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 user of this range of the alloca.
159 AssertingVH<Instruction> User;
161 /// \brief The particular pointer value derived from this alloca in use.
162 AssertingVH<Instruction> Ptr;
164 PartitionUse() : ByteRange(), User(), Ptr() {}
165 PartitionUse(uint64_t BeginOffset, uint64_t EndOffset,
166 Instruction *User, Instruction *Ptr)
167 : ByteRange(BeginOffset, EndOffset), User(User), Ptr(Ptr) {}
170 /// \brief Construct a partitioning of a particular alloca.
172 /// Construction does most of the work for partitioning the alloca. This
173 /// performs the necessary walks of users and builds a partitioning from it.
174 AllocaPartitioning(const TargetData &TD, AllocaInst &AI);
176 /// \brief Test whether a pointer to the allocation escapes our analysis.
178 /// If this is true, the partitioning is never fully built and should be
180 bool isEscaped() const { return PointerEscapingInstr; }
182 /// \brief Support for iterating over the partitions.
184 typedef SmallVectorImpl<Partition>::iterator iterator;
185 iterator begin() { return Partitions.begin(); }
186 iterator end() { return Partitions.end(); }
188 typedef SmallVectorImpl<Partition>::const_iterator const_iterator;
189 const_iterator begin() const { return Partitions.begin(); }
190 const_iterator end() const { return Partitions.end(); }
193 /// \brief Support for iterating over and manipulating a particular
194 /// partition's uses.
196 /// The iteration support provided for uses is more limited, but also
197 /// includes some manipulation routines to support rewriting the uses of
198 /// partitions during SROA.
200 typedef SmallVectorImpl<PartitionUse>::iterator use_iterator;
201 use_iterator use_begin(unsigned Idx) { return Uses[Idx].begin(); }
202 use_iterator use_begin(const_iterator I) { return Uses[I - begin()].begin(); }
203 use_iterator use_end(unsigned Idx) { return Uses[Idx].end(); }
204 use_iterator use_end(const_iterator I) { return Uses[I - begin()].end(); }
205 void use_push_back(unsigned Idx, const PartitionUse &U) {
206 Uses[Idx].push_back(U);
208 void use_push_back(const_iterator I, const PartitionUse &U) {
209 Uses[I - begin()].push_back(U);
211 void use_erase(unsigned Idx, use_iterator UI) { Uses[Idx].erase(UI); }
212 void use_erase(const_iterator I, use_iterator UI) {
213 Uses[I - begin()].erase(UI);
216 typedef SmallVectorImpl<PartitionUse>::const_iterator const_use_iterator;
217 const_use_iterator use_begin(unsigned Idx) const { return Uses[Idx].begin(); }
218 const_use_iterator use_begin(const_iterator I) const {
219 return Uses[I - begin()].begin();
221 const_use_iterator use_end(unsigned Idx) const { return Uses[Idx].end(); }
222 const_use_iterator use_end(const_iterator I) const {
223 return Uses[I - begin()].end();
227 /// \brief Allow iterating the dead users for this alloca.
229 /// These are instructions which will never actually use the alloca as they
230 /// are outside the allocated range. They are safe to replace with undef and
233 typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator;
234 dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); }
235 dead_user_iterator dead_user_end() const { return DeadUsers.end(); }
238 /// \brief Allow iterating the dead expressions referring to this alloca.
240 /// These are operands which have cannot actually be used to refer to the
241 /// alloca as they are outside its range and the user doesn't correct for
242 /// that. These mostly consist of PHI node inputs and the like which we just
243 /// need to replace with undef.
245 typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator;
246 dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); }
247 dead_op_iterator dead_op_end() const { return DeadOperands.end(); }
250 /// \brief MemTransferInst auxiliary data.
251 /// This struct provides some auxiliary data about memory transfer
252 /// intrinsics such as memcpy and memmove. These intrinsics can use two
253 /// different ranges within the same alloca, and provide other challenges to
254 /// correctly represent. We stash extra data to help us untangle this
255 /// after the partitioning is complete.
256 struct MemTransferOffsets {
257 uint64_t DestBegin, DestEnd;
258 uint64_t SourceBegin, SourceEnd;
261 MemTransferOffsets getMemTransferOffsets(MemTransferInst &II) const {
262 return MemTransferInstData.lookup(&II);
265 /// \brief Map from a PHI or select operand back to a partition.
267 /// When manipulating PHI nodes or selects, they can use more than one
268 /// partition of an alloca. We store a special mapping to allow finding the
269 /// partition referenced by each of these operands, if any.
270 iterator findPartitionForPHIOrSelectOperand(Instruction &I, Value *Op) {
271 SmallDenseMap<std::pair<Instruction *, Value *>,
272 std::pair<unsigned, unsigned> >::const_iterator MapIt
273 = PHIOrSelectOpMap.find(std::make_pair(&I, Op));
274 if (MapIt == PHIOrSelectOpMap.end())
277 return begin() + MapIt->second.first;
280 /// \brief Map from a PHI or select operand back to the specific use of
283 /// Similar to mapping these operands back to the partitions, this maps
284 /// directly to the use structure of that partition.
285 use_iterator findPartitionUseForPHIOrSelectOperand(Instruction &I,
287 SmallDenseMap<std::pair<Instruction *, Value *>,
288 std::pair<unsigned, unsigned> >::const_iterator MapIt
289 = PHIOrSelectOpMap.find(std::make_pair(&I, Op));
290 assert(MapIt != PHIOrSelectOpMap.end());
291 return Uses[MapIt->second.first].begin() + MapIt->second.second;
294 /// \brief Compute a common type among the uses of a particular partition.
296 /// This routines walks all of the uses of a particular partition and tries
297 /// to find a common type between them. Untyped operations such as memset and
298 /// memcpy are ignored.
299 Type *getCommonType(iterator I) const;
301 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
302 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
303 void printUsers(raw_ostream &OS, const_iterator I,
304 StringRef Indent = " ") const;
305 void print(raw_ostream &OS) const;
306 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump(const_iterator I) const;
307 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump() const;
311 template <typename DerivedT, typename RetT = void> class BuilderBase;
312 class PartitionBuilder;
313 friend class AllocaPartitioning::PartitionBuilder;
315 friend class AllocaPartitioning::UseBuilder;
318 /// \brief Handle to alloca instruction to simplify method interfaces.
322 /// \brief The instruction responsible for this alloca having no partitioning.
324 /// When an instruction (potentially) escapes the pointer to the alloca, we
325 /// store a pointer to that here and abort trying to partition the alloca.
326 /// This will be null if the alloca is partitioned successfully.
327 Instruction *PointerEscapingInstr;
329 /// \brief The partitions of the alloca.
331 /// We store a vector of the partitions over the alloca here. This vector is
332 /// sorted by increasing begin offset, and then by decreasing end offset. See
333 /// the Partition inner class for more details. Initially (during
334 /// construction) there are overlaps, but we form a disjoint sequence of
335 /// partitions while finishing construction and a fully constructed object is
336 /// expected to always have this as a disjoint space.
337 SmallVector<Partition, 8> Partitions;
339 /// \brief The uses of the partitions.
341 /// This is essentially a mapping from each partition to a list of uses of
342 /// that partition. The mapping is done with a Uses vector that has the exact
343 /// same number of entries as the partition vector. Each entry is itself
344 /// a vector of the uses.
345 SmallVector<SmallVector<PartitionUse, 2>, 8> Uses;
347 /// \brief Instructions which will become dead if we rewrite the alloca.
349 /// Note that these are not separated by partition. This is because we expect
350 /// a partitioned alloca to be completely rewritten or not rewritten at all.
351 /// If rewritten, all these instructions can simply be removed and replaced
352 /// with undef as they come from outside of the allocated space.
353 SmallVector<Instruction *, 8> DeadUsers;
355 /// \brief Operands which will become dead if we rewrite the alloca.
357 /// These are operands that in their particular use can be replaced with
358 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
359 /// to PHI nodes and the like. They aren't entirely dead (there might be
360 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
361 /// want to swap this particular input for undef to simplify the use lists of
363 SmallVector<Use *, 8> DeadOperands;
365 /// \brief The underlying storage for auxiliary memcpy and memset info.
366 SmallDenseMap<MemTransferInst *, MemTransferOffsets, 4> MemTransferInstData;
368 /// \brief A side datastructure used when building up the partitions and uses.
370 /// This mapping is only really used during the initial building of the
371 /// partitioning so that we can retain information about PHI and select nodes
373 SmallDenseMap<Instruction *, std::pair<uint64_t, bool> > PHIOrSelectSizes;
375 /// \brief Auxiliary information for particular PHI or select operands.
376 SmallDenseMap<std::pair<Instruction *, Value *>,
377 std::pair<unsigned, unsigned>, 4> PHIOrSelectOpMap;
379 /// \brief A utility routine called from the constructor.
381 /// This does what it says on the tin. It is the key of the alloca partition
382 /// splitting and merging. After it is called we have the desired disjoint
383 /// collection of partitions.
384 void splitAndMergePartitions();
388 template <typename DerivedT, typename RetT>
389 class AllocaPartitioning::BuilderBase
390 : public InstVisitor<DerivedT, RetT> {
392 BuilderBase(const TargetData &TD, AllocaInst &AI, AllocaPartitioning &P)
394 AllocSize(TD.getTypeAllocSize(AI.getAllocatedType())),
400 const TargetData &TD;
401 const uint64_t AllocSize;
402 AllocaPartitioning &P;
408 SmallVector<OffsetUse, 8> Queue;
410 // The active offset and use while visiting.
414 void enqueueUsers(Instruction &I, int64_t UserOffset) {
415 SmallPtrSet<User *, 8> UserSet;
416 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
418 if (!UserSet.insert(*UI))
421 OffsetUse OU = { &UI.getUse(), UserOffset };
426 bool computeConstantGEPOffset(GetElementPtrInst &GEPI, int64_t &GEPOffset) {
428 for (gep_type_iterator GTI = gep_type_begin(GEPI), GTE = gep_type_end(GEPI);
430 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
436 // Handle a struct index, which adds its field offset to the pointer.
437 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
438 unsigned ElementIdx = OpC->getZExtValue();
439 const StructLayout *SL = TD.getStructLayout(STy);
440 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
441 // Check that we can continue to model this GEP in a signed 64-bit offset.
442 if (ElementOffset > INT64_MAX ||
444 ((uint64_t)GEPOffset + ElementOffset) > INT64_MAX)) {
445 DEBUG(dbgs() << "WARNING: Encountered a cumulative offset exceeding "
446 << "what can be represented in an int64_t!\n"
447 << " alloca: " << P.AI << "\n");
451 GEPOffset = ElementOffset + (uint64_t)-GEPOffset;
453 GEPOffset += ElementOffset;
457 APInt Index = OpC->getValue().sextOrTrunc(TD.getPointerSizeInBits());
458 Index *= APInt(Index.getBitWidth(),
459 TD.getTypeAllocSize(GTI.getIndexedType()));
460 Index += APInt(Index.getBitWidth(), (uint64_t)GEPOffset,
462 // Check if the result can be stored in our int64_t offset.
463 if (!Index.isSignedIntN(sizeof(GEPOffset) * 8)) {
464 DEBUG(dbgs() << "WARNING: Encountered a cumulative offset exceeding "
465 << "what can be represented in an int64_t!\n"
466 << " alloca: " << P.AI << "\n");
470 GEPOffset = Index.getSExtValue();
475 Value *foldSelectInst(SelectInst &SI) {
476 // If the condition being selected on is a constant or the same value is
477 // being selected between, fold the select. Yes this does (rarely) happen
479 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
480 return SI.getOperand(1+CI->isZero());
481 if (SI.getOperand(1) == SI.getOperand(2)) {
482 assert(*U == SI.getOperand(1));
483 return SI.getOperand(1);
489 /// \brief Builder for the alloca partitioning.
491 /// This class builds an alloca partitioning by recursively visiting the uses
492 /// of an alloca and splitting the partitions for each load and store at each
494 class AllocaPartitioning::PartitionBuilder
495 : public BuilderBase<PartitionBuilder, bool> {
496 friend class InstVisitor<PartitionBuilder, bool>;
498 SmallDenseMap<Instruction *, unsigned> MemTransferPartitionMap;
501 PartitionBuilder(const TargetData &TD, AllocaInst &AI, AllocaPartitioning &P)
502 : BuilderBase<PartitionBuilder, bool>(TD, AI, P) {}
504 /// \brief Run the builder over the allocation.
506 // Note that we have to re-evaluate size on each trip through the loop as
507 // the queue grows at the tail.
508 for (unsigned Idx = 0; Idx < Queue.size(); ++Idx) {
510 Offset = Queue[Idx].Offset;
511 if (!visit(cast<Instruction>(U->getUser())))
518 bool markAsEscaping(Instruction &I) {
519 P.PointerEscapingInstr = &I;
523 void insertUse(Instruction &I, int64_t Offset, uint64_t Size,
524 bool IsSplittable = false) {
525 // Completely skip uses which have a zero size or don't overlap the
528 (Offset >= 0 && (uint64_t)Offset >= AllocSize) ||
529 (Offset < 0 && (uint64_t)-Offset >= Size)) {
530 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
531 << " which starts past the end of the " << AllocSize
533 << " alloca: " << P.AI << "\n"
534 << " use: " << I << "\n");
538 // Clamp the start to the beginning of the allocation.
540 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
541 << " to start at the beginning of the alloca:\n"
542 << " alloca: " << P.AI << "\n"
543 << " use: " << I << "\n");
544 Size -= (uint64_t)-Offset;
548 uint64_t BeginOffset = Offset, EndOffset = BeginOffset + Size;
550 // Clamp the end offset to the end of the allocation. Note that this is
551 // formulated to handle even the case where "BeginOffset + Size" overflows.
552 assert(AllocSize >= BeginOffset); // Established above.
553 if (Size > AllocSize - BeginOffset) {
554 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
555 << " to remain within the " << AllocSize << " byte alloca:\n"
556 << " alloca: " << P.AI << "\n"
557 << " use: " << I << "\n");
558 EndOffset = AllocSize;
561 // See if we can just add a user onto the last slot currently occupied.
562 if (!P.Partitions.empty() &&
563 P.Partitions.back().BeginOffset == BeginOffset &&
564 P.Partitions.back().EndOffset == EndOffset) {
565 P.Partitions.back().IsSplittable &= IsSplittable;
569 Partition New(BeginOffset, EndOffset, IsSplittable);
570 P.Partitions.push_back(New);
573 bool handleLoadOrStore(Type *Ty, Instruction &I, int64_t Offset) {
574 uint64_t Size = TD.getTypeStoreSize(Ty);
576 // If this memory access can be shown to *statically* extend outside the
577 // bounds of of the allocation, it's behavior is undefined, so simply
578 // ignore it. Note that this is more strict than the generic clamping
579 // behavior of insertUse. We also try to handle cases which might run the
581 // FIXME: We should instead consider the pointer to have escaped if this
582 // function is being instrumented for addressing bugs or race conditions.
583 if (Offset < 0 || (uint64_t)Offset >= AllocSize ||
584 Size > (AllocSize - (uint64_t)Offset)) {
585 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte "
586 << (isa<LoadInst>(I) ? "load" : "store") << " @" << Offset
587 << " which extends past the end of the " << AllocSize
589 << " alloca: " << P.AI << "\n"
590 << " use: " << I << "\n");
594 insertUse(I, Offset, Size);
598 bool visitBitCastInst(BitCastInst &BC) {
599 enqueueUsers(BC, Offset);
603 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
605 if (!computeConstantGEPOffset(GEPI, GEPOffset))
606 return markAsEscaping(GEPI);
608 enqueueUsers(GEPI, GEPOffset);
612 bool visitLoadInst(LoadInst &LI) {
613 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
614 "All simple FCA loads should have been pre-split");
615 return handleLoadOrStore(LI.getType(), LI, Offset);
618 bool visitStoreInst(StoreInst &SI) {
619 Value *ValOp = SI.getValueOperand();
621 return markAsEscaping(SI);
623 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
624 "All simple FCA stores should have been pre-split");
625 return handleLoadOrStore(ValOp->getType(), SI, Offset);
629 bool visitMemSetInst(MemSetInst &II) {
630 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
631 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
632 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
633 insertUse(II, Offset, Size, Length);
637 bool visitMemTransferInst(MemTransferInst &II) {
638 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
639 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
641 // Zero-length mem transfer intrinsics can be ignored entirely.
644 MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
646 // Only intrinsics with a constant length can be split.
647 Offsets.IsSplittable = Length;
649 if (*U != II.getRawDest()) {
650 assert(*U == II.getRawSource());
651 Offsets.SourceBegin = Offset;
652 Offsets.SourceEnd = Offset + Size;
654 Offsets.DestBegin = Offset;
655 Offsets.DestEnd = Offset + Size;
658 insertUse(II, Offset, Size, Offsets.IsSplittable);
659 unsigned NewIdx = P.Partitions.size() - 1;
661 SmallDenseMap<Instruction *, unsigned>::const_iterator PMI;
662 bool Inserted = false;
663 llvm::tie(PMI, Inserted)
664 = MemTransferPartitionMap.insert(std::make_pair(&II, NewIdx));
665 if (Offsets.IsSplittable &&
666 (!Inserted || II.getRawSource() == II.getRawDest())) {
667 // We've found a memory transfer intrinsic which refers to the alloca as
668 // both a source and dest. This is detected either by direct equality of
669 // the operand values, or when we visit the intrinsic twice due to two
670 // different chains of values leading to it. We refuse to split these to
671 // simplify splitting logic. If possible, SROA will still split them into
672 // separate allocas and then re-analyze.
673 Offsets.IsSplittable = false;
674 P.Partitions[PMI->second].IsSplittable = false;
675 P.Partitions[NewIdx].IsSplittable = false;
681 // Disable SRoA for any intrinsics except for lifetime invariants.
682 // FIXME: What about debug instrinsics? This matches old behavior, but
683 // doesn't make sense.
684 bool visitIntrinsicInst(IntrinsicInst &II) {
685 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
686 II.getIntrinsicID() == Intrinsic::lifetime_end) {
687 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
688 uint64_t Size = std::min(AllocSize - Offset, Length->getLimitedValue());
689 insertUse(II, Offset, Size, true);
693 return markAsEscaping(II);
696 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
697 // We consider any PHI or select that results in a direct load or store of
698 // the same offset to be a viable use for partitioning purposes. These uses
699 // are considered unsplittable and the size is the maximum loaded or stored
701 SmallPtrSet<Instruction *, 4> Visited;
702 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
703 Visited.insert(Root);
704 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
705 // If there are no loads or stores, the access is dead. We mark that as
706 // a size zero access.
709 Instruction *I, *UsedI;
710 llvm::tie(UsedI, I) = Uses.pop_back_val();
712 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
713 Size = std::max(Size, TD.getTypeStoreSize(LI->getType()));
716 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
717 Value *Op = SI->getOperand(0);
720 Size = std::max(Size, TD.getTypeStoreSize(Op->getType()));
724 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
725 if (!GEP->hasAllZeroIndices())
727 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
728 !isa<SelectInst>(I)) {
732 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
734 if (Visited.insert(cast<Instruction>(*UI)))
735 Uses.push_back(std::make_pair(I, cast<Instruction>(*UI)));
736 } while (!Uses.empty());
741 bool visitPHINode(PHINode &PN) {
742 // See if we already have computed info on this node.
743 std::pair<uint64_t, bool> &PHIInfo = P.PHIOrSelectSizes[&PN];
745 PHIInfo.second = true;
746 insertUse(PN, Offset, PHIInfo.first);
750 // Check for an unsafe use of the PHI node.
751 if (Instruction *EscapingI = hasUnsafePHIOrSelectUse(&PN, PHIInfo.first))
752 return markAsEscaping(*EscapingI);
754 insertUse(PN, Offset, PHIInfo.first);
758 bool visitSelectInst(SelectInst &SI) {
759 if (Value *Result = foldSelectInst(SI)) {
761 // If the result of the constant fold will be the pointer, recurse
762 // through the select as if we had RAUW'ed it.
763 enqueueUsers(SI, Offset);
768 // See if we already have computed info on this node.
769 std::pair<uint64_t, bool> &SelectInfo = P.PHIOrSelectSizes[&SI];
770 if (SelectInfo.first) {
771 SelectInfo.second = true;
772 insertUse(SI, Offset, SelectInfo.first);
776 // Check for an unsafe use of the PHI node.
777 if (Instruction *EscapingI = hasUnsafePHIOrSelectUse(&SI, SelectInfo.first))
778 return markAsEscaping(*EscapingI);
780 insertUse(SI, Offset, SelectInfo.first);
784 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
785 bool visitInstruction(Instruction &I) { return markAsEscaping(I); }
789 /// \brief Use adder for the alloca partitioning.
791 /// This class adds the uses of an alloca to all of the partitions which they
792 /// use. For splittable partitions, this can end up doing essentially a linear
793 /// walk of the partitions, but the number of steps remains bounded by the
794 /// total result instruction size:
795 /// - The number of partitions is a result of the number unsplittable
796 /// instructions using the alloca.
797 /// - The number of users of each partition is at worst the total number of
798 /// splittable instructions using the alloca.
799 /// Thus we will produce N * M instructions in the end, where N are the number
800 /// of unsplittable uses and M are the number of splittable. This visitor does
801 /// the exact same number of updates to the partitioning.
803 /// In the more common case, this visitor will leverage the fact that the
804 /// partition space is pre-sorted, and do a logarithmic search for the
805 /// partition needed, making the total visit a classical ((N + M) * log(N))
806 /// complexity operation.
807 class AllocaPartitioning::UseBuilder : public BuilderBase<UseBuilder> {
808 friend class InstVisitor<UseBuilder>;
810 /// \brief Set to de-duplicate dead instructions found in the use walk.
811 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
814 UseBuilder(const TargetData &TD, AllocaInst &AI, AllocaPartitioning &P)
815 : BuilderBase<UseBuilder>(TD, AI, P) {}
817 /// \brief Run the builder over the allocation.
819 // Note that we have to re-evaluate size on each trip through the loop as
820 // the queue grows at the tail.
821 for (unsigned Idx = 0; Idx < Queue.size(); ++Idx) {
823 Offset = Queue[Idx].Offset;
824 this->visit(cast<Instruction>(U->getUser()));
829 void markAsDead(Instruction &I) {
830 if (VisitedDeadInsts.insert(&I))
831 P.DeadUsers.push_back(&I);
834 void insertUse(Instruction &User, int64_t Offset, uint64_t Size) {
835 // If the use has a zero size or extends outside of the allocation, record
836 // it as a dead use for elimination later.
837 if (Size == 0 || (uint64_t)Offset >= AllocSize ||
838 (Offset < 0 && (uint64_t)-Offset >= Size))
839 return markAsDead(User);
841 // Clamp the start to the beginning of the allocation.
843 Size -= (uint64_t)-Offset;
847 uint64_t BeginOffset = Offset, EndOffset = BeginOffset + Size;
849 // Clamp the end offset to the end of the allocation. Note that this is
850 // formulated to handle even the case where "BeginOffset + Size" overflows.
851 assert(AllocSize >= BeginOffset); // Established above.
852 if (Size > AllocSize - BeginOffset)
853 EndOffset = AllocSize;
855 // NB: This only works if we have zero overlapping partitions.
856 iterator B = std::lower_bound(P.begin(), P.end(), BeginOffset);
857 if (B != P.begin() && llvm::prior(B)->EndOffset > BeginOffset)
859 for (iterator I = B, E = P.end(); I != E && I->BeginOffset < EndOffset;
861 PartitionUse NewUse(std::max(I->BeginOffset, BeginOffset),
862 std::min(I->EndOffset, EndOffset),
863 &User, cast<Instruction>(*U));
864 P.use_push_back(I, NewUse);
865 if (isa<PHINode>(U->getUser()) || isa<SelectInst>(U->getUser()))
866 P.PHIOrSelectOpMap[std::make_pair(&User, U->get())]
867 = std::make_pair(I - P.begin(), P.Uses[I - P.begin()].size() - 1);
871 void handleLoadOrStore(Type *Ty, Instruction &I, int64_t Offset) {
872 uint64_t Size = TD.getTypeStoreSize(Ty);
874 // If this memory access can be shown to *statically* extend outside the
875 // bounds of of the allocation, it's behavior is undefined, so simply
876 // ignore it. Note that this is more strict than the generic clamping
877 // behavior of insertUse.
878 if (Offset < 0 || (uint64_t)Offset >= AllocSize ||
879 Size > (AllocSize - (uint64_t)Offset))
880 return markAsDead(I);
882 insertUse(I, Offset, Size);
885 void visitBitCastInst(BitCastInst &BC) {
887 return markAsDead(BC);
889 enqueueUsers(BC, Offset);
892 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
893 if (GEPI.use_empty())
894 return markAsDead(GEPI);
897 if (!computeConstantGEPOffset(GEPI, GEPOffset))
898 llvm_unreachable("Unable to compute constant offset for use");
900 enqueueUsers(GEPI, GEPOffset);
903 void visitLoadInst(LoadInst &LI) {
904 handleLoadOrStore(LI.getType(), LI, Offset);
907 void visitStoreInst(StoreInst &SI) {
908 handleLoadOrStore(SI.getOperand(0)->getType(), SI, Offset);
911 void visitMemSetInst(MemSetInst &II) {
912 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
913 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
914 insertUse(II, Offset, Size);
917 void visitMemTransferInst(MemTransferInst &II) {
918 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
919 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
920 insertUse(II, Offset, Size);
923 void visitIntrinsicInst(IntrinsicInst &II) {
924 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
925 II.getIntrinsicID() == Intrinsic::lifetime_end);
927 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
928 insertUse(II, Offset,
929 std::min(AllocSize - Offset, Length->getLimitedValue()));
932 void insertPHIOrSelect(Instruction &User, uint64_t Offset) {
933 uint64_t Size = P.PHIOrSelectSizes.lookup(&User).first;
935 // For PHI and select operands outside the alloca, we can't nuke the entire
936 // phi or select -- the other side might still be relevant, so we special
937 // case them here and use a separate structure to track the operands
938 // themselves which should be replaced with undef.
939 if (Offset >= AllocSize) {
940 P.DeadOperands.push_back(U);
944 insertUse(User, Offset, Size);
946 void visitPHINode(PHINode &PN) {
948 return markAsDead(PN);
950 insertPHIOrSelect(PN, Offset);
952 void visitSelectInst(SelectInst &SI) {
954 return markAsDead(SI);
956 if (Value *Result = foldSelectInst(SI)) {
958 // If the result of the constant fold will be the pointer, recurse
959 // through the select as if we had RAUW'ed it.
960 enqueueUsers(SI, Offset);
962 // Otherwise the operand to the select is dead, and we can replace it
964 P.DeadOperands.push_back(U);
969 insertPHIOrSelect(SI, Offset);
972 /// \brief Unreachable, we've already visited the alloca once.
973 void visitInstruction(Instruction &I) {
974 llvm_unreachable("Unhandled instruction in use builder.");
978 void AllocaPartitioning::splitAndMergePartitions() {
979 size_t NumDeadPartitions = 0;
981 // Track the range of splittable partitions that we pass when accumulating
982 // overlapping unsplittable partitions.
983 uint64_t SplitEndOffset = 0ull;
985 Partition New(0ull, 0ull, false);
987 for (unsigned i = 0, j = i, e = Partitions.size(); i != e; i = j) {
990 if (!Partitions[i].IsSplittable || New.BeginOffset == New.EndOffset) {
991 assert(New.BeginOffset == New.EndOffset);
994 assert(New.IsSplittable);
995 New.EndOffset = std::max(New.EndOffset, Partitions[i].EndOffset);
997 assert(New.BeginOffset != New.EndOffset);
999 // Scan the overlapping partitions.
1000 while (j != e && New.EndOffset > Partitions[j].BeginOffset) {
1001 // If the new partition we are forming is splittable, stop at the first
1002 // unsplittable partition.
1003 if (New.IsSplittable && !Partitions[j].IsSplittable)
1006 // Grow the new partition to include any equally splittable range. 'j' is
1007 // always equally splittable when New is splittable, but when New is not
1008 // splittable, we may subsume some (or part of some) splitable partition
1009 // without growing the new one.
1010 if (New.IsSplittable == Partitions[j].IsSplittable) {
1011 New.EndOffset = std::max(New.EndOffset, Partitions[j].EndOffset);
1013 assert(!New.IsSplittable);
1014 assert(Partitions[j].IsSplittable);
1015 SplitEndOffset = std::max(SplitEndOffset, Partitions[j].EndOffset);
1018 Partitions[j].BeginOffset = Partitions[j].EndOffset = UINT64_MAX;
1019 ++NumDeadPartitions;
1023 // If the new partition is splittable, chop off the end as soon as the
1024 // unsplittable subsequent partition starts and ensure we eventually cover
1025 // the splittable area.
1026 if (j != e && New.IsSplittable) {
1027 SplitEndOffset = std::max(SplitEndOffset, New.EndOffset);
1028 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
1031 // Add the new partition if it differs from the original one and is
1032 // non-empty. We can end up with an empty partition here if it was
1033 // splittable but there is an unsplittable one that starts at the same
1035 if (New != Partitions[i]) {
1036 if (New.BeginOffset != New.EndOffset)
1037 Partitions.push_back(New);
1038 // Mark the old one for removal.
1039 Partitions[i].BeginOffset = Partitions[i].EndOffset = UINT64_MAX;
1040 ++NumDeadPartitions;
1043 New.BeginOffset = New.EndOffset;
1044 if (!New.IsSplittable) {
1045 New.EndOffset = std::max(New.EndOffset, SplitEndOffset);
1046 if (j != e && !Partitions[j].IsSplittable)
1047 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
1048 New.IsSplittable = true;
1049 // If there is a trailing splittable partition which won't be fused into
1050 // the next splittable partition go ahead and add it onto the partitions
1052 if (New.BeginOffset < New.EndOffset &&
1053 (j == e || !Partitions[j].IsSplittable ||
1054 New.EndOffset < Partitions[j].BeginOffset)) {
1055 Partitions.push_back(New);
1056 New.BeginOffset = New.EndOffset = 0ull;
1061 // Re-sort the partitions now that they have been split and merged into
1062 // disjoint set of partitions. Also remove any of the dead partitions we've
1063 // replaced in the process.
1064 std::sort(Partitions.begin(), Partitions.end());
1065 if (NumDeadPartitions) {
1066 assert(Partitions.back().BeginOffset == UINT64_MAX);
1067 assert(Partitions.back().EndOffset == UINT64_MAX);
1068 assert((ptrdiff_t)NumDeadPartitions ==
1069 std::count(Partitions.begin(), Partitions.end(), Partitions.back()));
1071 Partitions.erase(Partitions.end() - NumDeadPartitions, Partitions.end());
1074 AllocaPartitioning::AllocaPartitioning(const TargetData &TD, AllocaInst &AI)
1079 PointerEscapingInstr(0) {
1080 PartitionBuilder PB(TD, AI, *this);
1084 if (Partitions.size() > 1) {
1085 // Sort the uses. This arranges for the offsets to be in ascending order,
1086 // and the sizes to be in descending order.
1087 std::sort(Partitions.begin(), Partitions.end());
1089 // Intersect splittability for all partitions with equal offsets and sizes.
1090 // Then remove all but the first so that we have a sequence of non-equal but
1091 // potentially overlapping partitions.
1092 for (iterator I = Partitions.begin(), J = I, E = Partitions.end(); I != E;
1095 while (J != E && *I == *J) {
1096 I->IsSplittable &= J->IsSplittable;
1100 Partitions.erase(std::unique(Partitions.begin(), Partitions.end()),
1103 // Split splittable and merge unsplittable partitions into a disjoint set
1104 // of partitions over the used space of the allocation.
1105 splitAndMergePartitions();
1108 // Now build up the user lists for each of these disjoint partitions by
1109 // re-walking the recursive users of the alloca.
1110 Uses.resize(Partitions.size());
1111 UseBuilder UB(TD, AI, *this);
1115 Type *AllocaPartitioning::getCommonType(iterator I) const {
1117 for (const_use_iterator UI = use_begin(I), UE = use_end(I); UI != UE; ++UI) {
1118 if (isa<IntrinsicInst>(*UI->User))
1120 if (UI->BeginOffset != I->BeginOffset || UI->EndOffset != I->EndOffset)
1124 if (LoadInst *LI = dyn_cast<LoadInst>(&*UI->User)) {
1125 UserTy = LI->getType();
1126 } else if (StoreInst *SI = dyn_cast<StoreInst>(&*UI->User)) {
1127 UserTy = SI->getValueOperand()->getType();
1128 } else if (SelectInst *SI = dyn_cast<SelectInst>(&*UI->User)) {
1129 if (PointerType *PtrTy = dyn_cast<PointerType>(SI->getType()))
1130 UserTy = PtrTy->getElementType();
1131 } else if (PHINode *PN = dyn_cast<PHINode>(&*UI->User)) {
1132 if (PointerType *PtrTy = dyn_cast<PointerType>(PN->getType()))
1133 UserTy = PtrTy->getElementType();
1136 if (Ty && Ty != UserTy)
1144 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1146 void AllocaPartitioning::print(raw_ostream &OS, const_iterator I,
1147 StringRef Indent) const {
1148 OS << Indent << "partition #" << (I - begin())
1149 << " [" << I->BeginOffset << "," << I->EndOffset << ")"
1150 << (I->IsSplittable ? " (splittable)" : "")
1151 << (Uses[I - begin()].empty() ? " (zero uses)" : "")
1155 void AllocaPartitioning::printUsers(raw_ostream &OS, const_iterator I,
1156 StringRef Indent) const {
1157 for (const_use_iterator UI = use_begin(I), UE = use_end(I);
1159 OS << Indent << " [" << UI->BeginOffset << "," << UI->EndOffset << ") "
1160 << "used by: " << *UI->User << "\n";
1161 if (MemTransferInst *II = dyn_cast<MemTransferInst>(&*UI->User)) {
1162 const MemTransferOffsets &MTO = MemTransferInstData.lookup(II);
1164 if (!MTO.IsSplittable)
1165 IsDest = UI->BeginOffset == MTO.DestBegin;
1167 IsDest = MTO.DestBegin != 0u;
1168 OS << Indent << " (original " << (IsDest ? "dest" : "source") << ": "
1169 << "[" << (IsDest ? MTO.DestBegin : MTO.SourceBegin)
1170 << "," << (IsDest ? MTO.DestEnd : MTO.SourceEnd) << ")\n";
1175 void AllocaPartitioning::print(raw_ostream &OS) const {
1176 if (PointerEscapingInstr) {
1177 OS << "No partitioning for alloca: " << AI << "\n"
1178 << " A pointer to this alloca escaped by:\n"
1179 << " " << *PointerEscapingInstr << "\n";
1183 OS << "Partitioning of alloca: " << AI << "\n";
1185 for (const_iterator I = begin(), E = end(); I != E; ++I, ++Num) {
1191 void AllocaPartitioning::dump(const_iterator I) const { print(dbgs(), I); }
1192 void AllocaPartitioning::dump() const { print(dbgs()); }
1194 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1198 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
1200 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
1201 /// the loads and stores of an alloca instruction, as well as updating its
1202 /// debug information. This is used when a domtree is unavailable and thus
1203 /// mem2reg in its full form can't be used to handle promotion of allocas to
1205 class AllocaPromoter : public LoadAndStorePromoter {
1209 SmallVector<DbgDeclareInst *, 4> DDIs;
1210 SmallVector<DbgValueInst *, 4> DVIs;
1213 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
1214 AllocaInst &AI, DIBuilder &DIB)
1215 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
1217 void run(const SmallVectorImpl<Instruction*> &Insts) {
1218 // Remember which alloca we're promoting (for isInstInList).
1219 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
1220 for (Value::use_iterator UI = DebugNode->use_begin(),
1221 UE = DebugNode->use_end();
1223 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
1224 DDIs.push_back(DDI);
1225 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
1226 DVIs.push_back(DVI);
1229 LoadAndStorePromoter::run(Insts);
1230 AI.eraseFromParent();
1231 while (!DDIs.empty())
1232 DDIs.pop_back_val()->eraseFromParent();
1233 while (!DVIs.empty())
1234 DVIs.pop_back_val()->eraseFromParent();
1237 virtual bool isInstInList(Instruction *I,
1238 const SmallVectorImpl<Instruction*> &Insts) const {
1239 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1240 return LI->getOperand(0) == &AI;
1241 return cast<StoreInst>(I)->getPointerOperand() == &AI;
1244 virtual void updateDebugInfo(Instruction *Inst) const {
1245 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(),
1246 E = DDIs.end(); I != E; ++I) {
1247 DbgDeclareInst *DDI = *I;
1248 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
1249 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1250 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
1251 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1253 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(),
1254 E = DVIs.end(); I != E; ++I) {
1255 DbgValueInst *DVI = *I;
1257 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1258 // If an argument is zero extended then use argument directly. The ZExt
1259 // may be zapped by an optimization pass in future.
1260 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1261 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
1262 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1263 Arg = dyn_cast<Argument>(SExt->getOperand(0));
1265 Arg = SI->getOperand(0);
1266 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
1267 Arg = LI->getOperand(0);
1271 Instruction *DbgVal =
1272 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
1274 DbgVal->setDebugLoc(DVI->getDebugLoc());
1278 } // end anon namespace
1282 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
1284 /// This pass takes allocations which can be completely analyzed (that is, they
1285 /// don't escape) and tries to turn them into scalar SSA values. There are
1286 /// a few steps to this process.
1288 /// 1) It takes allocations of aggregates and analyzes the ways in which they
1289 /// are used to try to split them into smaller allocations, ideally of
1290 /// a single scalar data type. It will split up memcpy and memset accesses
1291 /// as necessary and try to isolate invidual scalar accesses.
1292 /// 2) It will transform accesses into forms which are suitable for SSA value
1293 /// promotion. This can be replacing a memset with a scalar store of an
1294 /// integer value, or it can involve speculating operations on a PHI or
1295 /// select to be a PHI or select of the results.
1296 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
1297 /// onto insert and extract operations on a vector value, and convert them to
1298 /// this form. By doing so, it will enable promotion of vector aggregates to
1299 /// SSA vector values.
1300 class SROA : public FunctionPass {
1301 const bool RequiresDomTree;
1304 const TargetData *TD;
1307 /// \brief Worklist of alloca instructions to simplify.
1309 /// Each alloca in the function is added to this. Each new alloca formed gets
1310 /// added to it as well to recursively simplify unless that alloca can be
1311 /// directly promoted. Finally, each time we rewrite a use of an alloca other
1312 /// the one being actively rewritten, we add it back onto the list if not
1313 /// already present to ensure it is re-visited.
1314 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
1316 /// \brief A collection of instructions to delete.
1317 /// We try to batch deletions to simplify code and make things a bit more
1319 SmallVector<Instruction *, 8> DeadInsts;
1321 /// \brief A set to prevent repeatedly marking an instruction split into many
1322 /// uses as dead. Only used to guard insertion into DeadInsts.
1323 SmallPtrSet<Instruction *, 4> DeadSplitInsts;
1325 /// \brief A collection of alloca instructions we can directly promote.
1326 std::vector<AllocaInst *> PromotableAllocas;
1329 SROA(bool RequiresDomTree = true)
1330 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
1331 C(0), TD(0), DT(0) {
1332 initializeSROAPass(*PassRegistry::getPassRegistry());
1334 bool runOnFunction(Function &F);
1335 void getAnalysisUsage(AnalysisUsage &AU) const;
1337 const char *getPassName() const { return "SROA"; }
1341 friend class AllocaPartitionRewriter;
1342 friend class AllocaPartitionVectorRewriter;
1344 bool rewriteAllocaPartition(AllocaInst &AI,
1345 AllocaPartitioning &P,
1346 AllocaPartitioning::iterator PI);
1347 bool splitAlloca(AllocaInst &AI, AllocaPartitioning &P);
1348 bool runOnAlloca(AllocaInst &AI);
1349 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
1350 bool promoteAllocas(Function &F);
1356 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
1357 return new SROA(RequiresDomTree);
1360 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
1362 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
1363 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
1366 /// \brief Accumulate the constant offsets in a GEP into a single APInt offset.
1368 /// If the provided GEP is all-constant, the total byte offset formed by the
1369 /// GEP is computed and Offset is set to it. If the GEP has any non-constant
1370 /// operands, the function returns false and the value of Offset is unmodified.
1371 static bool accumulateGEPOffsets(const TargetData &TD, GEPOperator &GEP,
1373 APInt GEPOffset(Offset.getBitWidth(), 0);
1374 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1375 GTI != GTE; ++GTI) {
1376 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
1379 if (OpC->isZero()) continue;
1381 // Handle a struct index, which adds its field offset to the pointer.
1382 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1383 unsigned ElementIdx = OpC->getZExtValue();
1384 const StructLayout *SL = TD.getStructLayout(STy);
1385 GEPOffset += APInt(Offset.getBitWidth(),
1386 SL->getElementOffset(ElementIdx));
1390 APInt TypeSize(Offset.getBitWidth(),
1391 TD.getTypeAllocSize(GTI.getIndexedType()));
1392 if (VectorType *VTy = dyn_cast<VectorType>(*GTI)) {
1393 assert((VTy->getScalarSizeInBits() % 8) == 0 &&
1394 "vector element size is not a multiple of 8, cannot GEP over it");
1395 TypeSize = VTy->getScalarSizeInBits() / 8;
1398 GEPOffset += OpC->getValue().sextOrTrunc(Offset.getBitWidth()) * TypeSize;
1404 /// \brief Build a GEP out of a base pointer and indices.
1406 /// This will return the BasePtr if that is valid, or build a new GEP
1407 /// instruction using the IRBuilder if GEP-ing is needed.
1408 static Value *buildGEP(IRBuilder<> &IRB, Value *BasePtr,
1409 SmallVectorImpl<Value *> &Indices,
1410 const Twine &Prefix) {
1411 if (Indices.empty())
1414 // A single zero index is a no-op, so check for this and avoid building a GEP
1416 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1419 return IRB.CreateInBoundsGEP(BasePtr, Indices, Prefix + ".idx");
1422 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1423 /// TargetTy without changing the offset of the pointer.
1425 /// This routine assumes we've already established a properly offset GEP with
1426 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1427 /// zero-indices down through type layers until we find one the same as
1428 /// TargetTy. If we can't find one with the same type, we at least try to use
1429 /// one with the same size. If none of that works, we just produce the GEP as
1430 /// indicated by Indices to have the correct offset.
1431 static Value *getNaturalGEPWithType(IRBuilder<> &IRB, const TargetData &TD,
1432 Value *BasePtr, Type *Ty, Type *TargetTy,
1433 SmallVectorImpl<Value *> &Indices,
1434 const Twine &Prefix) {
1436 return buildGEP(IRB, BasePtr, Indices, Prefix);
1438 // See if we can descend into a struct and locate a field with the correct
1440 unsigned NumLayers = 0;
1441 Type *ElementTy = Ty;
1443 if (ElementTy->isPointerTy())
1445 if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
1446 ElementTy = SeqTy->getElementType();
1447 Indices.push_back(IRB.getInt(APInt(TD.getPointerSizeInBits(), 0)));
1448 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1449 ElementTy = *STy->element_begin();
1450 Indices.push_back(IRB.getInt32(0));
1455 } while (ElementTy != TargetTy);
1456 if (ElementTy != TargetTy)
1457 Indices.erase(Indices.end() - NumLayers, Indices.end());
1459 return buildGEP(IRB, BasePtr, Indices, Prefix);
1462 /// \brief Recursively compute indices for a natural GEP.
1464 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1465 /// element types adding appropriate indices for the GEP.
1466 static Value *getNaturalGEPRecursively(IRBuilder<> &IRB, const TargetData &TD,
1467 Value *Ptr, Type *Ty, APInt &Offset,
1469 SmallVectorImpl<Value *> &Indices,
1470 const Twine &Prefix) {
1472 return getNaturalGEPWithType(IRB, TD, Ptr, Ty, TargetTy, Indices, Prefix);
1474 // We can't recurse through pointer types.
1475 if (Ty->isPointerTy())
1478 // We try to analyze GEPs over vectors here, but note that these GEPs are
1479 // extremely poorly defined currently. The long-term goal is to remove GEPing
1480 // over a vector from the IR completely.
1481 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1482 unsigned ElementSizeInBits = VecTy->getScalarSizeInBits();
1483 if (ElementSizeInBits % 8)
1484 return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
1485 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1486 APInt NumSkippedElements = Offset.udiv(ElementSize);
1487 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1489 Offset -= NumSkippedElements * ElementSize;
1490 Indices.push_back(IRB.getInt(NumSkippedElements));
1491 return getNaturalGEPRecursively(IRB, TD, Ptr, VecTy->getElementType(),
1492 Offset, TargetTy, Indices, Prefix);
1495 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1496 Type *ElementTy = ArrTy->getElementType();
1497 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1498 APInt NumSkippedElements = Offset.udiv(ElementSize);
1499 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1502 Offset -= NumSkippedElements * ElementSize;
1503 Indices.push_back(IRB.getInt(NumSkippedElements));
1504 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1508 StructType *STy = dyn_cast<StructType>(Ty);
1512 const StructLayout *SL = TD.getStructLayout(STy);
1513 uint64_t StructOffset = Offset.getZExtValue();
1514 if (StructOffset >= SL->getSizeInBytes())
1516 unsigned Index = SL->getElementContainingOffset(StructOffset);
1517 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1518 Type *ElementTy = STy->getElementType(Index);
1519 if (Offset.uge(TD.getTypeAllocSize(ElementTy)))
1520 return 0; // The offset points into alignment padding.
1522 Indices.push_back(IRB.getInt32(Index));
1523 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1527 /// \brief Get a natural GEP from a base pointer to a particular offset and
1528 /// resulting in a particular type.
1530 /// The goal is to produce a "natural" looking GEP that works with the existing
1531 /// composite types to arrive at the appropriate offset and element type for
1532 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1533 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1534 /// Indices, and setting Ty to the result subtype.
1536 /// If no natural GEP can be constructed, this function returns null.
1537 static Value *getNaturalGEPWithOffset(IRBuilder<> &IRB, const TargetData &TD,
1538 Value *Ptr, APInt Offset, Type *TargetTy,
1539 SmallVectorImpl<Value *> &Indices,
1540 const Twine &Prefix) {
1541 PointerType *Ty = cast<PointerType>(Ptr->getType());
1543 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1545 if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
1548 Type *ElementTy = Ty->getElementType();
1549 if (!ElementTy->isSized())
1550 return 0; // We can't GEP through an unsized element.
1551 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1552 if (ElementSize == 0)
1553 return 0; // Zero-length arrays can't help us build a natural GEP.
1554 APInt NumSkippedElements = Offset.udiv(ElementSize);
1556 Offset -= NumSkippedElements * ElementSize;
1557 Indices.push_back(IRB.getInt(NumSkippedElements));
1558 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1562 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1563 /// resulting pointer has PointerTy.
1565 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1566 /// and produces the pointer type desired. Where it cannot, it will try to use
1567 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1568 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1569 /// bitcast to the type.
1571 /// The strategy for finding the more natural GEPs is to peel off layers of the
1572 /// pointer, walking back through bit casts and GEPs, searching for a base
1573 /// pointer from which we can compute a natural GEP with the desired
1574 /// properities. The algorithm tries to fold as many constant indices into
1575 /// a single GEP as possible, thus making each GEP more independent of the
1576 /// surrounding code.
1577 static Value *getAdjustedPtr(IRBuilder<> &IRB, const TargetData &TD,
1578 Value *Ptr, APInt Offset, Type *PointerTy,
1579 const Twine &Prefix) {
1580 // Even though we don't look through PHI nodes, we could be called on an
1581 // instruction in an unreachable block, which may be on a cycle.
1582 SmallPtrSet<Value *, 4> Visited;
1583 Visited.insert(Ptr);
1584 SmallVector<Value *, 4> Indices;
1586 // We may end up computing an offset pointer that has the wrong type. If we
1587 // never are able to compute one directly that has the correct type, we'll
1588 // fall back to it, so keep it around here.
1589 Value *OffsetPtr = 0;
1591 // Remember any i8 pointer we come across to re-use if we need to do a raw
1594 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1596 Type *TargetTy = PointerTy->getPointerElementType();
1599 // First fold any existing GEPs into the offset.
1600 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1601 APInt GEPOffset(Offset.getBitWidth(), 0);
1602 if (!accumulateGEPOffsets(TD, *GEP, GEPOffset))
1604 Offset += GEPOffset;
1605 Ptr = GEP->getPointerOperand();
1606 if (!Visited.insert(Ptr))
1610 // See if we can perform a natural GEP here.
1612 if (Value *P = getNaturalGEPWithOffset(IRB, TD, Ptr, Offset, TargetTy,
1614 if (P->getType() == PointerTy) {
1615 // Zap any offset pointer that we ended up computing in previous rounds.
1616 if (OffsetPtr && OffsetPtr->use_empty())
1617 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1618 I->eraseFromParent();
1626 // Stash this pointer if we've found an i8*.
1627 if (Ptr->getType()->isIntegerTy(8)) {
1629 Int8PtrOffset = Offset;
1632 // Peel off a layer of the pointer and update the offset appropriately.
1633 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1634 Ptr = cast<Operator>(Ptr)->getOperand(0);
1635 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1636 if (GA->mayBeOverridden())
1638 Ptr = GA->getAliasee();
1642 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1643 } while (Visited.insert(Ptr));
1647 Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
1648 Prefix + ".raw_cast");
1649 Int8PtrOffset = Offset;
1652 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1653 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1654 Prefix + ".raw_idx");
1658 // On the off chance we were targeting i8*, guard the bitcast here.
1659 if (Ptr->getType() != PointerTy)
1660 Ptr = IRB.CreateBitCast(Ptr, PointerTy, Prefix + ".cast");
1665 /// \brief Test whether the given alloca partition can be promoted to a vector.
1667 /// This is a quick test to check whether we can rewrite a particular alloca
1668 /// partition (and its newly formed alloca) into a vector alloca with only
1669 /// whole-vector loads and stores such that it could be promoted to a vector
1670 /// SSA value. We only can ensure this for a limited set of operations, and we
1671 /// don't want to do the rewrites unless we are confident that the result will
1672 /// be promotable, so we have an early test here.
1673 static bool isVectorPromotionViable(const TargetData &TD,
1675 AllocaPartitioning &P,
1676 uint64_t PartitionBeginOffset,
1677 uint64_t PartitionEndOffset,
1678 AllocaPartitioning::const_use_iterator I,
1679 AllocaPartitioning::const_use_iterator E) {
1680 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
1684 uint64_t VecSize = TD.getTypeSizeInBits(Ty);
1685 uint64_t ElementSize = Ty->getScalarSizeInBits();
1687 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1688 // that aren't byte sized.
1689 if (ElementSize % 8)
1691 assert((VecSize % 8) == 0 && "vector size not a multiple of element size?");
1695 for (; I != E; ++I) {
1696 uint64_t BeginOffset = I->BeginOffset - PartitionBeginOffset;
1697 uint64_t BeginIndex = BeginOffset / ElementSize;
1698 if (BeginIndex * ElementSize != BeginOffset ||
1699 BeginIndex >= Ty->getNumElements())
1701 uint64_t EndOffset = I->EndOffset - PartitionBeginOffset;
1702 uint64_t EndIndex = EndOffset / ElementSize;
1703 if (EndIndex * ElementSize != EndOffset ||
1704 EndIndex > Ty->getNumElements())
1707 // FIXME: We should build shuffle vector instructions to handle
1708 // non-element-sized accesses.
1709 if ((EndOffset - BeginOffset) != ElementSize &&
1710 (EndOffset - BeginOffset) != VecSize)
1713 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(&*I->User)) {
1714 if (MI->isVolatile())
1716 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(&*I->User)) {
1717 const AllocaPartitioning::MemTransferOffsets &MTO
1718 = P.getMemTransferOffsets(*MTI);
1719 if (!MTO.IsSplittable)
1722 } else if (I->Ptr->getType()->getPointerElementType()->isStructTy()) {
1723 // Disable vector promotion when there are loads or stores of an FCA.
1725 } else if (!isa<LoadInst>(*I->User) && !isa<StoreInst>(*I->User)) {
1732 /// \brief Test whether the given alloca partition can be promoted to an int.
1734 /// This is a quick test to check whether we can rewrite a particular alloca
1735 /// partition (and its newly formed alloca) into an integer alloca suitable for
1736 /// promotion to an SSA value. We only can ensure this for a limited set of
1737 /// operations, and we don't want to do the rewrites unless we are confident
1738 /// that the result will be promotable, so we have an early test here.
1739 static bool isIntegerPromotionViable(const TargetData &TD,
1741 AllocaPartitioning &P,
1742 AllocaPartitioning::const_use_iterator I,
1743 AllocaPartitioning::const_use_iterator E) {
1744 IntegerType *Ty = dyn_cast<IntegerType>(AllocaTy);
1748 // Check the uses to ensure the uses are (likely) promoteable integer uses.
1749 // Also ensure that the alloca has a covering load or store. We don't want
1750 // promote because of some other unsplittable entry (which we may make
1751 // splittable later) and lose the ability to promote each element access.
1752 bool WholeAllocaOp = false;
1753 for (; I != E; ++I) {
1754 if (LoadInst *LI = dyn_cast<LoadInst>(&*I->User)) {
1755 if (LI->isVolatile() || !LI->getType()->isIntegerTy())
1757 if (LI->getType() == Ty)
1758 WholeAllocaOp = true;
1759 } else if (StoreInst *SI = dyn_cast<StoreInst>(&*I->User)) {
1760 if (SI->isVolatile() || !SI->getValueOperand()->getType()->isIntegerTy())
1762 if (SI->getValueOperand()->getType() == Ty)
1763 WholeAllocaOp = true;
1764 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(&*I->User)) {
1765 if (MI->isVolatile())
1767 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(&*I->User)) {
1768 const AllocaPartitioning::MemTransferOffsets &MTO
1769 = P.getMemTransferOffsets(*MTI);
1770 if (!MTO.IsSplittable)
1777 return WholeAllocaOp;
1781 /// \brief Visitor to rewrite instructions using a partition of an alloca to
1782 /// use a new alloca.
1784 /// Also implements the rewriting to vector-based accesses when the partition
1785 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
1787 class AllocaPartitionRewriter : public InstVisitor<AllocaPartitionRewriter,
1789 // Befriend the base class so it can delegate to private visit methods.
1790 friend class llvm::InstVisitor<AllocaPartitionRewriter, bool>;
1792 const TargetData &TD;
1793 AllocaPartitioning &P;
1795 AllocaInst &OldAI, &NewAI;
1796 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
1798 // If we are rewriting an alloca partition which can be written as pure
1799 // vector operations, we stash extra information here. When VecTy is
1800 // non-null, we have some strict guarantees about the rewriten alloca:
1801 // - The new alloca is exactly the size of the vector type here.
1802 // - The accesses all either map to the entire vector or to a single
1804 // - The set of accessing instructions is only one of those handled above
1805 // in isVectorPromotionViable. Generally these are the same access kinds
1806 // which are promotable via mem2reg.
1809 uint64_t ElementSize;
1811 // This is a convenience and flag variable that will be null unless the new
1812 // alloca has a promotion-targeted integer type due to passing
1813 // isIntegerPromotionViable above. If it is non-null does, the desired
1814 // integer type will be stored here for easy access during rewriting.
1815 IntegerType *IntPromotionTy;
1817 // The offset of the partition user currently being rewritten.
1818 uint64_t BeginOffset, EndOffset;
1819 Instruction *OldPtr;
1821 // The name prefix to use when rewriting instructions for this alloca.
1822 std::string NamePrefix;
1825 AllocaPartitionRewriter(const TargetData &TD, AllocaPartitioning &P,
1826 AllocaPartitioning::iterator PI,
1827 SROA &Pass, AllocaInst &OldAI, AllocaInst &NewAI,
1828 uint64_t NewBeginOffset, uint64_t NewEndOffset)
1829 : TD(TD), P(P), Pass(Pass),
1830 OldAI(OldAI), NewAI(NewAI),
1831 NewAllocaBeginOffset(NewBeginOffset),
1832 NewAllocaEndOffset(NewEndOffset),
1833 VecTy(), ElementTy(), ElementSize(), IntPromotionTy(),
1834 BeginOffset(), EndOffset() {
1837 /// \brief Visit the users of the alloca partition and rewrite them.
1838 bool visitUsers(AllocaPartitioning::const_use_iterator I,
1839 AllocaPartitioning::const_use_iterator E) {
1840 if (isVectorPromotionViable(TD, NewAI.getAllocatedType(), P,
1841 NewAllocaBeginOffset, NewAllocaEndOffset,
1844 VecTy = cast<VectorType>(NewAI.getAllocatedType());
1845 ElementTy = VecTy->getElementType();
1846 assert((VecTy->getScalarSizeInBits() % 8) == 0 &&
1847 "Only multiple-of-8 sized vector elements are viable");
1848 ElementSize = VecTy->getScalarSizeInBits() / 8;
1849 } else if (isIntegerPromotionViable(TD, NewAI.getAllocatedType(),
1851 IntPromotionTy = cast<IntegerType>(NewAI.getAllocatedType());
1853 bool CanSROA = true;
1854 for (; I != E; ++I) {
1855 BeginOffset = I->BeginOffset;
1856 EndOffset = I->EndOffset;
1858 NamePrefix = (Twine(NewAI.getName()) + "." + Twine(BeginOffset)).str();
1859 CanSROA &= visit(I->User);
1871 // Every instruction which can end up as a user must have a rewrite rule.
1872 bool visitInstruction(Instruction &I) {
1873 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
1874 llvm_unreachable("No rewrite rule for this instruction!");
1877 Twine getName(const Twine &Suffix) {
1878 return NamePrefix + Suffix;
1881 Value *getAdjustedAllocaPtr(IRBuilder<> &IRB, Type *PointerTy) {
1882 assert(BeginOffset >= NewAllocaBeginOffset);
1883 APInt Offset(TD.getPointerSizeInBits(), BeginOffset - NewAllocaBeginOffset);
1884 return getAdjustedPtr(IRB, TD, &NewAI, Offset, PointerTy, getName(""));
1887 ConstantInt *getIndex(IRBuilder<> &IRB, uint64_t Offset) {
1888 assert(VecTy && "Can only call getIndex when rewriting a vector");
1889 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
1890 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
1891 uint32_t Index = RelOffset / ElementSize;
1892 assert(Index * ElementSize == RelOffset);
1893 return IRB.getInt32(Index);
1896 Value *extractInteger(IRBuilder<> &IRB, IntegerType *TargetTy,
1898 assert(IntPromotionTy && "Alloca is not an integer we can extract from");
1899 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
1901 assert(Offset >= NewAllocaBeginOffset && "Out of bounds offset");
1902 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
1904 V = IRB.CreateLShr(V, RelOffset*8, getName(".shift"));
1905 if (TargetTy != IntPromotionTy) {
1906 assert(TargetTy->getBitWidth() < IntPromotionTy->getBitWidth() &&
1907 "Cannot extract to a larger integer!");
1908 V = IRB.CreateTrunc(V, TargetTy, getName(".trunc"));
1913 StoreInst *insertInteger(IRBuilder<> &IRB, Value *V, uint64_t Offset) {
1914 IntegerType *Ty = cast<IntegerType>(V->getType());
1915 if (Ty == IntPromotionTy)
1916 return IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
1918 assert(Ty->getBitWidth() < IntPromotionTy->getBitWidth() &&
1919 "Cannot insert a larger integer!");
1920 V = IRB.CreateZExt(V, IntPromotionTy, getName(".ext"));
1921 assert(Offset >= NewAllocaBeginOffset && "Out of bounds offset");
1922 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
1924 V = IRB.CreateShl(V, RelOffset*8, getName(".shift"));
1926 APInt Mask = ~Ty->getMask().zext(IntPromotionTy->getBitWidth())
1928 Value *Old = IRB.CreateAnd(IRB.CreateAlignedLoad(&NewAI,
1929 NewAI.getAlignment(),
1930 getName(".oldload")),
1931 Mask, getName(".mask"));
1932 return IRB.CreateAlignedStore(IRB.CreateOr(Old, V, getName(".insert")),
1933 &NewAI, NewAI.getAlignment());
1936 void deleteIfTriviallyDead(Value *V) {
1937 Instruction *I = cast<Instruction>(V);
1938 if (isInstructionTriviallyDead(I))
1939 Pass.DeadInsts.push_back(I);
1942 Value *getValueCast(IRBuilder<> &IRB, Value *V, Type *Ty) {
1943 if (V->getType()->isIntegerTy() && Ty->isPointerTy())
1944 return IRB.CreateIntToPtr(V, Ty);
1945 if (V->getType()->isPointerTy() && Ty->isIntegerTy())
1946 return IRB.CreatePtrToInt(V, Ty);
1948 return IRB.CreateBitCast(V, Ty);
1951 bool rewriteVectorizedLoadInst(IRBuilder<> &IRB, LoadInst &LI, Value *OldOp) {
1953 if (LI.getType() == VecTy->getElementType() ||
1954 BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
1955 Result = IRB.CreateExtractElement(
1956 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), getName(".load")),
1957 getIndex(IRB, BeginOffset), getName(".extract"));
1959 Result = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
1962 if (Result->getType() != LI.getType())
1963 Result = getValueCast(IRB, Result, LI.getType());
1964 LI.replaceAllUsesWith(Result);
1965 Pass.DeadInsts.push_back(&LI);
1967 DEBUG(dbgs() << " to: " << *Result << "\n");
1971 bool rewriteIntegerLoad(IRBuilder<> &IRB, LoadInst &LI) {
1972 assert(!LI.isVolatile());
1973 Value *Result = extractInteger(IRB, cast<IntegerType>(LI.getType()),
1975 LI.replaceAllUsesWith(Result);
1976 Pass.DeadInsts.push_back(&LI);
1977 DEBUG(dbgs() << " to: " << *Result << "\n");
1981 bool visitLoadInst(LoadInst &LI) {
1982 DEBUG(dbgs() << " original: " << LI << "\n");
1983 Value *OldOp = LI.getOperand(0);
1984 assert(OldOp == OldPtr);
1985 IRBuilder<> IRB(&LI);
1988 return rewriteVectorizedLoadInst(IRB, LI, OldOp);
1990 return rewriteIntegerLoad(IRB, LI);
1992 Value *NewPtr = getAdjustedAllocaPtr(IRB,
1993 LI.getPointerOperand()->getType());
1994 LI.setOperand(0, NewPtr);
1995 DEBUG(dbgs() << " to: " << LI << "\n");
1997 deleteIfTriviallyDead(OldOp);
1998 return NewPtr == &NewAI && !LI.isVolatile();
2001 bool rewriteVectorizedStoreInst(IRBuilder<> &IRB, StoreInst &SI,
2003 Value *V = SI.getValueOperand();
2004 if (V->getType() == ElementTy ||
2005 BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
2006 if (V->getType() != ElementTy)
2007 V = getValueCast(IRB, V, ElementTy);
2008 LoadInst *LI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2010 V = IRB.CreateInsertElement(LI, V, getIndex(IRB, BeginOffset),
2011 getName(".insert"));
2012 } else if (V->getType() != VecTy) {
2013 V = getValueCast(IRB, V, VecTy);
2015 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2016 Pass.DeadInsts.push_back(&SI);
2019 DEBUG(dbgs() << " to: " << *Store << "\n");
2023 bool rewriteIntegerStore(IRBuilder<> &IRB, StoreInst &SI) {
2024 assert(!SI.isVolatile());
2025 StoreInst *Store = insertInteger(IRB, SI.getValueOperand(), BeginOffset);
2026 Pass.DeadInsts.push_back(&SI);
2028 DEBUG(dbgs() << " to: " << *Store << "\n");
2032 bool visitStoreInst(StoreInst &SI) {
2033 DEBUG(dbgs() << " original: " << SI << "\n");
2034 Value *OldOp = SI.getOperand(1);
2035 assert(OldOp == OldPtr);
2036 IRBuilder<> IRB(&SI);
2039 return rewriteVectorizedStoreInst(IRB, SI, OldOp);
2041 return rewriteIntegerStore(IRB, SI);
2043 Value *NewPtr = getAdjustedAllocaPtr(IRB,
2044 SI.getPointerOperand()->getType());
2045 SI.setOperand(1, NewPtr);
2046 DEBUG(dbgs() << " to: " << SI << "\n");
2048 deleteIfTriviallyDead(OldOp);
2049 return NewPtr == &NewAI && !SI.isVolatile();
2052 bool visitMemSetInst(MemSetInst &II) {
2053 DEBUG(dbgs() << " original: " << II << "\n");
2054 IRBuilder<> IRB(&II);
2055 assert(II.getRawDest() == OldPtr);
2057 // If the memset has a variable size, it cannot be split, just adjust the
2058 // pointer to the new alloca.
2059 if (!isa<Constant>(II.getLength())) {
2060 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2061 deleteIfTriviallyDead(OldPtr);
2065 // Record this instruction for deletion.
2066 if (Pass.DeadSplitInsts.insert(&II))
2067 Pass.DeadInsts.push_back(&II);
2069 Type *AllocaTy = NewAI.getAllocatedType();
2070 Type *ScalarTy = AllocaTy->getScalarType();
2072 // If this doesn't map cleanly onto the alloca type, and that type isn't
2073 // a single value type, just emit a memset.
2074 if (!VecTy && (BeginOffset != NewAllocaBeginOffset ||
2075 EndOffset != NewAllocaEndOffset ||
2076 !AllocaTy->isSingleValueType() ||
2077 !TD.isLegalInteger(TD.getTypeSizeInBits(ScalarTy)))) {
2078 Type *SizeTy = II.getLength()->getType();
2079 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2081 if (NewAI.getAlignment())
2082 Align = MinAlign(NewAI.getAlignment(),
2083 BeginOffset - NewAllocaBeginOffset);
2086 = IRB.CreateMemSet(getAdjustedAllocaPtr(IRB,
2087 II.getRawDest()->getType()),
2088 II.getValue(), Size, Align,
2091 DEBUG(dbgs() << " to: " << *New << "\n");
2095 // If we can represent this as a simple value, we have to build the actual
2096 // value to store, which requires expanding the byte present in memset to
2097 // a sensible representation for the alloca type. This is essentially
2098 // splatting the byte to a sufficiently wide integer, bitcasting to the
2099 // desired scalar type, and splatting it across any desired vector type.
2100 Value *V = II.getValue();
2101 IntegerType *VTy = cast<IntegerType>(V->getType());
2102 Type *IntTy = Type::getIntNTy(VTy->getContext(),
2103 TD.getTypeSizeInBits(ScalarTy));
2104 if (TD.getTypeSizeInBits(ScalarTy) > VTy->getBitWidth())
2105 V = IRB.CreateMul(IRB.CreateZExt(V, IntTy, getName(".zext")),
2106 ConstantExpr::getUDiv(
2107 Constant::getAllOnesValue(IntTy),
2108 ConstantExpr::getZExt(
2109 Constant::getAllOnesValue(V->getType()),
2111 getName(".isplat"));
2112 if (V->getType() != ScalarTy) {
2113 if (ScalarTy->isPointerTy())
2114 V = IRB.CreateIntToPtr(V, ScalarTy);
2115 else if (ScalarTy->isPrimitiveType() || ScalarTy->isVectorTy())
2116 V = IRB.CreateBitCast(V, ScalarTy);
2117 else if (ScalarTy->isIntegerTy())
2118 llvm_unreachable("Computed different integer types with equal widths");
2120 llvm_unreachable("Invalid scalar type");
2123 // If this is an element-wide memset of a vectorizable alloca, insert it.
2124 if (VecTy && (BeginOffset > NewAllocaBeginOffset ||
2125 EndOffset < NewAllocaEndOffset)) {
2126 StoreInst *Store = IRB.CreateAlignedStore(
2127 IRB.CreateInsertElement(IRB.CreateAlignedLoad(&NewAI,
2128 NewAI.getAlignment(),
2130 V, getIndex(IRB, BeginOffset),
2131 getName(".insert")),
2132 &NewAI, NewAI.getAlignment());
2134 DEBUG(dbgs() << " to: " << *Store << "\n");
2138 // Splat to a vector if needed.
2139 if (VectorType *VecTy = dyn_cast<VectorType>(AllocaTy)) {
2140 VectorType *SplatSourceTy = VectorType::get(V->getType(), 1);
2141 V = IRB.CreateShuffleVector(
2142 IRB.CreateInsertElement(UndefValue::get(SplatSourceTy), V,
2143 IRB.getInt32(0), getName(".vsplat.insert")),
2144 UndefValue::get(SplatSourceTy),
2145 ConstantVector::getSplat(VecTy->getNumElements(), IRB.getInt32(0)),
2146 getName(".vsplat.shuffle"));
2147 assert(V->getType() == VecTy);
2150 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2153 DEBUG(dbgs() << " to: " << *New << "\n");
2154 return !II.isVolatile();
2157 bool visitMemTransferInst(MemTransferInst &II) {
2158 // Rewriting of memory transfer instructions can be a bit tricky. We break
2159 // them into two categories: split intrinsics and unsplit intrinsics.
2161 DEBUG(dbgs() << " original: " << II << "\n");
2162 IRBuilder<> IRB(&II);
2164 assert(II.getRawSource() == OldPtr || II.getRawDest() == OldPtr);
2165 bool IsDest = II.getRawDest() == OldPtr;
2167 const AllocaPartitioning::MemTransferOffsets &MTO
2168 = P.getMemTransferOffsets(II);
2170 // For unsplit intrinsics, we simply modify the source and destination
2171 // pointers in place. This isn't just an optimization, it is a matter of
2172 // correctness. With unsplit intrinsics we may be dealing with transfers
2173 // within a single alloca before SROA ran, or with transfers that have
2174 // a variable length. We may also be dealing with memmove instead of
2175 // memcpy, and so simply updating the pointers is the necessary for us to
2176 // update both source and dest of a single call.
2177 if (!MTO.IsSplittable) {
2178 Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource();
2180 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2182 II.setSource(getAdjustedAllocaPtr(IRB, II.getRawSource()->getType()));
2184 DEBUG(dbgs() << " to: " << II << "\n");
2185 deleteIfTriviallyDead(OldOp);
2188 // For split transfer intrinsics we have an incredibly useful assurance:
2189 // the source and destination do not reside within the same alloca, and at
2190 // least one of them does not escape. This means that we can replace
2191 // memmove with memcpy, and we don't need to worry about all manner of
2192 // downsides to splitting and transforming the operations.
2194 // Compute the relative offset within the transfer.
2195 unsigned IntPtrWidth = TD.getPointerSizeInBits();
2196 APInt RelOffset(IntPtrWidth, BeginOffset - (IsDest ? MTO.DestBegin
2197 : MTO.SourceBegin));
2199 // If this doesn't map cleanly onto the alloca type, and that type isn't
2200 // a single value type, just emit a memcpy.
2202 = !VecTy && (BeginOffset != NewAllocaBeginOffset ||
2203 EndOffset != NewAllocaEndOffset ||
2204 !NewAI.getAllocatedType()->isSingleValueType());
2206 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2207 // size hasn't been shrunk based on analysis of the viable range, this is
2209 if (EmitMemCpy && &OldAI == &NewAI) {
2210 uint64_t OrigBegin = IsDest ? MTO.DestBegin : MTO.SourceBegin;
2211 uint64_t OrigEnd = IsDest ? MTO.DestEnd : MTO.SourceEnd;
2212 // Ensure the start lines up.
2213 assert(BeginOffset == OrigBegin);
2216 // Rewrite the size as needed.
2217 if (EndOffset != OrigEnd)
2218 II.setLength(ConstantInt::get(II.getLength()->getType(),
2219 EndOffset - BeginOffset));
2222 // Record this instruction for deletion.
2223 if (Pass.DeadSplitInsts.insert(&II))
2224 Pass.DeadInsts.push_back(&II);
2226 bool IsVectorElement = VecTy && (BeginOffset > NewAllocaBeginOffset ||
2227 EndOffset < NewAllocaEndOffset);
2229 Type *OtherPtrTy = IsDest ? II.getRawSource()->getType()
2230 : II.getRawDest()->getType();
2232 OtherPtrTy = IsVectorElement ? VecTy->getElementType()->getPointerTo()
2235 // Compute the other pointer, folding as much as possible to produce
2236 // a single, simple GEP in most cases.
2237 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2238 OtherPtr = getAdjustedPtr(IRB, TD, OtherPtr, RelOffset, OtherPtrTy,
2239 getName("." + OtherPtr->getName()));
2241 unsigned Align = II.getAlignment();
2243 Align = MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(),
2244 MinAlign(II.getAlignment(), NewAI.getAlignment()));
2246 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2247 // alloca that should be re-examined after rewriting this instruction.
2249 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets()))
2250 Pass.Worklist.insert(AI);
2254 = getAdjustedAllocaPtr(IRB, IsDest ? II.getRawDest()->getType()
2255 : II.getRawSource()->getType());
2256 Type *SizeTy = II.getLength()->getType();
2257 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2259 CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
2260 IsDest ? OtherPtr : OurPtr,
2261 Size, Align, II.isVolatile());
2263 DEBUG(dbgs() << " to: " << *New << "\n");
2267 Value *SrcPtr = OtherPtr;
2268 Value *DstPtr = &NewAI;
2270 std::swap(SrcPtr, DstPtr);
2273 if (IsVectorElement && !IsDest) {
2274 // We have to extract rather than load.
2275 Src = IRB.CreateExtractElement(
2276 IRB.CreateAlignedLoad(SrcPtr, Align, getName(".copyload")),
2277 getIndex(IRB, BeginOffset),
2278 getName(".copyextract"));
2280 Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(),
2281 getName(".copyload"));
2284 if (IsVectorElement && IsDest) {
2285 // We have to insert into a loaded copy before storing.
2286 Src = IRB.CreateInsertElement(
2287 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), getName(".load")),
2288 Src, getIndex(IRB, BeginOffset),
2289 getName(".insert"));
2292 StoreInst *Store = cast<StoreInst>(
2293 IRB.CreateAlignedStore(Src, DstPtr, Align, II.isVolatile()));
2295 DEBUG(dbgs() << " to: " << *Store << "\n");
2296 return !II.isVolatile();
2299 bool visitIntrinsicInst(IntrinsicInst &II) {
2300 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2301 II.getIntrinsicID() == Intrinsic::lifetime_end);
2302 DEBUG(dbgs() << " original: " << II << "\n");
2303 IRBuilder<> IRB(&II);
2304 assert(II.getArgOperand(1) == OldPtr);
2306 // Record this instruction for deletion.
2307 if (Pass.DeadSplitInsts.insert(&II))
2308 Pass.DeadInsts.push_back(&II);
2311 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2312 EndOffset - BeginOffset);
2313 Value *Ptr = getAdjustedAllocaPtr(IRB, II.getArgOperand(1)->getType());
2315 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2316 New = IRB.CreateLifetimeStart(Ptr, Size);
2318 New = IRB.CreateLifetimeEnd(Ptr, Size);
2320 DEBUG(dbgs() << " to: " << *New << "\n");
2324 /// PHI instructions that use an alloca and are subsequently loaded can be
2325 /// rewritten to load both input pointers in the pred blocks and then PHI the
2326 /// results, allowing the load of the alloca to be promoted.
2328 /// %P2 = phi [i32* %Alloca, i32* %Other]
2329 /// %V = load i32* %P2
2331 /// %V1 = load i32* %Alloca -> will be mem2reg'd
2333 /// %V2 = load i32* %Other
2335 /// %V = phi [i32 %V1, i32 %V2]
2337 /// We can do this to a select if its only uses are loads and if the operand
2338 /// to the select can be loaded unconditionally.
2340 /// FIXME: This should be hoisted into a generic utility, likely in
2341 /// Transforms/Util/Local.h
2342 bool isSafePHIToSpeculate(PHINode &PN, SmallVectorImpl<LoadInst *> &Loads) {
2343 // For now, we can only do this promotion if the load is in the same block
2344 // as the PHI, and if there are no stores between the phi and load.
2345 // TODO: Allow recursive phi users.
2346 // TODO: Allow stores.
2347 BasicBlock *BB = PN.getParent();
2348 unsigned MaxAlign = 0;
2349 for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end();
2351 LoadInst *LI = dyn_cast<LoadInst>(*UI);
2352 if (LI == 0 || !LI->isSimple()) return false;
2354 // For now we only allow loads in the same block as the PHI. This is
2355 // a common case that happens when instcombine merges two loads through
2357 if (LI->getParent() != BB) return false;
2359 // Ensure that there are no instructions between the PHI and the load that
2361 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
2362 if (BBI->mayWriteToMemory())
2365 MaxAlign = std::max(MaxAlign, LI->getAlignment());
2366 Loads.push_back(LI);
2369 // We can only transform this if it is safe to push the loads into the
2370 // predecessor blocks. The only thing to watch out for is that we can't put
2371 // a possibly trapping load in the predecessor if it is a critical edge.
2372 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num;
2374 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
2375 Value *InVal = PN.getIncomingValue(Idx);
2377 // If the value is produced by the terminator of the predecessor (an
2378 // invoke) or it has side-effects, there is no valid place to put a load
2379 // in the predecessor.
2380 if (TI == InVal || TI->mayHaveSideEffects())
2383 // If the predecessor has a single successor, then the edge isn't
2385 if (TI->getNumSuccessors() == 1)
2388 // If this pointer is always safe to load, or if we can prove that there
2389 // is already a load in the block, then we can move the load to the pred
2391 if (InVal->isDereferenceablePointer() ||
2392 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, &TD))
2401 bool visitPHINode(PHINode &PN) {
2402 DEBUG(dbgs() << " original: " << PN << "\n");
2403 // We would like to compute a new pointer in only one place, but have it be
2404 // as local as possible to the PHI. To do that, we re-use the location of
2405 // the old pointer, which necessarily must be in the right position to
2406 // dominate the PHI.
2407 IRBuilder<> PtrBuilder(cast<Instruction>(OldPtr));
2409 SmallVector<LoadInst *, 4> Loads;
2410 if (!isSafePHIToSpeculate(PN, Loads)) {
2411 Value *NewPtr = getAdjustedAllocaPtr(PtrBuilder, OldPtr->getType());
2412 // Replace the operands which were using the old pointer.
2413 User::op_iterator OI = PN.op_begin(), OE = PN.op_end();
2414 for (; OI != OE; ++OI)
2418 DEBUG(dbgs() << " to: " << PN << "\n");
2419 deleteIfTriviallyDead(OldPtr);
2422 assert(!Loads.empty());
2424 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
2425 IRBuilder<> PHIBuilder(&PN);
2426 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues());
2427 NewPN->takeName(&PN);
2429 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
2430 // matter which one we get and if any differ, it doesn't matter.
2431 LoadInst *SomeLoad = cast<LoadInst>(Loads.back());
2432 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
2433 unsigned Align = SomeLoad->getAlignment();
2434 Value *NewPtr = getAdjustedAllocaPtr(PtrBuilder, OldPtr->getType());
2436 // Rewrite all loads of the PN to use the new PHI.
2438 LoadInst *LI = Loads.pop_back_val();
2439 LI->replaceAllUsesWith(NewPN);
2440 Pass.DeadInsts.push_back(LI);
2441 } while (!Loads.empty());
2443 // Inject loads into all of the pred blocks.
2444 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
2445 BasicBlock *Pred = PN.getIncomingBlock(Idx);
2446 TerminatorInst *TI = Pred->getTerminator();
2447 Value *InVal = PN.getIncomingValue(Idx);
2448 IRBuilder<> PredBuilder(TI);
2450 // Map the value to the new alloca pointer if this was the old alloca
2452 bool ThisOperand = InVal == OldPtr;
2457 = PredBuilder.CreateLoad(InVal, getName(".sroa.speculate." +
2459 ++NumLoadsSpeculated;
2460 Load->setAlignment(Align);
2462 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
2463 NewPN->addIncoming(Load, Pred);
2467 Instruction *OtherPtr = dyn_cast<Instruction>(InVal);
2469 // No uses to rewrite.
2472 // Try to lookup and rewrite any partition uses corresponding to this phi
2474 AllocaPartitioning::iterator PI
2475 = P.findPartitionForPHIOrSelectOperand(PN, OtherPtr);
2476 if (PI != P.end()) {
2477 // If the other pointer is within the partitioning, replace the PHI in
2478 // its uses with the load we just speculated, or add another load for
2479 // it to rewrite if we've already replaced the PHI.
2480 AllocaPartitioning::use_iterator UI
2481 = P.findPartitionUseForPHIOrSelectOperand(PN, OtherPtr);
2482 if (isa<PHINode>(*UI->User))
2485 AllocaPartitioning::PartitionUse OtherUse = *UI;
2486 OtherUse.User = Load;
2487 P.use_push_back(PI, OtherUse);
2491 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
2492 return NewPtr == &NewAI;
2495 /// Select instructions that use an alloca and are subsequently loaded can be
2496 /// rewritten to load both input pointers and then select between the result,
2497 /// allowing the load of the alloca to be promoted.
2499 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
2500 /// %V = load i32* %P2
2502 /// %V1 = load i32* %Alloca -> will be mem2reg'd
2503 /// %V2 = load i32* %Other
2504 /// %V = select i1 %cond, i32 %V1, i32 %V2
2506 /// We can do this to a select if its only uses are loads and if the operand
2507 /// to the select can be loaded unconditionally.
2508 bool isSafeSelectToSpeculate(SelectInst &SI,
2509 SmallVectorImpl<LoadInst *> &Loads) {
2510 Value *TValue = SI.getTrueValue();
2511 Value *FValue = SI.getFalseValue();
2512 bool TDerefable = TValue->isDereferenceablePointer();
2513 bool FDerefable = FValue->isDereferenceablePointer();
2515 for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end();
2517 LoadInst *LI = dyn_cast<LoadInst>(*UI);
2518 if (LI == 0 || !LI->isSimple()) return false;
2520 // Both operands to the select need to be dereferencable, either
2521 // absolutely (e.g. allocas) or at this point because we can see other
2523 if (!TDerefable && !isSafeToLoadUnconditionally(TValue, LI,
2524 LI->getAlignment(), &TD))
2526 if (!FDerefable && !isSafeToLoadUnconditionally(FValue, LI,
2527 LI->getAlignment(), &TD))
2529 Loads.push_back(LI);
2535 bool visitSelectInst(SelectInst &SI) {
2536 DEBUG(dbgs() << " original: " << SI << "\n");
2537 IRBuilder<> IRB(&SI);
2539 // Find the operand we need to rewrite here.
2540 bool IsTrueVal = SI.getTrueValue() == OldPtr;
2542 assert(SI.getFalseValue() != OldPtr && "Pointer is both operands!");
2544 assert(SI.getFalseValue() == OldPtr && "Pointer isn't an operand!");
2545 Value *NewPtr = getAdjustedAllocaPtr(IRB, OldPtr->getType());
2547 // If the select isn't safe to speculate, just use simple logic to emit it.
2548 SmallVector<LoadInst *, 4> Loads;
2549 if (!isSafeSelectToSpeculate(SI, Loads)) {
2550 SI.setOperand(IsTrueVal ? 1 : 2, NewPtr);
2551 DEBUG(dbgs() << " to: " << SI << "\n");
2552 deleteIfTriviallyDead(OldPtr);
2556 Value *OtherPtr = IsTrueVal ? SI.getFalseValue() : SI.getTrueValue();
2557 AllocaPartitioning::iterator PI
2558 = P.findPartitionForPHIOrSelectOperand(SI, OtherPtr);
2559 AllocaPartitioning::PartitionUse OtherUse;
2560 if (PI != P.end()) {
2561 // If the other pointer is within the partitioning, remove the select
2562 // from its uses. We'll add in the new loads below.
2563 AllocaPartitioning::use_iterator UI
2564 = P.findPartitionUseForPHIOrSelectOperand(SI, OtherPtr);
2566 P.use_erase(PI, UI);
2569 Value *TV = IsTrueVal ? NewPtr : SI.getTrueValue();
2570 Value *FV = IsTrueVal ? SI.getFalseValue() : NewPtr;
2571 // Replace the loads of the select with a select of two loads.
2572 while (!Loads.empty()) {
2573 LoadInst *LI = Loads.pop_back_val();
2575 IRB.SetInsertPoint(LI);
2577 IRB.CreateLoad(TV, getName("." + LI->getName() + ".true"));
2579 IRB.CreateLoad(FV, getName("." + LI->getName() + ".false"));
2580 NumLoadsSpeculated += 2;
2581 if (PI != P.end()) {
2582 LoadInst *OtherLoad = IsTrueVal ? FL : TL;
2583 assert(OtherUse.Ptr == OtherLoad->getOperand(0));
2584 OtherUse.User = OtherLoad;
2585 P.use_push_back(PI, OtherUse);
2588 // Transfer alignment and TBAA info if present.
2589 TL->setAlignment(LI->getAlignment());
2590 FL->setAlignment(LI->getAlignment());
2591 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
2592 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
2593 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
2596 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL);
2598 DEBUG(dbgs() << " speculated to: " << *V << "\n");
2599 LI->replaceAllUsesWith(V);
2600 Pass.DeadInsts.push_back(LI);
2603 deleteIfTriviallyDead(OldPtr);
2604 return NewPtr == &NewAI;
2611 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2613 /// This pass aggressively rewrites all aggregate loads and stores on
2614 /// a particular pointer (or any pointer derived from it which we can identify)
2615 /// with scalar loads and stores.
2616 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2617 // Befriend the base class so it can delegate to private visit methods.
2618 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2620 const TargetData &TD;
2622 /// Queue of pointer uses to analyze and potentially rewrite.
2623 SmallVector<Use *, 8> Queue;
2625 /// Set to prevent us from cycling with phi nodes and loops.
2626 SmallPtrSet<User *, 8> Visited;
2628 /// The current pointer use being rewritten. This is used to dig up the used
2629 /// value (as opposed to the user).
2633 AggLoadStoreRewriter(const TargetData &TD) : TD(TD) {}
2635 /// Rewrite loads and stores through a pointer and all pointers derived from
2637 bool rewrite(Instruction &I) {
2638 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2640 bool Changed = false;
2641 while (!Queue.empty()) {
2642 U = Queue.pop_back_val();
2643 Changed |= visit(cast<Instruction>(U->getUser()));
2649 /// Enqueue all the users of the given instruction for further processing.
2650 /// This uses a set to de-duplicate users.
2651 void enqueueUsers(Instruction &I) {
2652 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
2654 if (Visited.insert(*UI))
2655 Queue.push_back(&UI.getUse());
2658 // Conservative default is to not rewrite anything.
2659 bool visitInstruction(Instruction &I) { return false; }
2661 /// \brief Generic recursive split emission class.
2662 template <typename Derived>
2665 /// The builder used to form new instructions.
2667 /// The indices which to be used with insert- or extractvalue to select the
2668 /// appropriate value within the aggregate.
2669 SmallVector<unsigned, 4> Indices;
2670 /// The indices to a GEP instruction which will move Ptr to the correct slot
2671 /// within the aggregate.
2672 SmallVector<Value *, 4> GEPIndices;
2673 /// The base pointer of the original op, used as a base for GEPing the
2674 /// split operations.
2677 /// Initialize the splitter with an insertion point, Ptr and start with a
2678 /// single zero GEP index.
2679 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
2680 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
2683 /// \brief Generic recursive split emission routine.
2685 /// This method recursively splits an aggregate op (load or store) into
2686 /// scalar or vector ops. It splits recursively until it hits a single value
2687 /// and emits that single value operation via the template argument.
2689 /// The logic of this routine relies on GEPs and insertvalue and
2690 /// extractvalue all operating with the same fundamental index list, merely
2691 /// formatted differently (GEPs need actual values).
2693 /// \param Ty The type being split recursively into smaller ops.
2694 /// \param Agg The aggregate value being built up or stored, depending on
2695 /// whether this is splitting a load or a store respectively.
2696 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
2697 if (Ty->isSingleValueType())
2698 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
2700 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2701 unsigned OldSize = Indices.size();
2703 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
2705 assert(Indices.size() == OldSize && "Did not return to the old size");
2706 Indices.push_back(Idx);
2707 GEPIndices.push_back(IRB.getInt32(Idx));
2708 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
2709 GEPIndices.pop_back();
2715 if (StructType *STy = dyn_cast<StructType>(Ty)) {
2716 unsigned OldSize = Indices.size();
2718 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
2720 assert(Indices.size() == OldSize && "Did not return to the old size");
2721 Indices.push_back(Idx);
2722 GEPIndices.push_back(IRB.getInt32(Idx));
2723 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
2724 GEPIndices.pop_back();
2730 llvm_unreachable("Only arrays and structs are aggregate loadable types");
2734 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
2735 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2736 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
2738 /// Emit a leaf load of a single value. This is called at the leaves of the
2739 /// recursive emission to actually load values.
2740 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2741 assert(Ty->isSingleValueType());
2742 // Load the single value and insert it using the indices.
2743 Value *Load = IRB.CreateLoad(IRB.CreateInBoundsGEP(Ptr, GEPIndices,
2746 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
2747 DEBUG(dbgs() << " to: " << *Load << "\n");
2751 bool visitLoadInst(LoadInst &LI) {
2752 assert(LI.getPointerOperand() == *U);
2753 if (!LI.isSimple() || LI.getType()->isSingleValueType())
2756 // We have an aggregate being loaded, split it apart.
2757 DEBUG(dbgs() << " original: " << LI << "\n");
2758 LoadOpSplitter Splitter(&LI, *U);
2759 Value *V = UndefValue::get(LI.getType());
2760 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
2761 LI.replaceAllUsesWith(V);
2762 LI.eraseFromParent();
2766 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
2767 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2768 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
2770 /// Emit a leaf store of a single value. This is called at the leaves of the
2771 /// recursive emission to actually produce stores.
2772 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2773 assert(Ty->isSingleValueType());
2774 // Extract the single value and store it using the indices.
2775 Value *Store = IRB.CreateStore(
2776 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
2777 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
2779 DEBUG(dbgs() << " to: " << *Store << "\n");
2783 bool visitStoreInst(StoreInst &SI) {
2784 if (!SI.isSimple() || SI.getPointerOperand() != *U)
2786 Value *V = SI.getValueOperand();
2787 if (V->getType()->isSingleValueType())
2790 // We have an aggregate being stored, split it apart.
2791 DEBUG(dbgs() << " original: " << SI << "\n");
2792 StoreOpSplitter Splitter(&SI, *U);
2793 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
2794 SI.eraseFromParent();
2798 bool visitBitCastInst(BitCastInst &BC) {
2803 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
2808 bool visitPHINode(PHINode &PN) {
2813 bool visitSelectInst(SelectInst &SI) {
2820 /// \brief Try to find a partition of the aggregate type passed in for a given
2821 /// offset and size.
2823 /// This recurses through the aggregate type and tries to compute a subtype
2824 /// based on the offset and size. When the offset and size span a sub-section
2825 /// of an array, it will even compute a new array type for that sub-section,
2826 /// and the same for structs.
2828 /// Note that this routine is very strict and tries to find a partition of the
2829 /// type which produces the *exact* right offset and size. It is not forgiving
2830 /// when the size or offset cause either end of type-based partition to be off.
2831 /// Also, this is a best-effort routine. It is reasonable to give up and not
2832 /// return a type if necessary.
2833 static Type *getTypePartition(const TargetData &TD, Type *Ty,
2834 uint64_t Offset, uint64_t Size) {
2835 if (Offset == 0 && TD.getTypeAllocSize(Ty) == Size)
2838 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
2839 // We can't partition pointers...
2840 if (SeqTy->isPointerTy())
2843 Type *ElementTy = SeqTy->getElementType();
2844 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
2845 uint64_t NumSkippedElements = Offset / ElementSize;
2846 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy))
2847 if (NumSkippedElements >= ArrTy->getNumElements())
2849 if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy))
2850 if (NumSkippedElements >= VecTy->getNumElements())
2852 Offset -= NumSkippedElements * ElementSize;
2854 // First check if we need to recurse.
2855 if (Offset > 0 || Size < ElementSize) {
2856 // Bail if the partition ends in a different array element.
2857 if ((Offset + Size) > ElementSize)
2859 // Recurse through the element type trying to peel off offset bytes.
2860 return getTypePartition(TD, ElementTy, Offset, Size);
2862 assert(Offset == 0);
2864 if (Size == ElementSize)
2866 assert(Size > ElementSize);
2867 uint64_t NumElements = Size / ElementSize;
2868 if (NumElements * ElementSize != Size)
2870 return ArrayType::get(ElementTy, NumElements);
2873 StructType *STy = dyn_cast<StructType>(Ty);
2877 const StructLayout *SL = TD.getStructLayout(STy);
2878 if (Offset >= SL->getSizeInBytes())
2880 uint64_t EndOffset = Offset + Size;
2881 if (EndOffset > SL->getSizeInBytes())
2884 unsigned Index = SL->getElementContainingOffset(Offset);
2885 Offset -= SL->getElementOffset(Index);
2887 Type *ElementTy = STy->getElementType(Index);
2888 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
2889 if (Offset >= ElementSize)
2890 return 0; // The offset points into alignment padding.
2892 // See if any partition must be contained by the element.
2893 if (Offset > 0 || Size < ElementSize) {
2894 if ((Offset + Size) > ElementSize)
2896 return getTypePartition(TD, ElementTy, Offset, Size);
2898 assert(Offset == 0);
2900 if (Size == ElementSize)
2903 StructType::element_iterator EI = STy->element_begin() + Index,
2904 EE = STy->element_end();
2905 if (EndOffset < SL->getSizeInBytes()) {
2906 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
2907 if (Index == EndIndex)
2908 return 0; // Within a single element and its padding.
2910 // Don't try to form "natural" types if the elements don't line up with the
2912 // FIXME: We could potentially recurse down through the last element in the
2913 // sub-struct to find a natural end point.
2914 if (SL->getElementOffset(EndIndex) != EndOffset)
2917 assert(Index < EndIndex);
2918 EE = STy->element_begin() + EndIndex;
2921 // Try to build up a sub-structure.
2922 SmallVector<Type *, 4> ElementTys;
2924 ElementTys.push_back(*EI++);
2926 StructType *SubTy = StructType::get(STy->getContext(), ElementTys,
2928 const StructLayout *SubSL = TD.getStructLayout(SubTy);
2929 if (Size != SubSL->getSizeInBytes())
2930 return 0; // The sub-struct doesn't have quite the size needed.
2935 /// \brief Rewrite an alloca partition's users.
2937 /// This routine drives both of the rewriting goals of the SROA pass. It tries
2938 /// to rewrite uses of an alloca partition to be conducive for SSA value
2939 /// promotion. If the partition needs a new, more refined alloca, this will
2940 /// build that new alloca, preserving as much type information as possible, and
2941 /// rewrite the uses of the old alloca to point at the new one and have the
2942 /// appropriate new offsets. It also evaluates how successful the rewrite was
2943 /// at enabling promotion and if it was successful queues the alloca to be
2945 bool SROA::rewriteAllocaPartition(AllocaInst &AI,
2946 AllocaPartitioning &P,
2947 AllocaPartitioning::iterator PI) {
2948 uint64_t AllocaSize = PI->EndOffset - PI->BeginOffset;
2949 if (P.use_begin(PI) == P.use_end(PI))
2950 return false; // No live uses left of this partition.
2952 // Try to compute a friendly type for this partition of the alloca. This
2953 // won't always succeed, in which case we fall back to a legal integer type
2954 // or an i8 array of an appropriate size.
2956 if (Type *PartitionTy = P.getCommonType(PI))
2957 if (TD->getTypeAllocSize(PartitionTy) >= AllocaSize)
2958 AllocaTy = PartitionTy;
2960 if (Type *PartitionTy = getTypePartition(*TD, AI.getAllocatedType(),
2961 PI->BeginOffset, AllocaSize))
2962 AllocaTy = PartitionTy;
2964 (AllocaTy->isArrayTy() &&
2965 AllocaTy->getArrayElementType()->isIntegerTy())) &&
2966 TD->isLegalInteger(AllocaSize * 8))
2967 AllocaTy = Type::getIntNTy(*C, AllocaSize * 8);
2969 AllocaTy = ArrayType::get(Type::getInt8Ty(*C), AllocaSize);
2970 assert(TD->getTypeAllocSize(AllocaTy) >= AllocaSize);
2972 // Check for the case where we're going to rewrite to a new alloca of the
2973 // exact same type as the original, and with the same access offsets. In that
2974 // case, re-use the existing alloca, but still run through the rewriter to
2975 // performe phi and select speculation.
2977 if (AllocaTy == AI.getAllocatedType()) {
2978 assert(PI->BeginOffset == 0 &&
2979 "Non-zero begin offset but same alloca type");
2980 assert(PI == P.begin() && "Begin offset is zero on later partition");
2983 // FIXME: The alignment here is overly conservative -- we could in many
2984 // cases get away with much weaker alignment constraints.
2985 NewAI = new AllocaInst(AllocaTy, 0, AI.getAlignment(),
2986 AI.getName() + ".sroa." + Twine(PI - P.begin()),
2991 DEBUG(dbgs() << "Rewriting alloca partition "
2992 << "[" << PI->BeginOffset << "," << PI->EndOffset << ") to: "
2995 AllocaPartitionRewriter Rewriter(*TD, P, PI, *this, AI, *NewAI,
2996 PI->BeginOffset, PI->EndOffset);
2997 DEBUG(dbgs() << " rewriting ");
2998 DEBUG(P.print(dbgs(), PI, ""));
2999 if (Rewriter.visitUsers(P.use_begin(PI), P.use_end(PI))) {
3000 DEBUG(dbgs() << " and queuing for promotion\n");
3001 PromotableAllocas.push_back(NewAI);
3002 } else if (NewAI != &AI) {
3003 // If we can't promote the alloca, iterate on it to check for new
3004 // refinements exposed by splitting the current alloca. Don't iterate on an
3005 // alloca which didn't actually change and didn't get promoted.
3006 Worklist.insert(NewAI);
3011 /// \brief Walks the partitioning of an alloca rewriting uses of each partition.
3012 bool SROA::splitAlloca(AllocaInst &AI, AllocaPartitioning &P) {
3013 bool Changed = false;
3014 for (AllocaPartitioning::iterator PI = P.begin(), PE = P.end(); PI != PE;
3016 Changed |= rewriteAllocaPartition(AI, P, PI);
3021 /// \brief Analyze an alloca for SROA.
3023 /// This analyzes the alloca to ensure we can reason about it, builds
3024 /// a partitioning of the alloca, and then hands it off to be split and
3025 /// rewritten as needed.
3026 bool SROA::runOnAlloca(AllocaInst &AI) {
3027 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3028 ++NumAllocasAnalyzed;
3030 // Special case dead allocas, as they're trivial.
3031 if (AI.use_empty()) {
3032 AI.eraseFromParent();
3036 // Skip alloca forms that this analysis can't handle.
3037 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3038 TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
3041 // First check if this is a non-aggregate type that we should simply promote.
3042 if (!AI.getAllocatedType()->isAggregateType() && isAllocaPromotable(&AI)) {
3043 DEBUG(dbgs() << " Trivially scalar type, queuing for promotion...\n");
3044 PromotableAllocas.push_back(&AI);
3048 bool Changed = false;
3050 // First, split any FCA loads and stores touching this alloca to promote
3051 // better splitting and promotion opportunities.
3052 AggLoadStoreRewriter AggRewriter(*TD);
3053 Changed |= AggRewriter.rewrite(AI);
3055 // Build the partition set using a recursive instruction-visiting builder.
3056 AllocaPartitioning P(*TD, AI);
3057 DEBUG(P.print(dbgs()));
3061 // No partitions to split. Leave the dead alloca for a later pass to clean up.
3062 if (P.begin() == P.end())
3065 // Delete all the dead users of this alloca before splitting and rewriting it.
3066 for (AllocaPartitioning::dead_user_iterator DI = P.dead_user_begin(),
3067 DE = P.dead_user_end();
3070 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3071 DeadInsts.push_back(*DI);
3073 for (AllocaPartitioning::dead_op_iterator DO = P.dead_op_begin(),
3074 DE = P.dead_op_end();
3077 // Clobber the use with an undef value.
3078 **DO = UndefValue::get(OldV->getType());
3079 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3080 if (isInstructionTriviallyDead(OldI)) {
3082 DeadInsts.push_back(OldI);
3086 return splitAlloca(AI, P) || Changed;
3089 /// \brief Delete the dead instructions accumulated in this run.
3091 /// Recursively deletes the dead instructions we've accumulated. This is done
3092 /// at the very end to maximize locality of the recursive delete and to
3093 /// minimize the problems of invalidated instruction pointers as such pointers
3094 /// are used heavily in the intermediate stages of the algorithm.
3096 /// We also record the alloca instructions deleted here so that they aren't
3097 /// subsequently handed to mem2reg to promote.
3098 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3099 DeadSplitInsts.clear();
3100 while (!DeadInsts.empty()) {
3101 Instruction *I = DeadInsts.pop_back_val();
3102 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3104 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
3105 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
3106 // Zero out the operand and see if it becomes trivially dead.
3108 if (isInstructionTriviallyDead(U))
3109 DeadInsts.push_back(U);
3112 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3113 DeletedAllocas.insert(AI);
3116 I->eraseFromParent();
3120 /// \brief Promote the allocas, using the best available technique.
3122 /// This attempts to promote whatever allocas have been identified as viable in
3123 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3124 /// If there is a domtree available, we attempt to promote using the full power
3125 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3126 /// based on the SSAUpdater utilities. This function returns whether any
3127 /// promotion occured.
3128 bool SROA::promoteAllocas(Function &F) {
3129 if (PromotableAllocas.empty())
3132 NumPromoted += PromotableAllocas.size();
3134 if (DT && !ForceSSAUpdater) {
3135 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3136 PromoteMemToReg(PromotableAllocas, *DT);
3137 PromotableAllocas.clear();
3141 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3143 DIBuilder DIB(*F.getParent());
3144 SmallVector<Instruction*, 64> Insts;
3146 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3147 AllocaInst *AI = PromotableAllocas[Idx];
3148 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
3150 Instruction *I = cast<Instruction>(*UI++);
3151 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3152 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3153 // leading to them) here. Eventually it should use them to optimize the
3154 // scalar values produced.
3155 if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
3156 assert(onlyUsedByLifetimeMarkers(I) &&
3157 "Found a bitcast used outside of a lifetime marker.");
3158 while (!I->use_empty())
3159 cast<Instruction>(*I->use_begin())->eraseFromParent();
3160 I->eraseFromParent();
3163 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3164 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3165 II->getIntrinsicID() == Intrinsic::lifetime_end);
3166 II->eraseFromParent();
3172 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3176 PromotableAllocas.clear();
3181 /// \brief A predicate to test whether an alloca belongs to a set.
3182 class IsAllocaInSet {
3183 typedef SmallPtrSet<AllocaInst *, 4> SetType;
3187 IsAllocaInSet(const SetType &Set) : Set(Set) {}
3188 bool operator()(AllocaInst *AI) { return Set.count(AI); }
3192 bool SROA::runOnFunction(Function &F) {
3193 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3194 C = &F.getContext();
3195 TD = getAnalysisIfAvailable<TargetData>();
3197 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3200 DT = getAnalysisIfAvailable<DominatorTree>();
3202 BasicBlock &EntryBB = F.getEntryBlock();
3203 for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end());
3205 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3206 Worklist.insert(AI);
3208 bool Changed = false;
3209 // A set of deleted alloca instruction pointers which should be removed from
3210 // the list of promotable allocas.
3211 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3213 while (!Worklist.empty()) {
3214 Changed |= runOnAlloca(*Worklist.pop_back_val());
3215 deleteDeadInstructions(DeletedAllocas);
3216 if (!DeletedAllocas.empty()) {
3217 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3218 PromotableAllocas.end(),
3219 IsAllocaInSet(DeletedAllocas)),
3220 PromotableAllocas.end());
3221 DeletedAllocas.clear();
3225 Changed |= promoteAllocas(F);
3230 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3231 if (RequiresDomTree)
3232 AU.addRequired<DominatorTree>();
3233 AU.setPreservesCFG();