711 } else if (!isa<LoadInst>(I->U->getUser()) &&
1712 !isa<StoreInst>(I->U->getUser())) {
1719 /// \brief Test whether the given alloca partition can be promoted to an int.
1721 /// This is a quick test to check whether we can rewrite a particular alloca
1722 /// partition (and its newly formed alloca) into an integer alloca suitable for
1723 /// promotion to an SSA value. We only can ensure this for a limited set of
1724 /// operations, and we don't want to do the rewrites unless we are confident
1725 /// that the result will be promotable, so we have an early test here.
1726 static bool isIntegerPromotionViable(const TargetData &TD,
1728 AllocaPartitioning &P,
1729 AllocaPartitioning::const_use_iterator I,
1730 AllocaPartitioning::const_use_iterator E) {
1731 IntegerType *Ty = dyn_cast<IntegerType>(AllocaTy);
1735 // Check the uses to ensure the uses are (likely) promoteable integer uses.
1736 // Also ensure that the alloca has a covering load or store. We don't want
1737 // promote because of some other unsplittable entry (which we may make
1738 // splittable later) and lose the ability to promote each element access.
1739 bool WholeAllocaOp = false;
1740 for (; I != E; ++I) {
1741 if (LoadInst *LI = dyn_cast<LoadInst>(I->U->getUser())) {
1742 if (LI->isVolatile() || !LI->getType()->isIntegerTy())
1744 if (LI->getType() == Ty)
1745 WholeAllocaOp = true;
1746 } else if (StoreInst *SI = dyn_cast<StoreInst>(I->U->getUser())) {
1747 if (SI->isVolatile() || !SI->getValueOperand()->getType()->isIntegerTy())
1749 if (SI->getValueOperand()->getType() == Ty)
1750 WholeAllocaOp = true;
1751 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
1752 if (MI->isVolatile())
1754 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
1755 const AllocaPartitioning::MemTransferOffsets &MTO
1756 = P.getMemTransferOffsets(*MTI);
1757 if (!MTO.IsSplittable)
1764 return WholeAllocaOp;
1768 /// \brief Visitor to speculate PHIs and Selects where possible.
1769 class PHIOrSelectSpeculator : public InstVisitor<PHIOrSelectSpeculator> {
1770 // Befriend the base class so it can delegate to private visit methods.
1771 friend class llvm::InstVisitor<PHIOrSelectSpeculator>;
1773 const TargetData &TD;
1774 AllocaPartitioning &P;
1778 PHIOrSelectSpeculator(const TargetData &TD, AllocaPartitioning &P, SROA &Pass)
1779 : TD(TD), P(P), Pass(Pass) {}
1781 /// \brief Visit the users of the alloca partition and rewrite them.
1782 void visitUsers(AllocaPartitioning::const_use_iterator I,
1783 AllocaPartitioning::const_use_iterator E) {
1785 visit(cast<Instruction>(I->U->getUser()));
1789 // By default, skip this instruction.
1790 void visitInstruction(Instruction &I) {}
1792 /// PHI instructions that use an alloca and are subsequently loaded can be
1793 /// rewritten to load both input pointers in the pred blocks and then PHI the
1794 /// results, allowing the load of the alloca to be promoted.
1796 /// %P2 = phi [i32* %Alloca, i32* %Other]
1797 /// %V = load i32* %P2
1799 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1801 /// %V2 = load i32* %Other
1803 /// %V = phi [i32 %V1, i32 %V2]
1805 /// We can do this to a select if its only uses are loads and if the operand
1806 /// to the select can be loaded unconditionally.
1808 /// FIXME: This should be hoisted into a generic utility, likely in
1809 /// Transforms/Util/Local.h
1810 bool isSafePHIToSpeculate(PHINode &PN, SmallVectorImpl<LoadInst *> &Loads) {
1811 // For now, we can only do this promotion if the load is in the same block
1812 // as the PHI, and if there are no stores between the phi and load.
1813 // TODO: Allow recursive phi users.
1814 // TODO: Allow stores.
1815 BasicBlock *BB = PN.getParent();
1816 unsigned MaxAlign = 0;
1817 for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end();
1819 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1820 if (LI == 0 || !LI->isSimple()) return false;
1822 // For now we only allow loads in the same block as the PHI. This is
1823 // a common case that happens when instcombine merges two loads through
1825 if (LI->getParent() != BB) return false;
1827 // Ensure that there are no instructions between the PHI and the load that
1829 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1830 if (BBI->mayWriteToMemory())
1833 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1834 Loads.push_back(LI);
1837 // We can only transform this if it is safe to push the loads into the
1838 // predecessor blocks. The only thing to watch out for is that we can't put
1839 // a possibly trapping load in the predecessor if it is a critical edge.
1840 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num;
1842 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1843 Value *InVal = PN.getIncomingValue(Idx);
1845 // If the value is produced by the terminator of the predecessor (an
1846 // invoke) or it has side-effects, there is no valid place to put a load
1847 // in the predecessor.
1848 if (TI == InVal || TI->mayHaveSideEffects())
1851 // If the predecessor has a single successor, then the edge isn't
1853 if (TI->getNumSuccessors() == 1)
1856 // If this pointer is always safe to load, or if we can prove that there
1857 // is already a load in the block, then we can move the load to the pred
1859 if (InVal->isDereferenceablePointer() ||
1860 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, &TD))
1869 void visitPHINode(PHINode &PN) {
1870 DEBUG(dbgs() << " original: " << PN << "\n");
1872 SmallVector<LoadInst *, 4> Loads;
1873 if (!isSafePHIToSpeculate(PN, Loads))
1876 assert(!Loads.empty());
1878 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1879 IRBuilder<> PHIBuilder(&PN);
1880 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1881 PN.getName() + ".sroa.speculated");
1883 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1884 // matter which one we get and if any differ, it doesn't matter.
1885 LoadInst *SomeLoad = cast<LoadInst>(Loads.back());
1886 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1887 unsigned Align = SomeLoad->getAlignment();
1889 // Rewrite all loads of the PN to use the new PHI.
1891 LoadInst *LI = Loads.pop_back_val();
1892 LI->replaceAllUsesWith(NewPN);
1893 Pass.DeadInsts.push_back(LI);
1894 } while (!Loads.empty());
1896 // Inject loads into all of the pred blocks.
1897 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1898 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1899 TerminatorInst *TI = Pred->getTerminator();
1900 Use *InUse = &PN.getOperandUse(PN.getOperandNumForIncomingValue(Idx));
1901 Value *InVal = PN.getIncomingValue(Idx);
1902 IRBuilder<> PredBuilder(TI);
1905 = PredBuilder.CreateLoad(InVal, (PN.getName() + ".sroa.speculate.load." +
1907 ++NumLoadsSpeculated;
1908 Load->setAlignment(Align);
1910 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1911 NewPN->addIncoming(Load, Pred);
1913 Instruction *Ptr = dyn_cast<Instruction>(InVal);
1915 // No uses to rewrite.
1918 // Try to lookup and rewrite any partition uses corresponding to this phi
1920 AllocaPartitioning::iterator PI
1921 = P.findPartitionForPHIOrSelectOperand(InUse);
1925 // Replace the Use in the PartitionUse for this operand with the Use
1927 AllocaPartitioning::use_iterator UI
1928 = P.findPartitionUseForPHIOrSelectOperand(InUse);
1929 assert(isa<PHINode>(*UI->U->getUser()));
1930 UI->U = &Load->getOperandUse(Load->getPointerOperandIndex());
1932 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1935 /// Select instructions that use an alloca and are subsequently loaded can be
1936 /// rewritten to load both input pointers and then select between the result,
1937 /// allowing the load of the alloca to be promoted.
1939 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1940 /// %V = load i32* %P2
1942 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1943 /// %V2 = load i32* %Other
1944 /// %V = select i1 %cond, i32 %V1, i32 %V2
1946 /// We can do this to a select if its only uses are loads and if the operand
1947 /// to the select can be loaded unconditionally.
1948 bool isSafeSelectToSpeculate(SelectInst &SI,
1949 SmallVectorImpl<LoadInst *> &Loads) {
1950 Value *TValue = SI.getTrueValue();
1951 Value *FValue = SI.getFalseValue();
1952 bool TDerefable = TValue->isDereferenceablePointer();
1953 bool FDerefable = FValue->isDereferenceablePointer();
1955 for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end();
1957 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1958 if (LI == 0 || !LI->isSimple()) return false;
1960 // Both operands to the select need to be dereferencable, either
1961 // absolutely (e.g. allocas) or at this point because we can see other
1963 if (!TDerefable && !isSafeToLoadUnconditionally(TValue, LI,
1964 LI->getAlignment(), &TD))
1966 if (!FDerefable && !isSafeToLoadUnconditionally(FValue, LI,
1967 LI->getAlignment(), &TD))
1969 Loads.push_back(LI);
1975 void visitSelectInst(SelectInst &SI) {
1976 DEBUG(dbgs() << " original: " << SI << "\n");
1977 IRBuilder<> IRB(&SI);
1979 // If the select isn't safe to speculate, just use simple logic to emit it.
1980 SmallVector<LoadInst *, 4> Loads;
1981 if (!isSafeSelectToSpeculate(SI, Loads))
1984 Use *Ops[2] = { &SI.getOperandUse(1), &SI.getOperandUse(2) };
1985 AllocaPartitioning::iterator PIs[2];
1986 AllocaPartitioning::PartitionUse PUs[2];
1987 for (unsigned i = 0, e = 2; i != e; ++i) {
1988 PIs[i] = P.findPartitionForPHIOrSelectOperand(Ops[i]);
1989 if (PIs[i] != P.end()) {
1990 // If the pointer is within the partitioning, remove the select from
1991 // its uses. We'll add in the new loads below.
1992 AllocaPartitioning::use_iterator UI
1993 = P.findPartitionUseForPHIOrSelectOperand(Ops[i]);
1995 P.use_erase(PIs[i], UI);
1999 Value *TV = SI.getTrueValue();
2000 Value *FV = SI.getFalseValue();
2001 // Replace the loads of the select with a select of two loads.
2002 while (!Loads.empty()) {
2003 LoadInst *LI = Loads.pop_back_val();
2005 IRB.SetInsertPoint(LI);
2007 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
2009 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
2010 NumLoadsSpeculated += 2;
2012 // Transfer alignment and TBAA info if present.
2013 TL->setAlignment(LI->getAlignment());
2014 FL->setAlignment(LI->getAlignment());
2015 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
2016 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
2017 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
2020 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
2021 LI->getName() + ".sroa.speculated");
2023 LoadInst *Loads[2] = { TL, FL };
2024 for (unsigned i = 0, e = 2; i != e; ++i) {
2025 if (PIs[i] != P.end()) {
2026 Use *LoadUse = &Loads[i]->getOperandUse(0);
2027 assert(PUs[i].U->get() == LoadUse->get());
2029 P.use_push_back(PIs[i], PUs[i]);
2033 DEBUG(dbgs() << " speculated to: " << *V << "\n");
2034 LI->replaceAllUsesWith(V);
2035 Pass.DeadInsts.push_back(LI);
2040 /// \brief Visitor to rewrite instructions using a partition of an alloca to
2041 /// use a new alloca.
2043 /// Also implements the rewriting to vector-based accesses when the partition
2044 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2046 class AllocaPartitionRewriter : public InstVisitor<AllocaPartitionRewriter,
2048 // Befriend the base class so it can delegate to private visit methods.
2049 friend class llvm::InstVisitor<AllocaPartitionRewriter, bool>;
2051 const TargetData &TD;
2052 AllocaPartitioning &P;
2054 AllocaInst &OldAI, &NewAI;
2055 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2057 // If we are rewriting an alloca partition which can be written as pure
2058 // vector operations, we stash extra information here. When VecTy is
2059 // non-null, we have some strict guarantees about the rewriten alloca:
2060 // - The new alloca is exactly the size of the vector type here.
2061 // - The accesses all either map to the entire vector or to a single
2063 // - The set of accessing instructions is only one of those handled above
2064 // in isVectorPromotionViable. Generally these are the same access kinds
2065 // which are promotable via mem2reg.
2068 uint64_t ElementSize;
2070 // This is a convenience and flag variable that will be null unless the new
2071 // alloca has a promotion-targeted integer type due to passing
2072 // isIntegerPromotionViable above. If it is non-null does, the desired
2073 // integer type will be stored here for easy access during rewriting.
2074 IntegerType *IntPromotionTy;
2076 // The offset of the partition user currently being rewritten.
2077 uint64_t BeginOffset, EndOffset;
2079 Instruction *OldPtr;
2081 // The name prefix to use when rewriting instructions for this alloca.
2082 std::string NamePrefix;
2085 AllocaPartitionRewriter(const TargetData &TD, AllocaPartitioning &P,
2086 AllocaPartitioning::iterator PI,
2087 SROA &Pass, AllocaInst &OldAI, AllocaInst &NewAI,
2088 uint64_t NewBeginOffset, uint64_t NewEndOffset)
2089 : TD(TD), P(P), Pass(Pass),
2090 OldAI(OldAI), NewAI(NewAI),
2091 NewAllocaBeginOffset(NewBeginOffset),
2092 NewAllocaEndOffset(NewEndOffset),
2093 VecTy(), ElementTy(), ElementSize(), IntPromotionTy(),
2094 BeginOffset(), EndOffset() {
2097 /// \brief Visit the users of the alloca partition and rewrite them.
2098 bool visitUsers(AllocaPartitioning::const_use_iterator I,
2099 AllocaPartitioning::const_use_iterator E) {
2100 if (isVectorPromotionViable(TD, NewAI.getAllocatedType(), P,
2101 NewAllocaBeginOffset, NewAllocaEndOffset,
2104 VecTy = cast<VectorType>(NewAI.getAllocatedType());
2105 ElementTy = VecTy->getElementType();
2106 assert((VecTy->getScalarSizeInBits() % 8) == 0 &&
2107 "Only multiple-of-8 sized vector elements are viable");
2108 ElementSize = VecTy->getScalarSizeInBits() / 8;
2109 } else if (isIntegerPromotionViable(TD, NewAI.getAllocatedType(),
2111 IntPromotionTy = cast<IntegerType>(NewAI.getAllocatedType());
2113 bool CanSROA = true;
2114 for (; I != E; ++I) {
2115 BeginOffset = I->BeginOffset;
2116 EndOffset = I->EndOffset;
2118 OldPtr = cast<Instruction>(I->U->get());
2119 NamePrefix = (Twine(NewAI.getName()) + "." + Twine(BeginOffset)).str();
2120 CanSROA &= visit(cast<Instruction>(I->U->getUser()));
2132 // Every instruction which can end up as a user must have a rewrite rule.
2133 bool visitInstruction(Instruction &I) {
2134 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2135 llvm_unreachable("No rewrite rule for this instruction!");
2138 Twine getName(const Twine &Suffix) {
2139 return NamePrefix + Suffix;
2142 Value *getAdjustedAllocaPtr(IRBuilder<> &IRB, Type *PointerTy) {
2143 assert(BeginOffset >= NewAllocaBeginOffset);
2144 APInt Offset(TD.getPointerSizeInBits(), BeginOffset - NewAllocaBeginOffset);
2145 return getAdjustedPtr(IRB, TD, &NewAI, Offset, PointerTy, getName(""));
2148 unsigned getAdjustedAlign(uint64_t Offset) {
2149 unsigned NewAIAlign = NewAI.getAlignment();
2151 NewAIAlign = TD.getABITypeAlignment(NewAI.getAllocatedType());
2152 return MinAlign(NewAIAlign, Offset);
2154 unsigned getAdjustedAlign() {
2155 return getAdjustedAlign(BeginOffset - NewAllocaBeginOffset);
2158 bool isTypeAlignSufficient(Type *Ty) {
2159 return TD.getABITypeAlignment(Ty) >= getAdjustedAlign();
2162 ConstantInt *getIndex(IRBuilder<> &IRB, uint64_t Offset) {
2163 assert(VecTy && "Can only call getIndex when rewriting a vector");
2164 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2165 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2166 uint32_t Index = RelOffset / ElementSize;
2167 assert(Index * ElementSize == RelOffset);
2168 return IRB.getInt32(Index);
2171 Value *extractInteger(IRBuilder<> &IRB, IntegerType *TargetTy,
2173 assert(IntPromotionTy && "Alloca is not an integer we can extract from");
2174 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2176 assert(Offset >= NewAllocaBeginOffset && "Out of bounds offset");
2177 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2179 V = IRB.CreateLShr(V, RelOffset*8, getName(".shift"));
2180 if (TargetTy != IntPromotionTy) {
2181 assert(TargetTy->getBitWidth() < IntPromotionTy->getBitWidth() &&
2182 "Cannot extract to a larger integer!");
2183 V = IRB.CreateTrunc(V, TargetTy, getName(".trunc"));
2188 StoreInst *insertInteger(IRBuilder<> &IRB, Value *V, uint64_t Offset) {
2189 IntegerType *Ty = cast<IntegerType>(V->getType());
2190 if (Ty == IntPromotionTy)
2191 return IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2193 assert(Ty->getBitWidth() < IntPromotionTy->getBitWidth() &&
2194 "Cannot insert a larger integer!");
2195 V = IRB.CreateZExt(V, IntPromotionTy, getName(".ext"));
2196 assert(Offset >= NewAllocaBeginOffset && "Out of bounds offset");
2197 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2199 V = IRB.CreateShl(V, RelOffset*8, getName(".shift"));
2201 APInt Mask = ~Ty->getMask().zext(IntPromotionTy->getBitWidth())
2203 Value *Old = IRB.CreateAnd(IRB.CreateAlignedLoad(&NewAI,
2204 NewAI.getAlignment(),
2205 getName(".oldload")),
2206 Mask, getName(".mask"));
2207 return IRB.CreateAlignedStore(IRB.CreateOr(Old, V, getName(".insert")),
2208 &NewAI, NewAI.getAlignment());
2211 void deleteIfTriviallyDead(Value *V) {
2212 Instruction *I = cast<Instruction>(V);
2213 if (isInstructionTriviallyDead(I))
2214 Pass.DeadInsts.push_back(I);
2217 Value *getValueCast(IRBuilder<> &IRB, Value *V, Type *Ty) {
2218 if (V->getType()->isIntegerTy() && Ty->isPointerTy())
2219 return IRB.CreateIntToPtr(V, Ty);
2220 if (V->getType()->isPointerTy() && Ty->isIntegerTy())
2221 return IRB.CreatePtrToInt(V, Ty);
2223 return IRB.CreateBitCast(V, Ty);
2226 bool rewriteVectorizedLoadInst(IRBuilder<> &IRB, LoadInst &LI, Value *OldOp) {
2228 if (LI.getType() == VecTy->getElementType() ||
2229 BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
2230 Result = IRB.CreateExtractElement(
2231 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), getName(".load")),
2232 getIndex(IRB, BeginOffset), getName(".extract"));
2234 Result = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2237 if (Result->getType() != LI.getType())
2238 Result = getValueCast(IRB, Result, LI.getType());
2239 LI.replaceAllUsesWith(Result);
2240 Pass.DeadInsts.push_back(&LI);
2242 DEBUG(dbgs() << " to: " << *Result << "\n");
2246 bool rewriteIntegerLoad(IRBuilder<> &IRB, LoadInst &LI) {
2247 assert(!LI.isVolatile());
2248 Value *Result = extractInteger(IRB, cast<IntegerType>(LI.getType()),
2250 LI.replaceAllUsesWith(Result);
2251 Pass.DeadInsts.push_back(&LI);
2252 DEBUG(dbgs() << " to: " << *Result << "\n");
2256 bool visitLoadInst(LoadInst &LI) {
2257 DEBUG(dbgs() << " original: " << LI << "\n");
2258 Value *OldOp = LI.getOperand(0);
2259 assert(OldOp == OldPtr);
2260 IRBuilder<> IRB(&LI);
2263 return rewriteVectorizedLoadInst(IRB, LI, OldOp);
2265 return rewriteIntegerLoad(IRB, LI);
2267 Value *NewPtr = getAdjustedAllocaPtr(IRB,
2268 LI.getPointerOperand()->getType());
2269 LI.setOperand(0, NewPtr);
2270 if (LI.getAlignment() || !isTypeAlignSufficient(LI.getType()))
2271 LI.setAlignment(getAdjustedAlign());
2272 DEBUG(dbgs() << " to: " << LI << "\n");
2274 deleteIfTriviallyDead(OldOp);
2275 return NewPtr == &NewAI && !LI.isVolatile();
2278 bool rewriteVectorizedStoreInst(IRBuilder<> &IRB, StoreInst &SI,
2280 Value *V = SI.getValueOperand();
2281 if (V->getType() == ElementTy ||
2282 BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
2283 if (V->getType() != ElementTy)
2284 V = getValueCast(IRB, V, ElementTy);
2285 LoadInst *LI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2287 V = IRB.CreateInsertElement(LI, V, getIndex(IRB, BeginOffset),
2288 getName(".insert"));
2289 } else if (V->getType() != VecTy) {
2290 V = getValueCast(IRB, V, VecTy);
2292 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2293 Pass.DeadInsts.push_back(&SI);
2296 DEBUG(dbgs() << " to: " << *Store << "\n");
2300 bool rewriteIntegerStore(IRBuilder<> &IRB, StoreInst &SI) {
2301 assert(!SI.isVolatile());
2302 StoreInst *Store = insertInteger(IRB, SI.getValueOperand(), BeginOffset);
2303 Pass.DeadInsts.push_back(&SI);
2305 DEBUG(dbgs() << " to: " << *Store << "\n");
2309 bool visitStoreInst(StoreInst &SI) {
2310 DEBUG(dbgs() << " original: " << SI << "\n");
2311 Value *OldOp = SI.getOperand(1);
2312 assert(OldOp == OldPtr);
2313 IRBuilder<> IRB(&SI);
2316 return rewriteVectorizedStoreInst(IRB, SI, OldOp);
2318 return rewriteIntegerStore(IRB, SI);
2320 Value *NewPtr = getAdjustedAllocaPtr(IRB,
2321 SI.getPointerOperand()->getType());
2322 SI.setOperand(1, NewPtr);
2323 if (SI.getAlignment() ||
2324 !isTypeAlignSufficient(SI.getValueOperand()->getType()))
2325 SI.setAlignment(getAdjustedAlign());
2326 if (SI.getAlignment())
2327 SI.setAlignment(MinAlign(NewAI.getAlignment(),
2328 BeginOffset - NewAllocaBeginOffset));
2329 DEBUG(dbgs() << " to: " << SI << "\n");
2331 deleteIfTriviallyDead(OldOp);
2332 return NewPtr == &NewAI && !SI.isVolatile();
2335 bool visitMemSetInst(MemSetInst &II) {
2336 DEBUG(dbgs() << " original: " << II << "\n");
2337 IRBuilder<> IRB(&II);
2338 assert(II.getRawDest() == OldPtr);
2340 // If the memset has a variable size, it cannot be split, just adjust the
2341 // pointer to the new alloca.
2342 if (!isa<Constant>(II.getLength())) {
2343 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2344 Type *CstTy = II.getAlignmentCst()->getType();
2345 II.setAlignment(ConstantInt::get(CstTy, getAdjustedAlign()));
2347 deleteIfTriviallyDead(OldPtr);
2351 // Record this instruction for deletion.
2352 if (Pass.DeadSplitInsts.insert(&II))
2353 Pass.DeadInsts.push_back(&II);
2355 Type *AllocaTy = NewAI.getAllocatedType();
2356 Type *ScalarTy = AllocaTy->getScalarType();
2358 // If this doesn't map cleanly onto the alloca type, and that type isn't
2359 // a single value type, just emit a memset.
2360 if (!VecTy && (BeginOffset != NewAllocaBeginOffset ||
2361 EndOffset != NewAllocaEndOffset ||
2362 !AllocaTy->isSingleValueType() ||
2363 !TD.isLegalInteger(TD.getTypeSizeInBits(ScalarTy)))) {
2364 Type *SizeTy = II.getLength()->getType();
2365 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2367 = IRB.CreateMemSet(getAdjustedAllocaPtr(IRB,
2368 II.getRawDest()->getType()),
2369 II.getValue(), Size, getAdjustedAlign(),
2372 DEBUG(dbgs() << " to: " << *New << "\n");
2376 // If we can represent this as a simple value, we have to build the actual
2377 // value to store, which requires expanding the byte present in memset to
2378 // a sensible representation for the alloca type. This is essentially
2379 // splatting the byte to a sufficiently wide integer, bitcasting to the
2380 // desired scalar type, and splatting it across any desired vector type.
2381 Value *V = II.getValue();
2382 IntegerType *VTy = cast<IntegerType>(V->getType());
2383 Type *IntTy = Type::getIntNTy(VTy->getContext(),
2384 TD.getTypeSizeInBits(ScalarTy));
2385 if (TD.getTypeSizeInBits(ScalarTy) > VTy->getBitWidth())
2386 V = IRB.CreateMul(IRB.CreateZExt(V, IntTy, getName(".zext")),
2387 ConstantExpr::getUDiv(
2388 Constant::getAllOnesValue(IntTy),
2389 ConstantExpr::getZExt(
2390 Constant::getAllOnesValue(V->getType()),
2392 getName(".isplat"));
2393 if (V->getType() != ScalarTy) {
2394 if (ScalarTy->isPointerTy())
2395 V = IRB.CreateIntToPtr(V, ScalarTy);
2396 else if (ScalarTy->isPrimitiveType() || ScalarTy->isVectorTy())
2397 V = IRB.CreateBitCast(V, ScalarTy);
2398 else if (ScalarTy->isIntegerTy())
2399 llvm_unreachable("Computed different integer types with equal widths");
2401 llvm_unreachable("Invalid scalar type");
2404 // If this is an element-wide memset of a vectorizable alloca, insert it.
2405 if (VecTy && (BeginOffset > NewAllocaBeginOffset ||
2406 EndOffset < NewAllocaEndOffset)) {
2407 StoreInst *Store = IRB.CreateAlignedStore(
2408 IRB.CreateInsertElement(IRB.CreateAlignedLoad(&NewAI,
2409 NewAI.getAlignment(),
2411 V, getIndex(IRB, BeginOffset),
2412 getName(".insert")),
2413 &NewAI, NewAI.getAlignment());
2415 DEBUG(dbgs() << " to: " << *Store << "\n");
2419 // Splat to a vector if needed.
2420 if (VectorType *VecTy = dyn_cast<VectorType>(AllocaTy)) {
2421 VectorType *SplatSourceTy = VectorType::get(V->getType(), 1);
2422 V = IRB.CreateShuffleVector(
2423 IRB.CreateInsertElement(UndefValue::get(SplatSourceTy), V,
2424 IRB.getInt32(0), getName(".vsplat.insert")),
2425 UndefValue::get(SplatSourceTy),
2426 ConstantVector::getSplat(VecTy->getNumElements(), IRB.getInt32(0)),
2427 getName(".vsplat.shuffle"));
2428 assert(V->getType() == VecTy);
2431 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2434 DEBUG(dbgs() << " to: " << *New << "\n");
2435 return !II.isVolatile();
2438 bool visitMemTransferInst(MemTransferInst &II) {
2439 // Rewriting of memory transfer instructions can be a bit tricky. We break
2440 // them into two categories: split intrinsics and unsplit intrinsics.
2442 DEBUG(dbgs() << " original: " << II << "\n");
2443 IRBuilder<> IRB(&II);
2445 assert(II.getRawSource() == OldPtr || II.getRawDest() == OldPtr);
2446 bool IsDest = II.getRawDest() == OldPtr;
2448 const AllocaPartitioning::MemTransferOffsets &MTO
2449 = P.getMemTransferOffsets(II);
2451 // Compute the relative offset within the transfer.
2452 unsigned IntPtrWidth = TD.getPointerSizeInBits();
2453 APInt RelOffset(IntPtrWidth, BeginOffset - (IsDest ? MTO.DestBegin
2454 : MTO.SourceBegin));
2456 unsigned Align = II.getAlignment();
2458 Align = MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(),
2459 MinAlign(II.getAlignment(), getAdjustedAlign()));
2461 // For unsplit intrinsics, we simply modify the source and destination
2462 // pointers in place. This isn't just an optimization, it is a matter of
2463 // correctness. With unsplit intrinsics we may be dealing with transfers
2464 // within a single alloca before SROA ran, or with transfers that have
2465 // a variable length. We may also be dealing with memmove instead of
2466 // memcpy, and so simply updating the pointers is the necessary for us to
2467 // update both source and dest of a single call.
2468 if (!MTO.IsSplittable) {
2469 Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource();
2471 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2473 II.setSource(getAdjustedAllocaPtr(IRB, II.getRawSource()->getType()));
2475 Type *CstTy = II.getAlignmentCst()->getType();
2476 II.setAlignment(ConstantInt::get(CstTy, Align));
2478 DEBUG(dbgs() << " to: " << II << "\n");
2479 deleteIfTriviallyDead(OldOp);
2482 // For split transfer intrinsics we have an incredibly useful assurance:
2483 // the source and destination do not reside within the same alloca, and at
2484 // least one of them does not escape. This means that we can replace
2485 // memmove with memcpy, and we don't need to worry about all manner of
2486 // downsides to splitting and transforming the operations.
2488 // If this doesn't map cleanly onto the alloca type, and that type isn't
2489 // a single value type, just emit a memcpy.
2491 = !VecTy && (BeginOffset != NewAllocaBeginOffset ||
2492 EndOffset != NewAllocaEndOffset ||
2493 !NewAI.getAllocatedType()->isSingleValueType());
2495 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2496 // size hasn't been shrunk based on analysis of the viable range, this is
2498 if (EmitMemCpy && &OldAI == &NewAI) {
2499 uint64_t OrigBegin = IsDest ? MTO.DestBegin : MTO.SourceBegin;
2500 uint64_t OrigEnd = IsDest ? MTO.DestEnd : MTO.SourceEnd;
2501 // Ensure the start lines up.
2502 assert(BeginOffset == OrigBegin);
2505 // Rewrite the size as needed.
2506 if (EndOffset != OrigEnd)
2507 II.setLength(ConstantInt::get(II.getLength()->getType(),
2508 EndOffset - BeginOffset));
2511 // Record this instruction for deletion.
2512 if (Pass.DeadSplitInsts.insert(&II))
2513 Pass.DeadInsts.push_back(&II);
2515 bool IsVectorElement = VecTy && (BeginOffset > NewAllocaBeginOffset ||
2516 EndOffset < NewAllocaEndOffset);
2518 Type *OtherPtrTy = IsDest ? II.getRawSource()->getType()
2519 : II.getRawDest()->getType();
2521 OtherPtrTy = IsVectorElement ? VecTy->getElementType()->getPointerTo()
2524 // Compute the other pointer, folding as much as possible to produce
2525 // a single, simple GEP in most cases.
2526 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2527 OtherPtr = getAdjustedPtr(IRB, TD, OtherPtr, RelOffset, OtherPtrTy,
2528 getName("." + OtherPtr->getName()));
2530 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2531 // alloca that should be re-examined after rewriting this instruction.
2533 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets()))
2534 Pass.Worklist.insert(AI);
2538 = getAdjustedAllocaPtr(IRB, IsDest ? II.getRawDest()->getType()
2539 : II.getRawSource()->getType());
2540 Type *SizeTy = II.getLength()->getType();
2541 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2543 CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
2544 IsDest ? OtherPtr : OurPtr,
2545 Size, Align, II.isVolatile());
2547 DEBUG(dbgs() << " to: " << *New << "\n");
2551 Value *SrcPtr = OtherPtr;
2552 Value *DstPtr = &NewAI;
2554 std::swap(SrcPtr, DstPtr);
2557 if (IsVectorElement && !IsDest) {
2558 // We have to extract rather than load.
2559 Src = IRB.CreateExtractElement(
2560 IRB.CreateAlignedLoad(SrcPtr, Align, getName(".copyload")),
2561 getIndex(IRB, BeginOffset),
2562 getName(".copyextract"));
2564 Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(),
2565 getName(".copyload"));
2568 if (IsVectorElement && IsDest) {
2569 // We have to insert into a loaded copy before storing.
2570 Src = IRB.CreateInsertElement(
2571 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), getName(".load")),
2572 Src, getIndex(IRB, BeginOffset),
2573 getName(".insert"));
2576 StoreInst *Store = cast<StoreInst>(
2577 IRB.CreateAlignedStore(Src, DstPtr, Align, II.isVolatile()));
2579 DEBUG(dbgs() << " to: " << *Store << "\n");
2580 return !II.isVolatile();
2583 bool visitIntrinsicInst(IntrinsicInst &II) {
2584 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2585 II.getIntrinsicID() == Intrinsic::lifetime_end);
2586 DEBUG(dbgs() << " original: " << II << "\n");
2587 IRBuilder<> IRB(&II);
2588 assert(II.getArgOperand(1) == OldPtr);
2590 // Record this instruction for deletion.
2591 if (Pass.DeadSplitInsts.insert(&II))
2592 Pass.DeadInsts.push_back(&II);
2595 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2596 EndOffset - BeginOffset);
2597 Value *Ptr = getAdjustedAllocaPtr(IRB, II.getArgOperand(1)->getType());
2599 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2600 New = IRB.CreateLifetimeStart(Ptr, Size);
2602 New = IRB.CreateLifetimeEnd(Ptr, Size);
2604 DEBUG(dbgs() << " to: " << *New << "\n");
2608 bool visitPHINode(PHINode &PN) {
2609 DEBUG(dbgs() << " original: " << PN << "\n");
2611 // We would like to compute a new pointer in only one place, but have it be
2612 // as local as possible to the PHI. To do that, we re-use the location of
2613 // the old pointer, which necessarily must be in the right position to
2614 // dominate the PHI.
2615 IRBuilder<> PtrBuilder(cast<Instruction>(OldPtr));
2617 Value *NewPtr = getAdjustedAllocaPtr(PtrBuilder, OldPtr->getType());
2618 // Replace the operands which were using the old pointer.
2619 User::op_iterator OI = PN.op_begin(), OE = PN.op_end();
2620 for (; OI != OE; ++OI)
2624 DEBUG(dbgs() << " to: " << PN << "\n");
2625 deleteIfTriviallyDead(OldPtr);
2629 bool visitSelectInst(SelectInst &SI) {
2630 DEBUG(dbgs() << " original: " << SI << "\n");
2631 IRBuilder<> IRB(&SI);
2633 // Find the operand we need to rewrite here.
2634 bool IsTrueVal = SI.getTrueValue() == OldPtr;
2636 assert(SI.getFalseValue() != OldPtr && "Pointer is both operands!");
2638 assert(SI.getFalseValue() == OldPtr && "Pointer isn't an operand!");
2640 Value *NewPtr = getAdjustedAllocaPtr(IRB, OldPtr->getType());
2641 SI.setOperand(IsTrueVal ? 1 : 2, NewPtr);
2642 DEBUG(dbgs() << " to: " << SI << "\n");
2643 deleteIfTriviallyDead(OldPtr);
2651 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2653 /// This pass aggressively rewrites all aggregate loads and stores on
2654 /// a particular pointer (or any pointer derived from it which we can identify)
2655 /// with scalar loads and stores.
2656 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2657 // Befriend the base class so it can delegate to private visit methods.
2658 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2660 const TargetData &TD;
2662 /// Queue of pointer uses to analyze and potentially rewrite.
2663 SmallVector<Use *, 8> Queue;
2665 /// Set to prevent us from cycling with phi nodes and loops.
2666 SmallPtrSet<User *, 8> Visited;
2668 /// The current pointer use being rewritten. This is used to dig up the used
2669 /// value (as opposed to the user).
2673 AggLoadStoreRewriter(const TargetData &TD) : TD(TD) {}
2675 /// Rewrite loads and stores through a pointer and all pointers derived from
2677 bool rewrite(Instruction &I) {
2678 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2680 bool Changed = false;
2681 while (!Queue.empty()) {
2682 U = Queue.pop_back_val();
2683 Changed |= visit(cast<Instruction>(U->getUser()));
2689 /// Enqueue all the users of the given instruction for further processing.
2690 /// This uses a set to de-duplicate users.
2691 void enqueueUsers(Instruction &I) {
2692 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
2694 if (Visited.insert(*UI))
2695 Queue.push_back(&UI.getUse());
2698 // Conservative default is to not rewrite anything.
2699 bool visitInstruction(Instruction &I) { return false; }
2701 /// \brief Generic recursive split emission class.
2702 template <typename Derived>
2705 /// The builder used to form new instructions.
2707 /// The indices which to be used with insert- or extractvalue to select the
2708 /// appropriate value within the aggregate.
2709 SmallVector<unsigned, 4> Indices;
2710 /// The indices to a GEP instruction which will move Ptr to the correct slot
2711 /// within the aggregate.
2712 SmallVector<Value *, 4> GEPIndices;
2713 /// The base pointer of the original op, used as a base for GEPing the
2714 /// split operations.
2717 /// Initialize the splitter with an insertion point, Ptr and start with a
2718 /// single zero GEP index.
2719 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
2720 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
2723 /// \brief Generic recursive split emission routine.
2725 /// This method recursively splits an aggregate op (load or store) into
2726 /// scalar or vector ops. It splits recursively until it hits a single value
2727 /// and emits that single value operation via the template argument.
2729 /// The logic of this routine relies on GEPs and insertvalue and
2730 /// extractvalue all operating with the same fundamental index list, merely
2731 /// formatted differently (GEPs need actual values).
2733 /// \param Ty The type being split recursively into smaller ops.
2734 /// \param Agg The aggregate value being built up or stored, depending on
2735 /// whether this is splitting a load or a store respectively.
2736 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
2737 if (Ty->isSingleValueType())
2738 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
2740 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2741 unsigned OldSize = Indices.size();
2743 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
2745 assert(Indices.size() == OldSize && "Did not return to the old size");
2746 Indices.push_back(Idx);
2747 GEPIndices.push_back(IRB.getInt32(Idx));
2748 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
2749 GEPIndices.pop_back();
2755 if (StructType *STy = dyn_cast<StructType>(Ty)) {
2756 unsigned OldSize = Indices.size();
2758 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
2760 assert(Indices.size() == OldSize && "Did not return to the old size");
2761 Indices.push_back(Idx);
2762 GEPIndices.push_back(IRB.getInt32(Idx));
2763 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
2764 GEPIndices.pop_back();
2770 llvm_unreachable("Only arrays and structs are aggregate loadable types");
2774 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
2775 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2776 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
2778 /// Emit a leaf load of a single value. This is called at the leaves of the
2779 /// recursive emission to actually load values.
2780 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2781 assert(Ty->isSingleValueType());
2782 // Load the single value and insert it using the indices.
2783 Value *Load = IRB.CreateLoad(IRB.CreateInBoundsGEP(Ptr, GEPIndices,
2786 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
2787 DEBUG(dbgs() << " to: " << *Load << "\n");
2791 bool visitLoadInst(LoadInst &LI) {
2792 assert(LI.getPointerOperand() == *U);
2793 if (!LI.isSimple() || LI.getType()->isSingleValueType())
2796 // We have an aggregate being loaded, split it apart.
2797 DEBUG(dbgs() << " original: " << LI << "\n");
2798 LoadOpSplitter Splitter(&LI, *U);
2799 Value *V = UndefValue::get(LI.getType());
2800 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
2801 LI.replaceAllUsesWith(V);
2802 LI.eraseFromParent();
2806 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
2807 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2808 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
2810 /// Emit a leaf store of a single value. This is called at the leaves of the
2811 /// recursive emission to actually produce stores.
2812 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2813 assert(Ty->isSingleValueType());
2814 // Extract the single value and store it using the indices.
2815 Value *Store = IRB.CreateStore(
2816 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
2817 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
2819 DEBUG(dbgs() << " to: " << *Store << "\n");
2823 bool visitStoreInst(StoreInst &SI) {
2824 if (!SI.isSimple() || SI.getPointerOperand() != *U)
2826 Value *V = SI.getValueOperand();
2827 if (V->getType()->isSingleValueType())
2830 // We have an aggregate being stored, split it apart.
2831 DEBUG(dbgs() << " original: " << SI << "\n");
2832 StoreOpSplitter Splitter(&SI, *U);
2833 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
2834 SI.eraseFromParent();
2838 bool visitBitCastInst(BitCastInst &BC) {
2843 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
2848 bool visitPHINode(PHINode &PN) {
2853 bool visitSelectInst(SelectInst &SI) {
2860 /// \brief Try to find a partition of the aggregate type passed in for a given
2861 /// offset and size.
2863 /// This recurses through the aggregate type and tries to compute a subtype
2864 /// based on the offset and size. When the offset and size span a sub-section
2865 /// of an array, it will even compute a new array type for that sub-section,
2866 /// and the same for structs.
2868 /// Note that this routine is very strict and tries to find a partition of the
2869 /// type which produces the *exact* right offset and size. It is not forgiving
2870 /// when the size or offset cause either end of type-based partition to be off.
2871 /// Also, this is a best-effort routine. It is reasonable to give up and not
2872 /// return a type if necessary.
2873 static Type *getTypePartition(const TargetData &TD, Type *Ty,
2874 uint64_t Offset, uint64_t Size) {
2875 if (Offset == 0 && TD.getTypeAllocSize(Ty) == Size)
2878 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
2879 // We can't partition pointers...
2880 if (SeqTy->isPointerTy())
2883 Type *ElementTy = SeqTy->getElementType();
2884 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
2885 uint64_t NumSkippedElements = Offset / ElementSize;
2886 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy))
2887 if (NumSkippedElements >= ArrTy->getNumElements())
2889 if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy))
2890 if (NumSkippedElements >= VecTy->getNumElements())
2892 Offset -= NumSkippedElements * ElementSize;
2894 // First check if we need to recurse.
2895 if (Offset > 0 || Size < ElementSize) {
2896 // Bail if the partition ends in a different array element.
2897 if ((Offset + Size) > ElementSize)
2899 // Recurse through the element type trying to peel off offset bytes.
2900 return getTypePartition(TD, ElementTy, Offset, Size);
2902 assert(Offset == 0);
2904 if (Size == ElementSize)
2906 assert(Size > ElementSize);
2907 uint64_t NumElements = Size / ElementSize;
2908 if (NumElements * ElementSize != Size)
2910 return ArrayType::get(ElementTy, NumElements);
2913 StructType *STy = dyn_cast<StructType>(Ty);
2917 const StructLayout *SL = TD.getStructLayout(STy);
2918 if (Offset >= SL->getSizeInBytes())
2920 uint64_t EndOffset = Offset + Size;
2921 if (EndOffset > SL->getSizeInBytes())
2924 unsigned Index = SL->getElementContainingOffset(Offset);
2925 Offset -= SL->getElementOffset(Index);
2927 Type *ElementTy = STy->getElementType(Index);
2928 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
2929 if (Offset >= ElementSize)
2930 return 0; // The offset points into alignment padding.
2932 // See if any partition must be contained by the element.
2933 if (Offset > 0 || Size < ElementSize) {
2934 if ((Offset + Size) > ElementSize)
2936 return getTypePartition(TD, ElementTy, Offset, Size);
2938 assert(Offset == 0);
2940 if (Size == ElementSize)
2943 StructType::element_iterator EI = STy->element_begin() + Index,
2944 EE = STy->element_end();
2945 if (EndOffset < SL->getSizeInBytes()) {
2946 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
2947 if (Index == EndIndex)
2948 return 0; // Within a single element and its padding.
2950 // Don't try to form "natural" types if the elements don't line up with the
2952 // FIXME: We could potentially recurse down through the last element in the
2953 // sub-struct to find a natural end point.
2954 if (SL->getElementOffset(EndIndex) != EndOffset)
2957 assert(Index < EndIndex);
2958 EE = STy->element_begin() + EndIndex;
2961 // Try to build up a sub-structure.
2962 SmallVector<Type *, 4> ElementTys;
2964 ElementTys.push_back(*EI++);
2966 StructType *SubTy = StructType::get(STy->getContext(), ElementTys,
2968 const StructLayout *SubSL = TD.getStructLayout(SubTy);
2969 if (Size != SubSL->getSizeInBytes())
2970 return 0; // The sub-struct doesn't have quite the size needed.
2975 /// \brief Rewrite an alloca partition's users.
2977 /// This routine drives both of the rewriting goals of the SROA pass. It tries
2978 /// to rewrite uses of an alloca partition to be conducive for SSA value
2979 /// promotion. If the partition needs a new, more refined alloca, this will
2980 /// build that new alloca, preserving as much type information as possible, and
2981 /// rewrite the uses of the old alloca to point at the new one and have the
2982 /// appropriate new offsets. It also evaluates how successful the rewrite was
2983 /// at enabling promotion and if it was successful queues the alloca to be
2985 bool SROA::rewriteAllocaPartition(AllocaInst &AI,
2986 AllocaPartitioning &P,
2987 AllocaPartitioning::iterator PI) {
2988 uint64_t AllocaSize = PI->EndOffset - PI->BeginOffset;
2989 if (P.use_begin(PI) == P.use_end(PI))
2990 return false; // No live uses left of this partition.
2992 DEBUG(dbgs() << "Speculating PHIs and selects in partition "
2993 << "[" << PI->BeginOffset << "," << PI->EndOffset << ")\n");
2995 PHIOrSelectSpeculator Speculator(*TD, P, *this);
2996 DEBUG(dbgs() << " speculating ");
2997 DEBUG(P.print(dbgs(), PI, ""));
2998 Speculator.visitUsers(P.use_begin(PI), P.use_end(PI));
3000 // Try to compute a friendly type for this partition of the alloca. This
3001 // won't always succeed, in which case we fall back to a legal integer type
3002 // or an i8 array of an appropriate size.
3004 if (Type *PartitionTy = P.getCommonType(PI))
3005 if (TD->getTypeAllocSize(PartitionTy) >= AllocaSize)
3006 AllocaTy = PartitionTy;
3008 if (Type *PartitionTy = getTypePartition(*TD, AI.getAllocatedType(),
3009 PI->BeginOffset, AllocaSize))
3010 AllocaTy = PartitionTy;
3012 (AllocaTy->isArrayTy() &&
3013 AllocaTy->getArrayElementType()->isIntegerTy())) &&
3014 TD->isLegalInteger(AllocaSize * 8))
3015 AllocaTy = Type::getIntNTy(*C, AllocaSize * 8);
3017 AllocaTy = ArrayType::get(Type::getInt8Ty(*C), AllocaSize);
3018 assert(TD->getTypeAllocSize(AllocaTy) >= AllocaSize);
3020 // Check for the case where we're going to rewrite to a new alloca of the
3021 // exact same type as the original, and with the same access offsets. In that
3022 // case, re-use the existing alloca, but still run through the rewriter to
3023 // performe phi and select speculation.
3025 if (AllocaTy == AI.getAllocatedType()) {
3026 assert(PI->BeginOffset == 0 &&
3027 "Non-zero begin offset but same alloca type");
3028 assert(PI == P.begin() && "Begin offset is zero on later partition");
3031 unsigned Alignment = AI.getAlignment();
3033 // The minimum alignment which users can rely on when the explicit
3034 // alignment is omitted or zero is that required by the ABI for this
3036 Alignment = TD->getABITypeAlignment(AI.getAllocatedType());
3038 Alignment = MinAlign(Alignment, PI->BeginOffset);
3039 // If we will get at least this much alignment from the type alone, leave
3040 // the alloca's alignment unconstrained.
3041 if (Alignment <= TD->getABITypeAlignment(AllocaTy))
3043 NewAI = new AllocaInst(AllocaTy, 0, Alignment,
3044 AI.getName() + ".sroa." + Twine(PI - P.begin()),
3049 DEBUG(dbgs() << "Rewriting alloca partition "
3050 << "[" << PI->BeginOffset << "," << PI->EndOffset << ") to: "
3053 AllocaPartitionRewriter Rewriter(*TD, P, PI, *this, AI, *NewAI,
3054 PI->BeginOffset, PI->EndOffset);
3055 DEBUG(dbgs() << " rewriting ");
3056 DEBUG(P.print(dbgs(), PI, ""));
3057 if (Rewriter.visitUsers(P.use_begin(PI), P.use_end(PI))) {
3058 DEBUG(dbgs() << " and queuing for promotion\n");
3059 PromotableAllocas.push_back(NewAI);
3060 } else if (NewAI != &AI) {
3061 // If we can't promote the alloca, iterate on it to check for new
3062 // refinements exposed by splitting the current alloca. Don't iterate on an
3063 // alloca which didn't actually change and didn't get promoted.
3064 Worklist.insert(NewAI);
3069 /// \brief Walks the partitioning of an alloca rewriting uses of each partition.
3070 bool SROA::splitAlloca(AllocaInst &AI, AllocaPartitioning &P) {
3071 bool Changed = false;
3072 for (AllocaPartitioning::iterator PI = P.begin(), PE = P.end(); PI != PE;
3074 Changed |= rewriteAllocaPartition(AI, P, PI);
3079 /// \brief Analyze an alloca for SROA.
3081 /// This analyzes the alloca to ensure we can reason about it, builds
3082 /// a partitioning of the alloca, and then hands it off to be split and
3083 /// rewritten as needed.
3084 bool SROA::runOnAlloca(AllocaInst &AI) {
3085 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3086 ++NumAllocasAnalyzed;
3088 // Special case dead allocas, as they're trivial.
3089 if (AI.use_empty()) {
3090 AI.eraseFromParent();
3094 // Skip alloca forms that this analysis can't handle.
3095 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3096 TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
3099 // First check if this is a non-aggregate type that we should simply promote.
3100 if (!AI.getAllocatedType()->isAggregateType() && isAllocaPromotable(&AI)) {
3101 DEBUG(dbgs() << " Trivially scalar type, queuing for promotion...\n");
3102 PromotableAllocas.push_back(&AI);
3106 bool Changed = false;
3108 // First, split any FCA loads and stores touching this alloca to promote
3109 // better splitting and promotion opportunities.
3110 AggLoadStoreRewriter AggRewriter(*TD);
3111 Changed |= AggRewriter.rewrite(AI);
3113 // Build the partition set using a recursive instruction-visiting builder.
3114 AllocaPartitioning P(*TD, AI);
3115 DEBUG(P.print(dbgs()));
3119 // No partitions to split. Leave the dead alloca for a later pass to clean up.
3120 if (P.begin() == P.end())
3123 // Delete all the dead users of this alloca before splitting and rewriting it.
3124 for (AllocaPartitioning::dead_user_iterator DI = P.dead_user_begin(),
3125 DE = P.dead_user_end();
3128 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3129 DeadInsts.push_back(*DI);
3131 for (AllocaPartitioning::dead_op_iterator DO = P.dead_op_begin(),
3132 DE = P.dead_op_end();
3135 // Clobber the use with an undef value.
3136 **DO = UndefValue::get(OldV->getType());
3137 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3138 if (isInstructionTriviallyDead(OldI)) {
3140 DeadInsts.push_back(OldI);
3144 return splitAlloca(AI, P) || Changed;
3147 /// \brief Delete the dead instructions accumulated in this run.
3149 /// Recursively deletes the dead instructions we've accumulated. This is done
3150 /// at the very end to maximize locality of the recursive delete and to
3151 /// minimize the problems of invalidated instruction pointers as such pointers
3152 /// are used heavily in the intermediate stages of the algorithm.
3154 /// We also record the alloca instructions deleted here so that they aren't
3155 /// subsequently handed to mem2reg to promote.
3156 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3157 DeadSplitInsts.clear();
3158 while (!DeadInsts.empty()) {
3159 Instruction *I = DeadInsts.pop_back_val();
3160 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3162 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
3163 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
3164 // Zero out the operand and see if it becomes trivially dead.
3166 if (isInstructionTriviallyDead(U))
3167 DeadInsts.push_back(U);
3170 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3171 DeletedAllocas.insert(AI);
3174 I->eraseFromParent();
3178 /// \brief Promote the allocas, using the best available technique.
3180 /// This attempts to promote whatever allocas have been identified as viable in
3181 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3182 /// If there is a domtree available, we attempt to promote using the full power
3183 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3184 /// based on the SSAUpdater utilities. This function returns whether any
3185 /// promotion occured.
3186 bool SROA::promoteAllocas(Function &F) {
3187 if (PromotableAllocas.empty())
3190 NumPromoted += PromotableAllocas.size();
3192 if (DT && !ForceSSAUpdater) {
3193 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3194 PromoteMemToReg(PromotableAllocas, *DT);
3195 PromotableAllocas.clear();
3199 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3201 DIBuilder DIB(*F.getParent());
3202 SmallVector<Instruction*, 64> Insts;
3204 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3205 AllocaInst *AI = PromotableAllocas[Idx];
3206 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
3208 Instruction *I = cast<Instruction>(*UI++);
3209 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3210 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3211 // leading to them) here. Eventually it should use them to optimize the
3212 // scalar values produced.
3213 if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
3214 assert(onlyUsedByLifetimeMarkers(I) &&
3215 "Found a bitcast used outside of a lifetime marker.");
3216 while (!I->use_empty())
3217 cast<Instruction>(*I->use_begin())->eraseFromParent();
3218 I->eraseFromParent();
3221 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3222 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3223 II->getIntrinsicID() == Intrinsic::lifetime_end);
3224 II->eraseFromParent();
3230 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3234 PromotableAllocas.clear();
3239 /// \brief A predicate to test whether an alloca belongs to a set.
3240 class IsAllocaInSet {
3241 typedef SmallPtrSet<AllocaInst *, 4> SetType;
3245 IsAllocaInSet(const SetType &Set) : Set(Set) {}
3246 bool operator()(AllocaInst *AI) { return Set.count(AI); }
3250 bool SROA::runOnFunction(Function &F) {
3251 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3252 C = &F.getContext();
3253 TD = getAnalysisIfAvailable<TargetData>();
3255 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3258 DT = getAnalysisIfAvailable<DominatorTree>();
3260 BasicBlock &EntryBB = F.getEntryBlock();
3261 for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end());
3263 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3264 Worklist.insert(AI);
3266 bool Changed = false;
3267 // A set of deleted alloca instruction pointers which should be removed from
3268 // the list of promotable allocas.
3269 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3271 while (!Worklist.empty()) {
3272 Changed |= runOnAlloca(*Worklist.pop_back_val());
3273 deleteDeadInstructions(DeletedAllocas);
3274 if (!DeletedAllocas.empty()) {
3275 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3276 PromotableAllocas.end(),
3277 IsAllocaInSet(DeletedAllocas)),
3278 PromotableAllocas.end());
3279 DeletedAllocas.clear();
3283 Changed |= promoteAllocas(F);
3288 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3289 if (RequiresDomTree)
3290 AU.addRequired<DominatorTree>();
3291 AU.setPreservesCFG();