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/ADT/STLExtras.h"
29 #include "llvm/ADT/SetVector.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/Analysis/Dominators.h"
33 #include "llvm/Analysis/Loads.h"
34 #include "llvm/Analysis/PtrUseVisitor.h"
35 #include "llvm/Analysis/ValueTracking.h"
36 #include "llvm/Constants.h"
37 #include "llvm/DIBuilder.h"
38 #include "llvm/DataLayout.h"
39 #include "llvm/DebugInfo.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/Function.h"
42 #include "llvm/IRBuilder.h"
43 #include "llvm/InstVisitor.h"
44 #include "llvm/Instructions.h"
45 #include "llvm/IntrinsicInst.h"
46 #include "llvm/LLVMContext.h"
47 #include "llvm/Module.h"
48 #include "llvm/Operator.h"
49 #include "llvm/Pass.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/MathExtras.h"
55 #include "llvm/Support/raw_ostream.h"
56 #include "llvm/Transforms/Utils/Local.h"
57 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
58 #include "llvm/Transforms/Utils/SSAUpdater.h"
61 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
62 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
63 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
64 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
65 STATISTIC(NumDeleted, "Number of instructions deleted");
66 STATISTIC(NumVectorized, "Number of vectorized aggregates");
68 /// Hidden option to force the pass to not use DomTree and mem2reg, instead
69 /// forming SSA values through the SSAUpdater infrastructure.
71 ForceSSAUpdater("force-ssa-updater", cl::init(false), cl::Hidden);
74 /// \brief Alloca partitioning representation.
76 /// This class represents a partitioning of an alloca into slices, and
77 /// information about the nature of uses of each slice of the alloca. The goal
78 /// is that this information is sufficient to decide if and how to split the
79 /// alloca apart and replace slices with scalars. It is also intended that this
80 /// structure can capture the relevant information needed both to decide about
81 /// and to enact these transformations.
82 class AllocaPartitioning {
84 /// \brief A common base class for representing a half-open byte range.
86 /// \brief The beginning offset of the range.
89 /// \brief The ending offset, not included in the range.
92 ByteRange() : BeginOffset(), EndOffset() {}
93 ByteRange(uint64_t BeginOffset, uint64_t EndOffset)
94 : BeginOffset(BeginOffset), EndOffset(EndOffset) {}
96 /// \brief Support for ordering ranges.
98 /// This provides an ordering over ranges such that start offsets are
99 /// always increasing, and within equal start offsets, the end offsets are
100 /// decreasing. Thus the spanning range comes first in a cluster with the
101 /// same start position.
102 bool operator<(const ByteRange &RHS) const {
103 if (BeginOffset < RHS.BeginOffset) return true;
104 if (BeginOffset > RHS.BeginOffset) return false;
105 if (EndOffset > RHS.EndOffset) return true;
109 /// \brief Support comparison with a single offset to allow binary searches.
110 friend bool operator<(const ByteRange &LHS, uint64_t RHSOffset) {
111 return LHS.BeginOffset < RHSOffset;
114 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
115 const ByteRange &RHS) {
116 return LHSOffset < RHS.BeginOffset;
119 bool operator==(const ByteRange &RHS) const {
120 return BeginOffset == RHS.BeginOffset && EndOffset == RHS.EndOffset;
122 bool operator!=(const ByteRange &RHS) const { return !operator==(RHS); }
125 /// \brief A partition of an alloca.
127 /// This structure represents a contiguous partition of the alloca. These are
128 /// formed by examining the uses of the alloca. During formation, they may
129 /// overlap but once an AllocaPartitioning is built, the Partitions within it
130 /// are all disjoint.
131 struct Partition : public ByteRange {
132 /// \brief Whether this partition is splittable into smaller partitions.
134 /// We flag partitions as splittable when they are formed entirely due to
135 /// accesses by trivially splittable operations such as memset and memcpy.
138 /// \brief Test whether a partition has been marked as dead.
139 bool isDead() const {
140 if (BeginOffset == UINT64_MAX) {
141 assert(EndOffset == UINT64_MAX);
147 /// \brief Kill a partition.
148 /// This is accomplished by setting both its beginning and end offset to
149 /// the maximum possible value.
151 assert(!isDead() && "He's Dead, Jim!");
152 BeginOffset = EndOffset = UINT64_MAX;
155 Partition() : ByteRange(), IsSplittable() {}
156 Partition(uint64_t BeginOffset, uint64_t EndOffset, bool IsSplittable)
157 : ByteRange(BeginOffset, EndOffset), IsSplittable(IsSplittable) {}
160 /// \brief A particular use of a partition of the alloca.
162 /// This structure is used to associate uses of a partition with it. They
163 /// mark the range of bytes which are referenced by a particular instruction,
164 /// and includes a handle to the user itself and the pointer value in use.
165 /// The bounds of these uses are determined by intersecting the bounds of the
166 /// memory use itself with a particular partition. As a consequence there is
167 /// intentionally overlap between various uses of the same partition.
168 struct PartitionUse : public ByteRange {
169 /// \brief The use in question. Provides access to both user and used value.
171 /// Note that this may be null if the partition use is *dead*, that is, it
172 /// should be ignored.
175 PartitionUse() : ByteRange(), U() {}
176 PartitionUse(uint64_t BeginOffset, uint64_t EndOffset, Use *U)
177 : ByteRange(BeginOffset, EndOffset), U(U) {}
180 /// \brief Construct a partitioning of a particular alloca.
182 /// Construction does most of the work for partitioning the alloca. This
183 /// performs the necessary walks of users and builds a partitioning from it.
184 AllocaPartitioning(const DataLayout &TD, AllocaInst &AI);
186 /// \brief Test whether a pointer to the allocation escapes our analysis.
188 /// If this is true, the partitioning is never fully built and should be
190 bool isEscaped() const { return PointerEscapingInstr; }
192 /// \brief Support for iterating over the partitions.
194 typedef SmallVectorImpl<Partition>::iterator iterator;
195 iterator begin() { return Partitions.begin(); }
196 iterator end() { return Partitions.end(); }
198 typedef SmallVectorImpl<Partition>::const_iterator const_iterator;
199 const_iterator begin() const { return Partitions.begin(); }
200 const_iterator end() const { return Partitions.end(); }
203 /// \brief Support for iterating over and manipulating a particular
204 /// partition's uses.
206 /// The iteration support provided for uses is more limited, but also
207 /// includes some manipulation routines to support rewriting the uses of
208 /// partitions during SROA.
210 typedef SmallVectorImpl<PartitionUse>::iterator use_iterator;
211 use_iterator use_begin(unsigned Idx) { return Uses[Idx].begin(); }
212 use_iterator use_begin(const_iterator I) { return Uses[I - begin()].begin(); }
213 use_iterator use_end(unsigned Idx) { return Uses[Idx].end(); }
214 use_iterator use_end(const_iterator I) { return Uses[I - begin()].end(); }
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();
226 unsigned use_size(unsigned Idx) const { return Uses[Idx].size(); }
227 unsigned use_size(const_iterator I) const { return Uses[I - begin()].size(); }
228 const PartitionUse &getUse(unsigned PIdx, unsigned UIdx) const {
229 return Uses[PIdx][UIdx];
231 const PartitionUse &getUse(const_iterator I, unsigned UIdx) const {
232 return Uses[I - begin()][UIdx];
235 void use_push_back(unsigned Idx, const PartitionUse &PU) {
236 Uses[Idx].push_back(PU);
238 void use_push_back(const_iterator I, const PartitionUse &PU) {
239 Uses[I - begin()].push_back(PU);
243 /// \brief Allow iterating the dead users for this alloca.
245 /// These are instructions which will never actually use the alloca as they
246 /// are outside the allocated range. They are safe to replace with undef and
249 typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator;
250 dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); }
251 dead_user_iterator dead_user_end() const { return DeadUsers.end(); }
254 /// \brief Allow iterating the dead expressions referring to this alloca.
256 /// These are operands which have cannot actually be used to refer to the
257 /// alloca as they are outside its range and the user doesn't correct for
258 /// that. These mostly consist of PHI node inputs and the like which we just
259 /// need to replace with undef.
261 typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator;
262 dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); }
263 dead_op_iterator dead_op_end() const { return DeadOperands.end(); }
266 /// \brief MemTransferInst auxiliary data.
267 /// This struct provides some auxiliary data about memory transfer
268 /// intrinsics such as memcpy and memmove. These intrinsics can use two
269 /// different ranges within the same alloca, and provide other challenges to
270 /// correctly represent. We stash extra data to help us untangle this
271 /// after the partitioning is complete.
272 struct MemTransferOffsets {
273 /// The destination begin and end offsets when the destination is within
274 /// this alloca. If the end offset is zero the destination is not within
276 uint64_t DestBegin, DestEnd;
278 /// The source begin and end offsets when the source is within this alloca.
279 /// If the end offset is zero, the source is not within this alloca.
280 uint64_t SourceBegin, SourceEnd;
282 /// Flag for whether an alloca is splittable.
285 MemTransferOffsets getMemTransferOffsets(MemTransferInst &II) const {
286 return MemTransferInstData.lookup(&II);
289 /// \brief Map from a PHI or select operand back to a partition.
291 /// When manipulating PHI nodes or selects, they can use more than one
292 /// partition of an alloca. We store a special mapping to allow finding the
293 /// partition referenced by each of these operands, if any.
294 iterator findPartitionForPHIOrSelectOperand(Use *U) {
295 SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt
296 = PHIOrSelectOpMap.find(U);
297 if (MapIt == PHIOrSelectOpMap.end())
300 return begin() + MapIt->second.first;
303 /// \brief Map from a PHI or select operand back to the specific use of
306 /// Similar to mapping these operands back to the partitions, this maps
307 /// directly to the use structure of that partition.
308 use_iterator findPartitionUseForPHIOrSelectOperand(Use *U) {
309 SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt
310 = PHIOrSelectOpMap.find(U);
311 assert(MapIt != PHIOrSelectOpMap.end());
312 return Uses[MapIt->second.first].begin() + MapIt->second.second;
315 /// \brief Compute a common type among the uses of a particular partition.
317 /// This routines walks all of the uses of a particular partition and tries
318 /// to find a common type between them. Untyped operations such as memset and
319 /// memcpy are ignored.
320 Type *getCommonType(iterator I) const;
322 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
323 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
324 void printUsers(raw_ostream &OS, const_iterator I,
325 StringRef Indent = " ") const;
326 void print(raw_ostream &OS) const;
327 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump(const_iterator I) const;
328 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump() const;
332 template <typename DerivedT, typename RetT = void> class BuilderBase;
333 class PartitionBuilder;
334 friend class AllocaPartitioning::PartitionBuilder;
336 friend class AllocaPartitioning::UseBuilder;
338 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
339 /// \brief Handle to alloca instruction to simplify method interfaces.
343 /// \brief The instruction responsible for this alloca having no partitioning.
345 /// When an instruction (potentially) escapes the pointer to the alloca, we
346 /// store a pointer to that here and abort trying to partition the alloca.
347 /// This will be null if the alloca is partitioned successfully.
348 Instruction *PointerEscapingInstr;
350 /// \brief The partitions of the alloca.
352 /// We store a vector of the partitions over the alloca here. This vector is
353 /// sorted by increasing begin offset, and then by decreasing end offset. See
354 /// the Partition inner class for more details. Initially (during
355 /// construction) there are overlaps, but we form a disjoint sequence of
356 /// partitions while finishing construction and a fully constructed object is
357 /// expected to always have this as a disjoint space.
358 SmallVector<Partition, 8> Partitions;
360 /// \brief The uses of the partitions.
362 /// This is essentially a mapping from each partition to a list of uses of
363 /// that partition. The mapping is done with a Uses vector that has the exact
364 /// same number of entries as the partition vector. Each entry is itself
365 /// a vector of the uses.
366 SmallVector<SmallVector<PartitionUse, 2>, 8> Uses;
368 /// \brief Instructions which will become dead if we rewrite the alloca.
370 /// Note that these are not separated by partition. This is because we expect
371 /// a partitioned alloca to be completely rewritten or not rewritten at all.
372 /// If rewritten, all these instructions can simply be removed and replaced
373 /// with undef as they come from outside of the allocated space.
374 SmallVector<Instruction *, 8> DeadUsers;
376 /// \brief Operands which will become dead if we rewrite the alloca.
378 /// These are operands that in their particular use can be replaced with
379 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
380 /// to PHI nodes and the like. They aren't entirely dead (there might be
381 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
382 /// want to swap this particular input for undef to simplify the use lists of
384 SmallVector<Use *, 8> DeadOperands;
386 /// \brief The underlying storage for auxiliary memcpy and memset info.
387 SmallDenseMap<MemTransferInst *, MemTransferOffsets, 4> MemTransferInstData;
389 /// \brief A side datastructure used when building up the partitions and uses.
391 /// This mapping is only really used during the initial building of the
392 /// partitioning so that we can retain information about PHI and select nodes
394 SmallDenseMap<Instruction *, std::pair<uint64_t, bool> > PHIOrSelectSizes;
396 /// \brief Auxiliary information for particular PHI or select operands.
397 SmallDenseMap<Use *, std::pair<unsigned, unsigned>, 4> PHIOrSelectOpMap;
399 /// \brief A utility routine called from the constructor.
401 /// This does what it says on the tin. It is the key of the alloca partition
402 /// splitting and merging. After it is called we have the desired disjoint
403 /// collection of partitions.
404 void splitAndMergePartitions();
408 static Value *foldSelectInst(SelectInst &SI) {
409 // If the condition being selected on is a constant or the same value is
410 // being selected between, fold the select. Yes this does (rarely) happen
412 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
413 return SI.getOperand(1+CI->isZero());
414 if (SI.getOperand(1) == SI.getOperand(2)) {
415 return SI.getOperand(1);
420 /// \brief Builder for the alloca partitioning.
422 /// This class builds an alloca partitioning by recursively visiting the uses
423 /// of an alloca and splitting the partitions for each load and store at each
425 class AllocaPartitioning::PartitionBuilder
426 : public PtrUseVisitor<PartitionBuilder> {
427 friend class PtrUseVisitor<PartitionBuilder>;
428 friend class InstVisitor<PartitionBuilder>;
429 typedef PtrUseVisitor<PartitionBuilder> Base;
431 const uint64_t AllocSize;
432 AllocaPartitioning &P;
434 SmallDenseMap<Instruction *, unsigned> MemTransferPartitionMap;
437 PartitionBuilder(const DataLayout &DL, AllocaInst &AI, AllocaPartitioning &P)
438 : PtrUseVisitor<PartitionBuilder>(DL),
439 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())),
443 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
444 bool IsSplittable = false) {
445 // Completely skip uses which have a zero size or start either before or
446 // past the end of the allocation.
447 if (Size == 0 || Offset.isNegative() || Offset.uge(AllocSize)) {
448 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
449 << " which has zero size or starts outside of the "
450 << AllocSize << " byte alloca:\n"
451 << " alloca: " << P.AI << "\n"
452 << " use: " << I << "\n");
456 uint64_t BeginOffset = Offset.getZExtValue();
457 uint64_t EndOffset = BeginOffset + Size;
459 // Clamp the end offset to the end of the allocation. Note that this is
460 // formulated to handle even the case where "BeginOffset + Size" overflows.
461 // NOTE! This may appear superficially to be something we could ignore
462 // entirely, but that is not so! There may be PHI-node uses where some
463 // instructions are dead but not others. We can't completely ignore the
464 // PHI node, and so have to record at least the information here.
465 assert(AllocSize >= BeginOffset); // Established above.
466 if (Size > AllocSize - BeginOffset) {
467 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
468 << " to remain within the " << AllocSize << " byte alloca:\n"
469 << " alloca: " << P.AI << "\n"
470 << " use: " << I << "\n");
471 EndOffset = AllocSize;
474 Partition New(BeginOffset, EndOffset, IsSplittable);
475 P.Partitions.push_back(New);
478 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
480 uint64_t Size = DL.getTypeStoreSize(Ty);
482 // If this memory access can be shown to *statically* extend outside the
483 // bounds of of the allocation, it's behavior is undefined, so simply
484 // ignore it. Note that this is more strict than the generic clamping
485 // behavior of insertUse. We also try to handle cases which might run the
487 // FIXME: We should instead consider the pointer to have escaped if this
488 // function is being instrumented for addressing bugs or race conditions.
489 if (Offset.isNegative() || Size > AllocSize ||
490 Offset.ugt(AllocSize - Size)) {
491 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte "
492 << (isa<LoadInst>(I) ? "load" : "store") << " @" << Offset
493 << " which extends past the end of the " << AllocSize
495 << " alloca: " << P.AI << "\n"
496 << " use: " << I << "\n");
500 // We allow splitting of loads and stores where the type is an integer type
501 // and which cover the entire alloca. Such integer loads and stores
502 // often require decomposition into fine grained loads and stores.
503 bool IsSplittable = false;
504 if (IntegerType *ITy = dyn_cast<IntegerType>(Ty))
505 IsSplittable = !IsVolatile && ITy->getBitWidth() == AllocSize*8;
507 insertUse(I, Offset, Size, IsSplittable);
510 void visitLoadInst(LoadInst &LI) {
511 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
512 "All simple FCA loads should have been pre-split");
515 return PI.setAborted(&LI);
517 return handleLoadOrStore(LI.getType(), LI, Offset, LI.isVolatile());
520 void visitStoreInst(StoreInst &SI) {
521 Value *ValOp = SI.getValueOperand();
523 return PI.setEscapedAndAborted(&SI);
525 return PI.setAborted(&SI);
527 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
528 "All simple FCA stores should have been pre-split");
529 handleLoadOrStore(ValOp->getType(), SI, Offset, SI.isVolatile());
533 void visitMemSetInst(MemSetInst &II) {
534 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
535 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
536 if ((Length && Length->getValue() == 0) ||
537 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
538 // Zero-length mem transfer intrinsics can be ignored entirely.
542 return PI.setAborted(&II);
544 insertUse(II, Offset,
545 Length ? Length->getLimitedValue()
546 : AllocSize - Offset.getLimitedValue(),
550 void visitMemTransferInst(MemTransferInst &II) {
551 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
552 if ((Length && Length->getValue() == 0) ||
553 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
554 // Zero-length mem transfer intrinsics can be ignored entirely.
558 return PI.setAborted(&II);
560 uint64_t RawOffset = Offset.getLimitedValue();
561 uint64_t Size = Length ? Length->getLimitedValue()
562 : AllocSize - RawOffset;
564 MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
566 // Only intrinsics with a constant length can be split.
567 Offsets.IsSplittable = Length;
569 if (*U == II.getRawDest()) {
570 Offsets.DestBegin = RawOffset;
571 Offsets.DestEnd = RawOffset + Size;
573 if (*U == II.getRawSource()) {
574 Offsets.SourceBegin = RawOffset;
575 Offsets.SourceEnd = RawOffset + Size;
578 // If we have set up end offsets for both the source and the destination,
579 // we have found both sides of this transfer pointing at the same alloca.
580 bool SeenBothEnds = Offsets.SourceEnd && Offsets.DestEnd;
581 if (SeenBothEnds && II.getRawDest() != II.getRawSource()) {
582 unsigned PrevIdx = MemTransferPartitionMap[&II];
584 // Check if the begin offsets match and this is a non-volatile transfer.
585 // In that case, we can completely elide the transfer.
586 if (!II.isVolatile() && Offsets.SourceBegin == Offsets.DestBegin) {
587 P.Partitions[PrevIdx].kill();
591 // Otherwise we have an offset transfer within the same alloca. We can't
593 P.Partitions[PrevIdx].IsSplittable = Offsets.IsSplittable = false;
594 } else if (SeenBothEnds) {
595 // Handle the case where this exact use provides both ends of the
597 assert(II.getRawDest() == II.getRawSource());
599 // For non-volatile transfers this is a no-op.
600 if (!II.isVolatile())
603 // Otherwise just suppress splitting.
604 Offsets.IsSplittable = false;
608 // Insert the use now that we've fixed up the splittable nature.
609 insertUse(II, Offset, Size, Offsets.IsSplittable);
611 // Setup the mapping from intrinsic to partition of we've not seen both
612 // ends of this transfer.
614 unsigned NewIdx = P.Partitions.size() - 1;
616 = MemTransferPartitionMap.insert(std::make_pair(&II, NewIdx)).second;
618 "Already have intrinsic in map but haven't seen both ends");
623 // Disable SRoA for any intrinsics except for lifetime invariants.
624 // FIXME: What about debug instrinsics? This matches old behavior, but
625 // doesn't make sense.
626 void visitIntrinsicInst(IntrinsicInst &II) {
628 return PI.setAborted(&II);
630 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
631 II.getIntrinsicID() == Intrinsic::lifetime_end) {
632 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
633 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
634 Length->getLimitedValue());
635 insertUse(II, Offset, Size, true);
639 Base::visitIntrinsicInst(II);
642 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
643 // We consider any PHI or select that results in a direct load or store of
644 // the same offset to be a viable use for partitioning purposes. These uses
645 // are considered unsplittable and the size is the maximum loaded or stored
647 SmallPtrSet<Instruction *, 4> Visited;
648 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
649 Visited.insert(Root);
650 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
651 // If there are no loads or stores, the access is dead. We mark that as
652 // a size zero access.
655 Instruction *I, *UsedI;
656 llvm::tie(UsedI, I) = Uses.pop_back_val();
658 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
659 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
662 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
663 Value *Op = SI->getOperand(0);
666 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
670 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
671 if (!GEP->hasAllZeroIndices())
673 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
674 !isa<SelectInst>(I)) {
678 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
680 if (Visited.insert(cast<Instruction>(*UI)))
681 Uses.push_back(std::make_pair(I, cast<Instruction>(*UI)));
682 } while (!Uses.empty());
687 void visitPHINode(PHINode &PN) {
691 return PI.setAborted(&PN);
693 // See if we already have computed info on this node.
694 std::pair<uint64_t, bool> &PHIInfo = P.PHIOrSelectSizes[&PN];
696 PHIInfo.second = true;
697 insertUse(PN, Offset, PHIInfo.first);
701 // Check for an unsafe use of the PHI node.
702 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&PN, PHIInfo.first))
703 return PI.setAborted(UnsafeI);
705 insertUse(PN, Offset, PHIInfo.first);
708 void visitSelectInst(SelectInst &SI) {
711 if (Value *Result = foldSelectInst(SI)) {
713 // If the result of the constant fold will be the pointer, recurse
714 // through the select as if we had RAUW'ed it.
720 return PI.setAborted(&SI);
722 // See if we already have computed info on this node.
723 std::pair<uint64_t, bool> &SelectInfo = P.PHIOrSelectSizes[&SI];
724 if (SelectInfo.first) {
725 SelectInfo.second = true;
726 insertUse(SI, Offset, SelectInfo.first);
730 // Check for an unsafe use of the PHI node.
731 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&SI, SelectInfo.first))
732 return PI.setAborted(UnsafeI);
734 insertUse(SI, Offset, SelectInfo.first);
737 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
738 void visitInstruction(Instruction &I) {
743 /// \brief Use adder for the alloca partitioning.
745 /// This class adds the uses of an alloca to all of the partitions which they
746 /// use. For splittable partitions, this can end up doing essentially a linear
747 /// walk of the partitions, but the number of steps remains bounded by the
748 /// total result instruction size:
749 /// - The number of partitions is a result of the number unsplittable
750 /// instructions using the alloca.
751 /// - The number of users of each partition is at worst the total number of
752 /// splittable instructions using the alloca.
753 /// Thus we will produce N * M instructions in the end, where N are the number
754 /// of unsplittable uses and M are the number of splittable. This visitor does
755 /// the exact same number of updates to the partitioning.
757 /// In the more common case, this visitor will leverage the fact that the
758 /// partition space is pre-sorted, and do a logarithmic search for the
759 /// partition needed, making the total visit a classical ((N + M) * log(N))
760 /// complexity operation.
761 class AllocaPartitioning::UseBuilder : public PtrUseVisitor<UseBuilder> {
762 friend class PtrUseVisitor<UseBuilder>;
763 friend class InstVisitor<UseBuilder>;
764 typedef PtrUseVisitor<UseBuilder> Base;
766 const uint64_t AllocSize;
767 AllocaPartitioning &P;
769 /// \brief Set to de-duplicate dead instructions found in the use walk.
770 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
773 UseBuilder(const DataLayout &TD, AllocaInst &AI, AllocaPartitioning &P)
774 : PtrUseVisitor<UseBuilder>(TD),
775 AllocSize(TD.getTypeAllocSize(AI.getAllocatedType())),
779 void markAsDead(Instruction &I) {
780 if (VisitedDeadInsts.insert(&I))
781 P.DeadUsers.push_back(&I);
784 void insertUse(Instruction &User, const APInt &Offset, uint64_t Size) {
785 // If the use has a zero size or extends outside of the allocation, record
786 // it as a dead use for elimination later.
787 if (Size == 0 || Offset.isNegative() || Offset.uge(AllocSize))
788 return markAsDead(User);
790 uint64_t BeginOffset = Offset.getZExtValue();
791 uint64_t EndOffset = BeginOffset + Size;
793 // Clamp the end offset to the end of the allocation. Note that this is
794 // formulated to handle even the case where "BeginOffset + Size" overflows.
795 assert(AllocSize >= BeginOffset); // Established above.
796 if (Size > AllocSize - BeginOffset)
797 EndOffset = AllocSize;
799 // NB: This only works if we have zero overlapping partitions.
800 iterator B = std::lower_bound(P.begin(), P.end(), BeginOffset);
801 if (B != P.begin() && llvm::prior(B)->EndOffset > BeginOffset)
803 for (iterator I = B, E = P.end(); I != E && I->BeginOffset < EndOffset;
805 PartitionUse NewPU(std::max(I->BeginOffset, BeginOffset),
806 std::min(I->EndOffset, EndOffset), U);
807 P.use_push_back(I, NewPU);
808 if (isa<PHINode>(U->getUser()) || isa<SelectInst>(U->getUser()))
809 P.PHIOrSelectOpMap[U]
810 = std::make_pair(I - P.begin(), P.Uses[I - P.begin()].size() - 1);
814 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset) {
815 uint64_t Size = DL.getTypeStoreSize(Ty);
817 // If this memory access can be shown to *statically* extend outside the
818 // bounds of of the allocation, it's behavior is undefined, so simply
819 // ignore it. Note that this is more strict than the generic clamping
820 // behavior of insertUse.
821 if (Offset.isNegative() || Size > AllocSize ||
822 Offset.ugt(AllocSize - Size))
823 return markAsDead(I);
825 insertUse(I, Offset, Size);
828 void visitBitCastInst(BitCastInst &BC) {
830 return markAsDead(BC);
832 return Base::visitBitCastInst(BC);
835 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
836 if (GEPI.use_empty())
837 return markAsDead(GEPI);
839 return Base::visitGetElementPtrInst(GEPI);
842 void visitLoadInst(LoadInst &LI) {
843 assert(IsOffsetKnown);
844 handleLoadOrStore(LI.getType(), LI, Offset);
847 void visitStoreInst(StoreInst &SI) {
848 assert(IsOffsetKnown);
849 handleLoadOrStore(SI.getOperand(0)->getType(), SI, Offset);
852 void visitMemSetInst(MemSetInst &II) {
853 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
854 if ((Length && Length->getValue() == 0) ||
855 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
856 return markAsDead(II);
858 assert(IsOffsetKnown);
859 insertUse(II, Offset, Length ? Length->getLimitedValue()
860 : AllocSize - Offset.getLimitedValue());
863 void visitMemTransferInst(MemTransferInst &II) {
864 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
865 if ((Length && Length->getValue() == 0) ||
866 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
867 return markAsDead(II);
869 assert(IsOffsetKnown);
870 uint64_t Size = Length ? Length->getLimitedValue()
871 : AllocSize - Offset.getLimitedValue();
873 MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
874 if (!II.isVolatile() && Offsets.DestEnd && Offsets.SourceEnd &&
875 Offsets.DestBegin == Offsets.SourceBegin)
876 return markAsDead(II); // Skip identity transfers without side-effects.
878 insertUse(II, Offset, Size);
881 void visitIntrinsicInst(IntrinsicInst &II) {
882 assert(IsOffsetKnown);
883 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
884 II.getIntrinsicID() == Intrinsic::lifetime_end);
886 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
887 insertUse(II, Offset, std::min(Length->getLimitedValue(),
888 AllocSize - Offset.getLimitedValue()));
891 void insertPHIOrSelect(Instruction &User, const APInt &Offset) {
892 uint64_t Size = P.PHIOrSelectSizes.lookup(&User).first;
894 // For PHI and select operands outside the alloca, we can't nuke the entire
895 // phi or select -- the other side might still be relevant, so we special
896 // case them here and use a separate structure to track the operands
897 // themselves which should be replaced with undef.
898 if ((Offset.isNegative() && Offset.uge(Size)) ||
899 (!Offset.isNegative() && Offset.uge(AllocSize))) {
900 P.DeadOperands.push_back(U);
904 insertUse(User, Offset, Size);
907 void visitPHINode(PHINode &PN) {
909 return markAsDead(PN);
911 assert(IsOffsetKnown);
912 insertPHIOrSelect(PN, Offset);
915 void visitSelectInst(SelectInst &SI) {
917 return markAsDead(SI);
919 if (Value *Result = foldSelectInst(SI)) {
921 // If the result of the constant fold will be the pointer, recurse
922 // through the select as if we had RAUW'ed it.
925 // Otherwise the operand to the select is dead, and we can replace it
927 P.DeadOperands.push_back(U);
932 assert(IsOffsetKnown);
933 insertPHIOrSelect(SI, Offset);
936 /// \brief Unreachable, we've already visited the alloca once.
937 void visitInstruction(Instruction &I) {
938 llvm_unreachable("Unhandled instruction in use builder.");
942 void AllocaPartitioning::splitAndMergePartitions() {
943 size_t NumDeadPartitions = 0;
945 // Track the range of splittable partitions that we pass when accumulating
946 // overlapping unsplittable partitions.
947 uint64_t SplitEndOffset = 0ull;
949 Partition New(0ull, 0ull, false);
951 for (unsigned i = 0, j = i, e = Partitions.size(); i != e; i = j) {
954 if (!Partitions[i].IsSplittable || New.BeginOffset == New.EndOffset) {
955 assert(New.BeginOffset == New.EndOffset);
958 assert(New.IsSplittable);
959 New.EndOffset = std::max(New.EndOffset, Partitions[i].EndOffset);
961 assert(New.BeginOffset != New.EndOffset);
963 // Scan the overlapping partitions.
964 while (j != e && New.EndOffset > Partitions[j].BeginOffset) {
965 // If the new partition we are forming is splittable, stop at the first
966 // unsplittable partition.
967 if (New.IsSplittable && !Partitions[j].IsSplittable)
970 // Grow the new partition to include any equally splittable range. 'j' is
971 // always equally splittable when New is splittable, but when New is not
972 // splittable, we may subsume some (or part of some) splitable partition
973 // without growing the new one.
974 if (New.IsSplittable == Partitions[j].IsSplittable) {
975 New.EndOffset = std::max(New.EndOffset, Partitions[j].EndOffset);
977 assert(!New.IsSplittable);
978 assert(Partitions[j].IsSplittable);
979 SplitEndOffset = std::max(SplitEndOffset, Partitions[j].EndOffset);
982 Partitions[j].kill();
987 // If the new partition is splittable, chop off the end as soon as the
988 // unsplittable subsequent partition starts and ensure we eventually cover
989 // the splittable area.
990 if (j != e && New.IsSplittable) {
991 SplitEndOffset = std::max(SplitEndOffset, New.EndOffset);
992 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
995 // Add the new partition if it differs from the original one and is
996 // non-empty. We can end up with an empty partition here if it was
997 // splittable but there is an unsplittable one that starts at the same
999 if (New != Partitions[i]) {
1000 if (New.BeginOffset != New.EndOffset)
1001 Partitions.push_back(New);
1002 // Mark the old one for removal.
1003 Partitions[i].kill();
1004 ++NumDeadPartitions;
1007 New.BeginOffset = New.EndOffset;
1008 if (!New.IsSplittable) {
1009 New.EndOffset = std::max(New.EndOffset, SplitEndOffset);
1010 if (j != e && !Partitions[j].IsSplittable)
1011 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
1012 New.IsSplittable = true;
1013 // If there is a trailing splittable partition which won't be fused into
1014 // the next splittable partition go ahead and add it onto the partitions
1016 if (New.BeginOffset < New.EndOffset &&
1017 (j == e || !Partitions[j].IsSplittable ||
1018 New.EndOffset < Partitions[j].BeginOffset)) {
1019 Partitions.push_back(New);
1020 New.BeginOffset = New.EndOffset = 0ull;
1025 // Re-sort the partitions now that they have been split and merged into
1026 // disjoint set of partitions. Also remove any of the dead partitions we've
1027 // replaced in the process.
1028 std::sort(Partitions.begin(), Partitions.end());
1029 if (NumDeadPartitions) {
1030 assert(Partitions.back().isDead());
1031 assert((ptrdiff_t)NumDeadPartitions ==
1032 std::count(Partitions.begin(), Partitions.end(), Partitions.back()));
1034 Partitions.erase(Partitions.end() - NumDeadPartitions, Partitions.end());
1037 AllocaPartitioning::AllocaPartitioning(const DataLayout &TD, AllocaInst &AI)
1039 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1042 PointerEscapingInstr(0) {
1043 PartitionBuilder PB(TD, AI, *this);
1044 PartitionBuilder::PtrInfo PtrI = PB.visitPtr(AI);
1045 if (PtrI.isEscaped() || PtrI.isAborted()) {
1046 // FIXME: We should sink the escape vs. abort info into the caller nicely,
1047 // possibly by just storing the PtrInfo in the AllocaPartitioning.
1048 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
1049 : PtrI.getAbortingInst();
1050 assert(PointerEscapingInstr && "Did not track a bad instruction");
1054 // Sort the uses. This arranges for the offsets to be in ascending order,
1055 // and the sizes to be in descending order.
1056 std::sort(Partitions.begin(), Partitions.end());
1058 // Remove any partitions from the back which are marked as dead.
1059 while (!Partitions.empty() && Partitions.back().isDead())
1060 Partitions.pop_back();
1062 if (Partitions.size() > 1) {
1063 // Intersect splittability for all partitions with equal offsets and sizes.
1064 // Then remove all but the first so that we have a sequence of non-equal but
1065 // potentially overlapping partitions.
1066 for (iterator I = Partitions.begin(), J = I, E = Partitions.end(); I != E;
1069 while (J != E && *I == *J) {
1070 I->IsSplittable &= J->IsSplittable;
1074 Partitions.erase(std::unique(Partitions.begin(), Partitions.end()),
1077 // Split splittable and merge unsplittable partitions into a disjoint set
1078 // of partitions over the used space of the allocation.
1079 splitAndMergePartitions();
1082 // Now build up the user lists for each of these disjoint partitions by
1083 // re-walking the recursive users of the alloca.
1084 Uses.resize(Partitions.size());
1085 UseBuilder UB(TD, AI, *this);
1086 PtrI = UB.visitPtr(AI);
1087 assert(!PtrI.isEscaped() && "Previously analyzed pointer now escapes!");
1088 assert(!PtrI.isAborted() && "Early aborted the visit of the pointer.");
1091 Type *AllocaPartitioning::getCommonType(iterator I) const {
1093 for (const_use_iterator UI = use_begin(I), UE = use_end(I); UI != UE; ++UI) {
1095 continue; // Skip dead uses.
1096 if (isa<IntrinsicInst>(*UI->U->getUser()))
1098 if (UI->BeginOffset != I->BeginOffset || UI->EndOffset != I->EndOffset)
1102 if (LoadInst *LI = dyn_cast<LoadInst>(UI->U->getUser())) {
1103 UserTy = LI->getType();
1104 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI->U->getUser())) {
1105 UserTy = SI->getValueOperand()->getType();
1107 return 0; // Bail if we have weird uses.
1110 if (IntegerType *ITy = dyn_cast<IntegerType>(UserTy)) {
1111 // If the type is larger than the partition, skip it. We only encounter
1112 // this for split integer operations where we want to use the type of the
1113 // entity causing the split.
1114 if (ITy->getBitWidth() > (I->EndOffset - I->BeginOffset)*8)
1117 // If we have found an integer type use covering the alloca, use that
1118 // regardless of the other types, as integers are often used for a "bucket
1123 if (Ty && Ty != UserTy)
1131 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1133 void AllocaPartitioning::print(raw_ostream &OS, const_iterator I,
1134 StringRef Indent) const {
1135 OS << Indent << "partition #" << (I - begin())
1136 << " [" << I->BeginOffset << "," << I->EndOffset << ")"
1137 << (I->IsSplittable ? " (splittable)" : "")
1138 << (Uses[I - begin()].empty() ? " (zero uses)" : "")
1142 void AllocaPartitioning::printUsers(raw_ostream &OS, const_iterator I,
1143 StringRef Indent) const {
1144 for (const_use_iterator UI = use_begin(I), UE = use_end(I);
1147 continue; // Skip dead uses.
1148 OS << Indent << " [" << UI->BeginOffset << "," << UI->EndOffset << ") "
1149 << "used by: " << *UI->U->getUser() << "\n";
1150 if (MemTransferInst *II = dyn_cast<MemTransferInst>(UI->U->getUser())) {
1151 const MemTransferOffsets &MTO = MemTransferInstData.lookup(II);
1153 if (!MTO.IsSplittable)
1154 IsDest = UI->BeginOffset == MTO.DestBegin;
1156 IsDest = MTO.DestBegin != 0u;
1157 OS << Indent << " (original " << (IsDest ? "dest" : "source") << ": "
1158 << "[" << (IsDest ? MTO.DestBegin : MTO.SourceBegin)
1159 << "," << (IsDest ? MTO.DestEnd : MTO.SourceEnd) << ")\n";
1164 void AllocaPartitioning::print(raw_ostream &OS) const {
1165 if (PointerEscapingInstr) {
1166 OS << "No partitioning for alloca: " << AI << "\n"
1167 << " A pointer to this alloca escaped by:\n"
1168 << " " << *PointerEscapingInstr << "\n";
1172 OS << "Partitioning of alloca: " << AI << "\n";
1174 for (const_iterator I = begin(), E = end(); I != E; ++I, ++Num) {
1180 void AllocaPartitioning::dump(const_iterator I) const { print(dbgs(), I); }
1181 void AllocaPartitioning::dump() const { print(dbgs()); }
1183 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1187 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
1189 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
1190 /// the loads and stores of an alloca instruction, as well as updating its
1191 /// debug information. This is used when a domtree is unavailable and thus
1192 /// mem2reg in its full form can't be used to handle promotion of allocas to
1194 class AllocaPromoter : public LoadAndStorePromoter {
1198 SmallVector<DbgDeclareInst *, 4> DDIs;
1199 SmallVector<DbgValueInst *, 4> DVIs;
1202 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
1203 AllocaInst &AI, DIBuilder &DIB)
1204 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
1206 void run(const SmallVectorImpl<Instruction*> &Insts) {
1207 // Remember which alloca we're promoting (for isInstInList).
1208 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
1209 for (Value::use_iterator UI = DebugNode->use_begin(),
1210 UE = DebugNode->use_end();
1212 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
1213 DDIs.push_back(DDI);
1214 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
1215 DVIs.push_back(DVI);
1218 LoadAndStorePromoter::run(Insts);
1219 AI.eraseFromParent();
1220 while (!DDIs.empty())
1221 DDIs.pop_back_val()->eraseFromParent();
1222 while (!DVIs.empty())
1223 DVIs.pop_back_val()->eraseFromParent();
1226 virtual bool isInstInList(Instruction *I,
1227 const SmallVectorImpl<Instruction*> &Insts) const {
1228 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1229 return LI->getOperand(0) == &AI;
1230 return cast<StoreInst>(I)->getPointerOperand() == &AI;
1233 virtual void updateDebugInfo(Instruction *Inst) const {
1234 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(),
1235 E = DDIs.end(); I != E; ++I) {
1236 DbgDeclareInst *DDI = *I;
1237 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
1238 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1239 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
1240 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1242 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(),
1243 E = DVIs.end(); I != E; ++I) {
1244 DbgValueInst *DVI = *I;
1246 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1247 // If an argument is zero extended then use argument directly. The ZExt
1248 // may be zapped by an optimization pass in future.
1249 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1250 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
1251 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1252 Arg = dyn_cast<Argument>(SExt->getOperand(0));
1254 Arg = SI->getOperand(0);
1255 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
1256 Arg = LI->getOperand(0);
1260 Instruction *DbgVal =
1261 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
1263 DbgVal->setDebugLoc(DVI->getDebugLoc());
1267 } // end anon namespace
1271 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
1273 /// This pass takes allocations which can be completely analyzed (that is, they
1274 /// don't escape) and tries to turn them into scalar SSA values. There are
1275 /// a few steps to this process.
1277 /// 1) It takes allocations of aggregates and analyzes the ways in which they
1278 /// are used to try to split them into smaller allocations, ideally of
1279 /// a single scalar data type. It will split up memcpy and memset accesses
1280 /// as necessary and try to isolate invidual scalar accesses.
1281 /// 2) It will transform accesses into forms which are suitable for SSA value
1282 /// promotion. This can be replacing a memset with a scalar store of an
1283 /// integer value, or it can involve speculating operations on a PHI or
1284 /// select to be a PHI or select of the results.
1285 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
1286 /// onto insert and extract operations on a vector value, and convert them to
1287 /// this form. By doing so, it will enable promotion of vector aggregates to
1288 /// SSA vector values.
1289 class SROA : public FunctionPass {
1290 const bool RequiresDomTree;
1293 const DataLayout *TD;
1296 /// \brief Worklist of alloca instructions to simplify.
1298 /// Each alloca in the function is added to this. Each new alloca formed gets
1299 /// added to it as well to recursively simplify unless that alloca can be
1300 /// directly promoted. Finally, each time we rewrite a use of an alloca other
1301 /// the one being actively rewritten, we add it back onto the list if not
1302 /// already present to ensure it is re-visited.
1303 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
1305 /// \brief A collection of instructions to delete.
1306 /// We try to batch deletions to simplify code and make things a bit more
1308 SetVector<Instruction *, SmallVector<Instruction *, 8> > DeadInsts;
1310 /// \brief Post-promotion worklist.
1312 /// Sometimes we discover an alloca which has a high probability of becoming
1313 /// viable for SROA after a round of promotion takes place. In those cases,
1314 /// the alloca is enqueued here for re-processing.
1316 /// Note that we have to be very careful to clear allocas out of this list in
1317 /// the event they are deleted.
1318 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist;
1320 /// \brief A collection of alloca instructions we can directly promote.
1321 std::vector<AllocaInst *> PromotableAllocas;
1324 SROA(bool RequiresDomTree = true)
1325 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
1326 C(0), TD(0), DT(0) {
1327 initializeSROAPass(*PassRegistry::getPassRegistry());
1329 bool runOnFunction(Function &F);
1330 void getAnalysisUsage(AnalysisUsage &AU) const;
1332 const char *getPassName() const { return "SROA"; }
1336 friend class PHIOrSelectSpeculator;
1337 friend class AllocaPartitionRewriter;
1338 friend class AllocaPartitionVectorRewriter;
1340 bool rewriteAllocaPartition(AllocaInst &AI,
1341 AllocaPartitioning &P,
1342 AllocaPartitioning::iterator PI);
1343 bool splitAlloca(AllocaInst &AI, AllocaPartitioning &P);
1344 bool runOnAlloca(AllocaInst &AI);
1345 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
1346 bool promoteAllocas(Function &F);
1352 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
1353 return new SROA(RequiresDomTree);
1356 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
1358 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
1359 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
1363 /// \brief Visitor to speculate PHIs and Selects where possible.
1364 class PHIOrSelectSpeculator : public InstVisitor<PHIOrSelectSpeculator> {
1365 // Befriend the base class so it can delegate to private visit methods.
1366 friend class llvm::InstVisitor<PHIOrSelectSpeculator>;
1368 const DataLayout &TD;
1369 AllocaPartitioning &P;
1373 PHIOrSelectSpeculator(const DataLayout &TD, AllocaPartitioning &P, SROA &Pass)
1374 : TD(TD), P(P), Pass(Pass) {}
1376 /// \brief Visit the users of an alloca partition and rewrite them.
1377 void visitUsers(AllocaPartitioning::const_iterator PI) {
1378 // Note that we need to use an index here as the underlying vector of uses
1379 // may be grown during speculation. However, we never need to re-visit the
1380 // new uses, and so we can use the initial size bound.
1381 for (unsigned Idx = 0, Size = P.use_size(PI); Idx != Size; ++Idx) {
1382 const AllocaPartitioning::PartitionUse &PU = P.getUse(PI, Idx);
1384 continue; // Skip dead use.
1386 visit(cast<Instruction>(PU.U->getUser()));
1391 // By default, skip this instruction.
1392 void visitInstruction(Instruction &I) {}
1394 /// PHI instructions that use an alloca and are subsequently loaded can be
1395 /// rewritten to load both input pointers in the pred blocks and then PHI the
1396 /// results, allowing the load of the alloca to be promoted.
1398 /// %P2 = phi [i32* %Alloca, i32* %Other]
1399 /// %V = load i32* %P2
1401 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1403 /// %V2 = load i32* %Other
1405 /// %V = phi [i32 %V1, i32 %V2]
1407 /// We can do this to a select if its only uses are loads and if the operands
1408 /// to the select can be loaded unconditionally.
1410 /// FIXME: This should be hoisted into a generic utility, likely in
1411 /// Transforms/Util/Local.h
1412 bool isSafePHIToSpeculate(PHINode &PN, SmallVectorImpl<LoadInst *> &Loads) {
1413 // For now, we can only do this promotion if the load is in the same block
1414 // as the PHI, and if there are no stores between the phi and load.
1415 // TODO: Allow recursive phi users.
1416 // TODO: Allow stores.
1417 BasicBlock *BB = PN.getParent();
1418 unsigned MaxAlign = 0;
1419 for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end();
1421 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1422 if (LI == 0 || !LI->isSimple()) return false;
1424 // For now we only allow loads in the same block as the PHI. This is
1425 // a common case that happens when instcombine merges two loads through
1427 if (LI->getParent() != BB) return false;
1429 // Ensure that there are no instructions between the PHI and the load that
1431 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1432 if (BBI->mayWriteToMemory())
1435 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1436 Loads.push_back(LI);
1439 // We can only transform this if it is safe to push the loads into the
1440 // predecessor blocks. The only thing to watch out for is that we can't put
1441 // a possibly trapping load in the predecessor if it is a critical edge.
1442 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num;
1444 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1445 Value *InVal = PN.getIncomingValue(Idx);
1447 // If the value is produced by the terminator of the predecessor (an
1448 // invoke) or it has side-effects, there is no valid place to put a load
1449 // in the predecessor.
1450 if (TI == InVal || TI->mayHaveSideEffects())
1453 // If the predecessor has a single successor, then the edge isn't
1455 if (TI->getNumSuccessors() == 1)
1458 // If this pointer is always safe to load, or if we can prove that there
1459 // is already a load in the block, then we can move the load to the pred
1461 if (InVal->isDereferenceablePointer() ||
1462 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, &TD))
1471 void visitPHINode(PHINode &PN) {
1472 DEBUG(dbgs() << " original: " << PN << "\n");
1474 SmallVector<LoadInst *, 4> Loads;
1475 if (!isSafePHIToSpeculate(PN, Loads))
1478 assert(!Loads.empty());
1480 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1481 IRBuilder<> PHIBuilder(&PN);
1482 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1483 PN.getName() + ".sroa.speculated");
1485 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1486 // matter which one we get and if any differ, it doesn't matter.
1487 LoadInst *SomeLoad = cast<LoadInst>(Loads.back());
1488 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1489 unsigned Align = SomeLoad->getAlignment();
1491 // Rewrite all loads of the PN to use the new PHI.
1493 LoadInst *LI = Loads.pop_back_val();
1494 LI->replaceAllUsesWith(NewPN);
1495 Pass.DeadInsts.insert(LI);
1496 } while (!Loads.empty());
1498 // Inject loads into all of the pred blocks.
1499 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1500 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1501 TerminatorInst *TI = Pred->getTerminator();
1502 Use *InUse = &PN.getOperandUse(PN.getOperandNumForIncomingValue(Idx));
1503 Value *InVal = PN.getIncomingValue(Idx);
1504 IRBuilder<> PredBuilder(TI);
1507 = PredBuilder.CreateLoad(InVal, (PN.getName() + ".sroa.speculate.load." +
1509 ++NumLoadsSpeculated;
1510 Load->setAlignment(Align);
1512 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1513 NewPN->addIncoming(Load, Pred);
1515 Instruction *Ptr = dyn_cast<Instruction>(InVal);
1517 // No uses to rewrite.
1520 // Try to lookup and rewrite any partition uses corresponding to this phi
1522 AllocaPartitioning::iterator PI
1523 = P.findPartitionForPHIOrSelectOperand(InUse);
1527 // Replace the Use in the PartitionUse for this operand with the Use
1529 AllocaPartitioning::use_iterator UI
1530 = P.findPartitionUseForPHIOrSelectOperand(InUse);
1531 assert(isa<PHINode>(*UI->U->getUser()));
1532 UI->U = &Load->getOperandUse(Load->getPointerOperandIndex());
1534 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1537 /// Select instructions that use an alloca and are subsequently loaded can be
1538 /// rewritten to load both input pointers and then select between the result,
1539 /// allowing the load of the alloca to be promoted.
1541 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1542 /// %V = load i32* %P2
1544 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1545 /// %V2 = load i32* %Other
1546 /// %V = select i1 %cond, i32 %V1, i32 %V2
1548 /// We can do this to a select if its only uses are loads and if the operand
1549 /// to the select can be loaded unconditionally.
1550 bool isSafeSelectToSpeculate(SelectInst &SI,
1551 SmallVectorImpl<LoadInst *> &Loads) {
1552 Value *TValue = SI.getTrueValue();
1553 Value *FValue = SI.getFalseValue();
1554 bool TDerefable = TValue->isDereferenceablePointer();
1555 bool FDerefable = FValue->isDereferenceablePointer();
1557 for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end();
1559 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1560 if (LI == 0 || !LI->isSimple()) return false;
1562 // Both operands to the select need to be dereferencable, either
1563 // absolutely (e.g. allocas) or at this point because we can see other
1565 if (!TDerefable && !isSafeToLoadUnconditionally(TValue, LI,
1566 LI->getAlignment(), &TD))
1568 if (!FDerefable && !isSafeToLoadUnconditionally(FValue, LI,
1569 LI->getAlignment(), &TD))
1571 Loads.push_back(LI);
1577 void visitSelectInst(SelectInst &SI) {
1578 DEBUG(dbgs() << " original: " << SI << "\n");
1579 IRBuilder<> IRB(&SI);
1581 // If the select isn't safe to speculate, just use simple logic to emit it.
1582 SmallVector<LoadInst *, 4> Loads;
1583 if (!isSafeSelectToSpeculate(SI, Loads))
1586 Use *Ops[2] = { &SI.getOperandUse(1), &SI.getOperandUse(2) };
1587 AllocaPartitioning::iterator PIs[2];
1588 AllocaPartitioning::PartitionUse PUs[2];
1589 for (unsigned i = 0, e = 2; i != e; ++i) {
1590 PIs[i] = P.findPartitionForPHIOrSelectOperand(Ops[i]);
1591 if (PIs[i] != P.end()) {
1592 // If the pointer is within the partitioning, remove the select from
1593 // its uses. We'll add in the new loads below.
1594 AllocaPartitioning::use_iterator UI
1595 = P.findPartitionUseForPHIOrSelectOperand(Ops[i]);
1597 // Clear out the use here so that the offsets into the use list remain
1598 // stable but this use is ignored when rewriting.
1603 Value *TV = SI.getTrueValue();
1604 Value *FV = SI.getFalseValue();
1605 // Replace the loads of the select with a select of two loads.
1606 while (!Loads.empty()) {
1607 LoadInst *LI = Loads.pop_back_val();
1609 IRB.SetInsertPoint(LI);
1611 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1613 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1614 NumLoadsSpeculated += 2;
1616 // Transfer alignment and TBAA info if present.
1617 TL->setAlignment(LI->getAlignment());
1618 FL->setAlignment(LI->getAlignment());
1619 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1620 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
1621 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
1624 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1625 LI->getName() + ".sroa.speculated");
1627 LoadInst *Loads[2] = { TL, FL };
1628 for (unsigned i = 0, e = 2; i != e; ++i) {
1629 if (PIs[i] != P.end()) {
1630 Use *LoadUse = &Loads[i]->getOperandUse(0);
1631 assert(PUs[i].U->get() == LoadUse->get());
1633 P.use_push_back(PIs[i], PUs[i]);
1637 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1638 LI->replaceAllUsesWith(V);
1639 Pass.DeadInsts.insert(LI);
1645 /// \brief Accumulate the constant offsets in a GEP into a single APInt offset.
1647 /// If the provided GEP is all-constant, the total byte offset formed by the
1648 /// GEP is computed and Offset is set to it. If the GEP has any non-constant
1649 /// operands, the function returns false and the value of Offset is unmodified.
1650 static bool accumulateGEPOffsets(const DataLayout &TD, GEPOperator &GEP,
1652 APInt GEPOffset(Offset.getBitWidth(), 0);
1653 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1654 GTI != GTE; ++GTI) {
1655 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
1658 if (OpC->isZero()) continue;
1660 // Handle a struct index, which adds its field offset to the pointer.
1661 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1662 unsigned ElementIdx = OpC->getZExtValue();
1663 const StructLayout *SL = TD.getStructLayout(STy);
1664 GEPOffset += APInt(Offset.getBitWidth(),
1665 SL->getElementOffset(ElementIdx));
1669 APInt TypeSize(Offset.getBitWidth(),
1670 TD.getTypeAllocSize(GTI.getIndexedType()));
1671 if (VectorType *VTy = dyn_cast<VectorType>(*GTI)) {
1672 assert((VTy->getScalarSizeInBits() % 8) == 0 &&
1673 "vector element size is not a multiple of 8, cannot GEP over it");
1674 TypeSize = VTy->getScalarSizeInBits() / 8;
1677 GEPOffset += OpC->getValue().sextOrTrunc(Offset.getBitWidth()) * TypeSize;
1683 /// \brief Build a GEP out of a base pointer and indices.
1685 /// This will return the BasePtr if that is valid, or build a new GEP
1686 /// instruction using the IRBuilder if GEP-ing is needed.
1687 static Value *buildGEP(IRBuilder<> &IRB, Value *BasePtr,
1688 SmallVectorImpl<Value *> &Indices,
1689 const Twine &Prefix) {
1690 if (Indices.empty())
1693 // A single zero index is a no-op, so check for this and avoid building a GEP
1695 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1698 return IRB.CreateInBoundsGEP(BasePtr, Indices, Prefix + ".idx");
1701 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1702 /// TargetTy without changing the offset of the pointer.
1704 /// This routine assumes we've already established a properly offset GEP with
1705 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1706 /// zero-indices down through type layers until we find one the same as
1707 /// TargetTy. If we can't find one with the same type, we at least try to use
1708 /// one with the same size. If none of that works, we just produce the GEP as
1709 /// indicated by Indices to have the correct offset.
1710 static Value *getNaturalGEPWithType(IRBuilder<> &IRB, const DataLayout &TD,
1711 Value *BasePtr, Type *Ty, Type *TargetTy,
1712 SmallVectorImpl<Value *> &Indices,
1713 const Twine &Prefix) {
1715 return buildGEP(IRB, BasePtr, Indices, Prefix);
1717 // See if we can descend into a struct and locate a field with the correct
1719 unsigned NumLayers = 0;
1720 Type *ElementTy = Ty;
1722 if (ElementTy->isPointerTy())
1724 if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
1725 ElementTy = SeqTy->getElementType();
1726 // Note that we use the default address space as this index is over an
1727 // array or a vector, not a pointer.
1728 Indices.push_back(IRB.getInt(APInt(TD.getPointerSizeInBits(0), 0)));
1729 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1730 if (STy->element_begin() == STy->element_end())
1731 break; // Nothing left to descend into.
1732 ElementTy = *STy->element_begin();
1733 Indices.push_back(IRB.getInt32(0));
1738 } while (ElementTy != TargetTy);
1739 if (ElementTy != TargetTy)
1740 Indices.erase(Indices.end() - NumLayers, Indices.end());
1742 return buildGEP(IRB, BasePtr, Indices, Prefix);
1745 /// \brief Recursively compute indices for a natural GEP.
1747 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1748 /// element types adding appropriate indices for the GEP.
1749 static Value *getNaturalGEPRecursively(IRBuilder<> &IRB, const DataLayout &TD,
1750 Value *Ptr, Type *Ty, APInt &Offset,
1752 SmallVectorImpl<Value *> &Indices,
1753 const Twine &Prefix) {
1755 return getNaturalGEPWithType(IRB, TD, Ptr, Ty, TargetTy, Indices, Prefix);
1757 // We can't recurse through pointer types.
1758 if (Ty->isPointerTy())
1761 // We try to analyze GEPs over vectors here, but note that these GEPs are
1762 // extremely poorly defined currently. The long-term goal is to remove GEPing
1763 // over a vector from the IR completely.
1764 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1765 unsigned ElementSizeInBits = VecTy->getScalarSizeInBits();
1766 if (ElementSizeInBits % 8)
1767 return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
1768 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1769 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1770 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1772 Offset -= NumSkippedElements * ElementSize;
1773 Indices.push_back(IRB.getInt(NumSkippedElements));
1774 return getNaturalGEPRecursively(IRB, TD, Ptr, VecTy->getElementType(),
1775 Offset, TargetTy, Indices, Prefix);
1778 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1779 Type *ElementTy = ArrTy->getElementType();
1780 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1781 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1782 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1785 Offset -= NumSkippedElements * ElementSize;
1786 Indices.push_back(IRB.getInt(NumSkippedElements));
1787 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1791 StructType *STy = dyn_cast<StructType>(Ty);
1795 const StructLayout *SL = TD.getStructLayout(STy);
1796 uint64_t StructOffset = Offset.getZExtValue();
1797 if (StructOffset >= SL->getSizeInBytes())
1799 unsigned Index = SL->getElementContainingOffset(StructOffset);
1800 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1801 Type *ElementTy = STy->getElementType(Index);
1802 if (Offset.uge(TD.getTypeAllocSize(ElementTy)))
1803 return 0; // The offset points into alignment padding.
1805 Indices.push_back(IRB.getInt32(Index));
1806 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1810 /// \brief Get a natural GEP from a base pointer to a particular offset and
1811 /// resulting in a particular type.
1813 /// The goal is to produce a "natural" looking GEP that works with the existing
1814 /// composite types to arrive at the appropriate offset and element type for
1815 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1816 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1817 /// Indices, and setting Ty to the result subtype.
1819 /// If no natural GEP can be constructed, this function returns null.
1820 static Value *getNaturalGEPWithOffset(IRBuilder<> &IRB, const DataLayout &TD,
1821 Value *Ptr, APInt Offset, Type *TargetTy,
1822 SmallVectorImpl<Value *> &Indices,
1823 const Twine &Prefix) {
1824 PointerType *Ty = cast<PointerType>(Ptr->getType());
1826 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1828 if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
1831 Type *ElementTy = Ty->getElementType();
1832 if (!ElementTy->isSized())
1833 return 0; // We can't GEP through an unsized element.
1834 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1835 if (ElementSize == 0)
1836 return 0; // Zero-length arrays can't help us build a natural GEP.
1837 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1839 Offset -= NumSkippedElements * ElementSize;
1840 Indices.push_back(IRB.getInt(NumSkippedElements));
1841 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1845 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1846 /// resulting pointer has PointerTy.
1848 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1849 /// and produces the pointer type desired. Where it cannot, it will try to use
1850 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1851 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1852 /// bitcast to the type.
1854 /// The strategy for finding the more natural GEPs is to peel off layers of the
1855 /// pointer, walking back through bit casts and GEPs, searching for a base
1856 /// pointer from which we can compute a natural GEP with the desired
1857 /// properities. The algorithm tries to fold as many constant indices into
1858 /// a single GEP as possible, thus making each GEP more independent of the
1859 /// surrounding code.
1860 static Value *getAdjustedPtr(IRBuilder<> &IRB, const DataLayout &TD,
1861 Value *Ptr, APInt Offset, Type *PointerTy,
1862 const Twine &Prefix) {
1863 // Even though we don't look through PHI nodes, we could be called on an
1864 // instruction in an unreachable block, which may be on a cycle.
1865 SmallPtrSet<Value *, 4> Visited;
1866 Visited.insert(Ptr);
1867 SmallVector<Value *, 4> Indices;
1869 // We may end up computing an offset pointer that has the wrong type. If we
1870 // never are able to compute one directly that has the correct type, we'll
1871 // fall back to it, so keep it around here.
1872 Value *OffsetPtr = 0;
1874 // Remember any i8 pointer we come across to re-use if we need to do a raw
1877 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1879 Type *TargetTy = PointerTy->getPointerElementType();
1882 // First fold any existing GEPs into the offset.
1883 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1884 APInt GEPOffset(Offset.getBitWidth(), 0);
1885 if (!accumulateGEPOffsets(TD, *GEP, GEPOffset))
1887 Offset += GEPOffset;
1888 Ptr = GEP->getPointerOperand();
1889 if (!Visited.insert(Ptr))
1893 // See if we can perform a natural GEP here.
1895 if (Value *P = getNaturalGEPWithOffset(IRB, TD, Ptr, Offset, TargetTy,
1897 if (P->getType() == PointerTy) {
1898 // Zap any offset pointer that we ended up computing in previous rounds.
1899 if (OffsetPtr && OffsetPtr->use_empty())
1900 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1901 I->eraseFromParent();
1909 // Stash this pointer if we've found an i8*.
1910 if (Ptr->getType()->isIntegerTy(8)) {
1912 Int8PtrOffset = Offset;
1915 // Peel off a layer of the pointer and update the offset appropriately.
1916 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1917 Ptr = cast<Operator>(Ptr)->getOperand(0);
1918 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1919 if (GA->mayBeOverridden())
1921 Ptr = GA->getAliasee();
1925 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1926 } while (Visited.insert(Ptr));
1930 Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
1931 Prefix + ".raw_cast");
1932 Int8PtrOffset = Offset;
1935 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1936 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1937 Prefix + ".raw_idx");
1941 // On the off chance we were targeting i8*, guard the bitcast here.
1942 if (Ptr->getType() != PointerTy)
1943 Ptr = IRB.CreateBitCast(Ptr, PointerTy, Prefix + ".cast");
1948 /// \brief Test whether we can convert a value from the old to the new type.
1950 /// This predicate should be used to guard calls to convertValue in order to
1951 /// ensure that we only try to convert viable values. The strategy is that we
1952 /// will peel off single element struct and array wrappings to get to an
1953 /// underlying value, and convert that value.
1954 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1957 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1959 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1962 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1963 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1965 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1973 /// \brief Generic routine to convert an SSA value to a value of a different
1976 /// This will try various different casting techniques, such as bitcasts,
1977 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1978 /// two types for viability with this routine.
1979 static Value *convertValue(const DataLayout &DL, IRBuilder<> &IRB, Value *V,
1981 assert(canConvertValue(DL, V->getType(), Ty) &&
1982 "Value not convertable to type");
1983 if (V->getType() == Ty)
1985 if (V->getType()->isIntegerTy() && Ty->isPointerTy())
1986 return IRB.CreateIntToPtr(V, Ty);
1987 if (V->getType()->isPointerTy() && Ty->isIntegerTy())
1988 return IRB.CreatePtrToInt(V, Ty);
1990 return IRB.CreateBitCast(V, Ty);
1993 /// \brief Test whether the given alloca partition can be promoted to a vector.
1995 /// This is a quick test to check whether we can rewrite a particular alloca
1996 /// partition (and its newly formed alloca) into a vector alloca with only
1997 /// whole-vector loads and stores such that it could be promoted to a vector
1998 /// SSA value. We only can ensure this for a limited set of operations, and we
1999 /// don't want to do the rewrites unless we are confident that the result will
2000 /// be promotable, so we have an early test here.
2001 static bool isVectorPromotionViable(const DataLayout &TD,
2003 AllocaPartitioning &P,
2004 uint64_t PartitionBeginOffset,
2005 uint64_t PartitionEndOffset,
2006 AllocaPartitioning::const_use_iterator I,
2007 AllocaPartitioning::const_use_iterator E) {
2008 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
2012 uint64_t VecSize = TD.getTypeSizeInBits(Ty);
2013 uint64_t ElementSize = Ty->getScalarSizeInBits();
2015 // While the definition of LLVM vectors is bitpacked, we don't support sizes
2016 // that aren't byte sized.
2017 if (ElementSize % 8)
2019 assert((VecSize % 8) == 0 && "vector size not a multiple of element size?");
2023 for (; I != E; ++I) {
2025 continue; // Skip dead use.
2027 uint64_t BeginOffset = I->BeginOffset - PartitionBeginOffset;
2028 uint64_t BeginIndex = BeginOffset / ElementSize;
2029 if (BeginIndex * ElementSize != BeginOffset ||
2030 BeginIndex >= Ty->getNumElements())
2032 uint64_t EndOffset = I->EndOffset - PartitionBeginOffset;
2033 uint64_t EndIndex = EndOffset / ElementSize;
2034 if (EndIndex * ElementSize != EndOffset ||
2035 EndIndex > Ty->getNumElements())
2038 assert(EndIndex > BeginIndex && "Empty vector!");
2039 uint64_t NumElements = EndIndex - BeginIndex;
2041 = (NumElements == 1) ? Ty->getElementType()
2042 : VectorType::get(Ty->getElementType(), NumElements);
2044 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
2045 if (MI->isVolatile())
2047 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
2048 const AllocaPartitioning::MemTransferOffsets &MTO
2049 = P.getMemTransferOffsets(*MTI);
2050 if (!MTO.IsSplittable)
2053 } else if (I->U->get()->getType()->getPointerElementType()->isStructTy()) {
2054 // Disable vector promotion when there are loads or stores of an FCA.
2056 } else if (LoadInst *LI = dyn_cast<LoadInst>(I->U->getUser())) {
2057 if (LI->isVolatile())
2059 if (!canConvertValue(TD, PartitionTy, LI->getType()))
2061 } else if (StoreInst *SI = dyn_cast<StoreInst>(I->U->getUser())) {
2062 if (SI->isVolatile())
2064 if (!canConvertValue(TD, SI->getValueOperand()->getType(), PartitionTy))
2073 /// \brief Test whether the given alloca partition's integer operations can be
2074 /// widened to promotable ones.
2076 /// This is a quick test to check whether we can rewrite the integer loads and
2077 /// stores to a particular alloca into wider loads and stores and be able to
2078 /// promote the resulting alloca.
2079 static bool isIntegerWideningViable(const DataLayout &TD,
2081 uint64_t AllocBeginOffset,
2082 AllocaPartitioning &P,
2083 AllocaPartitioning::const_use_iterator I,
2084 AllocaPartitioning::const_use_iterator E) {
2085 uint64_t SizeInBits = TD.getTypeSizeInBits(AllocaTy);
2086 // Don't create integer types larger than the maximum bitwidth.
2087 if (SizeInBits > IntegerType::MAX_INT_BITS)
2090 // Don't try to handle allocas with bit-padding.
2091 if (SizeInBits != TD.getTypeStoreSizeInBits(AllocaTy))
2094 // We need to ensure that an integer type with the appropriate bitwidth can
2095 // be converted to the alloca type, whatever that is. We don't want to force
2096 // the alloca itself to have an integer type if there is a more suitable one.
2097 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
2098 if (!canConvertValue(TD, AllocaTy, IntTy) ||
2099 !canConvertValue(TD, IntTy, AllocaTy))
2102 uint64_t Size = TD.getTypeStoreSize(AllocaTy);
2104 // Check the uses to ensure the uses are (likely) promoteable integer uses.
2105 // Also ensure that the alloca has a covering load or store. We don't want
2106 // to widen the integer operotains only to fail to promote due to some other
2107 // unsplittable entry (which we may make splittable later).
2108 bool WholeAllocaOp = false;
2109 for (; I != E; ++I) {
2111 continue; // Skip dead use.
2113 uint64_t RelBegin = I->BeginOffset - AllocBeginOffset;
2114 uint64_t RelEnd = I->EndOffset - AllocBeginOffset;
2116 // We can't reasonably handle cases where the load or store extends past
2117 // the end of the aloca's type and into its padding.
2121 if (LoadInst *LI = dyn_cast<LoadInst>(I->U->getUser())) {
2122 if (LI->isVolatile())
2124 if (RelBegin == 0 && RelEnd == Size)
2125 WholeAllocaOp = true;
2126 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
2127 if (ITy->getBitWidth() < TD.getTypeStoreSizeInBits(ITy))
2131 // Non-integer loads need to be convertible from the alloca type so that
2132 // they are promotable.
2133 if (RelBegin != 0 || RelEnd != Size ||
2134 !canConvertValue(TD, AllocaTy, LI->getType()))
2136 } else if (StoreInst *SI = dyn_cast<StoreInst>(I->U->getUser())) {
2137 Type *ValueTy = SI->getValueOperand()->getType();
2138 if (SI->isVolatile())
2140 if (RelBegin == 0 && RelEnd == Size)
2141 WholeAllocaOp = true;
2142 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
2143 if (ITy->getBitWidth() < TD.getTypeStoreSizeInBits(ITy))
2147 // Non-integer stores need to be convertible to the alloca type so that
2148 // they are promotable.
2149 if (RelBegin != 0 || RelEnd != Size ||
2150 !canConvertValue(TD, ValueTy, AllocaTy))
2152 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
2153 if (MI->isVolatile())
2155 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
2156 const AllocaPartitioning::MemTransferOffsets &MTO
2157 = P.getMemTransferOffsets(*MTI);
2158 if (!MTO.IsSplittable)
2161 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->U->getUser())) {
2162 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2163 II->getIntrinsicID() != Intrinsic::lifetime_end)
2169 return WholeAllocaOp;
2172 static Value *extractInteger(const DataLayout &DL, IRBuilder<> &IRB, Value *V,
2173 IntegerType *Ty, uint64_t Offset,
2174 const Twine &Name) {
2175 DEBUG(dbgs() << " start: " << *V << "\n");
2176 IntegerType *IntTy = cast<IntegerType>(V->getType());
2177 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2178 "Element extends past full value");
2179 uint64_t ShAmt = 8*Offset;
2180 if (DL.isBigEndian())
2181 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2183 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2184 DEBUG(dbgs() << " shifted: " << *V << "\n");
2186 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2187 "Cannot extract to a larger integer!");
2189 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2190 DEBUG(dbgs() << " trunced: " << *V << "\n");
2195 static Value *insertInteger(const DataLayout &DL, IRBuilder<> &IRB, Value *Old,
2196 Value *V, uint64_t Offset, const Twine &Name) {
2197 IntegerType *IntTy = cast<IntegerType>(Old->getType());
2198 IntegerType *Ty = cast<IntegerType>(V->getType());
2199 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2200 "Cannot insert a larger integer!");
2201 DEBUG(dbgs() << " start: " << *V << "\n");
2203 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2204 DEBUG(dbgs() << " extended: " << *V << "\n");
2206 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2207 "Element store outside of alloca store");
2208 uint64_t ShAmt = 8*Offset;
2209 if (DL.isBigEndian())
2210 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2212 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2213 DEBUG(dbgs() << " shifted: " << *V << "\n");
2216 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2217 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2218 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2219 DEBUG(dbgs() << " masked: " << *Old << "\n");
2220 V = IRB.CreateOr(Old, V, Name + ".insert");
2221 DEBUG(dbgs() << " inserted: " << *V << "\n");
2226 static Value *extractVector(IRBuilder<> &IRB, Value *V,
2227 unsigned BeginIndex, unsigned EndIndex,
2228 const Twine &Name) {
2229 VectorType *VecTy = cast<VectorType>(V->getType());
2230 unsigned NumElements = EndIndex - BeginIndex;
2231 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2233 if (NumElements == VecTy->getNumElements())
2236 if (NumElements == 1) {
2237 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2239 DEBUG(dbgs() << " extract: " << *V << "\n");
2243 SmallVector<Constant*, 8> Mask;
2244 Mask.reserve(NumElements);
2245 for (unsigned i = BeginIndex; i != EndIndex; ++i)
2246 Mask.push_back(IRB.getInt32(i));
2247 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2248 ConstantVector::get(Mask),
2250 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2254 static Value *insertVector(IRBuilder<> &IRB, Value *Old, Value *V,
2255 unsigned BeginIndex, const Twine &Name) {
2256 VectorType *VecTy = cast<VectorType>(Old->getType());
2257 assert(VecTy && "Can only insert a vector into a vector");
2259 VectorType *Ty = dyn_cast<VectorType>(V->getType());
2261 // Single element to insert.
2262 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2264 DEBUG(dbgs() << " insert: " << *V << "\n");
2268 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
2269 "Too many elements!");
2270 if (Ty->getNumElements() == VecTy->getNumElements()) {
2271 assert(V->getType() == VecTy && "Vector type mismatch");
2274 unsigned EndIndex = BeginIndex + Ty->getNumElements();
2276 // When inserting a smaller vector into the larger to store, we first
2277 // use a shuffle vector to widen it with undef elements, and then
2278 // a second shuffle vector to select between the loaded vector and the
2280 SmallVector<Constant*, 8> Mask;
2281 Mask.reserve(VecTy->getNumElements());
2282 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2283 if (i >= BeginIndex && i < EndIndex)
2284 Mask.push_back(IRB.getInt32(i - BeginIndex));
2286 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2287 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2288 ConstantVector::get(Mask),
2290 DEBUG(dbgs() << " shuffle1: " << *V << "\n");
2293 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2294 if (i >= BeginIndex && i < EndIndex)
2295 Mask.push_back(IRB.getInt32(i));
2297 Mask.push_back(IRB.getInt32(i + VecTy->getNumElements()));
2298 V = IRB.CreateShuffleVector(V, Old, ConstantVector::get(Mask),
2300 DEBUG(dbgs() << " shuffle2: " << *V << "\n");
2305 /// \brief Visitor to rewrite instructions using a partition of an alloca to
2306 /// use a new alloca.
2308 /// Also implements the rewriting to vector-based accesses when the partition
2309 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2311 class AllocaPartitionRewriter : public InstVisitor<AllocaPartitionRewriter,
2313 // Befriend the base class so it can delegate to private visit methods.
2314 friend class llvm::InstVisitor<AllocaPartitionRewriter, bool>;
2316 const DataLayout &TD;
2317 AllocaPartitioning &P;
2319 AllocaInst &OldAI, &NewAI;
2320 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2323 // If we are rewriting an alloca partition which can be written as pure
2324 // vector operations, we stash extra information here. When VecTy is
2325 // non-null, we have some strict guarantees about the rewriten alloca:
2326 // - The new alloca is exactly the size of the vector type here.
2327 // - The accesses all either map to the entire vector or to a single
2329 // - The set of accessing instructions is only one of those handled above
2330 // in isVectorPromotionViable. Generally these are the same access kinds
2331 // which are promotable via mem2reg.
2334 uint64_t ElementSize;
2336 // This is a convenience and flag variable that will be null unless the new
2337 // alloca's integer operations should be widened to this integer type due to
2338 // passing isIntegerWideningViable above. If it is non-null, the desired
2339 // integer type will be stored here for easy access during rewriting.
2342 // The offset of the partition user currently being rewritten.
2343 uint64_t BeginOffset, EndOffset;
2345 Instruction *OldPtr;
2347 // The name prefix to use when rewriting instructions for this alloca.
2348 std::string NamePrefix;
2351 AllocaPartitionRewriter(const DataLayout &TD, AllocaPartitioning &P,
2352 AllocaPartitioning::iterator PI,
2353 SROA &Pass, AllocaInst &OldAI, AllocaInst &NewAI,
2354 uint64_t NewBeginOffset, uint64_t NewEndOffset)
2355 : TD(TD), P(P), Pass(Pass),
2356 OldAI(OldAI), NewAI(NewAI),
2357 NewAllocaBeginOffset(NewBeginOffset),
2358 NewAllocaEndOffset(NewEndOffset),
2359 NewAllocaTy(NewAI.getAllocatedType()),
2360 VecTy(), ElementTy(), ElementSize(), IntTy(),
2361 BeginOffset(), EndOffset() {
2364 /// \brief Visit the users of the alloca partition and rewrite them.
2365 bool visitUsers(AllocaPartitioning::const_use_iterator I,
2366 AllocaPartitioning::const_use_iterator E) {
2367 if (isVectorPromotionViable(TD, NewAI.getAllocatedType(), P,
2368 NewAllocaBeginOffset, NewAllocaEndOffset,
2371 VecTy = cast<VectorType>(NewAI.getAllocatedType());
2372 ElementTy = VecTy->getElementType();
2373 assert((VecTy->getScalarSizeInBits() % 8) == 0 &&
2374 "Only multiple-of-8 sized vector elements are viable");
2375 ElementSize = VecTy->getScalarSizeInBits() / 8;
2376 } else if (isIntegerWideningViable(TD, NewAI.getAllocatedType(),
2377 NewAllocaBeginOffset, P, I, E)) {
2378 IntTy = Type::getIntNTy(NewAI.getContext(),
2379 TD.getTypeSizeInBits(NewAI.getAllocatedType()));
2381 bool CanSROA = true;
2382 for (; I != E; ++I) {
2384 continue; // Skip dead uses.
2385 BeginOffset = I->BeginOffset;
2386 EndOffset = I->EndOffset;
2388 OldPtr = cast<Instruction>(I->U->get());
2389 NamePrefix = (Twine(NewAI.getName()) + "." + Twine(BeginOffset)).str();
2390 CanSROA &= visit(cast<Instruction>(I->U->getUser()));
2406 // Every instruction which can end up as a user must have a rewrite rule.
2407 bool visitInstruction(Instruction &I) {
2408 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2409 llvm_unreachable("No rewrite rule for this instruction!");
2412 Twine getName(const Twine &Suffix) {
2413 return NamePrefix + Suffix;
2416 Value *getAdjustedAllocaPtr(IRBuilder<> &IRB, Type *PointerTy) {
2417 assert(BeginOffset >= NewAllocaBeginOffset);
2418 APInt Offset(TD.getPointerSizeInBits(), BeginOffset - NewAllocaBeginOffset);
2419 return getAdjustedPtr(IRB, TD, &NewAI, Offset, PointerTy, getName(""));
2422 /// \brief Compute suitable alignment to access an offset into the new alloca.
2423 unsigned getOffsetAlign(uint64_t Offset) {
2424 unsigned NewAIAlign = NewAI.getAlignment();
2426 NewAIAlign = TD.getABITypeAlignment(NewAI.getAllocatedType());
2427 return MinAlign(NewAIAlign, Offset);
2430 /// \brief Compute suitable alignment to access this partition of the new
2432 unsigned getPartitionAlign() {
2433 return getOffsetAlign(BeginOffset - NewAllocaBeginOffset);
2436 /// \brief Compute suitable alignment to access a type at an offset of the
2439 /// \returns zero if the type's ABI alignment is a suitable alignment,
2440 /// otherwise returns the maximal suitable alignment.
2441 unsigned getOffsetTypeAlign(Type *Ty, uint64_t Offset) {
2442 unsigned Align = getOffsetAlign(Offset);
2443 return Align == TD.getABITypeAlignment(Ty) ? 0 : Align;
2446 /// \brief Compute suitable alignment to access a type at the beginning of
2447 /// this partition of the new alloca.
2449 /// See \c getOffsetTypeAlign for details; this routine delegates to it.
2450 unsigned getPartitionTypeAlign(Type *Ty) {
2451 return getOffsetTypeAlign(Ty, BeginOffset - NewAllocaBeginOffset);
2454 unsigned getIndex(uint64_t Offset) {
2455 assert(VecTy && "Can only call getIndex when rewriting a vector");
2456 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2457 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2458 uint32_t Index = RelOffset / ElementSize;
2459 assert(Index * ElementSize == RelOffset);
2463 void deleteIfTriviallyDead(Value *V) {
2464 Instruction *I = cast<Instruction>(V);
2465 if (isInstructionTriviallyDead(I))
2466 Pass.DeadInsts.insert(I);
2469 Value *rewriteVectorizedLoadInst(IRBuilder<> &IRB) {
2470 unsigned BeginIndex = getIndex(BeginOffset);
2471 unsigned EndIndex = getIndex(EndOffset);
2472 assert(EndIndex > BeginIndex && "Empty vector!");
2474 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2476 return extractVector(IRB, V, BeginIndex, EndIndex, getName(".vec"));
2479 Value *rewriteIntegerLoad(IRBuilder<> &IRB, LoadInst &LI) {
2480 assert(IntTy && "We cannot insert an integer to the alloca");
2481 assert(!LI.isVolatile());
2482 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2484 V = convertValue(TD, IRB, V, IntTy);
2485 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2486 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2487 if (Offset > 0 || EndOffset < NewAllocaEndOffset)
2488 V = extractInteger(TD, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2489 getName(".extract"));
2493 bool visitLoadInst(LoadInst &LI) {
2494 DEBUG(dbgs() << " original: " << LI << "\n");
2495 Value *OldOp = LI.getOperand(0);
2496 assert(OldOp == OldPtr);
2497 IRBuilder<> IRB(&LI);
2499 uint64_t Size = EndOffset - BeginOffset;
2500 bool IsSplitIntLoad = Size < TD.getTypeStoreSize(LI.getType());
2502 // If this memory access can be shown to *statically* extend outside the
2503 // bounds of the original allocation it's behavior is undefined. Rather
2504 // than trying to transform it, just replace it with undef.
2505 // FIXME: We should do something more clever for functions being
2506 // instrumented by asan.
2507 // FIXME: Eventually, once ASan and friends can flush out bugs here, this
2508 // should be transformed to a load of null making it unreachable.
2509 uint64_t OldAllocSize = TD.getTypeAllocSize(OldAI.getAllocatedType());
2510 if (TD.getTypeStoreSize(LI.getType()) > OldAllocSize) {
2511 LI.replaceAllUsesWith(UndefValue::get(LI.getType()));
2512 Pass.DeadInsts.insert(&LI);
2513 deleteIfTriviallyDead(OldOp);
2514 DEBUG(dbgs() << " to: undef!!\n");
2518 Type *TargetTy = IsSplitIntLoad ? Type::getIntNTy(LI.getContext(), Size * 8)
2520 bool IsPtrAdjusted = false;
2523 V = rewriteVectorizedLoadInst(IRB);
2524 } else if (IntTy && LI.getType()->isIntegerTy()) {
2525 V = rewriteIntegerLoad(IRB, LI);
2526 } else if (BeginOffset == NewAllocaBeginOffset &&
2527 canConvertValue(TD, NewAllocaTy, LI.getType())) {
2528 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2529 LI.isVolatile(), getName(".load"));
2531 Type *LTy = TargetTy->getPointerTo();
2532 V = IRB.CreateAlignedLoad(getAdjustedAllocaPtr(IRB, LTy),
2533 getPartitionTypeAlign(TargetTy),
2534 LI.isVolatile(), getName(".load"));
2535 IsPtrAdjusted = true;
2537 V = convertValue(TD, IRB, V, TargetTy);
2539 if (IsSplitIntLoad) {
2540 assert(!LI.isVolatile());
2541 assert(LI.getType()->isIntegerTy() &&
2542 "Only integer type loads and stores are split");
2543 assert(LI.getType()->getIntegerBitWidth() ==
2544 TD.getTypeStoreSizeInBits(LI.getType()) &&
2545 "Non-byte-multiple bit width");
2546 assert(LI.getType()->getIntegerBitWidth() ==
2547 TD.getTypeAllocSizeInBits(OldAI.getAllocatedType()) &&
2548 "Only alloca-wide loads can be split and recomposed");
2549 // Move the insertion point just past the load so that we can refer to it.
2550 IRB.SetInsertPoint(llvm::next(BasicBlock::iterator(&LI)));
2551 // Create a placeholder value with the same type as LI to use as the
2552 // basis for the new value. This allows us to replace the uses of LI with
2553 // the computed value, and then replace the placeholder with LI, leaving
2554 // LI only used for this computation.
2556 = new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2557 V = insertInteger(TD, IRB, Placeholder, V, BeginOffset,
2558 getName(".insert"));
2559 LI.replaceAllUsesWith(V);
2560 Placeholder->replaceAllUsesWith(&LI);
2563 LI.replaceAllUsesWith(V);
2566 Pass.DeadInsts.insert(&LI);
2567 deleteIfTriviallyDead(OldOp);
2568 DEBUG(dbgs() << " to: " << *V << "\n");
2569 return !LI.isVolatile() && !IsPtrAdjusted;
2572 bool rewriteVectorizedStoreInst(IRBuilder<> &IRB, Value *V,
2573 StoreInst &SI, Value *OldOp) {
2574 unsigned BeginIndex = getIndex(BeginOffset);
2575 unsigned EndIndex = getIndex(EndOffset);
2576 assert(EndIndex > BeginIndex && "Empty vector!");
2577 unsigned NumElements = EndIndex - BeginIndex;
2578 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2580 = (NumElements == 1) ? ElementTy
2581 : VectorType::get(ElementTy, NumElements);
2582 if (V->getType() != PartitionTy)
2583 V = convertValue(TD, IRB, V, PartitionTy);
2585 // Mix in the existing elements.
2586 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2588 V = insertVector(IRB, Old, V, BeginIndex, getName(".vec"));
2590 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2591 Pass.DeadInsts.insert(&SI);
2594 DEBUG(dbgs() << " to: " << *Store << "\n");
2598 bool rewriteIntegerStore(IRBuilder<> &IRB, Value *V, StoreInst &SI) {
2599 assert(IntTy && "We cannot extract an integer from the alloca");
2600 assert(!SI.isVolatile());
2601 if (TD.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2602 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2603 getName(".oldload"));
2604 Old = convertValue(TD, IRB, Old, IntTy);
2605 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2606 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2607 V = insertInteger(TD, IRB, Old, SI.getValueOperand(), Offset,
2608 getName(".insert"));
2610 V = convertValue(TD, IRB, V, NewAllocaTy);
2611 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2612 Pass.DeadInsts.insert(&SI);
2614 DEBUG(dbgs() << " to: " << *Store << "\n");
2618 bool visitStoreInst(StoreInst &SI) {
2619 DEBUG(dbgs() << " original: " << SI << "\n");
2620 Value *OldOp = SI.getOperand(1);
2621 assert(OldOp == OldPtr);
2622 IRBuilder<> IRB(&SI);
2624 Value *V = SI.getValueOperand();
2626 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2627 // alloca that should be re-examined after promoting this alloca.
2628 if (V->getType()->isPointerTy())
2629 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2630 Pass.PostPromotionWorklist.insert(AI);
2632 uint64_t Size = EndOffset - BeginOffset;
2633 if (Size < TD.getTypeStoreSize(V->getType())) {
2634 assert(!SI.isVolatile());
2635 assert(V->getType()->isIntegerTy() &&
2636 "Only integer type loads and stores are split");
2637 assert(V->getType()->getIntegerBitWidth() ==
2638 TD.getTypeStoreSizeInBits(V->getType()) &&
2639 "Non-byte-multiple bit width");
2640 assert(V->getType()->getIntegerBitWidth() ==
2641 TD.getTypeAllocSizeInBits(OldAI.getAllocatedType()) &&
2642 "Only alloca-wide stores can be split and recomposed");
2643 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), Size * 8);
2644 V = extractInteger(TD, IRB, V, NarrowTy, BeginOffset,
2645 getName(".extract"));
2649 return rewriteVectorizedStoreInst(IRB, V, SI, OldOp);
2650 if (IntTy && V->getType()->isIntegerTy())
2651 return rewriteIntegerStore(IRB, V, SI);
2654 if (BeginOffset == NewAllocaBeginOffset &&
2655 canConvertValue(TD, V->getType(), NewAllocaTy)) {
2656 V = convertValue(TD, IRB, V, NewAllocaTy);
2657 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2660 Value *NewPtr = getAdjustedAllocaPtr(IRB, V->getType()->getPointerTo());
2661 NewSI = IRB.CreateAlignedStore(V, NewPtr,
2662 getPartitionTypeAlign(V->getType()),
2666 Pass.DeadInsts.insert(&SI);
2667 deleteIfTriviallyDead(OldOp);
2669 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2670 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2673 /// \brief Compute an integer value from splatting an i8 across the given
2674 /// number of bytes.
2676 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2677 /// call this routine.
2678 /// FIXME: Heed the abvice above.
2680 /// \param V The i8 value to splat.
2681 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2682 Value *getIntegerSplat(IRBuilder<> &IRB, Value *V, unsigned Size) {
2683 assert(Size > 0 && "Expected a positive number of bytes.");
2684 IntegerType *VTy = cast<IntegerType>(V->getType());
2685 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2689 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size*8);
2690 V = IRB.CreateMul(IRB.CreateZExt(V, SplatIntTy, getName(".zext")),
2691 ConstantExpr::getUDiv(
2692 Constant::getAllOnesValue(SplatIntTy),
2693 ConstantExpr::getZExt(
2694 Constant::getAllOnesValue(V->getType()),
2696 getName(".isplat"));
2700 /// \brief Compute a vector splat for a given element value.
2701 Value *getVectorSplat(IRBuilder<> &IRB, Value *V, unsigned NumElements) {
2702 assert(NumElements > 0 && "Cannot splat to an empty vector.");
2704 // First insert it into a one-element vector so we can shuffle it. It is
2705 // really silly that LLVM's IR requires this in order to form a splat.
2706 Value *Undef = UndefValue::get(VectorType::get(V->getType(), 1));
2707 V = IRB.CreateInsertElement(Undef, V, IRB.getInt32(0),
2708 getName(".splatinsert"));
2710 // Shuffle the value across the desired number of elements.
2711 SmallVector<Constant*, 8> Mask(NumElements, IRB.getInt32(0));
2712 V = IRB.CreateShuffleVector(V, Undef, ConstantVector::get(Mask),
2714 DEBUG(dbgs() << " splat: " << *V << "\n");
2718 bool visitMemSetInst(MemSetInst &II) {
2719 DEBUG(dbgs() << " original: " << II << "\n");
2720 IRBuilder<> IRB(&II);
2721 assert(II.getRawDest() == OldPtr);
2723 // If the memset has a variable size, it cannot be split, just adjust the
2724 // pointer to the new alloca.
2725 if (!isa<Constant>(II.getLength())) {
2726 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2727 Type *CstTy = II.getAlignmentCst()->getType();
2728 II.setAlignment(ConstantInt::get(CstTy, getPartitionAlign()));
2730 deleteIfTriviallyDead(OldPtr);
2734 // Record this instruction for deletion.
2735 Pass.DeadInsts.insert(&II);
2737 Type *AllocaTy = NewAI.getAllocatedType();
2738 Type *ScalarTy = AllocaTy->getScalarType();
2740 // If this doesn't map cleanly onto the alloca type, and that type isn't
2741 // a single value type, just emit a memset.
2742 if (!VecTy && !IntTy &&
2743 (BeginOffset != NewAllocaBeginOffset ||
2744 EndOffset != NewAllocaEndOffset ||
2745 !AllocaTy->isSingleValueType() ||
2746 !TD.isLegalInteger(TD.getTypeSizeInBits(ScalarTy)) ||
2747 TD.getTypeSizeInBits(ScalarTy)%8 != 0)) {
2748 Type *SizeTy = II.getLength()->getType();
2749 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2751 = IRB.CreateMemSet(getAdjustedAllocaPtr(IRB,
2752 II.getRawDest()->getType()),
2753 II.getValue(), Size, getPartitionAlign(),
2756 DEBUG(dbgs() << " to: " << *New << "\n");
2760 // If we can represent this as a simple value, we have to build the actual
2761 // value to store, which requires expanding the byte present in memset to
2762 // a sensible representation for the alloca type. This is essentially
2763 // splatting the byte to a sufficiently wide integer, splatting it across
2764 // any desired vector width, and bitcasting to the final type.
2765 uint64_t Size = EndOffset - BeginOffset;
2766 Value *V = getIntegerSplat(IRB, II.getValue(), Size);
2769 // If this is a memset of a vectorized alloca, insert it.
2770 assert(ElementTy == ScalarTy);
2772 unsigned BeginIndex = getIndex(BeginOffset);
2773 unsigned EndIndex = getIndex(EndOffset);
2774 assert(EndIndex > BeginIndex && "Empty vector!");
2775 unsigned NumElements = EndIndex - BeginIndex;
2776 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2778 Value *Splat = getIntegerSplat(IRB, II.getValue(),
2779 TD.getTypeSizeInBits(ElementTy)/8);
2780 if (NumElements > 1) {
2781 Splat = getVectorSplat(IRB, Splat, NumElements);
2783 Type *SplatVecTy = VectorType::get(ElementTy, NumElements);
2784 if (Splat->getType() != SplatVecTy)
2785 Splat = convertValue(TD, IRB, Splat, SplatVecTy);
2788 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2789 getName(".oldload"));
2790 V = insertVector(IRB, Old, Splat, BeginIndex, getName(".vec"));
2792 // If this is a memset on an alloca where we can widen stores, insert the
2794 assert(!II.isVolatile());
2796 V = getIntegerSplat(IRB, II.getValue(), Size);
2798 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2799 EndOffset != NewAllocaBeginOffset)) {
2800 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2801 getName(".oldload"));
2802 Old = convertValue(TD, IRB, Old, IntTy);
2803 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2804 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2805 V = insertInteger(TD, IRB, Old, V, Offset, getName(".insert"));
2807 assert(V->getType() == IntTy &&
2808 "Wrong type for an alloca wide integer!");
2810 V = convertValue(TD, IRB, V, AllocaTy);
2812 // Established these invariants above.
2813 assert(BeginOffset == NewAllocaBeginOffset);
2814 assert(EndOffset == NewAllocaEndOffset);
2816 V = getIntegerSplat(IRB, II.getValue(),
2817 TD.getTypeSizeInBits(ScalarTy)/8);
2818 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2819 V = getVectorSplat(IRB, V, AllocaVecTy->getNumElements());
2821 V = convertValue(TD, IRB, V, AllocaTy);
2824 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2827 DEBUG(dbgs() << " to: " << *New << "\n");
2828 return !II.isVolatile();
2831 bool visitMemTransferInst(MemTransferInst &II) {
2832 // Rewriting of memory transfer instructions can be a bit tricky. We break
2833 // them into two categories: split intrinsics and unsplit intrinsics.
2835 DEBUG(dbgs() << " original: " << II << "\n");
2836 IRBuilder<> IRB(&II);
2838 assert(II.getRawSource() == OldPtr || II.getRawDest() == OldPtr);
2839 bool IsDest = II.getRawDest() == OldPtr;
2841 const AllocaPartitioning::MemTransferOffsets &MTO
2842 = P.getMemTransferOffsets(II);
2844 // Compute the relative offset within the transfer.
2845 unsigned IntPtrWidth = TD.getPointerSizeInBits();
2846 APInt RelOffset(IntPtrWidth, BeginOffset - (IsDest ? MTO.DestBegin
2847 : MTO.SourceBegin));
2849 unsigned Align = II.getAlignment();
2851 Align = MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(),
2852 MinAlign(II.getAlignment(), getPartitionAlign()));
2854 // For unsplit intrinsics, we simply modify the source and destination
2855 // pointers in place. This isn't just an optimization, it is a matter of
2856 // correctness. With unsplit intrinsics we may be dealing with transfers
2857 // within a single alloca before SROA ran, or with transfers that have
2858 // a variable length. We may also be dealing with memmove instead of
2859 // memcpy, and so simply updating the pointers is the necessary for us to
2860 // update both source and dest of a single call.
2861 if (!MTO.IsSplittable) {
2862 Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource();
2864 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2866 II.setSource(getAdjustedAllocaPtr(IRB, II.getRawSource()->getType()));
2868 Type *CstTy = II.getAlignmentCst()->getType();
2869 II.setAlignment(ConstantInt::get(CstTy, Align));
2871 DEBUG(dbgs() << " to: " << II << "\n");
2872 deleteIfTriviallyDead(OldOp);
2875 // For split transfer intrinsics we have an incredibly useful assurance:
2876 // the source and destination do not reside within the same alloca, and at
2877 // least one of them does not escape. This means that we can replace
2878 // memmove with memcpy, and we don't need to worry about all manner of
2879 // downsides to splitting and transforming the operations.
2881 // If this doesn't map cleanly onto the alloca type, and that type isn't
2882 // a single value type, just emit a memcpy.
2884 = !VecTy && !IntTy && (BeginOffset != NewAllocaBeginOffset ||
2885 EndOffset != NewAllocaEndOffset ||
2886 !NewAI.getAllocatedType()->isSingleValueType());
2888 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2889 // size hasn't been shrunk based on analysis of the viable range, this is
2891 if (EmitMemCpy && &OldAI == &NewAI) {
2892 uint64_t OrigBegin = IsDest ? MTO.DestBegin : MTO.SourceBegin;
2893 uint64_t OrigEnd = IsDest ? MTO.DestEnd : MTO.SourceEnd;
2894 // Ensure the start lines up.
2895 assert(BeginOffset == OrigBegin);
2898 // Rewrite the size as needed.
2899 if (EndOffset != OrigEnd)
2900 II.setLength(ConstantInt::get(II.getLength()->getType(),
2901 EndOffset - BeginOffset));
2904 // Record this instruction for deletion.
2905 Pass.DeadInsts.insert(&II);
2907 bool IsWholeAlloca = BeginOffset == NewAllocaBeginOffset &&
2908 EndOffset == NewAllocaEndOffset;
2909 bool IsVectorElement = VecTy && !IsWholeAlloca;
2910 uint64_t Size = EndOffset - BeginOffset;
2911 IntegerType *SubIntTy
2912 = IntTy ? Type::getIntNTy(IntTy->getContext(), Size*8) : 0;
2914 Type *OtherPtrTy = IsDest ? II.getRawSource()->getType()
2915 : II.getRawDest()->getType();
2917 if (IsVectorElement)
2918 OtherPtrTy = VecTy->getElementType()->getPointerTo();
2919 else if (IntTy && !IsWholeAlloca)
2920 OtherPtrTy = SubIntTy->getPointerTo();
2922 OtherPtrTy = NewAI.getType();
2925 // Compute the other pointer, folding as much as possible to produce
2926 // a single, simple GEP in most cases.
2927 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2928 OtherPtr = getAdjustedPtr(IRB, TD, OtherPtr, RelOffset, OtherPtrTy,
2929 getName("." + OtherPtr->getName()));
2931 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2932 // alloca that should be re-examined after rewriting this instruction.
2934 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets()))
2935 Pass.Worklist.insert(AI);
2939 = getAdjustedAllocaPtr(IRB, IsDest ? II.getRawDest()->getType()
2940 : II.getRawSource()->getType());
2941 Type *SizeTy = II.getLength()->getType();
2942 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2944 CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
2945 IsDest ? OtherPtr : OurPtr,
2946 Size, Align, II.isVolatile());
2948 DEBUG(dbgs() << " to: " << *New << "\n");
2952 // Note that we clamp the alignment to 1 here as a 0 alignment for a memcpy
2953 // is equivalent to 1, but that isn't true if we end up rewriting this as
2958 Value *SrcPtr = OtherPtr;
2959 Value *DstPtr = &NewAI;
2961 std::swap(SrcPtr, DstPtr);
2964 if (IsVectorElement && !IsDest) {
2965 // We have to extract rather than load.
2966 Src = IRB.CreateExtractElement(
2967 IRB.CreateAlignedLoad(SrcPtr, Align, getName(".copyload")),
2968 IRB.getInt32(getIndex(BeginOffset)),
2969 getName(".copyextract"));
2970 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2971 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2973 Src = convertValue(TD, IRB, Src, IntTy);
2974 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2975 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2976 Src = extractInteger(TD, IRB, Src, SubIntTy, Offset, getName(".extract"));
2978 Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(),
2979 getName(".copyload"));
2982 if (IntTy && !IsWholeAlloca && IsDest) {
2983 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2984 getName(".oldload"));
2985 Old = convertValue(TD, IRB, Old, IntTy);
2986 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2987 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2988 Src = insertInteger(TD, IRB, Old, Src, Offset, getName(".insert"));
2989 Src = convertValue(TD, IRB, Src, NewAllocaTy);
2992 if (IsVectorElement && IsDest) {
2993 // We have to insert into a loaded copy before storing.
2994 Src = IRB.CreateInsertElement(
2995 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), getName(".load")),
2996 Src, IRB.getInt32(getIndex(BeginOffset)),
2997 getName(".insert"));
3000 StoreInst *Store = cast<StoreInst>(
3001 IRB.CreateAlignedStore(Src, DstPtr, Align, II.isVolatile()));
3003 DEBUG(dbgs() << " to: " << *Store << "\n");
3004 return !II.isVolatile();
3007 bool visitIntrinsicInst(IntrinsicInst &II) {
3008 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
3009 II.getIntrinsicID() == Intrinsic::lifetime_end);
3010 DEBUG(dbgs() << " original: " << II << "\n");
3011 IRBuilder<> IRB(&II);
3012 assert(II.getArgOperand(1) == OldPtr);
3014 // Record this instruction for deletion.
3015 Pass.DeadInsts.insert(&II);
3018 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
3019 EndOffset - BeginOffset);
3020 Value *Ptr = getAdjustedAllocaPtr(IRB, II.getArgOperand(1)->getType());
3022 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
3023 New = IRB.CreateLifetimeStart(Ptr, Size);
3025 New = IRB.CreateLifetimeEnd(Ptr, Size);
3027 DEBUG(dbgs() << " to: " << *New << "\n");
3031 bool visitPHINode(PHINode &PN) {
3032 DEBUG(dbgs() << " original: " << PN << "\n");
3034 // We would like to compute a new pointer in only one place, but have it be
3035 // as local as possible to the PHI. To do that, we re-use the location of
3036 // the old pointer, which necessarily must be in the right position to
3037 // dominate the PHI.
3038 IRBuilder<> PtrBuilder(cast<Instruction>(OldPtr));
3040 Value *NewPtr = getAdjustedAllocaPtr(PtrBuilder, OldPtr->getType());
3041 // Replace the operands which were using the old pointer.
3042 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
3044 DEBUG(dbgs() << " to: " << PN << "\n");
3045 deleteIfTriviallyDead(OldPtr);
3049 bool visitSelectInst(SelectInst &SI) {
3050 DEBUG(dbgs() << " original: " << SI << "\n");
3051 IRBuilder<> IRB(&SI);
3053 // Find the operand we need to rewrite here.
3054 bool IsTrueVal = SI.getTrueValue() == OldPtr;
3056 assert(SI.getFalseValue() != OldPtr && "Pointer is both operands!");
3058 assert(SI.getFalseValue() == OldPtr && "Pointer isn't an operand!");
3060 Value *NewPtr = getAdjustedAllocaPtr(IRB, OldPtr->getType());
3061 SI.setOperand(IsTrueVal ? 1 : 2, NewPtr);
3062 DEBUG(dbgs() << " to: " << SI << "\n");
3063 deleteIfTriviallyDead(OldPtr);
3071 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
3073 /// This pass aggressively rewrites all aggregate loads and stores on
3074 /// a particular pointer (or any pointer derived from it which we can identify)
3075 /// with scalar loads and stores.
3076 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
3077 // Befriend the base class so it can delegate to private visit methods.
3078 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
3080 const DataLayout &TD;
3082 /// Queue of pointer uses to analyze and potentially rewrite.
3083 SmallVector<Use *, 8> Queue;
3085 /// Set to prevent us from cycling with phi nodes and loops.
3086 SmallPtrSet<User *, 8> Visited;
3088 /// The current pointer use being rewritten. This is used to dig up the used
3089 /// value (as opposed to the user).
3093 AggLoadStoreRewriter(const DataLayout &TD) : TD(TD) {}
3095 /// Rewrite loads and stores through a pointer and all pointers derived from
3097 bool rewrite(Instruction &I) {
3098 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
3100 bool Changed = false;
3101 while (!Queue.empty()) {
3102 U = Queue.pop_back_val();
3103 Changed |= visit(cast<Instruction>(U->getUser()));
3109 /// Enqueue all the users of the given instruction for further processing.
3110 /// This uses a set to de-duplicate users.
3111 void enqueueUsers(Instruction &I) {
3112 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
3114 if (Visited.insert(*UI))
3115 Queue.push_back(&UI.getUse());
3118 // Conservative default is to not rewrite anything.
3119 bool visitInstruction(Instruction &I) { return false; }
3121 /// \brief Generic recursive split emission class.
3122 template <typename Derived>
3125 /// The builder used to form new instructions.
3127 /// The indices which to be used with insert- or extractvalue to select the
3128 /// appropriate value within the aggregate.
3129 SmallVector<unsigned, 4> Indices;
3130 /// The indices to a GEP instruction which will move Ptr to the correct slot
3131 /// within the aggregate.
3132 SmallVector<Value *, 4> GEPIndices;
3133 /// The base pointer of the original op, used as a base for GEPing the
3134 /// split operations.
3137 /// Initialize the splitter with an insertion point, Ptr and start with a
3138 /// single zero GEP index.
3139 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
3140 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
3143 /// \brief Generic recursive split emission routine.
3145 /// This method recursively splits an aggregate op (load or store) into
3146 /// scalar or vector ops. It splits recursively until it hits a single value
3147 /// and emits that single value operation via the template argument.
3149 /// The logic of this routine relies on GEPs and insertvalue and
3150 /// extractvalue all operating with the same fundamental index list, merely
3151 /// formatted differently (GEPs need actual values).
3153 /// \param Ty The type being split recursively into smaller ops.
3154 /// \param Agg The aggregate value being built up or stored, depending on
3155 /// whether this is splitting a load or a store respectively.
3156 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3157 if (Ty->isSingleValueType())
3158 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
3160 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3161 unsigned OldSize = Indices.size();
3163 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3165 assert(Indices.size() == OldSize && "Did not return to the old size");
3166 Indices.push_back(Idx);
3167 GEPIndices.push_back(IRB.getInt32(Idx));
3168 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3169 GEPIndices.pop_back();
3175 if (StructType *STy = dyn_cast<StructType>(Ty)) {
3176 unsigned OldSize = Indices.size();
3178 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3180 assert(Indices.size() == OldSize && "Did not return to the old size");
3181 Indices.push_back(Idx);
3182 GEPIndices.push_back(IRB.getInt32(Idx));
3183 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3184 GEPIndices.pop_back();
3190 llvm_unreachable("Only arrays and structs are aggregate loadable types");
3194 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3195 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3196 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
3198 /// Emit a leaf load of a single value. This is called at the leaves of the
3199 /// recursive emission to actually load values.
3200 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3201 assert(Ty->isSingleValueType());
3202 // Load the single value and insert it using the indices.
3203 Value *Load = IRB.CreateLoad(IRB.CreateInBoundsGEP(Ptr, GEPIndices,
3206 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3207 DEBUG(dbgs() << " to: " << *Load << "\n");
3211 bool visitLoadInst(LoadInst &LI) {
3212 assert(LI.getPointerOperand() == *U);
3213 if (!LI.isSimple() || LI.getType()->isSingleValueType())
3216 // We have an aggregate being loaded, split it apart.
3217 DEBUG(dbgs() << " original: " << LI << "\n");
3218 LoadOpSplitter Splitter(&LI, *U);
3219 Value *V = UndefValue::get(LI.getType());
3220 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3221 LI.replaceAllUsesWith(V);
3222 LI.eraseFromParent();
3226 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3227 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3228 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
3230 /// Emit a leaf store of a single value. This is called at the leaves of the
3231 /// recursive emission to actually produce stores.
3232 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3233 assert(Ty->isSingleValueType());
3234 // Extract the single value and store it using the indices.
3235 Value *Store = IRB.CreateStore(
3236 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
3237 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
3239 DEBUG(dbgs() << " to: " << *Store << "\n");
3243 bool visitStoreInst(StoreInst &SI) {
3244 if (!SI.isSimple() || SI.getPointerOperand() != *U)
3246 Value *V = SI.getValueOperand();
3247 if (V->getType()->isSingleValueType())
3250 // We have an aggregate being stored, split it apart.
3251 DEBUG(dbgs() << " original: " << SI << "\n");
3252 StoreOpSplitter Splitter(&SI, *U);
3253 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3254 SI.eraseFromParent();
3258 bool visitBitCastInst(BitCastInst &BC) {
3263 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3268 bool visitPHINode(PHINode &PN) {
3273 bool visitSelectInst(SelectInst &SI) {
3280 /// \brief Strip aggregate type wrapping.
3282 /// This removes no-op aggregate types wrapping an underlying type. It will
3283 /// strip as many layers of types as it can without changing either the type
3284 /// size or the allocated size.
3285 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
3286 if (Ty->isSingleValueType())
3289 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3290 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
3293 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3294 InnerTy = ArrTy->getElementType();
3295 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3296 const StructLayout *SL = DL.getStructLayout(STy);
3297 unsigned Index = SL->getElementContainingOffset(0);
3298 InnerTy = STy->getElementType(Index);
3303 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3304 TypeSize > DL.getTypeSizeInBits(InnerTy))
3307 return stripAggregateTypeWrapping(DL, InnerTy);
3310 /// \brief Try to find a partition of the aggregate type passed in for a given
3311 /// offset and size.
3313 /// This recurses through the aggregate type and tries to compute a subtype
3314 /// based on the offset and size. When the offset and size span a sub-section
3315 /// of an array, it will even compute a new array type for that sub-section,
3316 /// and the same for structs.
3318 /// Note that this routine is very strict and tries to find a partition of the
3319 /// type which produces the *exact* right offset and size. It is not forgiving
3320 /// when the size or offset cause either end of type-based partition to be off.
3321 /// Also, this is a best-effort routine. It is reasonable to give up and not
3322 /// return a type if necessary.
3323 static Type *getTypePartition(const DataLayout &TD, Type *Ty,
3324 uint64_t Offset, uint64_t Size) {
3325 if (Offset == 0 && TD.getTypeAllocSize(Ty) == Size)
3326 return stripAggregateTypeWrapping(TD, Ty);
3327 if (Offset > TD.getTypeAllocSize(Ty) ||
3328 (TD.getTypeAllocSize(Ty) - Offset) < Size)
3331 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3332 // We can't partition pointers...
3333 if (SeqTy->isPointerTy())
3336 Type *ElementTy = SeqTy->getElementType();
3337 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
3338 uint64_t NumSkippedElements = Offset / ElementSize;
3339 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy))
3340 if (NumSkippedElements >= ArrTy->getNumElements())
3342 if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy))
3343 if (NumSkippedElements >= VecTy->getNumElements())
3345 Offset -= NumSkippedElements * ElementSize;
3347 // First check if we need to recurse.
3348 if (Offset > 0 || Size < ElementSize) {
3349 // Bail if the partition ends in a different array element.
3350 if ((Offset + Size) > ElementSize)
3352 // Recurse through the element type trying to peel off offset bytes.
3353 return getTypePartition(TD, ElementTy, Offset, Size);
3355 assert(Offset == 0);
3357 if (Size == ElementSize)
3358 return stripAggregateTypeWrapping(TD, ElementTy);
3359 assert(Size > ElementSize);
3360 uint64_t NumElements = Size / ElementSize;
3361 if (NumElements * ElementSize != Size)
3363 return ArrayType::get(ElementTy, NumElements);
3366 StructType *STy = dyn_cast<StructType>(Ty);
3370 const StructLayout *SL = TD.getStructLayout(STy);
3371 if (Offset >= SL->getSizeInBytes())
3373 uint64_t EndOffset = Offset + Size;
3374 if (EndOffset > SL->getSizeInBytes())
3377 unsigned Index = SL->getElementContainingOffset(Offset);
3378 Offset -= SL->getElementOffset(Index);
3380 Type *ElementTy = STy->getElementType(Index);
3381 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
3382 if (Offset >= ElementSize)
3383 return 0; // The offset points into alignment padding.
3385 // See if any partition must be contained by the element.
3386 if (Offset > 0 || Size < ElementSize) {
3387 if ((Offset + Size) > ElementSize)
3389 return getTypePartition(TD, ElementTy, Offset, Size);
3391 assert(Offset == 0);
3393 if (Size == ElementSize)
3394 return stripAggregateTypeWrapping(TD, ElementTy);
3396 StructType::element_iterator EI = STy->element_begin() + Index,
3397 EE = STy->element_end();
3398 if (EndOffset < SL->getSizeInBytes()) {
3399 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3400 if (Index == EndIndex)
3401 return 0; // Within a single element and its padding.
3403 // Don't try to form "natural" types if the elements don't line up with the
3405 // FIXME: We could potentially recurse down through the last element in the
3406 // sub-struct to find a natural end point.
3407 if (SL->getElementOffset(EndIndex) != EndOffset)
3410 assert(Index < EndIndex);
3411 EE = STy->element_begin() + EndIndex;
3414 // Try to build up a sub-structure.
3415 StructType *SubTy = StructType::get(STy->getContext(), makeArrayRef(EI, EE),
3417 const StructLayout *SubSL = TD.getStructLayout(SubTy);
3418 if (Size != SubSL->getSizeInBytes())
3419 return 0; // The sub-struct doesn't have quite the size needed.
3424 /// \brief Rewrite an alloca partition's users.
3426 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3427 /// to rewrite uses of an alloca partition to be conducive for SSA value
3428 /// promotion. If the partition needs a new, more refined alloca, this will
3429 /// build that new alloca, preserving as much type information as possible, and
3430 /// rewrite the uses of the old alloca to point at the new one and have the
3431 /// appropriate new offsets. It also evaluates how successful the rewrite was
3432 /// at enabling promotion and if it was successful queues the alloca to be
3434 bool SROA::rewriteAllocaPartition(AllocaInst &AI,
3435 AllocaPartitioning &P,
3436 AllocaPartitioning::iterator PI) {
3437 uint64_t AllocaSize = PI->EndOffset - PI->BeginOffset;
3438 bool IsLive = false;
3439 for (AllocaPartitioning::use_iterator UI = P.use_begin(PI),
3441 UI != UE && !IsLive; ++UI)
3445 return false; // No live uses left of this partition.
3447 DEBUG(dbgs() << "Speculating PHIs and selects in partition "
3448 << "[" << PI->BeginOffset << "," << PI->EndOffset << ")\n");
3450 PHIOrSelectSpeculator Speculator(*TD, P, *this);
3451 DEBUG(dbgs() << " speculating ");
3452 DEBUG(P.print(dbgs(), PI, ""));
3453 Speculator.visitUsers(PI);
3455 // Try to compute a friendly type for this partition of the alloca. This
3456 // won't always succeed, in which case we fall back to a legal integer type
3457 // or an i8 array of an appropriate size.
3459 if (Type *PartitionTy = P.getCommonType(PI))
3460 if (TD->getTypeAllocSize(PartitionTy) >= AllocaSize)
3461 AllocaTy = PartitionTy;
3463 if (Type *PartitionTy = getTypePartition(*TD, AI.getAllocatedType(),
3464 PI->BeginOffset, AllocaSize))
3465 AllocaTy = PartitionTy;
3467 (AllocaTy->isArrayTy() &&
3468 AllocaTy->getArrayElementType()->isIntegerTy())) &&
3469 TD->isLegalInteger(AllocaSize * 8))
3470 AllocaTy = Type::getIntNTy(*C, AllocaSize * 8);
3472 AllocaTy = ArrayType::get(Type::getInt8Ty(*C), AllocaSize);
3473 assert(TD->getTypeAllocSize(AllocaTy) >= AllocaSize);
3475 // Check for the case where we're going to rewrite to a new alloca of the
3476 // exact same type as the original, and with the same access offsets. In that
3477 // case, re-use the existing alloca, but still run through the rewriter to
3478 // performe phi and select speculation.
3480 if (AllocaTy == AI.getAllocatedType()) {
3481 assert(PI->BeginOffset == 0 &&
3482 "Non-zero begin offset but same alloca type");
3483 assert(PI == P.begin() && "Begin offset is zero on later partition");
3486 unsigned Alignment = AI.getAlignment();
3488 // The minimum alignment which users can rely on when the explicit
3489 // alignment is omitted or zero is that required by the ABI for this
3491 Alignment = TD->getABITypeAlignment(AI.getAllocatedType());
3493 Alignment = MinAlign(Alignment, PI->BeginOffset);
3494 // If we will get at least this much alignment from the type alone, leave
3495 // the alloca's alignment unconstrained.
3496 if (Alignment <= TD->getABITypeAlignment(AllocaTy))
3498 NewAI = new AllocaInst(AllocaTy, 0, Alignment,
3499 AI.getName() + ".sroa." + Twine(PI - P.begin()),
3504 DEBUG(dbgs() << "Rewriting alloca partition "
3505 << "[" << PI->BeginOffset << "," << PI->EndOffset << ") to: "
3508 // Track the high watermark of the post-promotion worklist. We will reset it
3509 // to this point if the alloca is not in fact scheduled for promotion.
3510 unsigned PPWOldSize = PostPromotionWorklist.size();
3512 AllocaPartitionRewriter Rewriter(*TD, P, PI, *this, AI, *NewAI,
3513 PI->BeginOffset, PI->EndOffset);
3514 DEBUG(dbgs() << " rewriting ");
3515 DEBUG(P.print(dbgs(), PI, ""));
3516 bool Promotable = Rewriter.visitUsers(P.use_begin(PI), P.use_end(PI));
3518 DEBUG(dbgs() << " and queuing for promotion\n");
3519 PromotableAllocas.push_back(NewAI);
3520 } else if (NewAI != &AI) {
3521 // If we can't promote the alloca, iterate on it to check for new
3522 // refinements exposed by splitting the current alloca. Don't iterate on an
3523 // alloca which didn't actually change and didn't get promoted.
3524 Worklist.insert(NewAI);
3527 // Drop any post-promotion work items if promotion didn't happen.
3529 while (PostPromotionWorklist.size() > PPWOldSize)
3530 PostPromotionWorklist.pop_back();
3535 /// \brief Walks the partitioning of an alloca rewriting uses of each partition.
3536 bool SROA::splitAlloca(AllocaInst &AI, AllocaPartitioning &P) {
3537 bool Changed = false;
3538 for (AllocaPartitioning::iterator PI = P.begin(), PE = P.end(); PI != PE;
3540 Changed |= rewriteAllocaPartition(AI, P, PI);
3545 /// \brief Analyze an alloca for SROA.
3547 /// This analyzes the alloca to ensure we can reason about it, builds
3548 /// a partitioning of the alloca, and then hands it off to be split and
3549 /// rewritten as needed.
3550 bool SROA::runOnAlloca(AllocaInst &AI) {
3551 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3552 ++NumAllocasAnalyzed;
3554 // Special case dead allocas, as they're trivial.
3555 if (AI.use_empty()) {
3556 AI.eraseFromParent();
3560 // Skip alloca forms that this analysis can't handle.
3561 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3562 TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
3565 bool Changed = false;
3567 // First, split any FCA loads and stores touching this alloca to promote
3568 // better splitting and promotion opportunities.
3569 AggLoadStoreRewriter AggRewriter(*TD);
3570 Changed |= AggRewriter.rewrite(AI);
3572 // Build the partition set using a recursive instruction-visiting builder.
3573 AllocaPartitioning P(*TD, AI);
3574 DEBUG(P.print(dbgs()));
3578 // Delete all the dead users of this alloca before splitting and rewriting it.
3579 for (AllocaPartitioning::dead_user_iterator DI = P.dead_user_begin(),
3580 DE = P.dead_user_end();
3583 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3584 DeadInsts.insert(*DI);
3586 for (AllocaPartitioning::dead_op_iterator DO = P.dead_op_begin(),
3587 DE = P.dead_op_end();
3590 // Clobber the use with an undef value.
3591 **DO = UndefValue::get(OldV->getType());
3592 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3593 if (isInstructionTriviallyDead(OldI)) {
3595 DeadInsts.insert(OldI);
3599 // No partitions to split. Leave the dead alloca for a later pass to clean up.
3600 if (P.begin() == P.end())
3603 return splitAlloca(AI, P) || Changed;
3606 /// \brief Delete the dead instructions accumulated in this run.
3608 /// Recursively deletes the dead instructions we've accumulated. This is done
3609 /// at the very end to maximize locality of the recursive delete and to
3610 /// minimize the problems of invalidated instruction pointers as such pointers
3611 /// are used heavily in the intermediate stages of the algorithm.
3613 /// We also record the alloca instructions deleted here so that they aren't
3614 /// subsequently handed to mem2reg to promote.
3615 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3616 while (!DeadInsts.empty()) {
3617 Instruction *I = DeadInsts.pop_back_val();
3618 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3620 I->replaceAllUsesWith(UndefValue::get(I->getType()));
3622 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
3623 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
3624 // Zero out the operand and see if it becomes trivially dead.
3626 if (isInstructionTriviallyDead(U))
3627 DeadInsts.insert(U);
3630 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3631 DeletedAllocas.insert(AI);
3634 I->eraseFromParent();
3638 /// \brief Promote the allocas, using the best available technique.
3640 /// This attempts to promote whatever allocas have been identified as viable in
3641 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3642 /// If there is a domtree available, we attempt to promote using the full power
3643 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3644 /// based on the SSAUpdater utilities. This function returns whether any
3645 /// promotion occured.
3646 bool SROA::promoteAllocas(Function &F) {
3647 if (PromotableAllocas.empty())
3650 NumPromoted += PromotableAllocas.size();
3652 if (DT && !ForceSSAUpdater) {
3653 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3654 PromoteMemToReg(PromotableAllocas, *DT);
3655 PromotableAllocas.clear();
3659 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3661 DIBuilder DIB(*F.getParent());
3662 SmallVector<Instruction*, 64> Insts;
3664 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3665 AllocaInst *AI = PromotableAllocas[Idx];
3666 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
3668 Instruction *I = cast<Instruction>(*UI++);
3669 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3670 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3671 // leading to them) here. Eventually it should use them to optimize the
3672 // scalar values produced.
3673 if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
3674 assert(onlyUsedByLifetimeMarkers(I) &&
3675 "Found a bitcast used outside of a lifetime marker.");
3676 while (!I->use_empty())
3677 cast<Instruction>(*I->use_begin())->eraseFromParent();
3678 I->eraseFromParent();
3681 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3682 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3683 II->getIntrinsicID() == Intrinsic::lifetime_end);
3684 II->eraseFromParent();
3690 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3694 PromotableAllocas.clear();
3699 /// \brief A predicate to test whether an alloca belongs to a set.
3700 class IsAllocaInSet {
3701 typedef SmallPtrSet<AllocaInst *, 4> SetType;
3705 typedef AllocaInst *argument_type;
3707 IsAllocaInSet(const SetType &Set) : Set(Set) {}
3708 bool operator()(AllocaInst *AI) const { return Set.count(AI); }
3712 bool SROA::runOnFunction(Function &F) {
3713 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3714 C = &F.getContext();
3715 TD = getAnalysisIfAvailable<DataLayout>();
3717 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3720 DT = getAnalysisIfAvailable<DominatorTree>();
3722 BasicBlock &EntryBB = F.getEntryBlock();
3723 for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end());
3725 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3726 Worklist.insert(AI);
3728 bool Changed = false;
3729 // A set of deleted alloca instruction pointers which should be removed from
3730 // the list of promotable allocas.
3731 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3734 while (!Worklist.empty()) {
3735 Changed |= runOnAlloca(*Worklist.pop_back_val());
3736 deleteDeadInstructions(DeletedAllocas);
3738 // Remove the deleted allocas from various lists so that we don't try to
3739 // continue processing them.
3740 if (!DeletedAllocas.empty()) {
3741 Worklist.remove_if(IsAllocaInSet(DeletedAllocas));
3742 PostPromotionWorklist.remove_if(IsAllocaInSet(DeletedAllocas));
3743 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3744 PromotableAllocas.end(),
3745 IsAllocaInSet(DeletedAllocas)),
3746 PromotableAllocas.end());
3747 DeletedAllocas.clear();
3751 Changed |= promoteAllocas(F);
3753 Worklist = PostPromotionWorklist;
3754 PostPromotionWorklist.clear();
3755 } while (!Worklist.empty());
3760 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3761 if (RequiresDomTree)
3762 AU.addRequired<DominatorTree>();
3763 AU.setPreservesCFG();