Bug Summary

File:llvm/lib/Transforms/Scalar/SROA.cpp
Warning:line 2897, column 54
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -clear-ast-before-backend -disable-llvm-verifier -discard-value-names -main-file-name SROA.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mframe-pointer=none -fmath-errno -ffp-contract=on -fno-rounding-math -mconstructor-aliases -funwind-tables=2 -target-cpu x86-64 -tune-cpu generic -debugger-tuning=gdb -ffunction-sections -fdata-sections -fcoverage-compilation-dir=/build/llvm-toolchain-snapshot-14~++20220116100644+5f782d25a742/build-llvm -resource-dir /usr/lib/llvm-14/lib/clang/14.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-14~++20220116100644+5f782d25a742/llvm/lib/Transforms/Scalar -I include -I /build/llvm-toolchain-snapshot-14~++20220116100644+5f782d25a742/llvm/include -D _FORTIFY_SOURCE=2 -D NDEBUG -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/x86_64-linux-gnu/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/backward -internal-isystem /usr/lib/llvm-14/lib/clang/14.0.0/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../x86_64-linux-gnu/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -fmacro-prefix-map=/build/llvm-toolchain-snapshot-14~++20220116100644+5f782d25a742/build-llvm=build-llvm -fmacro-prefix-map=/build/llvm-toolchain-snapshot-14~++20220116100644+5f782d25a742/= -fcoverage-prefix-map=/build/llvm-toolchain-snapshot-14~++20220116100644+5f782d25a742/build-llvm=build-llvm -fcoverage-prefix-map=/build/llvm-toolchain-snapshot-14~++20220116100644+5f782d25a742/= -O3 -Wno-unused-command-line-argument -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-class-memaccess -Wno-redundant-move -Wno-pessimizing-move -Wno-noexcept-type -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir=/build/llvm-toolchain-snapshot-14~++20220116100644+5f782d25a742/build-llvm -fdebug-prefix-map=/build/llvm-toolchain-snapshot-14~++20220116100644+5f782d25a742/build-llvm=build-llvm -fdebug-prefix-map=/build/llvm-toolchain-snapshot-14~++20220116100644+5f782d25a742/= -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -fcolor-diagnostics -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /tmp/scan-build-2022-01-16-232930-107970-1 -x c++ /build/llvm-toolchain-snapshot-14~++20220116100644+5f782d25a742/llvm/lib/Transforms/Scalar/SROA.cpp
1//===- SROA.cpp - Scalar Replacement Of Aggregates ------------------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8/// \file
9/// This transformation implements the well known scalar replacement of
10/// aggregates transformation. It tries to identify promotable elements of an
11/// aggregate alloca, and promote them to registers. It will also try to
12/// convert uses of an element (or set of elements) of an alloca into a vector
13/// or bitfield-style integer scalar if appropriate.
14///
15/// It works to do this with minimal slicing of the alloca so that regions
16/// which are merely transferred in and out of external memory remain unchanged
17/// and are not decomposed to scalar code.
18///
19/// Because this also performs alloca promotion, it can be thought of as also
20/// serving the purpose of SSA formation. The algorithm iterates on the
21/// function until all opportunities for promotion have been realized.
22///
23//===----------------------------------------------------------------------===//
24
25#include "llvm/Transforms/Scalar/SROA.h"
26#include "llvm/ADT/APInt.h"
27#include "llvm/ADT/ArrayRef.h"
28#include "llvm/ADT/DenseMap.h"
29#include "llvm/ADT/PointerIntPair.h"
30#include "llvm/ADT/STLExtras.h"
31#include "llvm/ADT/SetVector.h"
32#include "llvm/ADT/SmallBitVector.h"
33#include "llvm/ADT/SmallPtrSet.h"
34#include "llvm/ADT/SmallVector.h"
35#include "llvm/ADT/Statistic.h"
36#include "llvm/ADT/StringRef.h"
37#include "llvm/ADT/Twine.h"
38#include "llvm/ADT/iterator.h"
39#include "llvm/ADT/iterator_range.h"
40#include "llvm/Analysis/AssumptionCache.h"
41#include "llvm/Analysis/GlobalsModRef.h"
42#include "llvm/Analysis/Loads.h"
43#include "llvm/Analysis/PtrUseVisitor.h"
44#include "llvm/Config/llvm-config.h"
45#include "llvm/IR/BasicBlock.h"
46#include "llvm/IR/Constant.h"
47#include "llvm/IR/ConstantFolder.h"
48#include "llvm/IR/Constants.h"
49#include "llvm/IR/DIBuilder.h"
50#include "llvm/IR/DataLayout.h"
51#include "llvm/IR/DebugInfoMetadata.h"
52#include "llvm/IR/DerivedTypes.h"
53#include "llvm/IR/Dominators.h"
54#include "llvm/IR/Function.h"
55#include "llvm/IR/GetElementPtrTypeIterator.h"
56#include "llvm/IR/GlobalAlias.h"
57#include "llvm/IR/IRBuilder.h"
58#include "llvm/IR/InstVisitor.h"
59#include "llvm/IR/InstrTypes.h"
60#include "llvm/IR/Instruction.h"
61#include "llvm/IR/Instructions.h"
62#include "llvm/IR/IntrinsicInst.h"
63#include "llvm/IR/Intrinsics.h"
64#include "llvm/IR/LLVMContext.h"
65#include "llvm/IR/Metadata.h"
66#include "llvm/IR/Module.h"
67#include "llvm/IR/Operator.h"
68#include "llvm/IR/PassManager.h"
69#include "llvm/IR/Type.h"
70#include "llvm/IR/Use.h"
71#include "llvm/IR/User.h"
72#include "llvm/IR/Value.h"
73#include "llvm/InitializePasses.h"
74#include "llvm/Pass.h"
75#include "llvm/Support/Casting.h"
76#include "llvm/Support/CommandLine.h"
77#include "llvm/Support/Compiler.h"
78#include "llvm/Support/Debug.h"
79#include "llvm/Support/ErrorHandling.h"
80#include "llvm/Support/MathExtras.h"
81#include "llvm/Support/raw_ostream.h"
82#include "llvm/Transforms/Scalar.h"
83#include "llvm/Transforms/Utils/Local.h"
84#include "llvm/Transforms/Utils/PromoteMemToReg.h"
85#include <algorithm>
86#include <cassert>
87#include <chrono>
88#include <cstddef>
89#include <cstdint>
90#include <cstring>
91#include <iterator>
92#include <string>
93#include <tuple>
94#include <utility>
95#include <vector>
96
97using namespace llvm;
98using namespace llvm::sroa;
99
100#define DEBUG_TYPE"sroa" "sroa"
101
102STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement")static llvm::Statistic NumAllocasAnalyzed = {"sroa", "NumAllocasAnalyzed"
, "Number of allocas analyzed for replacement"}
;
103STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed")static llvm::Statistic NumAllocaPartitions = {"sroa", "NumAllocaPartitions"
, "Number of alloca partitions formed"}
;
104STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions per alloca")static llvm::Statistic MaxPartitionsPerAlloca = {"sroa", "MaxPartitionsPerAlloca"
, "Maximum number of partitions per alloca"}
;
105STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten")static llvm::Statistic NumAllocaPartitionUses = {"sroa", "NumAllocaPartitionUses"
, "Number of alloca partition uses rewritten"}
;
106STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition")static llvm::Statistic MaxUsesPerAllocaPartition = {"sroa", "MaxUsesPerAllocaPartition"
, "Maximum number of uses of a partition"}
;
107STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced")static llvm::Statistic NumNewAllocas = {"sroa", "NumNewAllocas"
, "Number of new, smaller allocas introduced"}
;
108STATISTIC(NumPromoted, "Number of allocas promoted to SSA values")static llvm::Statistic NumPromoted = {"sroa", "NumPromoted", "Number of allocas promoted to SSA values"
}
;
109STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion")static llvm::Statistic NumLoadsSpeculated = {"sroa", "NumLoadsSpeculated"
, "Number of loads speculated to allow promotion"}
;
110STATISTIC(NumDeleted, "Number of instructions deleted")static llvm::Statistic NumDeleted = {"sroa", "NumDeleted", "Number of instructions deleted"
}
;
111STATISTIC(NumVectorized, "Number of vectorized aggregates")static llvm::Statistic NumVectorized = {"sroa", "NumVectorized"
, "Number of vectorized aggregates"}
;
112
113/// Hidden option to experiment with completely strict handling of inbounds
114/// GEPs.
115static cl::opt<bool> SROAStrictInbounds("sroa-strict-inbounds", cl::init(false),
116 cl::Hidden);
117
118namespace {
119
120/// A custom IRBuilder inserter which prefixes all names, but only in
121/// Assert builds.
122class IRBuilderPrefixedInserter final : public IRBuilderDefaultInserter {
123 std::string Prefix;
124
125 Twine getNameWithPrefix(const Twine &Name) const {
126 return Name.isTriviallyEmpty() ? Name : Prefix + Name;
127 }
128
129public:
130 void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
131
132 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
133 BasicBlock::iterator InsertPt) const override {
134 IRBuilderDefaultInserter::InsertHelper(I, getNameWithPrefix(Name), BB,
135 InsertPt);
136 }
137};
138
139/// Provide a type for IRBuilder that drops names in release builds.
140using IRBuilderTy = IRBuilder<ConstantFolder, IRBuilderPrefixedInserter>;
141
142/// A used slice of an alloca.
143///
144/// This structure represents a slice of an alloca used by some instruction. It
145/// stores both the begin and end offsets of this use, a pointer to the use
146/// itself, and a flag indicating whether we can classify the use as splittable
147/// or not when forming partitions of the alloca.
148class Slice {
149 /// The beginning offset of the range.
150 uint64_t BeginOffset = 0;
151
152 /// The ending offset, not included in the range.
153 uint64_t EndOffset = 0;
154
155 /// Storage for both the use of this slice and whether it can be
156 /// split.
157 PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
158
159public:
160 Slice() = default;
161
162 Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
163 : BeginOffset(BeginOffset), EndOffset(EndOffset),
164 UseAndIsSplittable(U, IsSplittable) {}
165
166 uint64_t beginOffset() const { return BeginOffset; }
167 uint64_t endOffset() const { return EndOffset; }
168
169 bool isSplittable() const { return UseAndIsSplittable.getInt(); }
170 void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
171
172 Use *getUse() const { return UseAndIsSplittable.getPointer(); }
173
174 bool isDead() const { return getUse() == nullptr; }
175 void kill() { UseAndIsSplittable.setPointer(nullptr); }
176
177 /// Support for ordering ranges.
178 ///
179 /// This provides an ordering over ranges such that start offsets are
180 /// always increasing, and within equal start offsets, the end offsets are
181 /// decreasing. Thus the spanning range comes first in a cluster with the
182 /// same start position.
183 bool operator<(const Slice &RHS) const {
184 if (beginOffset() < RHS.beginOffset())
185 return true;
186 if (beginOffset() > RHS.beginOffset())
187 return false;
188 if (isSplittable() != RHS.isSplittable())
189 return !isSplittable();
190 if (endOffset() > RHS.endOffset())
191 return true;
192 return false;
193 }
194
195 /// Support comparison with a single offset to allow binary searches.
196 friend LLVM_ATTRIBUTE_UNUSED__attribute__((__unused__)) bool operator<(const Slice &LHS,
197 uint64_t RHSOffset) {
198 return LHS.beginOffset() < RHSOffset;
199 }
200 friend LLVM_ATTRIBUTE_UNUSED__attribute__((__unused__)) bool operator<(uint64_t LHSOffset,
201 const Slice &RHS) {
202 return LHSOffset < RHS.beginOffset();
203 }
204
205 bool operator==(const Slice &RHS) const {
206 return isSplittable() == RHS.isSplittable() &&
207 beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
208 }
209 bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
210};
211
212} // end anonymous namespace
213
214/// Representation of the alloca slices.
215///
216/// This class represents the slices of an alloca which are formed by its
217/// various uses. If a pointer escapes, we can't fully build a representation
218/// for the slices used and we reflect that in this structure. The uses are
219/// stored, sorted by increasing beginning offset and with unsplittable slices
220/// starting at a particular offset before splittable slices.
221class llvm::sroa::AllocaSlices {
222public:
223 /// Construct the slices of a particular alloca.
224 AllocaSlices(const DataLayout &DL, AllocaInst &AI);
225
226 /// Test whether a pointer to the allocation escapes our analysis.
227 ///
228 /// If this is true, the slices are never fully built and should be
229 /// ignored.
230 bool isEscaped() const { return PointerEscapingInstr; }
231
232 /// Support for iterating over the slices.
233 /// @{
234 using iterator = SmallVectorImpl<Slice>::iterator;
235 using range = iterator_range<iterator>;
236
237 iterator begin() { return Slices.begin(); }
238 iterator end() { return Slices.end(); }
239
240 using const_iterator = SmallVectorImpl<Slice>::const_iterator;
241 using const_range = iterator_range<const_iterator>;
242
243 const_iterator begin() const { return Slices.begin(); }
244 const_iterator end() const { return Slices.end(); }
245 /// @}
246
247 /// Erase a range of slices.
248 void erase(iterator Start, iterator Stop) { Slices.erase(Start, Stop); }
249
250 /// Insert new slices for this alloca.
251 ///
252 /// This moves the slices into the alloca's slices collection, and re-sorts
253 /// everything so that the usual ordering properties of the alloca's slices
254 /// hold.
255 void insert(ArrayRef<Slice> NewSlices) {
256 int OldSize = Slices.size();
257 Slices.append(NewSlices.begin(), NewSlices.end());
258 auto SliceI = Slices.begin() + OldSize;
259 llvm::sort(SliceI, Slices.end());
260 std::inplace_merge(Slices.begin(), SliceI, Slices.end());
261 }
262
263 // Forward declare the iterator and range accessor for walking the
264 // partitions.
265 class partition_iterator;
266 iterator_range<partition_iterator> partitions();
267
268 /// Access the dead users for this alloca.
269 ArrayRef<Instruction *> getDeadUsers() const { return DeadUsers; }
270
271 /// Access Uses that should be dropped if the alloca is promotable.
272 ArrayRef<Use *> getDeadUsesIfPromotable() const {
273 return DeadUseIfPromotable;
274 }
275
276 /// Access the dead operands referring to this alloca.
277 ///
278 /// These are operands which have cannot actually be used to refer to the
279 /// alloca as they are outside its range and the user doesn't correct for
280 /// that. These mostly consist of PHI node inputs and the like which we just
281 /// need to replace with undef.
282 ArrayRef<Use *> getDeadOperands() const { return DeadOperands; }
283
284#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
285 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
286 void printSlice(raw_ostream &OS, const_iterator I,
287 StringRef Indent = " ") const;
288 void printUse(raw_ostream &OS, const_iterator I,
289 StringRef Indent = " ") const;
290 void print(raw_ostream &OS) const;
291 void dump(const_iterator I) const;
292 void dump() const;
293#endif
294
295private:
296 template <typename DerivedT, typename RetT = void> class BuilderBase;
297 class SliceBuilder;
298
299 friend class AllocaSlices::SliceBuilder;
300
301#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
302 /// Handle to alloca instruction to simplify method interfaces.
303 AllocaInst &AI;
304#endif
305
306 /// The instruction responsible for this alloca not having a known set
307 /// of slices.
308 ///
309 /// When an instruction (potentially) escapes the pointer to the alloca, we
310 /// store a pointer to that here and abort trying to form slices of the
311 /// alloca. This will be null if the alloca slices are analyzed successfully.
312 Instruction *PointerEscapingInstr;
313
314 /// The slices of the alloca.
315 ///
316 /// We store a vector of the slices formed by uses of the alloca here. This
317 /// vector is sorted by increasing begin offset, and then the unsplittable
318 /// slices before the splittable ones. See the Slice inner class for more
319 /// details.
320 SmallVector<Slice, 8> Slices;
321
322 /// Instructions which will become dead if we rewrite the alloca.
323 ///
324 /// Note that these are not separated by slice. This is because we expect an
325 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
326 /// all these instructions can simply be removed and replaced with poison as
327 /// they come from outside of the allocated space.
328 SmallVector<Instruction *, 8> DeadUsers;
329
330 /// Uses which will become dead if can promote the alloca.
331 SmallVector<Use *, 8> DeadUseIfPromotable;
332
333 /// Operands which will become dead if we rewrite the alloca.
334 ///
335 /// These are operands that in their particular use can be replaced with
336 /// poison when we rewrite the alloca. These show up in out-of-bounds inputs
337 /// to PHI nodes and the like. They aren't entirely dead (there might be
338 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
339 /// want to swap this particular input for poison to simplify the use lists of
340 /// the alloca.
341 SmallVector<Use *, 8> DeadOperands;
342};
343
344/// A partition of the slices.
345///
346/// An ephemeral representation for a range of slices which can be viewed as
347/// a partition of the alloca. This range represents a span of the alloca's
348/// memory which cannot be split, and provides access to all of the slices
349/// overlapping some part of the partition.
350///
351/// Objects of this type are produced by traversing the alloca's slices, but
352/// are only ephemeral and not persistent.
353class llvm::sroa::Partition {
354private:
355 friend class AllocaSlices;
356 friend class AllocaSlices::partition_iterator;
357
358 using iterator = AllocaSlices::iterator;
359
360 /// The beginning and ending offsets of the alloca for this
361 /// partition.
362 uint64_t BeginOffset = 0, EndOffset = 0;
363
364 /// The start and end iterators of this partition.
365 iterator SI, SJ;
366
367 /// A collection of split slice tails overlapping the partition.
368 SmallVector<Slice *, 4> SplitTails;
369
370 /// Raw constructor builds an empty partition starting and ending at
371 /// the given iterator.
372 Partition(iterator SI) : SI(SI), SJ(SI) {}
373
374public:
375 /// The start offset of this partition.
376 ///
377 /// All of the contained slices start at or after this offset.
378 uint64_t beginOffset() const { return BeginOffset; }
379
380 /// The end offset of this partition.
381 ///
382 /// All of the contained slices end at or before this offset.
383 uint64_t endOffset() const { return EndOffset; }
384
385 /// The size of the partition.
386 ///
387 /// Note that this can never be zero.
388 uint64_t size() const {
389 assert(BeginOffset < EndOffset && "Partitions must span some bytes!")(static_cast <bool> (BeginOffset < EndOffset &&
"Partitions must span some bytes!") ? void (0) : __assert_fail
("BeginOffset < EndOffset && \"Partitions must span some bytes!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 389, __extension__ __PRETTY_FUNCTION__
))
;
390 return EndOffset - BeginOffset;
391 }
392
393 /// Test whether this partition contains no slices, and merely spans
394 /// a region occupied by split slices.
395 bool empty() const { return SI == SJ; }
396
397 /// \name Iterate slices that start within the partition.
398 /// These may be splittable or unsplittable. They have a begin offset >= the
399 /// partition begin offset.
400 /// @{
401 // FIXME: We should probably define a "concat_iterator" helper and use that
402 // to stitch together pointee_iterators over the split tails and the
403 // contiguous iterators of the partition. That would give a much nicer
404 // interface here. We could then additionally expose filtered iterators for
405 // split, unsplit, and unsplittable splices based on the usage patterns.
406 iterator begin() const { return SI; }
407 iterator end() const { return SJ; }
408 /// @}
409
410 /// Get the sequence of split slice tails.
411 ///
412 /// These tails are of slices which start before this partition but are
413 /// split and overlap into the partition. We accumulate these while forming
414 /// partitions.
415 ArrayRef<Slice *> splitSliceTails() const { return SplitTails; }
416};
417
418/// An iterator over partitions of the alloca's slices.
419///
420/// This iterator implements the core algorithm for partitioning the alloca's
421/// slices. It is a forward iterator as we don't support backtracking for
422/// efficiency reasons, and re-use a single storage area to maintain the
423/// current set of split slices.
424///
425/// It is templated on the slice iterator type to use so that it can operate
426/// with either const or non-const slice iterators.
427class AllocaSlices::partition_iterator
428 : public iterator_facade_base<partition_iterator, std::forward_iterator_tag,
429 Partition> {
430 friend class AllocaSlices;
431
432 /// Most of the state for walking the partitions is held in a class
433 /// with a nice interface for examining them.
434 Partition P;
435
436 /// We need to keep the end of the slices to know when to stop.
437 AllocaSlices::iterator SE;
438
439 /// We also need to keep track of the maximum split end offset seen.
440 /// FIXME: Do we really?
441 uint64_t MaxSplitSliceEndOffset = 0;
442
443 /// Sets the partition to be empty at given iterator, and sets the
444 /// end iterator.
445 partition_iterator(AllocaSlices::iterator SI, AllocaSlices::iterator SE)
446 : P(SI), SE(SE) {
447 // If not already at the end, advance our state to form the initial
448 // partition.
449 if (SI != SE)
450 advance();
451 }
452
453 /// Advance the iterator to the next partition.
454 ///
455 /// Requires that the iterator not be at the end of the slices.
456 void advance() {
457 assert((P.SI != SE || !P.SplitTails.empty()) &&(static_cast <bool> ((P.SI != SE || !P.SplitTails.empty
()) && "Cannot advance past the end of the slices!") ?
void (0) : __assert_fail ("(P.SI != SE || !P.SplitTails.empty()) && \"Cannot advance past the end of the slices!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 458, __extension__ __PRETTY_FUNCTION__
))
458 "Cannot advance past the end of the slices!")(static_cast <bool> ((P.SI != SE || !P.SplitTails.empty
()) && "Cannot advance past the end of the slices!") ?
void (0) : __assert_fail ("(P.SI != SE || !P.SplitTails.empty()) && \"Cannot advance past the end of the slices!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 458, __extension__ __PRETTY_FUNCTION__
))
;
459
460 // Clear out any split uses which have ended.
461 if (!P.SplitTails.empty()) {
462 if (P.EndOffset >= MaxSplitSliceEndOffset) {
463 // If we've finished all splits, this is easy.
464 P.SplitTails.clear();
465 MaxSplitSliceEndOffset = 0;
466 } else {
467 // Remove the uses which have ended in the prior partition. This
468 // cannot change the max split slice end because we just checked that
469 // the prior partition ended prior to that max.
470 llvm::erase_if(P.SplitTails,
471 [&](Slice *S) { return S->endOffset() <= P.EndOffset; });
472 assert(llvm::any_of(P.SplitTails,(static_cast <bool> (llvm::any_of(P.SplitTails, [&]
(Slice *S) { return S->endOffset() == MaxSplitSliceEndOffset
; }) && "Could not find the current max split slice offset!"
) ? void (0) : __assert_fail ("llvm::any_of(P.SplitTails, [&](Slice *S) { return S->endOffset() == MaxSplitSliceEndOffset; }) && \"Could not find the current max split slice offset!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 476, __extension__ __PRETTY_FUNCTION__
))
473 [&](Slice *S) {(static_cast <bool> (llvm::any_of(P.SplitTails, [&]
(Slice *S) { return S->endOffset() == MaxSplitSliceEndOffset
; }) && "Could not find the current max split slice offset!"
) ? void (0) : __assert_fail ("llvm::any_of(P.SplitTails, [&](Slice *S) { return S->endOffset() == MaxSplitSliceEndOffset; }) && \"Could not find the current max split slice offset!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 476, __extension__ __PRETTY_FUNCTION__
))
474 return S->endOffset() == MaxSplitSliceEndOffset;(static_cast <bool> (llvm::any_of(P.SplitTails, [&]
(Slice *S) { return S->endOffset() == MaxSplitSliceEndOffset
; }) && "Could not find the current max split slice offset!"
) ? void (0) : __assert_fail ("llvm::any_of(P.SplitTails, [&](Slice *S) { return S->endOffset() == MaxSplitSliceEndOffset; }) && \"Could not find the current max split slice offset!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 476, __extension__ __PRETTY_FUNCTION__
))
475 }) &&(static_cast <bool> (llvm::any_of(P.SplitTails, [&]
(Slice *S) { return S->endOffset() == MaxSplitSliceEndOffset
; }) && "Could not find the current max split slice offset!"
) ? void (0) : __assert_fail ("llvm::any_of(P.SplitTails, [&](Slice *S) { return S->endOffset() == MaxSplitSliceEndOffset; }) && \"Could not find the current max split slice offset!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 476, __extension__ __PRETTY_FUNCTION__
))
476 "Could not find the current max split slice offset!")(static_cast <bool> (llvm::any_of(P.SplitTails, [&]
(Slice *S) { return S->endOffset() == MaxSplitSliceEndOffset
; }) && "Could not find the current max split slice offset!"
) ? void (0) : __assert_fail ("llvm::any_of(P.SplitTails, [&](Slice *S) { return S->endOffset() == MaxSplitSliceEndOffset; }) && \"Could not find the current max split slice offset!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 476, __extension__ __PRETTY_FUNCTION__
))
;
477 assert(llvm::all_of(P.SplitTails,(static_cast <bool> (llvm::all_of(P.SplitTails, [&]
(Slice *S) { return S->endOffset() <= MaxSplitSliceEndOffset
; }) && "Max split slice end offset is not actually the max!"
) ? void (0) : __assert_fail ("llvm::all_of(P.SplitTails, [&](Slice *S) { return S->endOffset() <= MaxSplitSliceEndOffset; }) && \"Max split slice end offset is not actually the max!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 481, __extension__ __PRETTY_FUNCTION__
))
478 [&](Slice *S) {(static_cast <bool> (llvm::all_of(P.SplitTails, [&]
(Slice *S) { return S->endOffset() <= MaxSplitSliceEndOffset
; }) && "Max split slice end offset is not actually the max!"
) ? void (0) : __assert_fail ("llvm::all_of(P.SplitTails, [&](Slice *S) { return S->endOffset() <= MaxSplitSliceEndOffset; }) && \"Max split slice end offset is not actually the max!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 481, __extension__ __PRETTY_FUNCTION__
))
479 return S->endOffset() <= MaxSplitSliceEndOffset;(static_cast <bool> (llvm::all_of(P.SplitTails, [&]
(Slice *S) { return S->endOffset() <= MaxSplitSliceEndOffset
; }) && "Max split slice end offset is not actually the max!"
) ? void (0) : __assert_fail ("llvm::all_of(P.SplitTails, [&](Slice *S) { return S->endOffset() <= MaxSplitSliceEndOffset; }) && \"Max split slice end offset is not actually the max!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 481, __extension__ __PRETTY_FUNCTION__
))
480 }) &&(static_cast <bool> (llvm::all_of(P.SplitTails, [&]
(Slice *S) { return S->endOffset() <= MaxSplitSliceEndOffset
; }) && "Max split slice end offset is not actually the max!"
) ? void (0) : __assert_fail ("llvm::all_of(P.SplitTails, [&](Slice *S) { return S->endOffset() <= MaxSplitSliceEndOffset; }) && \"Max split slice end offset is not actually the max!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 481, __extension__ __PRETTY_FUNCTION__
))
481 "Max split slice end offset is not actually the max!")(static_cast <bool> (llvm::all_of(P.SplitTails, [&]
(Slice *S) { return S->endOffset() <= MaxSplitSliceEndOffset
; }) && "Max split slice end offset is not actually the max!"
) ? void (0) : __assert_fail ("llvm::all_of(P.SplitTails, [&](Slice *S) { return S->endOffset() <= MaxSplitSliceEndOffset; }) && \"Max split slice end offset is not actually the max!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 481, __extension__ __PRETTY_FUNCTION__
))
;
482 }
483 }
484
485 // If P.SI is already at the end, then we've cleared the split tail and
486 // now have an end iterator.
487 if (P.SI == SE) {
488 assert(P.SplitTails.empty() && "Failed to clear the split slices!")(static_cast <bool> (P.SplitTails.empty() && "Failed to clear the split slices!"
) ? void (0) : __assert_fail ("P.SplitTails.empty() && \"Failed to clear the split slices!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 488, __extension__ __PRETTY_FUNCTION__
))
;
489 return;
490 }
491
492 // If we had a non-empty partition previously, set up the state for
493 // subsequent partitions.
494 if (P.SI != P.SJ) {
495 // Accumulate all the splittable slices which started in the old
496 // partition into the split list.
497 for (Slice &S : P)
498 if (S.isSplittable() && S.endOffset() > P.EndOffset) {
499 P.SplitTails.push_back(&S);
500 MaxSplitSliceEndOffset =
501 std::max(S.endOffset(), MaxSplitSliceEndOffset);
502 }
503
504 // Start from the end of the previous partition.
505 P.SI = P.SJ;
506
507 // If P.SI is now at the end, we at most have a tail of split slices.
508 if (P.SI == SE) {
509 P.BeginOffset = P.EndOffset;
510 P.EndOffset = MaxSplitSliceEndOffset;
511 return;
512 }
513
514 // If the we have split slices and the next slice is after a gap and is
515 // not splittable immediately form an empty partition for the split
516 // slices up until the next slice begins.
517 if (!P.SplitTails.empty() && P.SI->beginOffset() != P.EndOffset &&
518 !P.SI->isSplittable()) {
519 P.BeginOffset = P.EndOffset;
520 P.EndOffset = P.SI->beginOffset();
521 return;
522 }
523 }
524
525 // OK, we need to consume new slices. Set the end offset based on the
526 // current slice, and step SJ past it. The beginning offset of the
527 // partition is the beginning offset of the next slice unless we have
528 // pre-existing split slices that are continuing, in which case we begin
529 // at the prior end offset.
530 P.BeginOffset = P.SplitTails.empty() ? P.SI->beginOffset() : P.EndOffset;
531 P.EndOffset = P.SI->endOffset();
532 ++P.SJ;
533
534 // There are two strategies to form a partition based on whether the
535 // partition starts with an unsplittable slice or a splittable slice.
536 if (!P.SI->isSplittable()) {
537 // When we're forming an unsplittable region, it must always start at
538 // the first slice and will extend through its end.
539 assert(P.BeginOffset == P.SI->beginOffset())(static_cast <bool> (P.BeginOffset == P.SI->beginOffset
()) ? void (0) : __assert_fail ("P.BeginOffset == P.SI->beginOffset()"
, "llvm/lib/Transforms/Scalar/SROA.cpp", 539, __extension__ __PRETTY_FUNCTION__
))
;
540
541 // Form a partition including all of the overlapping slices with this
542 // unsplittable slice.
543 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
544 if (!P.SJ->isSplittable())
545 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
546 ++P.SJ;
547 }
548
549 // We have a partition across a set of overlapping unsplittable
550 // partitions.
551 return;
552 }
553
554 // If we're starting with a splittable slice, then we need to form
555 // a synthetic partition spanning it and any other overlapping splittable
556 // splices.
557 assert(P.SI->isSplittable() && "Forming a splittable partition!")(static_cast <bool> (P.SI->isSplittable() &&
"Forming a splittable partition!") ? void (0) : __assert_fail
("P.SI->isSplittable() && \"Forming a splittable partition!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 557, __extension__ __PRETTY_FUNCTION__
))
;
558
559 // Collect all of the overlapping splittable slices.
560 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset &&
561 P.SJ->isSplittable()) {
562 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
563 ++P.SJ;
564 }
565
566 // Back upiP.EndOffset if we ended the span early when encountering an
567 // unsplittable slice. This synthesizes the early end offset of
568 // a partition spanning only splittable slices.
569 if (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
570 assert(!P.SJ->isSplittable())(static_cast <bool> (!P.SJ->isSplittable()) ? void (
0) : __assert_fail ("!P.SJ->isSplittable()", "llvm/lib/Transforms/Scalar/SROA.cpp"
, 570, __extension__ __PRETTY_FUNCTION__))
;
571 P.EndOffset = P.SJ->beginOffset();
572 }
573 }
574
575public:
576 bool operator==(const partition_iterator &RHS) const {
577 assert(SE == RHS.SE &&(static_cast <bool> (SE == RHS.SE && "End iterators don't match between compared partition iterators!"
) ? void (0) : __assert_fail ("SE == RHS.SE && \"End iterators don't match between compared partition iterators!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 578, __extension__ __PRETTY_FUNCTION__
))
578 "End iterators don't match between compared partition iterators!")(static_cast <bool> (SE == RHS.SE && "End iterators don't match between compared partition iterators!"
) ? void (0) : __assert_fail ("SE == RHS.SE && \"End iterators don't match between compared partition iterators!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 578, __extension__ __PRETTY_FUNCTION__
))
;
579
580 // The observed positions of partitions is marked by the P.SI iterator and
581 // the emptiness of the split slices. The latter is only relevant when
582 // P.SI == SE, as the end iterator will additionally have an empty split
583 // slices list, but the prior may have the same P.SI and a tail of split
584 // slices.
585 if (P.SI == RHS.P.SI && P.SplitTails.empty() == RHS.P.SplitTails.empty()) {
586 assert(P.SJ == RHS.P.SJ &&(static_cast <bool> (P.SJ == RHS.P.SJ && "Same set of slices formed two different sized partitions!"
) ? void (0) : __assert_fail ("P.SJ == RHS.P.SJ && \"Same set of slices formed two different sized partitions!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 587, __extension__ __PRETTY_FUNCTION__
))
587 "Same set of slices formed two different sized partitions!")(static_cast <bool> (P.SJ == RHS.P.SJ && "Same set of slices formed two different sized partitions!"
) ? void (0) : __assert_fail ("P.SJ == RHS.P.SJ && \"Same set of slices formed two different sized partitions!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 587, __extension__ __PRETTY_FUNCTION__
))
;
588 assert(P.SplitTails.size() == RHS.P.SplitTails.size() &&(static_cast <bool> (P.SplitTails.size() == RHS.P.SplitTails
.size() && "Same slice position with differently sized non-empty split "
"slice tails!") ? void (0) : __assert_fail ("P.SplitTails.size() == RHS.P.SplitTails.size() && \"Same slice position with differently sized non-empty split \" \"slice tails!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 590, __extension__ __PRETTY_FUNCTION__
))
589 "Same slice position with differently sized non-empty split "(static_cast <bool> (P.SplitTails.size() == RHS.P.SplitTails
.size() && "Same slice position with differently sized non-empty split "
"slice tails!") ? void (0) : __assert_fail ("P.SplitTails.size() == RHS.P.SplitTails.size() && \"Same slice position with differently sized non-empty split \" \"slice tails!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 590, __extension__ __PRETTY_FUNCTION__
))
590 "slice tails!")(static_cast <bool> (P.SplitTails.size() == RHS.P.SplitTails
.size() && "Same slice position with differently sized non-empty split "
"slice tails!") ? void (0) : __assert_fail ("P.SplitTails.size() == RHS.P.SplitTails.size() && \"Same slice position with differently sized non-empty split \" \"slice tails!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 590, __extension__ __PRETTY_FUNCTION__
))
;
591 return true;
592 }
593 return false;
594 }
595
596 partition_iterator &operator++() {
597 advance();
598 return *this;
599 }
600
601 Partition &operator*() { return P; }
602};
603
604/// A forward range over the partitions of the alloca's slices.
605///
606/// This accesses an iterator range over the partitions of the alloca's
607/// slices. It computes these partitions on the fly based on the overlapping
608/// offsets of the slices and the ability to split them. It will visit "empty"
609/// partitions to cover regions of the alloca only accessed via split
610/// slices.
611iterator_range<AllocaSlices::partition_iterator> AllocaSlices::partitions() {
612 return make_range(partition_iterator(begin(), end()),
613 partition_iterator(end(), end()));
614}
615
616static Value *foldSelectInst(SelectInst &SI) {
617 // If the condition being selected on is a constant or the same value is
618 // being selected between, fold the select. Yes this does (rarely) happen
619 // early on.
620 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
621 return SI.getOperand(1 + CI->isZero());
622 if (SI.getOperand(1) == SI.getOperand(2))
623 return SI.getOperand(1);
624
625 return nullptr;
626}
627
628/// A helper that folds a PHI node or a select.
629static Value *foldPHINodeOrSelectInst(Instruction &I) {
630 if (PHINode *PN = dyn_cast<PHINode>(&I)) {
631 // If PN merges together the same value, return that value.
632 return PN->hasConstantValue();
633 }
634 return foldSelectInst(cast<SelectInst>(I));
635}
636
637/// Builder for the alloca slices.
638///
639/// This class builds a set of alloca slices by recursively visiting the uses
640/// of an alloca and making a slice for each load and store at each offset.
641class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
642 friend class PtrUseVisitor<SliceBuilder>;
643 friend class InstVisitor<SliceBuilder>;
644
645 using Base = PtrUseVisitor<SliceBuilder>;
646
647 const uint64_t AllocSize;
648 AllocaSlices &AS;
649
650 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
651 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
652
653 /// Set to de-duplicate dead instructions found in the use walk.
654 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
655
656public:
657 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &AS)
658 : PtrUseVisitor<SliceBuilder>(DL),
659 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType()).getFixedSize()),
660 AS(AS) {}
661
662private:
663 void markAsDead(Instruction &I) {
664 if (VisitedDeadInsts.insert(&I).second)
665 AS.DeadUsers.push_back(&I);
666 }
667
668 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
669 bool IsSplittable = false) {
670 // Completely skip uses which have a zero size or start either before or
671 // past the end of the allocation.
672 if (Size == 0 || Offset.uge(AllocSize)) {
673 LLVM_DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "WARNING: Ignoring " << Size
<< " byte use @" << Offset << " which has zero size or starts outside of the "
<< AllocSize << " byte alloca:\n" << " alloca: "
<< AS.AI << "\n" << " use: " <<
I << "\n"; } } while (false)
674 << Offsetdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "WARNING: Ignoring " << Size
<< " byte use @" << Offset << " which has zero size or starts outside of the "
<< AllocSize << " byte alloca:\n" << " alloca: "
<< AS.AI << "\n" << " use: " <<
I << "\n"; } } while (false)
675 << " which has zero size or starts outside of the "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "WARNING: Ignoring " << Size
<< " byte use @" << Offset << " which has zero size or starts outside of the "
<< AllocSize << " byte alloca:\n" << " alloca: "
<< AS.AI << "\n" << " use: " <<
I << "\n"; } } while (false)
676 << AllocSize << " byte alloca:\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "WARNING: Ignoring " << Size
<< " byte use @" << Offset << " which has zero size or starts outside of the "
<< AllocSize << " byte alloca:\n" << " alloca: "
<< AS.AI << "\n" << " use: " <<
I << "\n"; } } while (false)
677 << " alloca: " << AS.AI << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "WARNING: Ignoring " << Size
<< " byte use @" << Offset << " which has zero size or starts outside of the "
<< AllocSize << " byte alloca:\n" << " alloca: "
<< AS.AI << "\n" << " use: " <<
I << "\n"; } } while (false)
678 << " use: " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "WARNING: Ignoring " << Size
<< " byte use @" << Offset << " which has zero size or starts outside of the "
<< AllocSize << " byte alloca:\n" << " alloca: "
<< AS.AI << "\n" << " use: " <<
I << "\n"; } } while (false)
;
679 return markAsDead(I);
680 }
681
682 uint64_t BeginOffset = Offset.getZExtValue();
683 uint64_t EndOffset = BeginOffset + Size;
684
685 // Clamp the end offset to the end of the allocation. Note that this is
686 // formulated to handle even the case where "BeginOffset + Size" overflows.
687 // This may appear superficially to be something we could ignore entirely,
688 // but that is not so! There may be widened loads or PHI-node uses where
689 // some instructions are dead but not others. We can't completely ignore
690 // them, and so have to record at least the information here.
691 assert(AllocSize >= BeginOffset)(static_cast <bool> (AllocSize >= BeginOffset) ? void
(0) : __assert_fail ("AllocSize >= BeginOffset", "llvm/lib/Transforms/Scalar/SROA.cpp"
, 691, __extension__ __PRETTY_FUNCTION__))
; // Established above.
692 if (Size > AllocSize - BeginOffset) {
693 LLVM_DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "WARNING: Clamping a " << Size
<< " byte use @" << Offset << " to remain within the "
<< AllocSize << " byte alloca:\n" << " alloca: "
<< AS.AI << "\n" << " use: " <<
I << "\n"; } } while (false)
694 << Offset << " to remain within the " << AllocSizedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "WARNING: Clamping a " << Size
<< " byte use @" << Offset << " to remain within the "
<< AllocSize << " byte alloca:\n" << " alloca: "
<< AS.AI << "\n" << " use: " <<
I << "\n"; } } while (false)
695 << " byte alloca:\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "WARNING: Clamping a " << Size
<< " byte use @" << Offset << " to remain within the "
<< AllocSize << " byte alloca:\n" << " alloca: "
<< AS.AI << "\n" << " use: " <<
I << "\n"; } } while (false)
696 << " alloca: " << AS.AI << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "WARNING: Clamping a " << Size
<< " byte use @" << Offset << " to remain within the "
<< AllocSize << " byte alloca:\n" << " alloca: "
<< AS.AI << "\n" << " use: " <<
I << "\n"; } } while (false)
697 << " use: " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "WARNING: Clamping a " << Size
<< " byte use @" << Offset << " to remain within the "
<< AllocSize << " byte alloca:\n" << " alloca: "
<< AS.AI << "\n" << " use: " <<
I << "\n"; } } while (false)
;
698 EndOffset = AllocSize;
699 }
700
701 AS.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
702 }
703
704 void visitBitCastInst(BitCastInst &BC) {
705 if (BC.use_empty())
706 return markAsDead(BC);
707
708 return Base::visitBitCastInst(BC);
709 }
710
711 void visitAddrSpaceCastInst(AddrSpaceCastInst &ASC) {
712 if (ASC.use_empty())
713 return markAsDead(ASC);
714
715 return Base::visitAddrSpaceCastInst(ASC);
716 }
717
718 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
719 if (GEPI.use_empty())
720 return markAsDead(GEPI);
721
722 if (SROAStrictInbounds && GEPI.isInBounds()) {
723 // FIXME: This is a manually un-factored variant of the basic code inside
724 // of GEPs with checking of the inbounds invariant specified in the
725 // langref in a very strict sense. If we ever want to enable
726 // SROAStrictInbounds, this code should be factored cleanly into
727 // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
728 // by writing out the code here where we have the underlying allocation
729 // size readily available.
730 APInt GEPOffset = Offset;
731 const DataLayout &DL = GEPI.getModule()->getDataLayout();
732 for (gep_type_iterator GTI = gep_type_begin(GEPI),
733 GTE = gep_type_end(GEPI);
734 GTI != GTE; ++GTI) {
735 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
736 if (!OpC)
737 break;
738
739 // Handle a struct index, which adds its field offset to the pointer.
740 if (StructType *STy = GTI.getStructTypeOrNull()) {
741 unsigned ElementIdx = OpC->getZExtValue();
742 const StructLayout *SL = DL.getStructLayout(STy);
743 GEPOffset +=
744 APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
745 } else {
746 // For array or vector indices, scale the index by the size of the
747 // type.
748 APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
749 GEPOffset +=
750 Index *
751 APInt(Offset.getBitWidth(),
752 DL.getTypeAllocSize(GTI.getIndexedType()).getFixedSize());
753 }
754
755 // If this index has computed an intermediate pointer which is not
756 // inbounds, then the result of the GEP is a poison value and we can
757 // delete it and all uses.
758 if (GEPOffset.ugt(AllocSize))
759 return markAsDead(GEPI);
760 }
761 }
762
763 return Base::visitGetElementPtrInst(GEPI);
764 }
765
766 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
767 uint64_t Size, bool IsVolatile) {
768 // We allow splitting of non-volatile loads and stores where the type is an
769 // integer type. These may be used to implement 'memcpy' or other "transfer
770 // of bits" patterns.
771 bool IsSplittable =
772 Ty->isIntegerTy() && !IsVolatile && DL.typeSizeEqualsStoreSize(Ty);
773
774 insertUse(I, Offset, Size, IsSplittable);
775 }
776
777 void visitLoadInst(LoadInst &LI) {
778 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&(static_cast <bool> ((!LI.isSimple() || LI.getType()->
isSingleValueType()) && "All simple FCA loads should have been pre-split"
) ? void (0) : __assert_fail ("(!LI.isSimple() || LI.getType()->isSingleValueType()) && \"All simple FCA loads should have been pre-split\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 779, __extension__ __PRETTY_FUNCTION__
))
779 "All simple FCA loads should have been pre-split")(static_cast <bool> ((!LI.isSimple() || LI.getType()->
isSingleValueType()) && "All simple FCA loads should have been pre-split"
) ? void (0) : __assert_fail ("(!LI.isSimple() || LI.getType()->isSingleValueType()) && \"All simple FCA loads should have been pre-split\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 779, __extension__ __PRETTY_FUNCTION__
))
;
780
781 if (!IsOffsetKnown)
782 return PI.setAborted(&LI);
783
784 if (LI.isVolatile() &&
785 LI.getPointerAddressSpace() != DL.getAllocaAddrSpace())
786 return PI.setAborted(&LI);
787
788 if (isa<ScalableVectorType>(LI.getType()))
789 return PI.setAborted(&LI);
790
791 uint64_t Size = DL.getTypeStoreSize(LI.getType()).getFixedSize();
792 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
793 }
794
795 void visitStoreInst(StoreInst &SI) {
796 Value *ValOp = SI.getValueOperand();
797 if (ValOp == *U)
798 return PI.setEscapedAndAborted(&SI);
799 if (!IsOffsetKnown)
800 return PI.setAborted(&SI);
801
802 if (SI.isVolatile() &&
803 SI.getPointerAddressSpace() != DL.getAllocaAddrSpace())
804 return PI.setAborted(&SI);
805
806 if (isa<ScalableVectorType>(ValOp->getType()))
807 return PI.setAborted(&SI);
808
809 uint64_t Size = DL.getTypeStoreSize(ValOp->getType()).getFixedSize();
810
811 // If this memory access can be shown to *statically* extend outside the
812 // bounds of the allocation, it's behavior is undefined, so simply
813 // ignore it. Note that this is more strict than the generic clamping
814 // behavior of insertUse. We also try to handle cases which might run the
815 // risk of overflow.
816 // FIXME: We should instead consider the pointer to have escaped if this
817 // function is being instrumented for addressing bugs or race conditions.
818 if (Size > AllocSize || Offset.ugt(AllocSize - Size)) {
819 LLVM_DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "WARNING: Ignoring " << Size
<< " byte store @" << Offset << " which extends past the end of the "
<< AllocSize << " byte alloca:\n" << " alloca: "
<< AS.AI << "\n" << " use: " <<
SI << "\n"; } } while (false)
820 << Offset << " which extends past the end of the "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "WARNING: Ignoring " << Size
<< " byte store @" << Offset << " which extends past the end of the "
<< AllocSize << " byte alloca:\n" << " alloca: "
<< AS.AI << "\n" << " use: " <<
SI << "\n"; } } while (false)
821 << AllocSize << " byte alloca:\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "WARNING: Ignoring " << Size
<< " byte store @" << Offset << " which extends past the end of the "
<< AllocSize << " byte alloca:\n" << " alloca: "
<< AS.AI << "\n" << " use: " <<
SI << "\n"; } } while (false)
822 << " alloca: " << AS.AI << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "WARNING: Ignoring " << Size
<< " byte store @" << Offset << " which extends past the end of the "
<< AllocSize << " byte alloca:\n" << " alloca: "
<< AS.AI << "\n" << " use: " <<
SI << "\n"; } } while (false)
823 << " use: " << SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "WARNING: Ignoring " << Size
<< " byte store @" << Offset << " which extends past the end of the "
<< AllocSize << " byte alloca:\n" << " alloca: "
<< AS.AI << "\n" << " use: " <<
SI << "\n"; } } while (false)
;
824 return markAsDead(SI);
825 }
826
827 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&(static_cast <bool> ((!SI.isSimple() || ValOp->getType
()->isSingleValueType()) && "All simple FCA stores should have been pre-split"
) ? void (0) : __assert_fail ("(!SI.isSimple() || ValOp->getType()->isSingleValueType()) && \"All simple FCA stores should have been pre-split\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 828, __extension__ __PRETTY_FUNCTION__
))
828 "All simple FCA stores should have been pre-split")(static_cast <bool> ((!SI.isSimple() || ValOp->getType
()->isSingleValueType()) && "All simple FCA stores should have been pre-split"
) ? void (0) : __assert_fail ("(!SI.isSimple() || ValOp->getType()->isSingleValueType()) && \"All simple FCA stores should have been pre-split\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 828, __extension__ __PRETTY_FUNCTION__
))
;
829 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
830 }
831
832 void visitMemSetInst(MemSetInst &II) {
833 assert(II.getRawDest() == *U && "Pointer use is not the destination?")(static_cast <bool> (II.getRawDest() == *U && "Pointer use is not the destination?"
) ? void (0) : __assert_fail ("II.getRawDest() == *U && \"Pointer use is not the destination?\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 833, __extension__ __PRETTY_FUNCTION__
))
;
834 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
835 if ((Length && Length->getValue() == 0) ||
836 (IsOffsetKnown && Offset.uge(AllocSize)))
837 // Zero-length mem transfer intrinsics can be ignored entirely.
838 return markAsDead(II);
839
840 if (!IsOffsetKnown)
841 return PI.setAborted(&II);
842
843 // Don't replace this with a store with a different address space. TODO:
844 // Use a store with the casted new alloca?
845 if (II.isVolatile() && II.getDestAddressSpace() != DL.getAllocaAddrSpace())
846 return PI.setAborted(&II);
847
848 insertUse(II, Offset, Length ? Length->getLimitedValue()
849 : AllocSize - Offset.getLimitedValue(),
850 (bool)Length);
851 }
852
853 void visitMemTransferInst(MemTransferInst &II) {
854 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
855 if (Length && Length->getValue() == 0)
856 // Zero-length mem transfer intrinsics can be ignored entirely.
857 return markAsDead(II);
858
859 // Because we can visit these intrinsics twice, also check to see if the
860 // first time marked this instruction as dead. If so, skip it.
861 if (VisitedDeadInsts.count(&II))
862 return;
863
864 if (!IsOffsetKnown)
865 return PI.setAborted(&II);
866
867 // Don't replace this with a load/store with a different address space.
868 // TODO: Use a store with the casted new alloca?
869 if (II.isVolatile() &&
870 (II.getDestAddressSpace() != DL.getAllocaAddrSpace() ||
871 II.getSourceAddressSpace() != DL.getAllocaAddrSpace()))
872 return PI.setAborted(&II);
873
874 // This side of the transfer is completely out-of-bounds, and so we can
875 // nuke the entire transfer. However, we also need to nuke the other side
876 // if already added to our partitions.
877 // FIXME: Yet another place we really should bypass this when
878 // instrumenting for ASan.
879 if (Offset.uge(AllocSize)) {
880 SmallDenseMap<Instruction *, unsigned>::iterator MTPI =
881 MemTransferSliceMap.find(&II);
882 if (MTPI != MemTransferSliceMap.end())
883 AS.Slices[MTPI->second].kill();
884 return markAsDead(II);
885 }
886
887 uint64_t RawOffset = Offset.getLimitedValue();
888 uint64_t Size = Length ? Length->getLimitedValue() : AllocSize - RawOffset;
889
890 // Check for the special case where the same exact value is used for both
891 // source and dest.
892 if (*U == II.getRawDest() && *U == II.getRawSource()) {
893 // For non-volatile transfers this is a no-op.
894 if (!II.isVolatile())
895 return markAsDead(II);
896
897 return insertUse(II, Offset, Size, /*IsSplittable=*/false);
898 }
899
900 // If we have seen both source and destination for a mem transfer, then
901 // they both point to the same alloca.
902 bool Inserted;
903 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
904 std::tie(MTPI, Inserted) =
905 MemTransferSliceMap.insert(std::make_pair(&II, AS.Slices.size()));
906 unsigned PrevIdx = MTPI->second;
907 if (!Inserted) {
908 Slice &PrevP = AS.Slices[PrevIdx];
909
910 // Check if the begin offsets match and this is a non-volatile transfer.
911 // In that case, we can completely elide the transfer.
912 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
913 PrevP.kill();
914 return markAsDead(II);
915 }
916
917 // Otherwise we have an offset transfer within the same alloca. We can't
918 // split those.
919 PrevP.makeUnsplittable();
920 }
921
922 // Insert the use now that we've fixed up the splittable nature.
923 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
924
925 // Check that we ended up with a valid index in the map.
926 assert(AS.Slices[PrevIdx].getUse()->getUser() == &II &&(static_cast <bool> (AS.Slices[PrevIdx].getUse()->getUser
() == &II && "Map index doesn't point back to a slice with this user."
) ? void (0) : __assert_fail ("AS.Slices[PrevIdx].getUse()->getUser() == &II && \"Map index doesn't point back to a slice with this user.\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 927, __extension__ __PRETTY_FUNCTION__
))
927 "Map index doesn't point back to a slice with this user.")(static_cast <bool> (AS.Slices[PrevIdx].getUse()->getUser
() == &II && "Map index doesn't point back to a slice with this user."
) ? void (0) : __assert_fail ("AS.Slices[PrevIdx].getUse()->getUser() == &II && \"Map index doesn't point back to a slice with this user.\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 927, __extension__ __PRETTY_FUNCTION__
))
;
928 }
929
930 // Disable SRoA for any intrinsics except for lifetime invariants and
931 // invariant group.
932 // FIXME: What about debug intrinsics? This matches old behavior, but
933 // doesn't make sense.
934 void visitIntrinsicInst(IntrinsicInst &II) {
935 if (II.isDroppable()) {
936 AS.DeadUseIfPromotable.push_back(U);
937 return;
938 }
939
940 if (!IsOffsetKnown)
941 return PI.setAborted(&II);
942
943 if (II.isLifetimeStartOrEnd()) {
944 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
945 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
946 Length->getLimitedValue());
947 insertUse(II, Offset, Size, true);
948 return;
949 }
950
951 if (II.isLaunderOrStripInvariantGroup()) {
952 enqueueUsers(II);
953 return;
954 }
955
956 Base::visitIntrinsicInst(II);
957 }
958
959 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
960 // We consider any PHI or select that results in a direct load or store of
961 // the same offset to be a viable use for slicing purposes. These uses
962 // are considered unsplittable and the size is the maximum loaded or stored
963 // size.
964 SmallPtrSet<Instruction *, 4> Visited;
965 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
966 Visited.insert(Root);
967 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
968 const DataLayout &DL = Root->getModule()->getDataLayout();
969 // If there are no loads or stores, the access is dead. We mark that as
970 // a size zero access.
971 Size = 0;
972 do {
973 Instruction *I, *UsedI;
974 std::tie(UsedI, I) = Uses.pop_back_val();
975
976 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
977 Size = std::max(Size,
978 DL.getTypeStoreSize(LI->getType()).getFixedSize());
979 continue;
980 }
981 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
982 Value *Op = SI->getOperand(0);
983 if (Op == UsedI)
984 return SI;
985 Size = std::max(Size,
986 DL.getTypeStoreSize(Op->getType()).getFixedSize());
987 continue;
988 }
989
990 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
991 if (!GEP->hasAllZeroIndices())
992 return GEP;
993 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
994 !isa<SelectInst>(I) && !isa<AddrSpaceCastInst>(I)) {
995 return I;
996 }
997
998 for (User *U : I->users())
999 if (Visited.insert(cast<Instruction>(U)).second)
1000 Uses.push_back(std::make_pair(I, cast<Instruction>(U)));
1001 } while (!Uses.empty());
1002
1003 return nullptr;
1004 }
1005
1006 void visitPHINodeOrSelectInst(Instruction &I) {
1007 assert(isa<PHINode>(I) || isa<SelectInst>(I))(static_cast <bool> (isa<PHINode>(I) || isa<SelectInst
>(I)) ? void (0) : __assert_fail ("isa<PHINode>(I) || isa<SelectInst>(I)"
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1007, __extension__ __PRETTY_FUNCTION__
))
;
1008 if (I.use_empty())
1009 return markAsDead(I);
1010
1011 // If this is a PHI node before a catchswitch, we cannot insert any non-PHI
1012 // instructions in this BB, which may be required during rewriting. Bail out
1013 // on these cases.
1014 if (isa<PHINode>(I) &&
1015 I.getParent()->getFirstInsertionPt() == I.getParent()->end())
1016 return PI.setAborted(&I);
1017
1018 // TODO: We could use SimplifyInstruction here to fold PHINodes and
1019 // SelectInsts. However, doing so requires to change the current
1020 // dead-operand-tracking mechanism. For instance, suppose neither loading
1021 // from %U nor %other traps. Then "load (select undef, %U, %other)" does not
1022 // trap either. However, if we simply replace %U with undef using the
1023 // current dead-operand-tracking mechanism, "load (select undef, undef,
1024 // %other)" may trap because the select may return the first operand
1025 // "undef".
1026 if (Value *Result = foldPHINodeOrSelectInst(I)) {
1027 if (Result == *U)
1028 // If the result of the constant fold will be the pointer, recurse
1029 // through the PHI/select as if we had RAUW'ed it.
1030 enqueueUsers(I);
1031 else
1032 // Otherwise the operand to the PHI/select is dead, and we can replace
1033 // it with poison.
1034 AS.DeadOperands.push_back(U);
1035
1036 return;
1037 }
1038
1039 if (!IsOffsetKnown)
1040 return PI.setAborted(&I);
1041
1042 // See if we already have computed info on this node.
1043 uint64_t &Size = PHIOrSelectSizes[&I];
1044 if (!Size) {
1045 // This is a new PHI/Select, check for an unsafe use of it.
1046 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&I, Size))
1047 return PI.setAborted(UnsafeI);
1048 }
1049
1050 // For PHI and select operands outside the alloca, we can't nuke the entire
1051 // phi or select -- the other side might still be relevant, so we special
1052 // case them here and use a separate structure to track the operands
1053 // themselves which should be replaced with poison.
1054 // FIXME: This should instead be escaped in the event we're instrumenting
1055 // for address sanitization.
1056 if (Offset.uge(AllocSize)) {
1057 AS.DeadOperands.push_back(U);
1058 return;
1059 }
1060
1061 insertUse(I, Offset, Size);
1062 }
1063
1064 void visitPHINode(PHINode &PN) { visitPHINodeOrSelectInst(PN); }
1065
1066 void visitSelectInst(SelectInst &SI) { visitPHINodeOrSelectInst(SI); }
1067
1068 /// Disable SROA entirely if there are unhandled users of the alloca.
1069 void visitInstruction(Instruction &I) { PI.setAborted(&I); }
1070};
1071
1072AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
1073 :
1074#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1075 AI(AI),
1076#endif
1077 PointerEscapingInstr(nullptr) {
1078 SliceBuilder PB(DL, AI, *this);
1079 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
1080 if (PtrI.isEscaped() || PtrI.isAborted()) {
1081 // FIXME: We should sink the escape vs. abort info into the caller nicely,
1082 // possibly by just storing the PtrInfo in the AllocaSlices.
1083 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
1084 : PtrI.getAbortingInst();
1085 assert(PointerEscapingInstr && "Did not track a bad instruction")(static_cast <bool> (PointerEscapingInstr && "Did not track a bad instruction"
) ? void (0) : __assert_fail ("PointerEscapingInstr && \"Did not track a bad instruction\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1085, __extension__ __PRETTY_FUNCTION__
))
;
1086 return;
1087 }
1088
1089 llvm::erase_if(Slices, [](const Slice &S) { return S.isDead(); });
1090
1091 // Sort the uses. This arranges for the offsets to be in ascending order,
1092 // and the sizes to be in descending order.
1093 llvm::stable_sort(Slices);
1094}
1095
1096#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1097
1098void AllocaSlices::print(raw_ostream &OS, const_iterator I,
1099 StringRef Indent) const {
1100 printSlice(OS, I, Indent);
1101 OS << "\n";
1102 printUse(OS, I, Indent);
1103}
1104
1105void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
1106 StringRef Indent) const {
1107 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
1108 << " slice #" << (I - begin())
1109 << (I->isSplittable() ? " (splittable)" : "");
1110}
1111
1112void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
1113 StringRef Indent) const {
1114 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
1115}
1116
1117void AllocaSlices::print(raw_ostream &OS) const {
1118 if (PointerEscapingInstr) {
1119 OS << "Can't analyze slices for alloca: " << AI << "\n"
1120 << " A pointer to this alloca escaped by:\n"
1121 << " " << *PointerEscapingInstr << "\n";
1122 return;
1123 }
1124
1125 OS << "Slices of alloca: " << AI << "\n";
1126 for (const_iterator I = begin(), E = end(); I != E; ++I)
1127 print(OS, I);
1128}
1129
1130LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void AllocaSlices::dump(const_iterator I) const {
1131 print(dbgs(), I);
1132}
1133LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void AllocaSlices::dump() const { print(dbgs()); }
1134
1135#endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1136
1137/// Walk the range of a partitioning looking for a common type to cover this
1138/// sequence of slices.
1139static std::pair<Type *, IntegerType *>
1140findCommonType(AllocaSlices::const_iterator B, AllocaSlices::const_iterator E,
1141 uint64_t EndOffset) {
1142 Type *Ty = nullptr;
1143 bool TyIsCommon = true;
1144 IntegerType *ITy = nullptr;
1145
1146 // Note that we need to look at *every* alloca slice's Use to ensure we
1147 // always get consistent results regardless of the order of slices.
1148 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1149 Use *U = I->getUse();
1150 if (isa<IntrinsicInst>(*U->getUser()))
1151 continue;
1152 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1153 continue;
1154
1155 Type *UserTy = nullptr;
1156 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1157 UserTy = LI->getType();
1158 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1159 UserTy = SI->getValueOperand()->getType();
1160 }
1161
1162 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1163 // If the type is larger than the partition, skip it. We only encounter
1164 // this for split integer operations where we want to use the type of the
1165 // entity causing the split. Also skip if the type is not a byte width
1166 // multiple.
1167 if (UserITy->getBitWidth() % 8 != 0 ||
1168 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1169 continue;
1170
1171 // Track the largest bitwidth integer type used in this way in case there
1172 // is no common type.
1173 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1174 ITy = UserITy;
1175 }
1176
1177 // To avoid depending on the order of slices, Ty and TyIsCommon must not
1178 // depend on types skipped above.
1179 if (!UserTy || (Ty && Ty != UserTy))
1180 TyIsCommon = false; // Give up on anything but an iN type.
1181 else
1182 Ty = UserTy;
1183 }
1184
1185 return {TyIsCommon ? Ty : nullptr, ITy};
1186}
1187
1188/// PHI instructions that use an alloca and are subsequently loaded can be
1189/// rewritten to load both input pointers in the pred blocks and then PHI the
1190/// results, allowing the load of the alloca to be promoted.
1191/// From this:
1192/// %P2 = phi [i32* %Alloca, i32* %Other]
1193/// %V = load i32* %P2
1194/// to:
1195/// %V1 = load i32* %Alloca -> will be mem2reg'd
1196/// ...
1197/// %V2 = load i32* %Other
1198/// ...
1199/// %V = phi [i32 %V1, i32 %V2]
1200///
1201/// We can do this to a select if its only uses are loads and if the operands
1202/// to the select can be loaded unconditionally.
1203///
1204/// FIXME: This should be hoisted into a generic utility, likely in
1205/// Transforms/Util/Local.h
1206static bool isSafePHIToSpeculate(PHINode &PN) {
1207 const DataLayout &DL = PN.getModule()->getDataLayout();
1208
1209 // For now, we can only do this promotion if the load is in the same block
1210 // as the PHI, and if there are no stores between the phi and load.
1211 // TODO: Allow recursive phi users.
1212 // TODO: Allow stores.
1213 BasicBlock *BB = PN.getParent();
1214 Align MaxAlign;
1215 uint64_t APWidth = DL.getIndexTypeSizeInBits(PN.getType());
1216 APInt MaxSize(APWidth, 0);
1217 bool HaveLoad = false;
1218 for (User *U : PN.users()) {
1219 LoadInst *LI = dyn_cast<LoadInst>(U);
1220 if (!LI || !LI->isSimple())
1221 return false;
1222
1223 // For now we only allow loads in the same block as the PHI. This is
1224 // a common case that happens when instcombine merges two loads through
1225 // a PHI.
1226 if (LI->getParent() != BB)
1227 return false;
1228
1229 // Ensure that there are no instructions between the PHI and the load that
1230 // could store.
1231 for (BasicBlock::iterator BBI(PN); &*BBI != LI; ++BBI)
1232 if (BBI->mayWriteToMemory())
1233 return false;
1234
1235 uint64_t Size = DL.getTypeStoreSize(LI->getType()).getFixedSize();
1236 MaxAlign = std::max(MaxAlign, LI->getAlign());
1237 MaxSize = MaxSize.ult(Size) ? APInt(APWidth, Size) : MaxSize;
1238 HaveLoad = true;
1239 }
1240
1241 if (!HaveLoad)
1242 return false;
1243
1244 // We can only transform this if it is safe to push the loads into the
1245 // predecessor blocks. The only thing to watch out for is that we can't put
1246 // a possibly trapping load in the predecessor if it is a critical edge.
1247 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1248 Instruction *TI = PN.getIncomingBlock(Idx)->getTerminator();
1249 Value *InVal = PN.getIncomingValue(Idx);
1250
1251 // If the value is produced by the terminator of the predecessor (an
1252 // invoke) or it has side-effects, there is no valid place to put a load
1253 // in the predecessor.
1254 if (TI == InVal || TI->mayHaveSideEffects())
1255 return false;
1256
1257 // If the predecessor has a single successor, then the edge isn't
1258 // critical.
1259 if (TI->getNumSuccessors() == 1)
1260 continue;
1261
1262 // If this pointer is always safe to load, or if we can prove that there
1263 // is already a load in the block, then we can move the load to the pred
1264 // block.
1265 if (isSafeToLoadUnconditionally(InVal, MaxAlign, MaxSize, DL, TI))
1266 continue;
1267
1268 return false;
1269 }
1270
1271 return true;
1272}
1273
1274static void speculatePHINodeLoads(IRBuilderTy &IRB, PHINode &PN) {
1275 LLVM_DEBUG(dbgs() << " original: " << PN << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << PN <<
"\n"; } } while (false)
;
1276
1277 LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1278 Type *LoadTy = SomeLoad->getType();
1279 IRB.SetInsertPoint(&PN);
1280 PHINode *NewPN = IRB.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1281 PN.getName() + ".sroa.speculated");
1282
1283 // Get the AA tags and alignment to use from one of the loads. It does not
1284 // matter which one we get and if any differ.
1285 AAMDNodes AATags = SomeLoad->getAAMetadata();
1286 Align Alignment = SomeLoad->getAlign();
1287
1288 // Rewrite all loads of the PN to use the new PHI.
1289 while (!PN.use_empty()) {
1290 LoadInst *LI = cast<LoadInst>(PN.user_back());
1291 LI->replaceAllUsesWith(NewPN);
1292 LI->eraseFromParent();
1293 }
1294
1295 // Inject loads into all of the pred blocks.
1296 DenseMap<BasicBlock*, Value*> InjectedLoads;
1297 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1298 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1299 Value *InVal = PN.getIncomingValue(Idx);
1300
1301 // A PHI node is allowed to have multiple (duplicated) entries for the same
1302 // basic block, as long as the value is the same. So if we already injected
1303 // a load in the predecessor, then we should reuse the same load for all
1304 // duplicated entries.
1305 if (Value* V = InjectedLoads.lookup(Pred)) {
1306 NewPN->addIncoming(V, Pred);
1307 continue;
1308 }
1309
1310 Instruction *TI = Pred->getTerminator();
1311 IRB.SetInsertPoint(TI);
1312
1313 LoadInst *Load = IRB.CreateAlignedLoad(
1314 LoadTy, InVal, Alignment,
1315 (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1316 ++NumLoadsSpeculated;
1317 if (AATags)
1318 Load->setAAMetadata(AATags);
1319 NewPN->addIncoming(Load, Pred);
1320 InjectedLoads[Pred] = Load;
1321 }
1322
1323 LLVM_DEBUG(dbgs() << " speculated to: " << *NewPN << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " speculated to: " <<
*NewPN << "\n"; } } while (false)
;
1324 PN.eraseFromParent();
1325}
1326
1327/// Select instructions that use an alloca and are subsequently loaded can be
1328/// rewritten to load both input pointers and then select between the result,
1329/// allowing the load of the alloca to be promoted.
1330/// From this:
1331/// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1332/// %V = load i32* %P2
1333/// to:
1334/// %V1 = load i32* %Alloca -> will be mem2reg'd
1335/// %V2 = load i32* %Other
1336/// %V = select i1 %cond, i32 %V1, i32 %V2
1337///
1338/// We can do this to a select if its only uses are loads and if the operand
1339/// to the select can be loaded unconditionally. If found an intervening bitcast
1340/// with a single use of the load, allow the promotion.
1341static bool isSafeSelectToSpeculate(SelectInst &SI) {
1342 Value *TValue = SI.getTrueValue();
1343 Value *FValue = SI.getFalseValue();
1344 const DataLayout &DL = SI.getModule()->getDataLayout();
1345
1346 for (User *U : SI.users()) {
1347 LoadInst *LI;
1348 BitCastInst *BC = dyn_cast<BitCastInst>(U);
1349 if (BC && BC->hasOneUse())
1350 LI = dyn_cast<LoadInst>(*BC->user_begin());
1351 else
1352 LI = dyn_cast<LoadInst>(U);
1353
1354 if (!LI || !LI->isSimple())
1355 return false;
1356
1357 // Both operands to the select need to be dereferenceable, either
1358 // absolutely (e.g. allocas) or at this point because we can see other
1359 // accesses to it.
1360 if (!isSafeToLoadUnconditionally(TValue, LI->getType(),
1361 LI->getAlign(), DL, LI))
1362 return false;
1363 if (!isSafeToLoadUnconditionally(FValue, LI->getType(),
1364 LI->getAlign(), DL, LI))
1365 return false;
1366 }
1367
1368 return true;
1369}
1370
1371static void speculateSelectInstLoads(IRBuilderTy &IRB, SelectInst &SI) {
1372 LLVM_DEBUG(dbgs() << " original: " << SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << SI <<
"\n"; } } while (false)
;
1373
1374 IRB.SetInsertPoint(&SI);
1375 Value *TV = SI.getTrueValue();
1376 Value *FV = SI.getFalseValue();
1377 // Replace the loads of the select with a select of two loads.
1378 while (!SI.use_empty()) {
1379 LoadInst *LI;
1380 BitCastInst *BC = dyn_cast<BitCastInst>(SI.user_back());
1381 if (BC) {
1382 assert(BC->hasOneUse() && "Bitcast should have a single use.")(static_cast <bool> (BC->hasOneUse() && "Bitcast should have a single use."
) ? void (0) : __assert_fail ("BC->hasOneUse() && \"Bitcast should have a single use.\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1382, __extension__ __PRETTY_FUNCTION__
))
;
1383 LI = cast<LoadInst>(BC->user_back());
1384 } else {
1385 LI = cast<LoadInst>(SI.user_back());
1386 }
1387
1388 assert(LI->isSimple() && "We only speculate simple loads")(static_cast <bool> (LI->isSimple() && "We only speculate simple loads"
) ? void (0) : __assert_fail ("LI->isSimple() && \"We only speculate simple loads\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1388, __extension__ __PRETTY_FUNCTION__
))
;
1389
1390 IRB.SetInsertPoint(LI);
1391 Value *NewTV =
1392 BC ? IRB.CreateBitCast(TV, BC->getType(), TV->getName() + ".sroa.cast")
1393 : TV;
1394 Value *NewFV =
1395 BC ? IRB.CreateBitCast(FV, BC->getType(), FV->getName() + ".sroa.cast")
1396 : FV;
1397 LoadInst *TL = IRB.CreateLoad(LI->getType(), NewTV,
1398 LI->getName() + ".sroa.speculate.load.true");
1399 LoadInst *FL = IRB.CreateLoad(LI->getType(), NewFV,
1400 LI->getName() + ".sroa.speculate.load.false");
1401 NumLoadsSpeculated += 2;
1402
1403 // Transfer alignment and AA info if present.
1404 TL->setAlignment(LI->getAlign());
1405 FL->setAlignment(LI->getAlign());
1406
1407 AAMDNodes Tags = LI->getAAMetadata();
1408 if (Tags) {
1409 TL->setAAMetadata(Tags);
1410 FL->setAAMetadata(Tags);
1411 }
1412
1413 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1414 LI->getName() + ".sroa.speculated");
1415
1416 LLVM_DEBUG(dbgs() << " speculated to: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " speculated to: " <<
*V << "\n"; } } while (false)
;
1417 LI->replaceAllUsesWith(V);
1418 LI->eraseFromParent();
1419 if (BC)
1420 BC->eraseFromParent();
1421 }
1422 SI.eraseFromParent();
1423}
1424
1425/// Build a GEP out of a base pointer and indices.
1426///
1427/// This will return the BasePtr if that is valid, or build a new GEP
1428/// instruction using the IRBuilder if GEP-ing is needed.
1429static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1430 SmallVectorImpl<Value *> &Indices,
1431 const Twine &NamePrefix) {
1432 if (Indices.empty())
1433 return BasePtr;
1434
1435 // A single zero index is a no-op, so check for this and avoid building a GEP
1436 // in that case.
1437 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1438 return BasePtr;
1439
1440 return IRB.CreateInBoundsGEP(BasePtr->getType()->getPointerElementType(),
1441 BasePtr, Indices, NamePrefix + "sroa_idx");
1442}
1443
1444/// Get a natural GEP off of the BasePtr walking through Ty toward
1445/// TargetTy without changing the offset of the pointer.
1446///
1447/// This routine assumes we've already established a properly offset GEP with
1448/// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1449/// zero-indices down through type layers until we find one the same as
1450/// TargetTy. If we can't find one with the same type, we at least try to use
1451/// one with the same size. If none of that works, we just produce the GEP as
1452/// indicated by Indices to have the correct offset.
1453static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1454 Value *BasePtr, Type *Ty, Type *TargetTy,
1455 SmallVectorImpl<Value *> &Indices,
1456 const Twine &NamePrefix) {
1457 if (Ty == TargetTy)
1458 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1459
1460 // Offset size to use for the indices.
1461 unsigned OffsetSize = DL.getIndexTypeSizeInBits(BasePtr->getType());
1462
1463 // See if we can descend into a struct and locate a field with the correct
1464 // type.
1465 unsigned NumLayers = 0;
1466 Type *ElementTy = Ty;
1467 do {
1468 if (ElementTy->isPointerTy())
1469 break;
1470
1471 if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1472 ElementTy = ArrayTy->getElementType();
1473 Indices.push_back(IRB.getIntN(OffsetSize, 0));
1474 } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1475 ElementTy = VectorTy->getElementType();
1476 Indices.push_back(IRB.getInt32(0));
1477 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1478 if (STy->element_begin() == STy->element_end())
1479 break; // Nothing left to descend into.
1480 ElementTy = *STy->element_begin();
1481 Indices.push_back(IRB.getInt32(0));
1482 } else {
1483 break;
1484 }
1485 ++NumLayers;
1486 } while (ElementTy != TargetTy);
1487 if (ElementTy != TargetTy)
1488 Indices.erase(Indices.end() - NumLayers, Indices.end());
1489
1490 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1491}
1492
1493/// Get a natural GEP from a base pointer to a particular offset and
1494/// resulting in a particular type.
1495///
1496/// The goal is to produce a "natural" looking GEP that works with the existing
1497/// composite types to arrive at the appropriate offset and element type for
1498/// a pointer. TargetTy is the element type the returned GEP should point-to if
1499/// possible. We recurse by decreasing Offset, adding the appropriate index to
1500/// Indices, and setting Ty to the result subtype.
1501///
1502/// If no natural GEP can be constructed, this function returns null.
1503static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1504 Value *Ptr, APInt Offset, Type *TargetTy,
1505 SmallVectorImpl<Value *> &Indices,
1506 const Twine &NamePrefix) {
1507 PointerType *Ty = cast<PointerType>(Ptr->getType());
1508
1509 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1510 // an i8.
1511 if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1512 return nullptr;
1513
1514 Type *ElementTy = Ty->getElementType();
1515 if (!ElementTy->isSized())
1516 return nullptr; // We can't GEP through an unsized element.
1517
1518 SmallVector<APInt> IntIndices = DL.getGEPIndicesForOffset(ElementTy, Offset);
1519 if (Offset != 0)
1520 return nullptr;
1521
1522 for (const APInt &Index : IntIndices)
1523 Indices.push_back(IRB.getInt(Index));
1524 return getNaturalGEPWithType(IRB, DL, Ptr, ElementTy, TargetTy, Indices,
1525 NamePrefix);
1526}
1527
1528/// Compute an adjusted pointer from Ptr by Offset bytes where the
1529/// resulting pointer has PointerTy.
1530///
1531/// This tries very hard to compute a "natural" GEP which arrives at the offset
1532/// and produces the pointer type desired. Where it cannot, it will try to use
1533/// the natural GEP to arrive at the offset and bitcast to the type. Where that
1534/// fails, it will try to use an existing i8* and GEP to the byte offset and
1535/// bitcast to the type.
1536///
1537/// The strategy for finding the more natural GEPs is to peel off layers of the
1538/// pointer, walking back through bit casts and GEPs, searching for a base
1539/// pointer from which we can compute a natural GEP with the desired
1540/// properties. The algorithm tries to fold as many constant indices into
1541/// a single GEP as possible, thus making each GEP more independent of the
1542/// surrounding code.
1543static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1544 APInt Offset, Type *PointerTy,
1545 const Twine &NamePrefix) {
1546 // Create i8 GEP for opaque pointers.
1547 if (Ptr->getType()->isOpaquePointerTy()) {
1548 if (Offset != 0)
1549 Ptr = IRB.CreateInBoundsGEP(IRB.getInt8Ty(), Ptr, IRB.getInt(Offset),
1550 NamePrefix + "sroa_idx");
1551 return IRB.CreatePointerBitCastOrAddrSpaceCast(Ptr, PointerTy,
1552 NamePrefix + "sroa_cast");
1553 }
1554
1555 // Even though we don't look through PHI nodes, we could be called on an
1556 // instruction in an unreachable block, which may be on a cycle.
1557 SmallPtrSet<Value *, 4> Visited;
1558 Visited.insert(Ptr);
1559 SmallVector<Value *, 4> Indices;
1560
1561 // We may end up computing an offset pointer that has the wrong type. If we
1562 // never are able to compute one directly that has the correct type, we'll
1563 // fall back to it, so keep it and the base it was computed from around here.
1564 Value *OffsetPtr = nullptr;
1565 Value *OffsetBasePtr;
1566
1567 // Remember any i8 pointer we come across to re-use if we need to do a raw
1568 // byte offset.
1569 Value *Int8Ptr = nullptr;
1570 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1571
1572 PointerType *TargetPtrTy = cast<PointerType>(PointerTy);
1573 Type *TargetTy = TargetPtrTy->getElementType();
1574
1575 // As `addrspacecast` is , `Ptr` (the storage pointer) may have different
1576 // address space from the expected `PointerTy` (the pointer to be used).
1577 // Adjust the pointer type based the original storage pointer.
1578 auto AS = cast<PointerType>(Ptr->getType())->getAddressSpace();
1579 PointerTy = TargetTy->getPointerTo(AS);
1580
1581 do {
1582 // First fold any existing GEPs into the offset.
1583 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1584 APInt GEPOffset(Offset.getBitWidth(), 0);
1585 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1586 break;
1587 Offset += GEPOffset;
1588 Ptr = GEP->getPointerOperand();
1589 if (!Visited.insert(Ptr).second)
1590 break;
1591 }
1592
1593 // See if we can perform a natural GEP here.
1594 Indices.clear();
1595 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1596 Indices, NamePrefix)) {
1597 // If we have a new natural pointer at the offset, clear out any old
1598 // offset pointer we computed. Unless it is the base pointer or
1599 // a non-instruction, we built a GEP we don't need. Zap it.
1600 if (OffsetPtr && OffsetPtr != OffsetBasePtr)
1601 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr)) {
1602 assert(I->use_empty() && "Built a GEP with uses some how!")(static_cast <bool> (I->use_empty() && "Built a GEP with uses some how!"
) ? void (0) : __assert_fail ("I->use_empty() && \"Built a GEP with uses some how!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1602, __extension__ __PRETTY_FUNCTION__
))
;
1603 I->eraseFromParent();
1604 }
1605 OffsetPtr = P;
1606 OffsetBasePtr = Ptr;
1607 // If we also found a pointer of the right type, we're done.
1608 if (P->getType() == PointerTy)
1609 break;
1610 }
1611
1612 // Stash this pointer if we've found an i8*.
1613 if (Ptr->getType()->isIntegerTy(8)) {
1614 Int8Ptr = Ptr;
1615 Int8PtrOffset = Offset;
1616 }
1617
1618 // Peel off a layer of the pointer and update the offset appropriately.
1619 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1620 Ptr = cast<Operator>(Ptr)->getOperand(0);
1621 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1622 if (GA->isInterposable())
1623 break;
1624 Ptr = GA->getAliasee();
1625 } else {
1626 break;
1627 }
1628 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!")(static_cast <bool> (Ptr->getType()->isPointerTy(
) && "Unexpected operand type!") ? void (0) : __assert_fail
("Ptr->getType()->isPointerTy() && \"Unexpected operand type!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1628, __extension__ __PRETTY_FUNCTION__
))
;
1629 } while (Visited.insert(Ptr).second);
1630
1631 if (!OffsetPtr) {
1632 if (!Int8Ptr) {
1633 Int8Ptr = IRB.CreateBitCast(
1634 Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1635 NamePrefix + "sroa_raw_cast");
1636 Int8PtrOffset = Offset;
1637 }
1638
1639 OffsetPtr = Int8PtrOffset == 0
1640 ? Int8Ptr
1641 : IRB.CreateInBoundsGEP(IRB.getInt8Ty(), Int8Ptr,
1642 IRB.getInt(Int8PtrOffset),
1643 NamePrefix + "sroa_raw_idx");
1644 }
1645 Ptr = OffsetPtr;
1646
1647 // On the off chance we were targeting i8*, guard the bitcast here.
1648 if (cast<PointerType>(Ptr->getType()) != TargetPtrTy) {
1649 Ptr = IRB.CreatePointerBitCastOrAddrSpaceCast(Ptr,
1650 TargetPtrTy,
1651 NamePrefix + "sroa_cast");
1652 }
1653
1654 return Ptr;
1655}
1656
1657/// Compute the adjusted alignment for a load or store from an offset.
1658static Align getAdjustedAlignment(Instruction *I, uint64_t Offset) {
1659 return commonAlignment(getLoadStoreAlignment(I), Offset);
1660}
1661
1662/// Test whether we can convert a value from the old to the new type.
1663///
1664/// This predicate should be used to guard calls to convertValue in order to
1665/// ensure that we only try to convert viable values. The strategy is that we
1666/// will peel off single element struct and array wrappings to get to an
1667/// underlying value, and convert that value.
1668static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1669 if (OldTy == NewTy)
1670 return true;
1671
1672 // For integer types, we can't handle any bit-width differences. This would
1673 // break both vector conversions with extension and introduce endianness
1674 // issues when in conjunction with loads and stores.
1675 if (isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) {
1676 assert(cast<IntegerType>(OldTy)->getBitWidth() !=(static_cast <bool> (cast<IntegerType>(OldTy)->
getBitWidth() != cast<IntegerType>(NewTy)->getBitWidth
() && "We can't have the same bitwidth for different int types"
) ? void (0) : __assert_fail ("cast<IntegerType>(OldTy)->getBitWidth() != cast<IntegerType>(NewTy)->getBitWidth() && \"We can't have the same bitwidth for different int types\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1678, __extension__ __PRETTY_FUNCTION__
))
1677 cast<IntegerType>(NewTy)->getBitWidth() &&(static_cast <bool> (cast<IntegerType>(OldTy)->
getBitWidth() != cast<IntegerType>(NewTy)->getBitWidth
() && "We can't have the same bitwidth for different int types"
) ? void (0) : __assert_fail ("cast<IntegerType>(OldTy)->getBitWidth() != cast<IntegerType>(NewTy)->getBitWidth() && \"We can't have the same bitwidth for different int types\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1678, __extension__ __PRETTY_FUNCTION__
))
1678 "We can't have the same bitwidth for different int types")(static_cast <bool> (cast<IntegerType>(OldTy)->
getBitWidth() != cast<IntegerType>(NewTy)->getBitWidth
() && "We can't have the same bitwidth for different int types"
) ? void (0) : __assert_fail ("cast<IntegerType>(OldTy)->getBitWidth() != cast<IntegerType>(NewTy)->getBitWidth() && \"We can't have the same bitwidth for different int types\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1678, __extension__ __PRETTY_FUNCTION__
))
;
1679 return false;
1680 }
1681
1682 if (DL.getTypeSizeInBits(NewTy).getFixedSize() !=
1683 DL.getTypeSizeInBits(OldTy).getFixedSize())
1684 return false;
1685 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1686 return false;
1687
1688 // We can convert pointers to integers and vice-versa. Same for vectors
1689 // of pointers and integers.
1690 OldTy = OldTy->getScalarType();
1691 NewTy = NewTy->getScalarType();
1692 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1693 if (NewTy->isPointerTy() && OldTy->isPointerTy()) {
1694 unsigned OldAS = OldTy->getPointerAddressSpace();
1695 unsigned NewAS = NewTy->getPointerAddressSpace();
1696 // Convert pointers if they are pointers from the same address space or
1697 // different integral (not non-integral) address spaces with the same
1698 // pointer size.
1699 return OldAS == NewAS ||
1700 (!DL.isNonIntegralAddressSpace(OldAS) &&
1701 !DL.isNonIntegralAddressSpace(NewAS) &&
1702 DL.getPointerSize(OldAS) == DL.getPointerSize(NewAS));
1703 }
1704
1705 // We can convert integers to integral pointers, but not to non-integral
1706 // pointers.
1707 if (OldTy->isIntegerTy())
1708 return !DL.isNonIntegralPointerType(NewTy);
1709
1710 // We can convert integral pointers to integers, but non-integral pointers
1711 // need to remain pointers.
1712 if (!DL.isNonIntegralPointerType(OldTy))
1713 return NewTy->isIntegerTy();
1714
1715 return false;
1716 }
1717
1718 return true;
1719}
1720
1721/// Generic routine to convert an SSA value to a value of a different
1722/// type.
1723///
1724/// This will try various different casting techniques, such as bitcasts,
1725/// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1726/// two types for viability with this routine.
1727static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1728 Type *NewTy) {
1729 Type *OldTy = V->getType();
1730 assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type")(static_cast <bool> (canConvertValue(DL, OldTy, NewTy) &&
"Value not convertable to type") ? void (0) : __assert_fail (
"canConvertValue(DL, OldTy, NewTy) && \"Value not convertable to type\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1730, __extension__ __PRETTY_FUNCTION__
))
;
1731
1732 if (OldTy == NewTy)
1733 return V;
1734
1735 assert(!(isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) &&(static_cast <bool> (!(isa<IntegerType>(OldTy) &&
isa<IntegerType>(NewTy)) && "Integer types must be the exact same to convert."
) ? void (0) : __assert_fail ("!(isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) && \"Integer types must be the exact same to convert.\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1736, __extension__ __PRETTY_FUNCTION__
))
1736 "Integer types must be the exact same to convert.")(static_cast <bool> (!(isa<IntegerType>(OldTy) &&
isa<IntegerType>(NewTy)) && "Integer types must be the exact same to convert."
) ? void (0) : __assert_fail ("!(isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) && \"Integer types must be the exact same to convert.\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1736, __extension__ __PRETTY_FUNCTION__
))
;
1737
1738 // See if we need inttoptr for this type pair. May require additional bitcast.
1739 if (OldTy->isIntOrIntVectorTy() && NewTy->isPtrOrPtrVectorTy()) {
1740 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1741 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1742 // Expand <4 x i32> to <2 x i8*> --> <4 x i32> to <2 x i64> to <2 x i8*>
1743 // Directly handle i64 to i8*
1744 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1745 NewTy);
1746 }
1747
1748 // See if we need ptrtoint for this type pair. May require additional bitcast.
1749 if (OldTy->isPtrOrPtrVectorTy() && NewTy->isIntOrIntVectorTy()) {
1750 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1751 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1752 // Expand <2 x i8*> to <4 x i32> --> <2 x i8*> to <2 x i64> to <4 x i32>
1753 // Expand i8* to i64 --> i8* to i64 to i64
1754 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1755 NewTy);
1756 }
1757
1758 if (OldTy->isPtrOrPtrVectorTy() && NewTy->isPtrOrPtrVectorTy()) {
1759 unsigned OldAS = OldTy->getPointerAddressSpace();
1760 unsigned NewAS = NewTy->getPointerAddressSpace();
1761 // To convert pointers with different address spaces (they are already
1762 // checked convertible, i.e. they have the same pointer size), so far we
1763 // cannot use `bitcast` (which has restrict on the same address space) or
1764 // `addrspacecast` (which is not always no-op casting). Instead, use a pair
1765 // of no-op `ptrtoint`/`inttoptr` casts through an integer with the same bit
1766 // size.
1767 if (OldAS != NewAS) {
1768 assert(DL.getPointerSize(OldAS) == DL.getPointerSize(NewAS))(static_cast <bool> (DL.getPointerSize(OldAS) == DL.getPointerSize
(NewAS)) ? void (0) : __assert_fail ("DL.getPointerSize(OldAS) == DL.getPointerSize(NewAS)"
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1768, __extension__ __PRETTY_FUNCTION__
))
;
1769 return IRB.CreateIntToPtr(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1770 NewTy);
1771 }
1772 }
1773
1774 return IRB.CreateBitCast(V, NewTy);
1775}
1776
1777/// Test whether the given slice use can be promoted to a vector.
1778///
1779/// This function is called to test each entry in a partition which is slated
1780/// for a single slice.
1781static bool isVectorPromotionViableForSlice(Partition &P, const Slice &S,
1782 VectorType *Ty,
1783 uint64_t ElementSize,
1784 const DataLayout &DL) {
1785 // First validate the slice offsets.
1786 uint64_t BeginOffset =
1787 std::max(S.beginOffset(), P.beginOffset()) - P.beginOffset();
1788 uint64_t BeginIndex = BeginOffset / ElementSize;
1789 if (BeginIndex * ElementSize != BeginOffset ||
1790 BeginIndex >= cast<FixedVectorType>(Ty)->getNumElements())
1791 return false;
1792 uint64_t EndOffset =
1793 std::min(S.endOffset(), P.endOffset()) - P.beginOffset();
1794 uint64_t EndIndex = EndOffset / ElementSize;
1795 if (EndIndex * ElementSize != EndOffset ||
1796 EndIndex > cast<FixedVectorType>(Ty)->getNumElements())
1797 return false;
1798
1799 assert(EndIndex > BeginIndex && "Empty vector!")(static_cast <bool> (EndIndex > BeginIndex &&
"Empty vector!") ? void (0) : __assert_fail ("EndIndex > BeginIndex && \"Empty vector!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1799, __extension__ __PRETTY_FUNCTION__
))
;
1800 uint64_t NumElements = EndIndex - BeginIndex;
1801 Type *SliceTy = (NumElements == 1)
1802 ? Ty->getElementType()
1803 : FixedVectorType::get(Ty->getElementType(), NumElements);
1804
1805 Type *SplitIntTy =
1806 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1807
1808 Use *U = S.getUse();
1809
1810 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1811 if (MI->isVolatile())
1812 return false;
1813 if (!S.isSplittable())
1814 return false; // Skip any unsplittable intrinsics.
1815 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1816 if (!II->isLifetimeStartOrEnd() && !II->isDroppable())
1817 return false;
1818 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1819 if (LI->isVolatile())
1820 return false;
1821 Type *LTy = LI->getType();
1822 // Disable vector promotion when there are loads or stores of an FCA.
1823 if (LTy->isStructTy())
1824 return false;
1825 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1826 assert(LTy->isIntegerTy())(static_cast <bool> (LTy->isIntegerTy()) ? void (0) :
__assert_fail ("LTy->isIntegerTy()", "llvm/lib/Transforms/Scalar/SROA.cpp"
, 1826, __extension__ __PRETTY_FUNCTION__))
;
1827 LTy = SplitIntTy;
1828 }
1829 if (!canConvertValue(DL, SliceTy, LTy))
1830 return false;
1831 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1832 if (SI->isVolatile())
1833 return false;
1834 Type *STy = SI->getValueOperand()->getType();
1835 // Disable vector promotion when there are loads or stores of an FCA.
1836 if (STy->isStructTy())
1837 return false;
1838 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1839 assert(STy->isIntegerTy())(static_cast <bool> (STy->isIntegerTy()) ? void (0) :
__assert_fail ("STy->isIntegerTy()", "llvm/lib/Transforms/Scalar/SROA.cpp"
, 1839, __extension__ __PRETTY_FUNCTION__))
;
1840 STy = SplitIntTy;
1841 }
1842 if (!canConvertValue(DL, STy, SliceTy))
1843 return false;
1844 } else {
1845 return false;
1846 }
1847
1848 return true;
1849}
1850
1851/// Test whether the given alloca partitioning and range of slices can be
1852/// promoted to a vector.
1853///
1854/// This is a quick test to check whether we can rewrite a particular alloca
1855/// partition (and its newly formed alloca) into a vector alloca with only
1856/// whole-vector loads and stores such that it could be promoted to a vector
1857/// SSA value. We only can ensure this for a limited set of operations, and we
1858/// don't want to do the rewrites unless we are confident that the result will
1859/// be promotable, so we have an early test here.
1860static VectorType *isVectorPromotionViable(Partition &P, const DataLayout &DL) {
1861 // Collect the candidate types for vector-based promotion. Also track whether
1862 // we have different element types.
1863 SmallVector<VectorType *, 4> CandidateTys;
1864 Type *CommonEltTy = nullptr;
1865 bool HaveCommonEltTy = true;
1866 auto CheckCandidateType = [&](Type *Ty) {
1867 if (auto *VTy = dyn_cast<VectorType>(Ty)) {
1868 // Return if bitcast to vectors is different for total size in bits.
1869 if (!CandidateTys.empty()) {
1870 VectorType *V = CandidateTys[0];
1871 if (DL.getTypeSizeInBits(VTy).getFixedSize() !=
1872 DL.getTypeSizeInBits(V).getFixedSize()) {
1873 CandidateTys.clear();
1874 return;
1875 }
1876 }
1877 CandidateTys.push_back(VTy);
1878 if (!CommonEltTy)
1879 CommonEltTy = VTy->getElementType();
1880 else if (CommonEltTy != VTy->getElementType())
1881 HaveCommonEltTy = false;
1882 }
1883 };
1884 // Consider any loads or stores that are the exact size of the slice.
1885 for (const Slice &S : P)
1886 if (S.beginOffset() == P.beginOffset() &&
1887 S.endOffset() == P.endOffset()) {
1888 if (auto *LI = dyn_cast<LoadInst>(S.getUse()->getUser()))
1889 CheckCandidateType(LI->getType());
1890 else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser()))
1891 CheckCandidateType(SI->getValueOperand()->getType());
1892 }
1893
1894 // If we didn't find a vector type, nothing to do here.
1895 if (CandidateTys.empty())
1896 return nullptr;
1897
1898 // Remove non-integer vector types if we had multiple common element types.
1899 // FIXME: It'd be nice to replace them with integer vector types, but we can't
1900 // do that until all the backends are known to produce good code for all
1901 // integer vector types.
1902 if (!HaveCommonEltTy) {
1903 llvm::erase_if(CandidateTys, [](VectorType *VTy) {
1904 return !VTy->getElementType()->isIntegerTy();
1905 });
1906
1907 // If there were no integer vector types, give up.
1908 if (CandidateTys.empty())
1909 return nullptr;
1910
1911 // Rank the remaining candidate vector types. This is easy because we know
1912 // they're all integer vectors. We sort by ascending number of elements.
1913 auto RankVectorTypes = [&DL](VectorType *RHSTy, VectorType *LHSTy) {
1914 (void)DL;
1915 assert(DL.getTypeSizeInBits(RHSTy).getFixedSize() ==(static_cast <bool> (DL.getTypeSizeInBits(RHSTy).getFixedSize
() == DL.getTypeSizeInBits(LHSTy).getFixedSize() && "Cannot have vector types of different sizes!"
) ? void (0) : __assert_fail ("DL.getTypeSizeInBits(RHSTy).getFixedSize() == DL.getTypeSizeInBits(LHSTy).getFixedSize() && \"Cannot have vector types of different sizes!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1917, __extension__ __PRETTY_FUNCTION__
))
1916 DL.getTypeSizeInBits(LHSTy).getFixedSize() &&(static_cast <bool> (DL.getTypeSizeInBits(RHSTy).getFixedSize
() == DL.getTypeSizeInBits(LHSTy).getFixedSize() && "Cannot have vector types of different sizes!"
) ? void (0) : __assert_fail ("DL.getTypeSizeInBits(RHSTy).getFixedSize() == DL.getTypeSizeInBits(LHSTy).getFixedSize() && \"Cannot have vector types of different sizes!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1917, __extension__ __PRETTY_FUNCTION__
))
1917 "Cannot have vector types of different sizes!")(static_cast <bool> (DL.getTypeSizeInBits(RHSTy).getFixedSize
() == DL.getTypeSizeInBits(LHSTy).getFixedSize() && "Cannot have vector types of different sizes!"
) ? void (0) : __assert_fail ("DL.getTypeSizeInBits(RHSTy).getFixedSize() == DL.getTypeSizeInBits(LHSTy).getFixedSize() && \"Cannot have vector types of different sizes!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1917, __extension__ __PRETTY_FUNCTION__
))
;
1918 assert(RHSTy->getElementType()->isIntegerTy() &&(static_cast <bool> (RHSTy->getElementType()->isIntegerTy
() && "All non-integer types eliminated!") ? void (0)
: __assert_fail ("RHSTy->getElementType()->isIntegerTy() && \"All non-integer types eliminated!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1919, __extension__ __PRETTY_FUNCTION__
))
1919 "All non-integer types eliminated!")(static_cast <bool> (RHSTy->getElementType()->isIntegerTy
() && "All non-integer types eliminated!") ? void (0)
: __assert_fail ("RHSTy->getElementType()->isIntegerTy() && \"All non-integer types eliminated!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1919, __extension__ __PRETTY_FUNCTION__
))
;
1920 assert(LHSTy->getElementType()->isIntegerTy() &&(static_cast <bool> (LHSTy->getElementType()->isIntegerTy
() && "All non-integer types eliminated!") ? void (0)
: __assert_fail ("LHSTy->getElementType()->isIntegerTy() && \"All non-integer types eliminated!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1921, __extension__ __PRETTY_FUNCTION__
))
1921 "All non-integer types eliminated!")(static_cast <bool> (LHSTy->getElementType()->isIntegerTy
() && "All non-integer types eliminated!") ? void (0)
: __assert_fail ("LHSTy->getElementType()->isIntegerTy() && \"All non-integer types eliminated!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1921, __extension__ __PRETTY_FUNCTION__
))
;
1922 return cast<FixedVectorType>(RHSTy)->getNumElements() <
1923 cast<FixedVectorType>(LHSTy)->getNumElements();
1924 };
1925 llvm::sort(CandidateTys, RankVectorTypes);
1926 CandidateTys.erase(
1927 std::unique(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes),
1928 CandidateTys.end());
1929 } else {
1930// The only way to have the same element type in every vector type is to
1931// have the same vector type. Check that and remove all but one.
1932#ifndef NDEBUG
1933 for (VectorType *VTy : CandidateTys) {
1934 assert(VTy->getElementType() == CommonEltTy &&(static_cast <bool> (VTy->getElementType() == CommonEltTy
&& "Unaccounted for element type!") ? void (0) : __assert_fail
("VTy->getElementType() == CommonEltTy && \"Unaccounted for element type!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1935, __extension__ __PRETTY_FUNCTION__
))
1935 "Unaccounted for element type!")(static_cast <bool> (VTy->getElementType() == CommonEltTy
&& "Unaccounted for element type!") ? void (0) : __assert_fail
("VTy->getElementType() == CommonEltTy && \"Unaccounted for element type!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1935, __extension__ __PRETTY_FUNCTION__
))
;
1936 assert(VTy == CandidateTys[0] &&(static_cast <bool> (VTy == CandidateTys[0] && "Different vector types with the same element type!"
) ? void (0) : __assert_fail ("VTy == CandidateTys[0] && \"Different vector types with the same element type!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1937, __extension__ __PRETTY_FUNCTION__
))
1937 "Different vector types with the same element type!")(static_cast <bool> (VTy == CandidateTys[0] && "Different vector types with the same element type!"
) ? void (0) : __assert_fail ("VTy == CandidateTys[0] && \"Different vector types with the same element type!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1937, __extension__ __PRETTY_FUNCTION__
))
;
1938 }
1939#endif
1940 CandidateTys.resize(1);
1941 }
1942
1943 // Try each vector type, and return the one which works.
1944 auto CheckVectorTypeForPromotion = [&](VectorType *VTy) {
1945 uint64_t ElementSize =
1946 DL.getTypeSizeInBits(VTy->getElementType()).getFixedSize();
1947
1948 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1949 // that aren't byte sized.
1950 if (ElementSize % 8)
1951 return false;
1952 assert((DL.getTypeSizeInBits(VTy).getFixedSize() % 8) == 0 &&(static_cast <bool> ((DL.getTypeSizeInBits(VTy).getFixedSize
() % 8) == 0 && "vector size not a multiple of element size?"
) ? void (0) : __assert_fail ("(DL.getTypeSizeInBits(VTy).getFixedSize() % 8) == 0 && \"vector size not a multiple of element size?\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1953, __extension__ __PRETTY_FUNCTION__
))
1953 "vector size not a multiple of element size?")(static_cast <bool> ((DL.getTypeSizeInBits(VTy).getFixedSize
() % 8) == 0 && "vector size not a multiple of element size?"
) ? void (0) : __assert_fail ("(DL.getTypeSizeInBits(VTy).getFixedSize() % 8) == 0 && \"vector size not a multiple of element size?\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 1953, __extension__ __PRETTY_FUNCTION__
))
;
1954 ElementSize /= 8;
1955
1956 for (const Slice &S : P)
1957 if (!isVectorPromotionViableForSlice(P, S, VTy, ElementSize, DL))
1958 return false;
1959
1960 for (const Slice *S : P.splitSliceTails())
1961 if (!isVectorPromotionViableForSlice(P, *S, VTy, ElementSize, DL))
1962 return false;
1963
1964 return true;
1965 };
1966 for (VectorType *VTy : CandidateTys)
1967 if (CheckVectorTypeForPromotion(VTy))
1968 return VTy;
1969
1970 return nullptr;
1971}
1972
1973/// Test whether a slice of an alloca is valid for integer widening.
1974///
1975/// This implements the necessary checking for the \c isIntegerWideningViable
1976/// test below on a single slice of the alloca.
1977static bool isIntegerWideningViableForSlice(const Slice &S,
1978 uint64_t AllocBeginOffset,
1979 Type *AllocaTy,
1980 const DataLayout &DL,
1981 bool &WholeAllocaOp) {
1982 uint64_t Size = DL.getTypeStoreSize(AllocaTy).getFixedSize();
1983
1984 uint64_t RelBegin = S.beginOffset() - AllocBeginOffset;
1985 uint64_t RelEnd = S.endOffset() - AllocBeginOffset;
1986
1987 // We can't reasonably handle cases where the load or store extends past
1988 // the end of the alloca's type and into its padding.
1989 if (RelEnd > Size)
1990 return false;
1991
1992 Use *U = S.getUse();
1993
1994 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1995 if (LI->isVolatile())
1996 return false;
1997 // We can't handle loads that extend past the allocated memory.
1998 if (DL.getTypeStoreSize(LI->getType()).getFixedSize() > Size)
1999 return false;
2000 // So far, AllocaSliceRewriter does not support widening split slice tails
2001 // in rewriteIntegerLoad.
2002 if (S.beginOffset() < AllocBeginOffset)
2003 return false;
2004 // Note that we don't count vector loads or stores as whole-alloca
2005 // operations which enable integer widening because we would prefer to use
2006 // vector widening instead.
2007 if (!isa<VectorType>(LI->getType()) && RelBegin == 0 && RelEnd == Size)
2008 WholeAllocaOp = true;
2009 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
2010 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy).getFixedSize())
2011 return false;
2012 } else if (RelBegin != 0 || RelEnd != Size ||
2013 !canConvertValue(DL, AllocaTy, LI->getType())) {
2014 // Non-integer loads need to be convertible from the alloca type so that
2015 // they are promotable.
2016 return false;
2017 }
2018 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
2019 Type *ValueTy = SI->getValueOperand()->getType();
2020 if (SI->isVolatile())
2021 return false;
2022 // We can't handle stores that extend past the allocated memory.
2023 if (DL.getTypeStoreSize(ValueTy).getFixedSize() > Size)
2024 return false;
2025 // So far, AllocaSliceRewriter does not support widening split slice tails
2026 // in rewriteIntegerStore.
2027 if (S.beginOffset() < AllocBeginOffset)
2028 return false;
2029 // Note that we don't count vector loads or stores as whole-alloca
2030 // operations which enable integer widening because we would prefer to use
2031 // vector widening instead.
2032 if (!isa<VectorType>(ValueTy) && RelBegin == 0 && RelEnd == Size)
2033 WholeAllocaOp = true;
2034 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
2035 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy).getFixedSize())
2036 return false;
2037 } else if (RelBegin != 0 || RelEnd != Size ||
2038 !canConvertValue(DL, ValueTy, AllocaTy)) {
2039 // Non-integer stores need to be convertible to the alloca type so that
2040 // they are promotable.
2041 return false;
2042 }
2043 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
2044 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
2045 return false;
2046 if (!S.isSplittable())
2047 return false; // Skip any unsplittable intrinsics.
2048 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
2049 if (!II->isLifetimeStartOrEnd() && !II->isDroppable())
2050 return false;
2051 } else {
2052 return false;
2053 }
2054
2055 return true;
2056}
2057
2058/// Test whether the given alloca partition's integer operations can be
2059/// widened to promotable ones.
2060///
2061/// This is a quick test to check whether we can rewrite the integer loads and
2062/// stores to a particular alloca into wider loads and stores and be able to
2063/// promote the resulting alloca.
2064static bool isIntegerWideningViable(Partition &P, Type *AllocaTy,
2065 const DataLayout &DL) {
2066 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy).getFixedSize();
2067 // Don't create integer types larger than the maximum bitwidth.
2068 if (SizeInBits > IntegerType::MAX_INT_BITS)
2069 return false;
2070
2071 // Don't try to handle allocas with bit-padding.
2072 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy).getFixedSize())
2073 return false;
2074
2075 // We need to ensure that an integer type with the appropriate bitwidth can
2076 // be converted to the alloca type, whatever that is. We don't want to force
2077 // the alloca itself to have an integer type if there is a more suitable one.
2078 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
2079 if (!canConvertValue(DL, AllocaTy, IntTy) ||
2080 !canConvertValue(DL, IntTy, AllocaTy))
2081 return false;
2082
2083 // While examining uses, we ensure that the alloca has a covering load or
2084 // store. We don't want to widen the integer operations only to fail to
2085 // promote due to some other unsplittable entry (which we may make splittable
2086 // later). However, if there are only splittable uses, go ahead and assume
2087 // that we cover the alloca.
2088 // FIXME: We shouldn't consider split slices that happen to start in the
2089 // partition here...
2090 bool WholeAllocaOp = P.empty() && DL.isLegalInteger(SizeInBits);
2091
2092 for (const Slice &S : P)
2093 if (!isIntegerWideningViableForSlice(S, P.beginOffset(), AllocaTy, DL,
2094 WholeAllocaOp))
2095 return false;
2096
2097 for (const Slice *S : P.splitSliceTails())
2098 if (!isIntegerWideningViableForSlice(*S, P.beginOffset(), AllocaTy, DL,
2099 WholeAllocaOp))
2100 return false;
2101
2102 return WholeAllocaOp;
2103}
2104
2105static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
2106 IntegerType *Ty, uint64_t Offset,
2107 const Twine &Name) {
2108 LLVM_DEBUG(dbgs() << " start: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " start: " << *V <<
"\n"; } } while (false)
;
2109 IntegerType *IntTy = cast<IntegerType>(V->getType());
2110 assert(DL.getTypeStoreSize(Ty).getFixedSize() + Offset <=(static_cast <bool> (DL.getTypeStoreSize(Ty).getFixedSize
() + Offset <= DL.getTypeStoreSize(IntTy).getFixedSize() &&
"Element extends past full value") ? void (0) : __assert_fail
("DL.getTypeStoreSize(Ty).getFixedSize() + Offset <= DL.getTypeStoreSize(IntTy).getFixedSize() && \"Element extends past full value\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2112, __extension__ __PRETTY_FUNCTION__
))
2111 DL.getTypeStoreSize(IntTy).getFixedSize() &&(static_cast <bool> (DL.getTypeStoreSize(Ty).getFixedSize
() + Offset <= DL.getTypeStoreSize(IntTy).getFixedSize() &&
"Element extends past full value") ? void (0) : __assert_fail
("DL.getTypeStoreSize(Ty).getFixedSize() + Offset <= DL.getTypeStoreSize(IntTy).getFixedSize() && \"Element extends past full value\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2112, __extension__ __PRETTY_FUNCTION__
))
2112 "Element extends past full value")(static_cast <bool> (DL.getTypeStoreSize(Ty).getFixedSize
() + Offset <= DL.getTypeStoreSize(IntTy).getFixedSize() &&
"Element extends past full value") ? void (0) : __assert_fail
("DL.getTypeStoreSize(Ty).getFixedSize() + Offset <= DL.getTypeStoreSize(IntTy).getFixedSize() && \"Element extends past full value\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2112, __extension__ __PRETTY_FUNCTION__
))
;
2113 uint64_t ShAmt = 8 * Offset;
2114 if (DL.isBigEndian())
2115 ShAmt = 8 * (DL.getTypeStoreSize(IntTy).getFixedSize() -
2116 DL.getTypeStoreSize(Ty).getFixedSize() - Offset);
2117 if (ShAmt) {
2118 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2119 LLVM_DEBUG(dbgs() << " shifted: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " shifted: " << *V <<
"\n"; } } while (false)
;
2120 }
2121 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&(static_cast <bool> (Ty->getBitWidth() <= IntTy->
getBitWidth() && "Cannot extract to a larger integer!"
) ? void (0) : __assert_fail ("Ty->getBitWidth() <= IntTy->getBitWidth() && \"Cannot extract to a larger integer!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2122, __extension__ __PRETTY_FUNCTION__
))
2122 "Cannot extract to a larger integer!")(static_cast <bool> (Ty->getBitWidth() <= IntTy->
getBitWidth() && "Cannot extract to a larger integer!"
) ? void (0) : __assert_fail ("Ty->getBitWidth() <= IntTy->getBitWidth() && \"Cannot extract to a larger integer!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2122, __extension__ __PRETTY_FUNCTION__
))
;
2123 if (Ty != IntTy) {
2124 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2125 LLVM_DEBUG(dbgs() << " trunced: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " trunced: " << *V <<
"\n"; } } while (false)
;
2126 }
2127 return V;
2128}
2129
2130static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
2131 Value *V, uint64_t Offset, const Twine &Name) {
2132 IntegerType *IntTy = cast<IntegerType>(Old->getType());
2133 IntegerType *Ty = cast<IntegerType>(V->getType());
2134 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&(static_cast <bool> (Ty->getBitWidth() <= IntTy->
getBitWidth() && "Cannot insert a larger integer!") ?
void (0) : __assert_fail ("Ty->getBitWidth() <= IntTy->getBitWidth() && \"Cannot insert a larger integer!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2135, __extension__ __PRETTY_FUNCTION__
))
2135 "Cannot insert a larger integer!")(static_cast <bool> (Ty->getBitWidth() <= IntTy->
getBitWidth() && "Cannot insert a larger integer!") ?
void (0) : __assert_fail ("Ty->getBitWidth() <= IntTy->getBitWidth() && \"Cannot insert a larger integer!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2135, __extension__ __PRETTY_FUNCTION__
))
;
2136 LLVM_DEBUG(dbgs() << " start: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " start: " << *V <<
"\n"; } } while (false)
;
2137 if (Ty != IntTy) {
2138 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2139 LLVM_DEBUG(dbgs() << " extended: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " extended: " << *V <<
"\n"; } } while (false)
;
2140 }
2141 assert(DL.getTypeStoreSize(Ty).getFixedSize() + Offset <=(static_cast <bool> (DL.getTypeStoreSize(Ty).getFixedSize
() + Offset <= DL.getTypeStoreSize(IntTy).getFixedSize() &&
"Element store outside of alloca store") ? void (0) : __assert_fail
("DL.getTypeStoreSize(Ty).getFixedSize() + Offset <= DL.getTypeStoreSize(IntTy).getFixedSize() && \"Element store outside of alloca store\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2143, __extension__ __PRETTY_FUNCTION__
))
2142 DL.getTypeStoreSize(IntTy).getFixedSize() &&(static_cast <bool> (DL.getTypeStoreSize(Ty).getFixedSize
() + Offset <= DL.getTypeStoreSize(IntTy).getFixedSize() &&
"Element store outside of alloca store") ? void (0) : __assert_fail
("DL.getTypeStoreSize(Ty).getFixedSize() + Offset <= DL.getTypeStoreSize(IntTy).getFixedSize() && \"Element store outside of alloca store\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2143, __extension__ __PRETTY_FUNCTION__
))
2143 "Element store outside of alloca store")(static_cast <bool> (DL.getTypeStoreSize(Ty).getFixedSize
() + Offset <= DL.getTypeStoreSize(IntTy).getFixedSize() &&
"Element store outside of alloca store") ? void (0) : __assert_fail
("DL.getTypeStoreSize(Ty).getFixedSize() + Offset <= DL.getTypeStoreSize(IntTy).getFixedSize() && \"Element store outside of alloca store\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2143, __extension__ __PRETTY_FUNCTION__
))
;
2144 uint64_t ShAmt = 8 * Offset;
2145 if (DL.isBigEndian())
2146 ShAmt = 8 * (DL.getTypeStoreSize(IntTy).getFixedSize() -
2147 DL.getTypeStoreSize(Ty).getFixedSize() - Offset);
2148 if (ShAmt) {
2149 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2150 LLVM_DEBUG(dbgs() << " shifted: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " shifted: " << *V <<
"\n"; } } while (false)
;
2151 }
2152
2153 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2154 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2155 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2156 LLVM_DEBUG(dbgs() << " masked: " << *Old << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " masked: " << *Old <<
"\n"; } } while (false)
;
2157 V = IRB.CreateOr(Old, V, Name + ".insert");
2158 LLVM_DEBUG(dbgs() << " inserted: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " inserted: " << *V <<
"\n"; } } while (false)
;
2159 }
2160 return V;
2161}
2162
2163static Value *extractVector(IRBuilderTy &IRB, Value *V, unsigned BeginIndex,
2164 unsigned EndIndex, const Twine &Name) {
2165 auto *VecTy = cast<FixedVectorType>(V->getType());
2166 unsigned NumElements = EndIndex - BeginIndex;
2167 assert(NumElements <= VecTy->getNumElements() && "Too many elements!")(static_cast <bool> (NumElements <= VecTy->getNumElements
() && "Too many elements!") ? void (0) : __assert_fail
("NumElements <= VecTy->getNumElements() && \"Too many elements!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2167, __extension__ __PRETTY_FUNCTION__
))
;
2168
2169 if (NumElements == VecTy->getNumElements())
2170 return V;
2171
2172 if (NumElements == 1) {
2173 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2174 Name + ".extract");
2175 LLVM_DEBUG(dbgs() << " extract: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " extract: " << *V <<
"\n"; } } while (false)
;
2176 return V;
2177 }
2178
2179 SmallVector<int, 8> Mask;
2180 Mask.reserve(NumElements);
2181 for (unsigned i = BeginIndex; i != EndIndex; ++i)
2182 Mask.push_back(i);
2183 V = IRB.CreateShuffleVector(V, Mask, Name + ".extract");
2184 LLVM_DEBUG(dbgs() << " shuffle: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " shuffle: " << *V <<
"\n"; } } while (false)
;
2185 return V;
2186}
2187
2188static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
2189 unsigned BeginIndex, const Twine &Name) {
2190 VectorType *VecTy = cast<VectorType>(Old->getType());
2191 assert(VecTy && "Can only insert a vector into a vector")(static_cast <bool> (VecTy && "Can only insert a vector into a vector"
) ? void (0) : __assert_fail ("VecTy && \"Can only insert a vector into a vector\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2191, __extension__ __PRETTY_FUNCTION__
))
;
2192
2193 VectorType *Ty = dyn_cast<VectorType>(V->getType());
2194 if (!Ty) {
2195 // Single element to insert.
2196 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2197 Name + ".insert");
2198 LLVM_DEBUG(dbgs() << " insert: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " insert: " << *V <<
"\n"; } } while (false)
;
2199 return V;
2200 }
2201
2202 assert(cast<FixedVectorType>(Ty)->getNumElements() <=(static_cast <bool> (cast<FixedVectorType>(Ty)->
getNumElements() <= cast<FixedVectorType>(VecTy)->
getNumElements() && "Too many elements!") ? void (0) :
__assert_fail ("cast<FixedVectorType>(Ty)->getNumElements() <= cast<FixedVectorType>(VecTy)->getNumElements() && \"Too many elements!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2204, __extension__ __PRETTY_FUNCTION__
))
2203 cast<FixedVectorType>(VecTy)->getNumElements() &&(static_cast <bool> (cast<FixedVectorType>(Ty)->
getNumElements() <= cast<FixedVectorType>(VecTy)->
getNumElements() && "Too many elements!") ? void (0) :
__assert_fail ("cast<FixedVectorType>(Ty)->getNumElements() <= cast<FixedVectorType>(VecTy)->getNumElements() && \"Too many elements!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2204, __extension__ __PRETTY_FUNCTION__
))
2204 "Too many elements!")(static_cast <bool> (cast<FixedVectorType>(Ty)->
getNumElements() <= cast<FixedVectorType>(VecTy)->
getNumElements() && "Too many elements!") ? void (0) :
__assert_fail ("cast<FixedVectorType>(Ty)->getNumElements() <= cast<FixedVectorType>(VecTy)->getNumElements() && \"Too many elements!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2204, __extension__ __PRETTY_FUNCTION__
))
;
2205 if (cast<FixedVectorType>(Ty)->getNumElements() ==
2206 cast<FixedVectorType>(VecTy)->getNumElements()) {
2207 assert(V->getType() == VecTy && "Vector type mismatch")(static_cast <bool> (V->getType() == VecTy &&
"Vector type mismatch") ? void (0) : __assert_fail ("V->getType() == VecTy && \"Vector type mismatch\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2207, __extension__ __PRETTY_FUNCTION__
))
;
2208 return V;
2209 }
2210 unsigned EndIndex = BeginIndex + cast<FixedVectorType>(Ty)->getNumElements();
2211
2212 // When inserting a smaller vector into the larger to store, we first
2213 // use a shuffle vector to widen it with undef elements, and then
2214 // a second shuffle vector to select between the loaded vector and the
2215 // incoming vector.
2216 SmallVector<int, 8> Mask;
2217 Mask.reserve(cast<FixedVectorType>(VecTy)->getNumElements());
2218 for (unsigned i = 0; i != cast<FixedVectorType>(VecTy)->getNumElements(); ++i)
2219 if (i >= BeginIndex && i < EndIndex)
2220 Mask.push_back(i - BeginIndex);
2221 else
2222 Mask.push_back(-1);
2223 V = IRB.CreateShuffleVector(V, Mask, Name + ".expand");
2224 LLVM_DEBUG(dbgs() << " shuffle: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " shuffle: " << *V <<
"\n"; } } while (false)
;
2225
2226 SmallVector<Constant *, 8> Mask2;
2227 Mask2.reserve(cast<FixedVectorType>(VecTy)->getNumElements());
2228 for (unsigned i = 0; i != cast<FixedVectorType>(VecTy)->getNumElements(); ++i)
2229 Mask2.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
2230
2231 V = IRB.CreateSelect(ConstantVector::get(Mask2), V, Old, Name + "blend");
2232
2233 LLVM_DEBUG(dbgs() << " blend: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " blend: " << *V <<
"\n"; } } while (false)
;
2234 return V;
2235}
2236
2237/// Visitor to rewrite instructions using p particular slice of an alloca
2238/// to use a new alloca.
2239///
2240/// Also implements the rewriting to vector-based accesses when the partition
2241/// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2242/// lives here.
2243class llvm::sroa::AllocaSliceRewriter
2244 : public InstVisitor<AllocaSliceRewriter, bool> {
2245 // Befriend the base class so it can delegate to private visit methods.
2246 friend class InstVisitor<AllocaSliceRewriter, bool>;
2247
2248 using Base = InstVisitor<AllocaSliceRewriter, bool>;
2249
2250 const DataLayout &DL;
2251 AllocaSlices &AS;
2252 SROAPass &Pass;
2253 AllocaInst &OldAI, &NewAI;
2254 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2255 Type *NewAllocaTy;
2256
2257 // This is a convenience and flag variable that will be null unless the new
2258 // alloca's integer operations should be widened to this integer type due to
2259 // passing isIntegerWideningViable above. If it is non-null, the desired
2260 // integer type will be stored here for easy access during rewriting.
2261 IntegerType *IntTy;
2262
2263 // If we are rewriting an alloca partition which can be written as pure
2264 // vector operations, we stash extra information here. When VecTy is
2265 // non-null, we have some strict guarantees about the rewritten alloca:
2266 // - The new alloca is exactly the size of the vector type here.
2267 // - The accesses all either map to the entire vector or to a single
2268 // element.
2269 // - The set of accessing instructions is only one of those handled above
2270 // in isVectorPromotionViable. Generally these are the same access kinds
2271 // which are promotable via mem2reg.
2272 VectorType *VecTy;
2273 Type *ElementTy;
2274 uint64_t ElementSize;
2275
2276 // The original offset of the slice currently being rewritten relative to
2277 // the original alloca.
2278 uint64_t BeginOffset = 0;
2279 uint64_t EndOffset = 0;
2280
2281 // The new offsets of the slice currently being rewritten relative to the
2282 // original alloca.
2283 uint64_t NewBeginOffset = 0, NewEndOffset = 0;
2284
2285 uint64_t SliceSize = 0;
2286 bool IsSplittable = false;
2287 bool IsSplit = false;
2288 Use *OldUse = nullptr;
2289 Instruction *OldPtr = nullptr;
2290
2291 // Track post-rewrite users which are PHI nodes and Selects.
2292 SmallSetVector<PHINode *, 8> &PHIUsers;
2293 SmallSetVector<SelectInst *, 8> &SelectUsers;
2294
2295 // Utility IR builder, whose name prefix is setup for each visited use, and
2296 // the insertion point is set to point to the user.
2297 IRBuilderTy IRB;
2298
2299public:
2300 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &AS, SROAPass &Pass,
2301 AllocaInst &OldAI, AllocaInst &NewAI,
2302 uint64_t NewAllocaBeginOffset,
2303 uint64_t NewAllocaEndOffset, bool IsIntegerPromotable,
2304 VectorType *PromotableVecTy,
2305 SmallSetVector<PHINode *, 8> &PHIUsers,
2306 SmallSetVector<SelectInst *, 8> &SelectUsers)
2307 : DL(DL), AS(AS), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2308 NewAllocaBeginOffset(NewAllocaBeginOffset),
2309 NewAllocaEndOffset(NewAllocaEndOffset),
2310 NewAllocaTy(NewAI.getAllocatedType()),
2311 IntTy(
2312 IsIntegerPromotable
2313 ? Type::getIntNTy(NewAI.getContext(),
2314 DL.getTypeSizeInBits(NewAI.getAllocatedType())
2315 .getFixedSize())
2316 : nullptr),
2317 VecTy(PromotableVecTy),
2318 ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2319 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy).getFixedSize() / 8
2320 : 0),
2321 PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2322 IRB(NewAI.getContext(), ConstantFolder()) {
2323 if (VecTy) {
2324 assert((DL.getTypeSizeInBits(ElementTy).getFixedSize() % 8) == 0 &&(static_cast <bool> ((DL.getTypeSizeInBits(ElementTy).getFixedSize
() % 8) == 0 && "Only multiple-of-8 sized vector elements are viable"
) ? void (0) : __assert_fail ("(DL.getTypeSizeInBits(ElementTy).getFixedSize() % 8) == 0 && \"Only multiple-of-8 sized vector elements are viable\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2325, __extension__ __PRETTY_FUNCTION__
))
2325 "Only multiple-of-8 sized vector elements are viable")(static_cast <bool> ((DL.getTypeSizeInBits(ElementTy).getFixedSize
() % 8) == 0 && "Only multiple-of-8 sized vector elements are viable"
) ? void (0) : __assert_fail ("(DL.getTypeSizeInBits(ElementTy).getFixedSize() % 8) == 0 && \"Only multiple-of-8 sized vector elements are viable\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2325, __extension__ __PRETTY_FUNCTION__
))
;
2326 ++NumVectorized;
2327 }
2328 assert((!IntTy && !VecTy) || (IntTy && !VecTy) || (!IntTy && VecTy))(static_cast <bool> ((!IntTy && !VecTy) || (IntTy
&& !VecTy) || (!IntTy && VecTy)) ? void (0) :
__assert_fail ("(!IntTy && !VecTy) || (IntTy && !VecTy) || (!IntTy && VecTy)"
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2328, __extension__ __PRETTY_FUNCTION__
))
;
2329 }
2330
2331 bool visit(AllocaSlices::const_iterator I) {
2332 bool CanSROA = true;
2333 BeginOffset = I->beginOffset();
2334 EndOffset = I->endOffset();
2335 IsSplittable = I->isSplittable();
2336 IsSplit =
2337 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2338 LLVM_DEBUG(dbgs() << " rewriting " << (IsSplit ? "split " : ""))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " rewriting " << (IsSplit ?
"split " : ""); } } while (false)
;
2339 LLVM_DEBUG(AS.printSlice(dbgs(), I, ""))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { AS.printSlice(dbgs(), I, ""); } } while (false)
;
2340 LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "\n"; } } while (false)
;
2341
2342 // Compute the intersecting offset range.
2343 assert(BeginOffset < NewAllocaEndOffset)(static_cast <bool> (BeginOffset < NewAllocaEndOffset
) ? void (0) : __assert_fail ("BeginOffset < NewAllocaEndOffset"
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2343, __extension__ __PRETTY_FUNCTION__
))
;
2344 assert(EndOffset > NewAllocaBeginOffset)(static_cast <bool> (EndOffset > NewAllocaBeginOffset
) ? void (0) : __assert_fail ("EndOffset > NewAllocaBeginOffset"
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2344, __extension__ __PRETTY_FUNCTION__
))
;
2345 NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2346 NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2347
2348 SliceSize = NewEndOffset - NewBeginOffset;
2349
2350 OldUse = I->getUse();
2351 OldPtr = cast<Instruction>(OldUse->get());
2352
2353 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2354 IRB.SetInsertPoint(OldUserI);
2355 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2356 IRB.getInserter().SetNamePrefix(
2357 Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2358
2359 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2360 if (VecTy || IntTy)
2361 assert(CanSROA)(static_cast <bool> (CanSROA) ? void (0) : __assert_fail
("CanSROA", "llvm/lib/Transforms/Scalar/SROA.cpp", 2361, __extension__
__PRETTY_FUNCTION__))
;
2362 return CanSROA;
2363 }
2364
2365private:
2366 // Make sure the other visit overloads are visible.
2367 using Base::visit;
2368
2369 // Every instruction which can end up as a user must have a rewrite rule.
2370 bool visitInstruction(Instruction &I) {
2371 LLVM_DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " !!!! Cannot rewrite: " <<
I << "\n"; } } while (false)
;
2372 llvm_unreachable("No rewrite rule for this instruction!")::llvm::llvm_unreachable_internal("No rewrite rule for this instruction!"
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2372)
;
2373 }
2374
2375 Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2376 // Note that the offset computation can use BeginOffset or NewBeginOffset
2377 // interchangeably for unsplit slices.
2378 assert(IsSplit || BeginOffset == NewBeginOffset)(static_cast <bool> (IsSplit || BeginOffset == NewBeginOffset
) ? void (0) : __assert_fail ("IsSplit || BeginOffset == NewBeginOffset"
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2378, __extension__ __PRETTY_FUNCTION__
))
;
2379 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2380
2381#ifndef NDEBUG
2382 StringRef OldName = OldPtr->getName();
2383 // Skip through the last '.sroa.' component of the name.
2384 size_t LastSROAPrefix = OldName.rfind(".sroa.");
2385 if (LastSROAPrefix != StringRef::npos) {
2386 OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2387 // Look for an SROA slice index.
2388 size_t IndexEnd = OldName.find_first_not_of("0123456789");
2389 if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2390 // Strip the index and look for the offset.
2391 OldName = OldName.substr(IndexEnd + 1);
2392 size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2393 if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2394 // Strip the offset.
2395 OldName = OldName.substr(OffsetEnd + 1);
2396 }
2397 }
2398 // Strip any SROA suffixes as well.
2399 OldName = OldName.substr(0, OldName.find(".sroa_"));
2400#endif
2401
2402 return getAdjustedPtr(IRB, DL, &NewAI,
2403 APInt(DL.getIndexTypeSizeInBits(PointerTy), Offset),
2404 PointerTy,
2405#ifndef NDEBUG
2406 Twine(OldName) + "."
2407#else
2408 Twine()
2409#endif
2410 );
2411 }
2412
2413 /// Compute suitable alignment to access this slice of the *new*
2414 /// alloca.
2415 ///
2416 /// You can optionally pass a type to this routine and if that type's ABI
2417 /// alignment is itself suitable, this will return zero.
2418 Align getSliceAlign() {
2419 return commonAlignment(NewAI.getAlign(),
2420 NewBeginOffset - NewAllocaBeginOffset);
2421 }
2422
2423 unsigned getIndex(uint64_t Offset) {
2424 assert(VecTy && "Can only call getIndex when rewriting a vector")(static_cast <bool> (VecTy && "Can only call getIndex when rewriting a vector"
) ? void (0) : __assert_fail ("VecTy && \"Can only call getIndex when rewriting a vector\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2424, __extension__ __PRETTY_FUNCTION__
))
;
2425 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2426 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds")(static_cast <bool> (RelOffset / ElementSize < (4294967295U
) && "Index out of bounds") ? void (0) : __assert_fail
("RelOffset / ElementSize < UINT32_MAX && \"Index out of bounds\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2426, __extension__ __PRETTY_FUNCTION__
))
;
2427 uint32_t Index = RelOffset / ElementSize;
2428 assert(Index * ElementSize == RelOffset)(static_cast <bool> (Index * ElementSize == RelOffset) ?
void (0) : __assert_fail ("Index * ElementSize == RelOffset"
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2428, __extension__ __PRETTY_FUNCTION__
))
;
2429 return Index;
2430 }
2431
2432 void deleteIfTriviallyDead(Value *V) {
2433 Instruction *I = cast<Instruction>(V);
2434 if (isInstructionTriviallyDead(I))
2435 Pass.DeadInsts.push_back(I);
2436 }
2437
2438 Value *rewriteVectorizedLoadInst(LoadInst &LI) {
2439 unsigned BeginIndex = getIndex(NewBeginOffset);
2440 unsigned EndIndex = getIndex(NewEndOffset);
2441 assert(EndIndex > BeginIndex && "Empty vector!")(static_cast <bool> (EndIndex > BeginIndex &&
"Empty vector!") ? void (0) : __assert_fail ("EndIndex > BeginIndex && \"Empty vector!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2441, __extension__ __PRETTY_FUNCTION__
))
;
2442
2443 LoadInst *Load = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2444 NewAI.getAlign(), "load");
2445
2446 Load->copyMetadata(LI, {LLVMContext::MD_mem_parallel_loop_access,
2447 LLVMContext::MD_access_group});
2448 return extractVector(IRB, Load, BeginIndex, EndIndex, "vec");
2449 }
2450
2451 Value *rewriteIntegerLoad(LoadInst &LI) {
2452 assert(IntTy && "We cannot insert an integer to the alloca")(static_cast <bool> (IntTy && "We cannot insert an integer to the alloca"
) ? void (0) : __assert_fail ("IntTy && \"We cannot insert an integer to the alloca\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2452, __extension__ __PRETTY_FUNCTION__
))
;
2453 assert(!LI.isVolatile())(static_cast <bool> (!LI.isVolatile()) ? void (0) : __assert_fail
("!LI.isVolatile()", "llvm/lib/Transforms/Scalar/SROA.cpp", 2453
, __extension__ __PRETTY_FUNCTION__))
;
2454 Value *V = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2455 NewAI.getAlign(), "load");
2456 V = convertValue(DL, IRB, V, IntTy);
2457 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset")(static_cast <bool> (NewBeginOffset >= NewAllocaBeginOffset
&& "Out of bounds offset") ? void (0) : __assert_fail
("NewBeginOffset >= NewAllocaBeginOffset && \"Out of bounds offset\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2457, __extension__ __PRETTY_FUNCTION__
))
;
2458 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2459 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset) {
2460 IntegerType *ExtractTy = Type::getIntNTy(LI.getContext(), SliceSize * 8);
2461 V = extractInteger(DL, IRB, V, ExtractTy, Offset, "extract");
2462 }
2463 // It is possible that the extracted type is not the load type. This
2464 // happens if there is a load past the end of the alloca, and as
2465 // a consequence the slice is narrower but still a candidate for integer
2466 // lowering. To handle this case, we just zero extend the extracted
2467 // integer.
2468 assert(cast<IntegerType>(LI.getType())->getBitWidth() >= SliceSize * 8 &&(static_cast <bool> (cast<IntegerType>(LI.getType
())->getBitWidth() >= SliceSize * 8 && "Can only handle an extract for an overly wide load"
) ? void (0) : __assert_fail ("cast<IntegerType>(LI.getType())->getBitWidth() >= SliceSize * 8 && \"Can only handle an extract for an overly wide load\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2469, __extension__ __PRETTY_FUNCTION__
))
2469 "Can only handle an extract for an overly wide load")(static_cast <bool> (cast<IntegerType>(LI.getType
())->getBitWidth() >= SliceSize * 8 && "Can only handle an extract for an overly wide load"
) ? void (0) : __assert_fail ("cast<IntegerType>(LI.getType())->getBitWidth() >= SliceSize * 8 && \"Can only handle an extract for an overly wide load\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2469, __extension__ __PRETTY_FUNCTION__
))
;
2470 if (cast<IntegerType>(LI.getType())->getBitWidth() > SliceSize * 8)
2471 V = IRB.CreateZExt(V, LI.getType());
2472 return V;
2473 }
2474
2475 bool visitLoadInst(LoadInst &LI) {
2476 LLVM_DEBUG(dbgs() << " original: " << LI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << LI <<
"\n"; } } while (false)
;
2477 Value *OldOp = LI.getOperand(0);
2478 assert(OldOp == OldPtr)(static_cast <bool> (OldOp == OldPtr) ? void (0) : __assert_fail
("OldOp == OldPtr", "llvm/lib/Transforms/Scalar/SROA.cpp", 2478
, __extension__ __PRETTY_FUNCTION__))
;
2479
2480 AAMDNodes AATags = LI.getAAMetadata();
2481
2482 unsigned AS = LI.getPointerAddressSpace();
2483
2484 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2485 : LI.getType();
2486 const bool IsLoadPastEnd =
2487 DL.getTypeStoreSize(TargetTy).getFixedSize() > SliceSize;
2488 bool IsPtrAdjusted = false;
2489 Value *V;
2490 if (VecTy) {
2491 V = rewriteVectorizedLoadInst(LI);
2492 } else if (IntTy && LI.getType()->isIntegerTy()) {
2493 V = rewriteIntegerLoad(LI);
2494 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2495 NewEndOffset == NewAllocaEndOffset &&
2496 (canConvertValue(DL, NewAllocaTy, TargetTy) ||
2497 (IsLoadPastEnd && NewAllocaTy->isIntegerTy() &&
2498 TargetTy->isIntegerTy()))) {
2499 LoadInst *NewLI = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2500 NewAI.getAlign(), LI.isVolatile(),
2501 LI.getName());
2502 if (AATags)
2503 NewLI->setAAMetadata(AATags.shift(NewBeginOffset - BeginOffset));
2504 if (LI.isVolatile())
2505 NewLI->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
2506 if (NewLI->isAtomic())
2507 NewLI->setAlignment(LI.getAlign());
2508
2509 // Any !nonnull metadata or !range metadata on the old load is also valid
2510 // on the new load. This is even true in some cases even when the loads
2511 // are different types, for example by mapping !nonnull metadata to
2512 // !range metadata by modeling the null pointer constant converted to the
2513 // integer type.
2514 // FIXME: Add support for range metadata here. Currently the utilities
2515 // for this don't propagate range metadata in trivial cases from one
2516 // integer load to another, don't handle non-addrspace-0 null pointers
2517 // correctly, and don't have any support for mapping ranges as the
2518 // integer type becomes winder or narrower.
2519 if (MDNode *N = LI.getMetadata(LLVMContext::MD_nonnull))
2520 copyNonnullMetadata(LI, N, *NewLI);
2521
2522 // Try to preserve nonnull metadata
2523 V = NewLI;
2524
2525 // If this is an integer load past the end of the slice (which means the
2526 // bytes outside the slice are undef or this load is dead) just forcibly
2527 // fix the integer size with correct handling of endianness.
2528 if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2529 if (auto *TITy = dyn_cast<IntegerType>(TargetTy))
2530 if (AITy->getBitWidth() < TITy->getBitWidth()) {
2531 V = IRB.CreateZExt(V, TITy, "load.ext");
2532 if (DL.isBigEndian())
2533 V = IRB.CreateShl(V, TITy->getBitWidth() - AITy->getBitWidth(),
2534 "endian_shift");
2535 }
2536 } else {
2537 Type *LTy = TargetTy->getPointerTo(AS);
2538 LoadInst *NewLI =
2539 IRB.CreateAlignedLoad(TargetTy, getNewAllocaSlicePtr(IRB, LTy),
2540 getSliceAlign(), LI.isVolatile(), LI.getName());
2541 if (AATags)
2542 NewLI->setAAMetadata(AATags.shift(NewBeginOffset - BeginOffset));
2543 if (LI.isVolatile())
2544 NewLI->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
2545 NewLI->copyMetadata(LI, {LLVMContext::MD_mem_parallel_loop_access,
2546 LLVMContext::MD_access_group});
2547
2548 V = NewLI;
2549 IsPtrAdjusted = true;
2550 }
2551 V = convertValue(DL, IRB, V, TargetTy);
2552
2553 if (IsSplit) {
2554 assert(!LI.isVolatile())(static_cast <bool> (!LI.isVolatile()) ? void (0) : __assert_fail
("!LI.isVolatile()", "llvm/lib/Transforms/Scalar/SROA.cpp", 2554
, __extension__ __PRETTY_FUNCTION__))
;
2555 assert(LI.getType()->isIntegerTy() &&(static_cast <bool> (LI.getType()->isIntegerTy() &&
"Only integer type loads and stores are split") ? void (0) :
__assert_fail ("LI.getType()->isIntegerTy() && \"Only integer type loads and stores are split\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2556, __extension__ __PRETTY_FUNCTION__
))
2556 "Only integer type loads and stores are split")(static_cast <bool> (LI.getType()->isIntegerTy() &&
"Only integer type loads and stores are split") ? void (0) :
__assert_fail ("LI.getType()->isIntegerTy() && \"Only integer type loads and stores are split\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2556, __extension__ __PRETTY_FUNCTION__
))
;
2557 assert(SliceSize < DL.getTypeStoreSize(LI.getType()).getFixedSize() &&(static_cast <bool> (SliceSize < DL.getTypeStoreSize
(LI.getType()).getFixedSize() && "Split load isn't smaller than original load"
) ? void (0) : __assert_fail ("SliceSize < DL.getTypeStoreSize(LI.getType()).getFixedSize() && \"Split load isn't smaller than original load\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2558, __extension__ __PRETTY_FUNCTION__
))
2558 "Split load isn't smaller than original load")(static_cast <bool> (SliceSize < DL.getTypeStoreSize
(LI.getType()).getFixedSize() && "Split load isn't smaller than original load"
) ? void (0) : __assert_fail ("SliceSize < DL.getTypeStoreSize(LI.getType()).getFixedSize() && \"Split load isn't smaller than original load\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2558, __extension__ __PRETTY_FUNCTION__
))
;
2559 assert(DL.typeSizeEqualsStoreSize(LI.getType()) &&(static_cast <bool> (DL.typeSizeEqualsStoreSize(LI.getType
()) && "Non-byte-multiple bit width") ? void (0) : __assert_fail
("DL.typeSizeEqualsStoreSize(LI.getType()) && \"Non-byte-multiple bit width\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2560, __extension__ __PRETTY_FUNCTION__
))
2560 "Non-byte-multiple bit width")(static_cast <bool> (DL.typeSizeEqualsStoreSize(LI.getType
()) && "Non-byte-multiple bit width") ? void (0) : __assert_fail
("DL.typeSizeEqualsStoreSize(LI.getType()) && \"Non-byte-multiple bit width\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2560, __extension__ __PRETTY_FUNCTION__
))
;
2561 // Move the insertion point just past the load so that we can refer to it.
2562 IRB.SetInsertPoint(&*std::next(BasicBlock::iterator(&LI)));
2563 // Create a placeholder value with the same type as LI to use as the
2564 // basis for the new value. This allows us to replace the uses of LI with
2565 // the computed value, and then replace the placeholder with LI, leaving
2566 // LI only used for this computation.
2567 Value *Placeholder = new LoadInst(
2568 LI.getType(), PoisonValue::get(LI.getType()->getPointerTo(AS)), "",
2569 false, Align(1));
2570 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset - BeginOffset,
2571 "insert");
2572 LI.replaceAllUsesWith(V);
2573 Placeholder->replaceAllUsesWith(&LI);
2574 Placeholder->deleteValue();
2575 } else {
2576 LI.replaceAllUsesWith(V);
2577 }
2578
2579 Pass.DeadInsts.push_back(&LI);
2580 deleteIfTriviallyDead(OldOp);
2581 LLVM_DEBUG(dbgs() << " to: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *V <<
"\n"; } } while (false)
;
2582 return !LI.isVolatile() && !IsPtrAdjusted;
2583 }
2584
2585 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp,
2586 AAMDNodes AATags) {
2587 if (V->getType() != VecTy) {
2588 unsigned BeginIndex = getIndex(NewBeginOffset);
2589 unsigned EndIndex = getIndex(NewEndOffset);
2590 assert(EndIndex > BeginIndex && "Empty vector!")(static_cast <bool> (EndIndex > BeginIndex &&
"Empty vector!") ? void (0) : __assert_fail ("EndIndex > BeginIndex && \"Empty vector!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2590, __extension__ __PRETTY_FUNCTION__
))
;
2591 unsigned NumElements = EndIndex - BeginIndex;
2592 assert(NumElements <= cast<FixedVectorType>(VecTy)->getNumElements() &&(static_cast <bool> (NumElements <= cast<FixedVectorType
>(VecTy)->getNumElements() && "Too many elements!"
) ? void (0) : __assert_fail ("NumElements <= cast<FixedVectorType>(VecTy)->getNumElements() && \"Too many elements!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2593, __extension__ __PRETTY_FUNCTION__
))
2593 "Too many elements!")(static_cast <bool> (NumElements <= cast<FixedVectorType
>(VecTy)->getNumElements() && "Too many elements!"
) ? void (0) : __assert_fail ("NumElements <= cast<FixedVectorType>(VecTy)->getNumElements() && \"Too many elements!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2593, __extension__ __PRETTY_FUNCTION__
))
;
2594 Type *SliceTy = (NumElements == 1)
2595 ? ElementTy
2596 : FixedVectorType::get(ElementTy, NumElements);
2597 if (V->getType() != SliceTy)
2598 V = convertValue(DL, IRB, V, SliceTy);
2599
2600 // Mix in the existing elements.
2601 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2602 NewAI.getAlign(), "load");
2603 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2604 }
2605 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlign());
2606 Store->copyMetadata(SI, {LLVMContext::MD_mem_parallel_loop_access,
2607 LLVMContext::MD_access_group});
2608 if (AATags)
2609 Store->setAAMetadata(AATags.shift(NewBeginOffset - BeginOffset));
2610 Pass.DeadInsts.push_back(&SI);
2611
2612 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *Store <<
"\n"; } } while (false)
;
2613 return true;
2614 }
2615
2616 bool rewriteIntegerStore(Value *V, StoreInst &SI, AAMDNodes AATags) {
2617 assert(IntTy && "We cannot extract an integer from the alloca")(static_cast <bool> (IntTy && "We cannot extract an integer from the alloca"
) ? void (0) : __assert_fail ("IntTy && \"We cannot extract an integer from the alloca\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2617, __extension__ __PRETTY_FUNCTION__
))
;
2618 assert(!SI.isVolatile())(static_cast <bool> (!SI.isVolatile()) ? void (0) : __assert_fail
("!SI.isVolatile()", "llvm/lib/Transforms/Scalar/SROA.cpp", 2618
, __extension__ __PRETTY_FUNCTION__))
;
2619 if (DL.getTypeSizeInBits(V->getType()).getFixedSize() !=
2620 IntTy->getBitWidth()) {
2621 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2622 NewAI.getAlign(), "oldload");
2623 Old = convertValue(DL, IRB, Old, IntTy);
2624 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset")(static_cast <bool> (BeginOffset >= NewAllocaBeginOffset
&& "Out of bounds offset") ? void (0) : __assert_fail
("BeginOffset >= NewAllocaBeginOffset && \"Out of bounds offset\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2624, __extension__ __PRETTY_FUNCTION__
))
;
2625 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2626 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset, "insert");
2627 }
2628 V = convertValue(DL, IRB, V, NewAllocaTy);
2629 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlign());
2630 Store->copyMetadata(SI, {LLVMContext::MD_mem_parallel_loop_access,
2631 LLVMContext::MD_access_group});
2632 if (AATags)
2633 Store->setAAMetadata(AATags.shift(NewBeginOffset - BeginOffset));
2634 Pass.DeadInsts.push_back(&SI);
2635 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *Store <<
"\n"; } } while (false)
;
2636 return true;
2637 }
2638
2639 bool visitStoreInst(StoreInst &SI) {
2640 LLVM_DEBUG(dbgs() << " original: " << SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << SI <<
"\n"; } } while (false)
;
2641 Value *OldOp = SI.getOperand(1);
2642 assert(OldOp == OldPtr)(static_cast <bool> (OldOp == OldPtr) ? void (0) : __assert_fail
("OldOp == OldPtr", "llvm/lib/Transforms/Scalar/SROA.cpp", 2642
, __extension__ __PRETTY_FUNCTION__))
;
2643
2644 AAMDNodes AATags = SI.getAAMetadata();
2645 Value *V = SI.getValueOperand();
2646
2647 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2648 // alloca that should be re-examined after promoting this alloca.
2649 if (V->getType()->isPointerTy())
2650 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2651 Pass.PostPromotionWorklist.insert(AI);
2652
2653 if (SliceSize < DL.getTypeStoreSize(V->getType()).getFixedSize()) {
2654 assert(!SI.isVolatile())(static_cast <bool> (!SI.isVolatile()) ? void (0) : __assert_fail
("!SI.isVolatile()", "llvm/lib/Transforms/Scalar/SROA.cpp", 2654
, __extension__ __PRETTY_FUNCTION__))
;
2655 assert(V->getType()->isIntegerTy() &&(static_cast <bool> (V->getType()->isIntegerTy() &&
"Only integer type loads and stores are split") ? void (0) :
__assert_fail ("V->getType()->isIntegerTy() && \"Only integer type loads and stores are split\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2656, __extension__ __PRETTY_FUNCTION__
))
2656 "Only integer type loads and stores are split")(static_cast <bool> (V->getType()->isIntegerTy() &&
"Only integer type loads and stores are split") ? void (0) :
__assert_fail ("V->getType()->isIntegerTy() && \"Only integer type loads and stores are split\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2656, __extension__ __PRETTY_FUNCTION__
))
;
2657 assert(DL.typeSizeEqualsStoreSize(V->getType()) &&(static_cast <bool> (DL.typeSizeEqualsStoreSize(V->getType
()) && "Non-byte-multiple bit width") ? void (0) : __assert_fail
("DL.typeSizeEqualsStoreSize(V->getType()) && \"Non-byte-multiple bit width\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2658, __extension__ __PRETTY_FUNCTION__
))
2658 "Non-byte-multiple bit width")(static_cast <bool> (DL.typeSizeEqualsStoreSize(V->getType
()) && "Non-byte-multiple bit width") ? void (0) : __assert_fail
("DL.typeSizeEqualsStoreSize(V->getType()) && \"Non-byte-multiple bit width\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2658, __extension__ __PRETTY_FUNCTION__
))
;
2659 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2660 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset - BeginOffset,
2661 "extract");
2662 }
2663
2664 if (VecTy)
2665 return rewriteVectorizedStoreInst(V, SI, OldOp, AATags);
2666 if (IntTy && V->getType()->isIntegerTy())
2667 return rewriteIntegerStore(V, SI, AATags);
2668
2669 const bool IsStorePastEnd =
2670 DL.getTypeStoreSize(V->getType()).getFixedSize() > SliceSize;
2671 StoreInst *NewSI;
2672 if (NewBeginOffset == NewAllocaBeginOffset &&
2673 NewEndOffset == NewAllocaEndOffset &&
2674 (canConvertValue(DL, V->getType(), NewAllocaTy) ||
2675 (IsStorePastEnd && NewAllocaTy->isIntegerTy() &&
2676 V->getType()->isIntegerTy()))) {
2677 // If this is an integer store past the end of slice (and thus the bytes
2678 // past that point are irrelevant or this is unreachable), truncate the
2679 // value prior to storing.
2680 if (auto *VITy = dyn_cast<IntegerType>(V->getType()))
2681 if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2682 if (VITy->getBitWidth() > AITy->getBitWidth()) {
2683 if (DL.isBigEndian())
2684 V = IRB.CreateLShr(V, VITy->getBitWidth() - AITy->getBitWidth(),
2685 "endian_shift");
2686 V = IRB.CreateTrunc(V, AITy, "load.trunc");
2687 }
2688
2689 V = convertValue(DL, IRB, V, NewAllocaTy);
2690 NewSI =
2691 IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlign(), SI.isVolatile());
2692 } else {
2693 unsigned AS = SI.getPointerAddressSpace();
2694 Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo(AS));
2695 NewSI =
2696 IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(), SI.isVolatile());
2697 }
2698 NewSI->copyMetadata(SI, {LLVMContext::MD_mem_parallel_loop_access,
2699 LLVMContext::MD_access_group});
2700 if (AATags)
2701 NewSI->setAAMetadata(AATags.shift(NewBeginOffset - BeginOffset));
2702 if (SI.isVolatile())
2703 NewSI->setAtomic(SI.getOrdering(), SI.getSyncScopeID());
2704 if (NewSI->isAtomic())
2705 NewSI->setAlignment(SI.getAlign());
2706 Pass.DeadInsts.push_back(&SI);
2707 deleteIfTriviallyDead(OldOp);
2708
2709 LLVM_DEBUG(dbgs() << " to: " << *NewSI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *NewSI <<
"\n"; } } while (false)
;
2710 return NewSI->getPointerOperand() == &NewAI &&
2711 NewSI->getValueOperand()->getType() == NewAllocaTy &&
2712 !SI.isVolatile();
2713 }
2714
2715 /// Compute an integer value from splatting an i8 across the given
2716 /// number of bytes.
2717 ///
2718 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2719 /// call this routine.
2720 /// FIXME: Heed the advice above.
2721 ///
2722 /// \param V The i8 value to splat.
2723 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2724 Value *getIntegerSplat(Value *V, unsigned Size) {
2725 assert(Size > 0 && "Expected a positive number of bytes.")(static_cast <bool> (Size > 0 && "Expected a positive number of bytes."
) ? void (0) : __assert_fail ("Size > 0 && \"Expected a positive number of bytes.\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2725, __extension__ __PRETTY_FUNCTION__
))
;
2726 IntegerType *VTy = cast<IntegerType>(V->getType());
2727 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte")(static_cast <bool> (VTy->getBitWidth() == 8 &&
"Expected an i8 value for the byte") ? void (0) : __assert_fail
("VTy->getBitWidth() == 8 && \"Expected an i8 value for the byte\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2727, __extension__ __PRETTY_FUNCTION__
))
;
2728 if (Size == 1)
2729 return V;
2730
2731 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size * 8);
2732 V = IRB.CreateMul(
2733 IRB.CreateZExt(V, SplatIntTy, "zext"),
2734 ConstantExpr::getUDiv(
2735 Constant::getAllOnesValue(SplatIntTy),
2736 ConstantExpr::getZExt(Constant::getAllOnesValue(V->getType()),
2737 SplatIntTy)),
2738 "isplat");
2739 return V;
2740 }
2741
2742 /// Compute a vector splat for a given element value.
2743 Value *getVectorSplat(Value *V, unsigned NumElements) {
2744 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2745 LLVM_DEBUG(dbgs() << " splat: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " splat: " << *V <<
"\n"; } } while (false)
;
2746 return V;
2747 }
2748
2749 bool visitMemSetInst(MemSetInst &II) {
2750 LLVM_DEBUG(dbgs() << " original: " << II << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << II <<
"\n"; } } while (false)
;
2751 assert(II.getRawDest() == OldPtr)(static_cast <bool> (II.getRawDest() == OldPtr) ? void (
0) : __assert_fail ("II.getRawDest() == OldPtr", "llvm/lib/Transforms/Scalar/SROA.cpp"
, 2751, __extension__ __PRETTY_FUNCTION__))
;
2752
2753 AAMDNodes AATags = II.getAAMetadata();
2754
2755 // If the memset has a variable size, it cannot be split, just adjust the
2756 // pointer to the new alloca.
2757 if (!isa<ConstantInt>(II.getLength())) {
2758 assert(!IsSplit)(static_cast <bool> (!IsSplit) ? void (0) : __assert_fail
("!IsSplit", "llvm/lib/Transforms/Scalar/SROA.cpp", 2758, __extension__
__PRETTY_FUNCTION__))
;
2759 assert(NewBeginOffset == BeginOffset)(static_cast <bool> (NewBeginOffset == BeginOffset) ? void
(0) : __assert_fail ("NewBeginOffset == BeginOffset", "llvm/lib/Transforms/Scalar/SROA.cpp"
, 2759, __extension__ __PRETTY_FUNCTION__))
;
2760 II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2761 II.setDestAlignment(getSliceAlign());
2762
2763 deleteIfTriviallyDead(OldPtr);
2764 return false;
2765 }
2766
2767 // Record this instruction for deletion.
2768 Pass.DeadInsts.push_back(&II);
2769
2770 Type *AllocaTy = NewAI.getAllocatedType();
2771 Type *ScalarTy = AllocaTy->getScalarType();
2772
2773 const bool CanContinue = [&]() {
2774 if (VecTy || IntTy)
2775 return true;
2776 if (BeginOffset > NewAllocaBeginOffset ||
2777 EndOffset < NewAllocaEndOffset)
2778 return false;
2779 // Length must be in range for FixedVectorType.
2780 auto *C = cast<ConstantInt>(II.getLength());
2781 const uint64_t Len = C->getLimitedValue();
2782 if (Len > std::numeric_limits<unsigned>::max())
2783 return false;
2784 auto *Int8Ty = IntegerType::getInt8Ty(NewAI.getContext());
2785 auto *SrcTy = FixedVectorType::get(Int8Ty, Len);
2786 return canConvertValue(DL, SrcTy, AllocaTy) &&
2787 DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy).getFixedSize());
2788 }();
2789
2790 // If this doesn't map cleanly onto the alloca type, and that type isn't
2791 // a single value type, just emit a memset.
2792 if (!CanContinue) {
2793 Type *SizeTy = II.getLength()->getType();
2794 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2795 CallInst *New = IRB.CreateMemSet(
2796 getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2797 MaybeAlign(getSliceAlign()), II.isVolatile());
2798 if (AATags)
2799 New->setAAMetadata(AATags.shift(NewBeginOffset - BeginOffset));
2800 LLVM_DEBUG(dbgs() << " to: " << *New << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *New <<
"\n"; } } while (false)
;
2801 return false;
2802 }
2803
2804 // If we can represent this as a simple value, we have to build the actual
2805 // value to store, which requires expanding the byte present in memset to
2806 // a sensible representation for the alloca type. This is essentially
2807 // splatting the byte to a sufficiently wide integer, splatting it across
2808 // any desired vector width, and bitcasting to the final type.
2809 Value *V;
2810
2811 if (VecTy) {
2812 // If this is a memset of a vectorized alloca, insert it.
2813 assert(ElementTy == ScalarTy)(static_cast <bool> (ElementTy == ScalarTy) ? void (0) :
__assert_fail ("ElementTy == ScalarTy", "llvm/lib/Transforms/Scalar/SROA.cpp"
, 2813, __extension__ __PRETTY_FUNCTION__))
;
2814
2815 unsigned BeginIndex = getIndex(NewBeginOffset);
2816 unsigned EndIndex = getIndex(NewEndOffset);
2817 assert(EndIndex > BeginIndex && "Empty vector!")(static_cast <bool> (EndIndex > BeginIndex &&
"Empty vector!") ? void (0) : __assert_fail ("EndIndex > BeginIndex && \"Empty vector!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2817, __extension__ __PRETTY_FUNCTION__
))
;
2818 unsigned NumElements = EndIndex - BeginIndex;
2819 assert(NumElements <= cast<FixedVectorType>(VecTy)->getNumElements() &&(static_cast <bool> (NumElements <= cast<FixedVectorType
>(VecTy)->getNumElements() && "Too many elements!"
) ? void (0) : __assert_fail ("NumElements <= cast<FixedVectorType>(VecTy)->getNumElements() && \"Too many elements!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2820, __extension__ __PRETTY_FUNCTION__
))
2820 "Too many elements!")(static_cast <bool> (NumElements <= cast<FixedVectorType
>(VecTy)->getNumElements() && "Too many elements!"
) ? void (0) : __assert_fail ("NumElements <= cast<FixedVectorType>(VecTy)->getNumElements() && \"Too many elements!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2820, __extension__ __PRETTY_FUNCTION__
))
;
2821
2822 Value *Splat = getIntegerSplat(
2823 II.getValue(), DL.getTypeSizeInBits(ElementTy).getFixedSize() / 8);
2824 Splat = convertValue(DL, IRB, Splat, ElementTy);
2825 if (NumElements > 1)
2826 Splat = getVectorSplat(Splat, NumElements);
2827
2828 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2829 NewAI.getAlign(), "oldload");
2830 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2831 } else if (IntTy) {
2832 // If this is a memset on an alloca where we can widen stores, insert the
2833 // set integer.
2834 assert(!II.isVolatile())(static_cast <bool> (!II.isVolatile()) ? void (0) : __assert_fail
("!II.isVolatile()", "llvm/lib/Transforms/Scalar/SROA.cpp", 2834
, __extension__ __PRETTY_FUNCTION__))
;
2835
2836 uint64_t Size = NewEndOffset - NewBeginOffset;
2837 V = getIntegerSplat(II.getValue(), Size);
2838
2839 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2840 EndOffset != NewAllocaBeginOffset)) {
2841 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2842 NewAI.getAlign(), "oldload");
2843 Old = convertValue(DL, IRB, Old, IntTy);
2844 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2845 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2846 } else {
2847 assert(V->getType() == IntTy &&(static_cast <bool> (V->getType() == IntTy &&
"Wrong type for an alloca wide integer!") ? void (0) : __assert_fail
("V->getType() == IntTy && \"Wrong type for an alloca wide integer!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2848, __extension__ __PRETTY_FUNCTION__
))
2848 "Wrong type for an alloca wide integer!")(static_cast <bool> (V->getType() == IntTy &&
"Wrong type for an alloca wide integer!") ? void (0) : __assert_fail
("V->getType() == IntTy && \"Wrong type for an alloca wide integer!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2848, __extension__ __PRETTY_FUNCTION__
))
;
2849 }
2850 V = convertValue(DL, IRB, V, AllocaTy);
2851 } else {
2852 // Established these invariants above.
2853 assert(NewBeginOffset == NewAllocaBeginOffset)(static_cast <bool> (NewBeginOffset == NewAllocaBeginOffset
) ? void (0) : __assert_fail ("NewBeginOffset == NewAllocaBeginOffset"
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2853, __extension__ __PRETTY_FUNCTION__
))
;
2854 assert(NewEndOffset == NewAllocaEndOffset)(static_cast <bool> (NewEndOffset == NewAllocaEndOffset
) ? void (0) : __assert_fail ("NewEndOffset == NewAllocaEndOffset"
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2854, __extension__ __PRETTY_FUNCTION__
))
;
2855
2856 V = getIntegerSplat(II.getValue(),
2857 DL.getTypeSizeInBits(ScalarTy).getFixedSize() / 8);
2858 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2859 V = getVectorSplat(
2860 V, cast<FixedVectorType>(AllocaVecTy)->getNumElements());
2861
2862 V = convertValue(DL, IRB, V, AllocaTy);
2863 }
2864
2865 StoreInst *New =
2866 IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlign(), II.isVolatile());
2867 New->copyMetadata(II, {LLVMContext::MD_mem_parallel_loop_access,
2868 LLVMContext::MD_access_group});
2869 if (AATags)
2870 New->setAAMetadata(AATags.shift(NewBeginOffset - BeginOffset));
2871 LLVM_DEBUG(dbgs() << " to: " << *New << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *New <<
"\n"; } } while (false)
;
2872 return !II.isVolatile();
2873 }
2874
2875 bool visitMemTransferInst(MemTransferInst &II) {
2876 // Rewriting of memory transfer instructions can be a bit tricky. We break
2877 // them into two categories: split intrinsics and unsplit intrinsics.
2878
2879 LLVM_DEBUG(dbgs() << " original: " << II << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << II <<
"\n"; } } while (false)
;
1
Assuming 'DebugFlag' is false
2
Loop condition is false. Exiting loop
2880
2881 AAMDNodes AATags = II.getAAMetadata();
2882
2883 bool IsDest = &II.getRawDestUse() == OldUse;
2884 assert((IsDest && II.getRawDest() == OldPtr) ||(static_cast <bool> ((IsDest && II.getRawDest()
== OldPtr) || (!IsDest && II.getRawSource() == OldPtr
)) ? void (0) : __assert_fail ("(IsDest && II.getRawDest() == OldPtr) || (!IsDest && II.getRawSource() == OldPtr)"
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2885, __extension__ __PRETTY_FUNCTION__
))
3
Assuming 'IsDest' is false
4
Assuming pointer value is null
5
'?' condition is true
2885 (!IsDest && II.getRawSource() == OldPtr))(static_cast <bool> ((IsDest && II.getRawDest()
== OldPtr) || (!IsDest && II.getRawSource() == OldPtr
)) ? void (0) : __assert_fail ("(IsDest && II.getRawDest() == OldPtr) || (!IsDest && II.getRawSource() == OldPtr)"
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2885, __extension__ __PRETTY_FUNCTION__
))
;
2886
2887 MaybeAlign SliceAlign = getSliceAlign();
2888
2889 // For unsplit intrinsics, we simply modify the source and destination
2890 // pointers in place. This isn't just an optimization, it is a matter of
2891 // correctness. With unsplit intrinsics we may be dealing with transfers
2892 // within a single alloca before SROA ran, or with transfers that have
2893 // a variable length. We may also be dealing with memmove instead of
2894 // memcpy, and so simply updating the pointers is the necessary for us to
2895 // update both source and dest of a single call.
2896 if (!IsSplittable) {
6
Assuming field 'IsSplittable' is false
7
Taking true branch
2897 Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
8
Called C++ object pointer is null
2898 if (IsDest) {
2899 II.setDest(AdjustedPtr);
2900 II.setDestAlignment(SliceAlign);
2901 }
2902 else {
2903 II.setSource(AdjustedPtr);
2904 II.setSourceAlignment(SliceAlign);
2905 }
2906
2907 LLVM_DEBUG(dbgs() << " to: " << II << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << II <<
"\n"; } } while (false)
;
2908 deleteIfTriviallyDead(OldPtr);
2909 return false;
2910 }
2911 // For split transfer intrinsics we have an incredibly useful assurance:
2912 // the source and destination do not reside within the same alloca, and at
2913 // least one of them does not escape. This means that we can replace
2914 // memmove with memcpy, and we don't need to worry about all manner of
2915 // downsides to splitting and transforming the operations.
2916
2917 // If this doesn't map cleanly onto the alloca type, and that type isn't
2918 // a single value type, just emit a memcpy.
2919 bool EmitMemCpy =
2920 !VecTy && !IntTy &&
2921 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2922 SliceSize !=
2923 DL.getTypeStoreSize(NewAI.getAllocatedType()).getFixedSize() ||
2924 !NewAI.getAllocatedType()->isSingleValueType());
2925
2926 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2927 // size hasn't been shrunk based on analysis of the viable range, this is
2928 // a no-op.
2929 if (EmitMemCpy && &OldAI == &NewAI) {
2930 // Ensure the start lines up.
2931 assert(NewBeginOffset == BeginOffset)(static_cast <bool> (NewBeginOffset == BeginOffset) ? void
(0) : __assert_fail ("NewBeginOffset == BeginOffset", "llvm/lib/Transforms/Scalar/SROA.cpp"
, 2931, __extension__ __PRETTY_FUNCTION__))
;
2932
2933 // Rewrite the size as needed.
2934 if (NewEndOffset != EndOffset)
2935 II.setLength(ConstantInt::get(II.getLength()->getType(),
2936 NewEndOffset - NewBeginOffset));
2937 return false;
2938 }
2939 // Record this instruction for deletion.
2940 Pass.DeadInsts.push_back(&II);
2941
2942 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2943 // alloca that should be re-examined after rewriting this instruction.
2944 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2945 if (AllocaInst *AI =
2946 dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2947 assert(AI != &OldAI && AI != &NewAI &&(static_cast <bool> (AI != &OldAI && AI != &
NewAI && "Splittable transfers cannot reach the same alloca on both ends."
) ? void (0) : __assert_fail ("AI != &OldAI && AI != &NewAI && \"Splittable transfers cannot reach the same alloca on both ends.\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2948, __extension__ __PRETTY_FUNCTION__
))
2948 "Splittable transfers cannot reach the same alloca on both ends.")(static_cast <bool> (AI != &OldAI && AI != &
NewAI && "Splittable transfers cannot reach the same alloca on both ends."
) ? void (0) : __assert_fail ("AI != &OldAI && AI != &NewAI && \"Splittable transfers cannot reach the same alloca on both ends.\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 2948, __extension__ __PRETTY_FUNCTION__
))
;
2949 Pass.Worklist.insert(AI);
2950 }
2951
2952 Type *OtherPtrTy = OtherPtr->getType();
2953 unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2954
2955 // Compute the relative offset for the other pointer within the transfer.
2956 unsigned OffsetWidth = DL.getIndexSizeInBits(OtherAS);
2957 APInt OtherOffset(OffsetWidth, NewBeginOffset - BeginOffset);
2958 Align OtherAlign =
2959 (IsDest ? II.getSourceAlign() : II.getDestAlign()).valueOrOne();
2960 OtherAlign =
2961 commonAlignment(OtherAlign, OtherOffset.zextOrTrunc(64).getZExtValue());
2962
2963 if (EmitMemCpy) {
2964 // Compute the other pointer, folding as much as possible to produce
2965 // a single, simple GEP in most cases.
2966 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2967 OtherPtr->getName() + ".");
2968
2969 Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2970 Type *SizeTy = II.getLength()->getType();
2971 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2972
2973 Value *DestPtr, *SrcPtr;
2974 MaybeAlign DestAlign, SrcAlign;
2975 // Note: IsDest is true iff we're copying into the new alloca slice
2976 if (IsDest) {
2977 DestPtr = OurPtr;
2978 DestAlign = SliceAlign;
2979 SrcPtr = OtherPtr;
2980 SrcAlign = OtherAlign;
2981 } else {
2982 DestPtr = OtherPtr;
2983 DestAlign = OtherAlign;
2984 SrcPtr = OurPtr;
2985 SrcAlign = SliceAlign;
2986 }
2987 CallInst *New = IRB.CreateMemCpy(DestPtr, DestAlign, SrcPtr, SrcAlign,
2988 Size, II.isVolatile());
2989 if (AATags)
2990 New->setAAMetadata(AATags.shift(NewBeginOffset - BeginOffset));
2991 LLVM_DEBUG(dbgs() << " to: " << *New << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *New <<
"\n"; } } while (false)
;
2992 return false;
2993 }
2994
2995 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2996 NewEndOffset == NewAllocaEndOffset;
2997 uint64_t Size = NewEndOffset - NewBeginOffset;
2998 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2999 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
3000 unsigned NumElements = EndIndex - BeginIndex;
3001 IntegerType *SubIntTy =
3002 IntTy ? Type::getIntNTy(IntTy->getContext(), Size * 8) : nullptr;
3003
3004 // Reset the other pointer type to match the register type we're going to
3005 // use, but using the address space of the original other pointer.
3006 Type *OtherTy;
3007 if (VecTy && !IsWholeAlloca) {
3008 if (NumElements == 1)
3009 OtherTy = VecTy->getElementType();
3010 else
3011 OtherTy = FixedVectorType::get(VecTy->getElementType(), NumElements);
3012 } else if (IntTy && !IsWholeAlloca) {
3013 OtherTy = SubIntTy;
3014 } else {
3015 OtherTy = NewAllocaTy;
3016 }
3017 OtherPtrTy = OtherTy->getPointerTo(OtherAS);
3018
3019 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
3020 OtherPtr->getName() + ".");
3021 MaybeAlign SrcAlign = OtherAlign;
3022 Value *DstPtr = &NewAI;
3023 MaybeAlign DstAlign = SliceAlign;
3024 if (!IsDest) {
3025 std::swap(SrcPtr, DstPtr);
3026 std::swap(SrcAlign, DstAlign);
3027 }
3028
3029 Value *Src;
3030 if (VecTy && !IsWholeAlloca && !IsDest) {
3031 Src = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
3032 NewAI.getAlign(), "load");
3033 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
3034 } else if (IntTy && !IsWholeAlloca && !IsDest) {
3035 Src = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
3036 NewAI.getAlign(), "load");
3037 Src = convertValue(DL, IRB, Src, IntTy);
3038 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
3039 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
3040 } else {
3041 LoadInst *Load = IRB.CreateAlignedLoad(OtherTy, SrcPtr, SrcAlign,
3042 II.isVolatile(), "copyload");
3043 Load->copyMetadata(II, {LLVMContext::MD_mem_parallel_loop_access,
3044 LLVMContext::MD_access_group});
3045 if (AATags)
3046 Load->setAAMetadata(AATags.shift(NewBeginOffset - BeginOffset));
3047 Src = Load;
3048 }
3049
3050 if (VecTy && !IsWholeAlloca && IsDest) {
3051 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
3052 NewAI.getAlign(), "oldload");
3053 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
3054 } else if (IntTy && !IsWholeAlloca && IsDest) {
3055 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
3056 NewAI.getAlign(), "oldload");
3057 Old = convertValue(DL, IRB, Old, IntTy);
3058 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
3059 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
3060 Src = convertValue(DL, IRB, Src, NewAllocaTy);
3061 }
3062
3063 StoreInst *Store = cast<StoreInst>(
3064 IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
3065 Store->copyMetadata(II, {LLVMContext::MD_mem_parallel_loop_access,
3066 LLVMContext::MD_access_group});
3067 if (AATags)
3068 Store->setAAMetadata(AATags.shift(NewBeginOffset - BeginOffset));
3069 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *Store <<
"\n"; } } while (false)
;
3070 return !II.isVolatile();
3071 }
3072
3073 bool visitIntrinsicInst(IntrinsicInst &II) {
3074 assert((II.isLifetimeStartOrEnd() || II.isDroppable()) &&(static_cast <bool> ((II.isLifetimeStartOrEnd() || II.isDroppable
()) && "Unexpected intrinsic!") ? void (0) : __assert_fail
("(II.isLifetimeStartOrEnd() || II.isDroppable()) && \"Unexpected intrinsic!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3075, __extension__ __PRETTY_FUNCTION__
))
3075 "Unexpected intrinsic!")(static_cast <bool> ((II.isLifetimeStartOrEnd() || II.isDroppable
()) && "Unexpected intrinsic!") ? void (0) : __assert_fail
("(II.isLifetimeStartOrEnd() || II.isDroppable()) && \"Unexpected intrinsic!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3075, __extension__ __PRETTY_FUNCTION__
))
;
3076 LLVM_DEBUG(dbgs() << " original: " << II << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << II <<
"\n"; } } while (false)
;
3077
3078 // Record this instruction for deletion.
3079 Pass.DeadInsts.push_back(&II);
3080
3081 if (II.isDroppable()) {
3082 assert(II.getIntrinsicID() == Intrinsic::assume && "Expected assume")(static_cast <bool> (II.getIntrinsicID() == Intrinsic::
assume && "Expected assume") ? void (0) : __assert_fail
("II.getIntrinsicID() == Intrinsic::assume && \"Expected assume\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3082, __extension__ __PRETTY_FUNCTION__
))
;
3083 // TODO For now we forget assumed information, this can be improved.
3084 OldPtr->dropDroppableUsesIn(II);
3085 return true;
3086 }
3087
3088 assert(II.getArgOperand(1) == OldPtr)(static_cast <bool> (II.getArgOperand(1) == OldPtr) ? void
(0) : __assert_fail ("II.getArgOperand(1) == OldPtr", "llvm/lib/Transforms/Scalar/SROA.cpp"
, 3088, __extension__ __PRETTY_FUNCTION__))
;
3089 // Lifetime intrinsics are only promotable if they cover the whole alloca.
3090 // Therefore, we drop lifetime intrinsics which don't cover the whole
3091 // alloca.
3092 // (In theory, intrinsics which partially cover an alloca could be
3093 // promoted, but PromoteMemToReg doesn't handle that case.)
3094 // FIXME: Check whether the alloca is promotable before dropping the
3095 // lifetime intrinsics?
3096 if (NewBeginOffset != NewAllocaBeginOffset ||
3097 NewEndOffset != NewAllocaEndOffset)
3098 return true;
3099
3100 ConstantInt *Size =
3101 ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
3102 NewEndOffset - NewBeginOffset);
3103 // Lifetime intrinsics always expect an i8* so directly get such a pointer
3104 // for the new alloca slice.
3105 Type *PointerTy = IRB.getInt8PtrTy(OldPtr->getType()->getPointerAddressSpace());
3106 Value *Ptr = getNewAllocaSlicePtr(IRB, PointerTy);
3107 Value *New;
3108 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
3109 New = IRB.CreateLifetimeStart(Ptr, Size);
3110 else
3111 New = IRB.CreateLifetimeEnd(Ptr, Size);
3112
3113 (void)New;
3114 LLVM_DEBUG(dbgs() << " to: " << *New << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *New <<
"\n"; } } while (false)
;
3115
3116 return true;
3117 }
3118
3119 void fixLoadStoreAlign(Instruction &Root) {
3120 // This algorithm implements the same visitor loop as
3121 // hasUnsafePHIOrSelectUse, and fixes the alignment of each load
3122 // or store found.
3123 SmallPtrSet<Instruction *, 4> Visited;
3124 SmallVector<Instruction *, 4> Uses;
3125 Visited.insert(&Root);
3126 Uses.push_back(&Root);
3127 do {
3128 Instruction *I = Uses.pop_back_val();
3129
3130 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3131 LI->setAlignment(std::min(LI->getAlign(), getSliceAlign()));
3132 continue;
3133 }
3134 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3135 SI->setAlignment(std::min(SI->getAlign(), getSliceAlign()));
3136 continue;
3137 }
3138
3139 assert(isa<BitCastInst>(I) || isa<AddrSpaceCastInst>(I) ||(static_cast <bool> (isa<BitCastInst>(I) || isa<
AddrSpaceCastInst>(I) || isa<PHINode>(I) || isa<SelectInst
>(I) || isa<GetElementPtrInst>(I)) ? void (0) : __assert_fail
("isa<BitCastInst>(I) || isa<AddrSpaceCastInst>(I) || isa<PHINode>(I) || isa<SelectInst>(I) || isa<GetElementPtrInst>(I)"
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3141, __extension__ __PRETTY_FUNCTION__
))
3140 isa<PHINode>(I) || isa<SelectInst>(I) ||(static_cast <bool> (isa<BitCastInst>(I) || isa<
AddrSpaceCastInst>(I) || isa<PHINode>(I) || isa<SelectInst
>(I) || isa<GetElementPtrInst>(I)) ? void (0) : __assert_fail
("isa<BitCastInst>(I) || isa<AddrSpaceCastInst>(I) || isa<PHINode>(I) || isa<SelectInst>(I) || isa<GetElementPtrInst>(I)"
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3141, __extension__ __PRETTY_FUNCTION__
))
3141 isa<GetElementPtrInst>(I))(static_cast <bool> (isa<BitCastInst>(I) || isa<
AddrSpaceCastInst>(I) || isa<PHINode>(I) || isa<SelectInst
>(I) || isa<GetElementPtrInst>(I)) ? void (0) : __assert_fail
("isa<BitCastInst>(I) || isa<AddrSpaceCastInst>(I) || isa<PHINode>(I) || isa<SelectInst>(I) || isa<GetElementPtrInst>(I)"
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3141, __extension__ __PRETTY_FUNCTION__
))
;
3142 for (User *U : I->users())
3143 if (Visited.insert(cast<Instruction>(U)).second)
3144 Uses.push_back(cast<Instruction>(U));
3145 } while (!Uses.empty());
3146 }
3147
3148 bool visitPHINode(PHINode &PN) {
3149 LLVM_DEBUG(dbgs() << " original: " << PN << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << PN <<
"\n"; } } while (false)
;
3150 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable")(static_cast <bool> (BeginOffset >= NewAllocaBeginOffset
&& "PHIs are unsplittable") ? void (0) : __assert_fail
("BeginOffset >= NewAllocaBeginOffset && \"PHIs are unsplittable\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3150, __extension__ __PRETTY_FUNCTION__
))
;
3151 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable")(static_cast <bool> (EndOffset <= NewAllocaEndOffset
&& "PHIs are unsplittable") ? void (0) : __assert_fail
("EndOffset <= NewAllocaEndOffset && \"PHIs are unsplittable\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3151, __extension__ __PRETTY_FUNCTION__
))
;
3152
3153 // We would like to compute a new pointer in only one place, but have it be
3154 // as local as possible to the PHI. To do that, we re-use the location of
3155 // the old pointer, which necessarily must be in the right position to
3156 // dominate the PHI.
3157 IRBuilderBase::InsertPointGuard Guard(IRB);
3158 if (isa<PHINode>(OldPtr))
3159 IRB.SetInsertPoint(&*OldPtr->getParent()->getFirstInsertionPt());
3160 else
3161 IRB.SetInsertPoint(OldPtr);
3162 IRB.SetCurrentDebugLocation(OldPtr->getDebugLoc());
3163
3164 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3165 // Replace the operands which were using the old pointer.
3166 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
3167
3168 LLVM_DEBUG(dbgs() << " to: " << PN << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << PN <<
"\n"; } } while (false)
;
3169 deleteIfTriviallyDead(OldPtr);
3170
3171 // Fix the alignment of any loads or stores using this PHI node.
3172 fixLoadStoreAlign(PN);
3173
3174 // PHIs can't be promoted on their own, but often can be speculated. We
3175 // check the speculation outside of the rewriter so that we see the
3176 // fully-rewritten alloca.
3177 PHIUsers.insert(&PN);
3178 return true;
3179 }
3180
3181 bool visitSelectInst(SelectInst &SI) {
3182 LLVM_DEBUG(dbgs() << " original: " << SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << SI <<
"\n"; } } while (false)
;
3183 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&(static_cast <bool> ((SI.getTrueValue() == OldPtr || SI
.getFalseValue() == OldPtr) && "Pointer isn't an operand!"
) ? void (0) : __assert_fail ("(SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) && \"Pointer isn't an operand!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3184, __extension__ __PRETTY_FUNCTION__
))
3184 "Pointer isn't an operand!")(static_cast <bool> ((SI.getTrueValue() == OldPtr || SI
.getFalseValue() == OldPtr) && "Pointer isn't an operand!"
) ? void (0) : __assert_fail ("(SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) && \"Pointer isn't an operand!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3184, __extension__ __PRETTY_FUNCTION__
))
;
3185 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable")(static_cast <bool> (BeginOffset >= NewAllocaBeginOffset
&& "Selects are unsplittable") ? void (0) : __assert_fail
("BeginOffset >= NewAllocaBeginOffset && \"Selects are unsplittable\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3185, __extension__ __PRETTY_FUNCTION__
))
;
3186 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable")(static_cast <bool> (EndOffset <= NewAllocaEndOffset
&& "Selects are unsplittable") ? void (0) : __assert_fail
("EndOffset <= NewAllocaEndOffset && \"Selects are unsplittable\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3186, __extension__ __PRETTY_FUNCTION__
))
;
3187
3188 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3189 // Replace the operands which were using the old pointer.
3190 if (SI.getOperand(1) == OldPtr)
3191 SI.setOperand(1, NewPtr);
3192 if (SI.getOperand(2) == OldPtr)
3193 SI.setOperand(2, NewPtr);
3194
3195 LLVM_DEBUG(dbgs() << " to: " << SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << SI <<
"\n"; } } while (false)
;
3196 deleteIfTriviallyDead(OldPtr);
3197
3198 // Fix the alignment of any loads or stores using this select.
3199 fixLoadStoreAlign(SI);
3200
3201 // Selects can't be promoted on their own, but often can be speculated. We
3202 // check the speculation outside of the rewriter so that we see the
3203 // fully-rewritten alloca.
3204 SelectUsers.insert(&SI);
3205 return true;
3206 }
3207};
3208
3209namespace {
3210
3211/// Visitor to rewrite aggregate loads and stores as scalar.
3212///
3213/// This pass aggressively rewrites all aggregate loads and stores on
3214/// a particular pointer (or any pointer derived from it which we can identify)
3215/// with scalar loads and stores.
3216class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
3217 // Befriend the base class so it can delegate to private visit methods.
3218 friend class InstVisitor<AggLoadStoreRewriter, bool>;
3219
3220 /// Queue of pointer uses to analyze and potentially rewrite.
3221 SmallVector<Use *, 8> Queue;
3222
3223 /// Set to prevent us from cycling with phi nodes and loops.
3224 SmallPtrSet<User *, 8> Visited;
3225
3226 /// The current pointer use being rewritten. This is used to dig up the used
3227 /// value (as opposed to the user).
3228 Use *U = nullptr;
3229
3230 /// Used to calculate offsets, and hence alignment, of subobjects.
3231 const DataLayout &DL;
3232
3233 IRBuilderTy &IRB;
3234
3235public:
3236 AggLoadStoreRewriter(const DataLayout &DL, IRBuilderTy &IRB)
3237 : DL(DL), IRB(IRB) {}
3238
3239 /// Rewrite loads and stores through a pointer and all pointers derived from
3240 /// it.
3241 bool rewrite(Instruction &I) {
3242 LLVM_DEBUG(dbgs() << " Rewriting FCA loads and stores...\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Rewriting FCA loads and stores...\n"
; } } while (false)
;
3243 enqueueUsers(I);
3244 bool Changed = false;
3245 while (!Queue.empty()) {
3246 U = Queue.pop_back_val();
3247 Changed |= visit(cast<Instruction>(U->getUser()));
3248 }
3249 return Changed;
3250 }
3251
3252private:
3253 /// Enqueue all the users of the given instruction for further processing.
3254 /// This uses a set to de-duplicate users.
3255 void enqueueUsers(Instruction &I) {
3256 for (Use &U : I.uses())
3257 if (Visited.insert(U.getUser()).second)
3258 Queue.push_back(&U);
3259 }
3260
3261 // Conservative default is to not rewrite anything.
3262 bool visitInstruction(Instruction &I) { return false; }
3263
3264 /// Generic recursive split emission class.
3265 template <typename Derived> class OpSplitter {
3266 protected:
3267 /// The builder used to form new instructions.
3268 IRBuilderTy &IRB;
3269
3270 /// The indices which to be used with insert- or extractvalue to select the
3271 /// appropriate value within the aggregate.
3272 SmallVector<unsigned, 4> Indices;
3273
3274 /// The indices to a GEP instruction which will move Ptr to the correct slot
3275 /// within the aggregate.
3276 SmallVector<Value *, 4> GEPIndices;
3277
3278 /// The base pointer of the original op, used as a base for GEPing the
3279 /// split operations.
3280 Value *Ptr;
3281
3282 /// The base pointee type being GEPed into.
3283 Type *BaseTy;
3284
3285 /// Known alignment of the base pointer.
3286 Align BaseAlign;
3287
3288 /// To calculate offset of each component so we can correctly deduce
3289 /// alignments.
3290 const DataLayout &DL;
3291
3292 /// Initialize the splitter with an insertion point, Ptr and start with a
3293 /// single zero GEP index.
3294 OpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
3295 Align BaseAlign, const DataLayout &DL, IRBuilderTy &IRB)
3296 : IRB(IRB), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr), BaseTy(BaseTy),
3297 BaseAlign(BaseAlign), DL(DL) {
3298 IRB.SetInsertPoint(InsertionPoint);
3299 }
3300
3301 public:
3302 /// Generic recursive split emission routine.
3303 ///
3304 /// This method recursively splits an aggregate op (load or store) into
3305 /// scalar or vector ops. It splits recursively until it hits a single value
3306 /// and emits that single value operation via the template argument.
3307 ///
3308 /// The logic of this routine relies on GEPs and insertvalue and
3309 /// extractvalue all operating with the same fundamental index list, merely
3310 /// formatted differently (GEPs need actual values).
3311 ///
3312 /// \param Ty The type being split recursively into smaller ops.
3313 /// \param Agg The aggregate value being built up or stored, depending on
3314 /// whether this is splitting a load or a store respectively.
3315 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3316 if (Ty->isSingleValueType()) {
3317 unsigned Offset = DL.getIndexedOffsetInType(BaseTy, GEPIndices);
3318 return static_cast<Derived *>(this)->emitFunc(
3319 Ty, Agg, commonAlignment(BaseAlign, Offset), Name);
3320 }
3321
3322 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3323 unsigned OldSize = Indices.size();
3324 (void)OldSize;
3325 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3326 ++Idx) {
3327 assert(Indices.size() == OldSize && "Did not return to the old size")(static_cast <bool> (Indices.size() == OldSize &&
"Did not return to the old size") ? void (0) : __assert_fail
("Indices.size() == OldSize && \"Did not return to the old size\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3327, __extension__ __PRETTY_FUNCTION__
))
;
3328 Indices.push_back(Idx);
3329 GEPIndices.push_back(IRB.getInt32(Idx));
3330 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3331 GEPIndices.pop_back();
3332 Indices.pop_back();
3333 }
3334 return;
3335 }
3336
3337 if (StructType *STy = dyn_cast<StructType>(Ty)) {
3338 unsigned OldSize = Indices.size();
3339 (void)OldSize;
3340 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3341 ++Idx) {
3342 assert(Indices.size() == OldSize && "Did not return to the old size")(static_cast <bool> (Indices.size() == OldSize &&
"Did not return to the old size") ? void (0) : __assert_fail
("Indices.size() == OldSize && \"Did not return to the old size\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3342, __extension__ __PRETTY_FUNCTION__
))
;
3343 Indices.push_back(Idx);
3344 GEPIndices.push_back(IRB.getInt32(Idx));
3345 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3346 GEPIndices.pop_back();
3347 Indices.pop_back();
3348 }
3349 return;
3350 }
3351
3352 llvm_unreachable("Only arrays and structs are aggregate loadable types")::llvm::llvm_unreachable_internal("Only arrays and structs are aggregate loadable types"
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3352)
;
3353 }
3354 };
3355
3356 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3357 AAMDNodes AATags;
3358
3359 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
3360 AAMDNodes AATags, Align BaseAlign, const DataLayout &DL,
3361 IRBuilderTy &IRB)
3362 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr, BaseTy, BaseAlign, DL,
3363 IRB),
3364 AATags(AATags) {}
3365
3366 /// Emit a leaf load of a single value. This is called at the leaves of the
3367 /// recursive emission to actually load values.
3368 void emitFunc(Type *Ty, Value *&Agg, Align Alignment, const Twine &Name) {
3369 assert(Ty->isSingleValueType())(static_cast <bool> (Ty->isSingleValueType()) ? void
(0) : __assert_fail ("Ty->isSingleValueType()", "llvm/lib/Transforms/Scalar/SROA.cpp"
, 3369, __extension__ __PRETTY_FUNCTION__))
;
3370 // Load the single value and insert it using the indices.
3371 Value *GEP =
3372 IRB.CreateInBoundsGEP(BaseTy, Ptr, GEPIndices, Name + ".gep");
3373 LoadInst *Load =
3374 IRB.CreateAlignedLoad(Ty, GEP, Alignment, Name + ".load");
3375
3376 APInt Offset(
3377 DL.getIndexSizeInBits(Ptr->getType()->getPointerAddressSpace()), 0);
3378 if (AATags &&
3379 GEPOperator::accumulateConstantOffset(BaseTy, GEPIndices, DL, Offset))
3380 Load->setAAMetadata(AATags.shift(Offset.getZExtValue()));
3381
3382 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3383 LLVM_DEBUG(dbgs() << " to: " << *Load << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *Load <<
"\n"; } } while (false)
;
3384 }
3385 };
3386
3387 bool visitLoadInst(LoadInst &LI) {
3388 assert(LI.getPointerOperand() == *U)(static_cast <bool> (LI.getPointerOperand() == *U) ? void
(0) : __assert_fail ("LI.getPointerOperand() == *U", "llvm/lib/Transforms/Scalar/SROA.cpp"
, 3388, __extension__ __PRETTY_FUNCTION__))
;
3389 if (!LI.isSimple() || LI.getType()->isSingleValueType())
3390 return false;
3391
3392 // We have an aggregate being loaded, split it apart.
3393 LLVM_DEBUG(dbgs() << " original: " << LI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << LI <<
"\n"; } } while (false)
;
3394 LoadOpSplitter Splitter(&LI, *U, LI.getType(), LI.getAAMetadata(),
3395 getAdjustedAlignment(&LI, 0), DL, IRB);
3396 Value *V = PoisonValue::get(LI.getType());
3397 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3398 Visited.erase(&LI);
3399 LI.replaceAllUsesWith(V);
3400 LI.eraseFromParent();
3401 return true;
3402 }
3403
3404 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3405 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
3406 AAMDNodes AATags, Align BaseAlign, const DataLayout &DL,
3407 IRBuilderTy &IRB)
3408 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr, BaseTy, BaseAlign,
3409 DL, IRB),
3410 AATags(AATags) {}
3411 AAMDNodes AATags;
3412 /// Emit a leaf store of a single value. This is called at the leaves of the
3413 /// recursive emission to actually produce stores.
3414 void emitFunc(Type *Ty, Value *&Agg, Align Alignment, const Twine &Name) {
3415 assert(Ty->isSingleValueType())(static_cast <bool> (Ty->isSingleValueType()) ? void
(0) : __assert_fail ("Ty->isSingleValueType()", "llvm/lib/Transforms/Scalar/SROA.cpp"
, 3415, __extension__ __PRETTY_FUNCTION__))
;
3416 // Extract the single value and store it using the indices.
3417 //
3418 // The gep and extractvalue values are factored out of the CreateStore
3419 // call to make the output independent of the argument evaluation order.
3420 Value *ExtractValue =
3421 IRB.CreateExtractValue(Agg, Indices, Name + ".extract");
3422 Value *InBoundsGEP =
3423 IRB.CreateInBoundsGEP(BaseTy, Ptr, GEPIndices, Name + ".gep");
3424 StoreInst *Store =
3425 IRB.CreateAlignedStore(ExtractValue, InBoundsGEP, Alignment);
3426
3427 APInt Offset(
3428 DL.getIndexSizeInBits(Ptr->getType()->getPointerAddressSpace()), 0);
3429 if (AATags &&
3430 GEPOperator::accumulateConstantOffset(BaseTy, GEPIndices, DL, Offset))
3431 Store->setAAMetadata(AATags.shift(Offset.getZExtValue()));
3432
3433 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *Store <<
"\n"; } } while (false)
;
3434 }
3435 };
3436
3437 bool visitStoreInst(StoreInst &SI) {
3438 if (!SI.isSimple() || SI.getPointerOperand() != *U)
3439 return false;
3440 Value *V = SI.getValueOperand();
3441 if (V->getType()->isSingleValueType())
3442 return false;
3443
3444 // We have an aggregate being stored, split it apart.
3445 LLVM_DEBUG(dbgs() << " original: " << SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << SI <<
"\n"; } } while (false)
;
3446 StoreOpSplitter Splitter(&SI, *U, V->getType(), SI.getAAMetadata(),
3447 getAdjustedAlignment(&SI, 0), DL, IRB);
3448 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3449 Visited.erase(&SI);
3450 SI.eraseFromParent();
3451 return true;
3452 }
3453
3454 bool visitBitCastInst(BitCastInst &BC) {
3455 enqueueUsers(BC);
3456 return false;
3457 }
3458
3459 bool visitAddrSpaceCastInst(AddrSpaceCastInst &ASC) {
3460 enqueueUsers(ASC);
3461 return false;
3462 }
3463
3464 // Fold gep (select cond, ptr1, ptr2) => select cond, gep(ptr1), gep(ptr2)
3465 bool foldGEPSelect(GetElementPtrInst &GEPI) {
3466 if (!GEPI.hasAllConstantIndices())
3467 return false;
3468
3469 SelectInst *Sel = cast<SelectInst>(GEPI.getPointerOperand());
3470
3471 LLVM_DEBUG(dbgs() << " Rewriting gep(select) -> select(gep):"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Rewriting gep(select) -> select(gep):"
<< "\n original: " << *Sel << "\n "
<< GEPI; } } while (false)
3472 << "\n original: " << *Seldo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Rewriting gep(select) -> select(gep):"
<< "\n original: " << *Sel << "\n "
<< GEPI; } } while (false)
3473 << "\n " << GEPI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Rewriting gep(select) -> select(gep):"
<< "\n original: " << *Sel << "\n "
<< GEPI; } } while (false)
;
3474
3475 IRB.SetInsertPoint(&GEPI);
3476 SmallVector<Value *, 4> Index(GEPI.indices());
3477 bool IsInBounds = GEPI.isInBounds();
3478
3479 Type *Ty = GEPI.getSourceElementType();
3480 Value *True = Sel->getTrueValue();
3481 Value *NTrue =
3482 IsInBounds
3483 ? IRB.CreateInBoundsGEP(Ty, True, Index,
3484 True->getName() + ".sroa.gep")
3485 : IRB.CreateGEP(Ty, True, Index, True->getName() + ".sroa.gep");
3486
3487 Value *False = Sel->getFalseValue();
3488
3489 Value *NFalse =
3490 IsInBounds
3491 ? IRB.CreateInBoundsGEP(Ty, False, Index,
3492 False->getName() + ".sroa.gep")
3493 : IRB.CreateGEP(Ty, False, Index, False->getName() + ".sroa.gep");
3494
3495 Value *NSel = IRB.CreateSelect(Sel->getCondition(), NTrue, NFalse,
3496 Sel->getName() + ".sroa.sel");
3497 Visited.erase(&GEPI);
3498 GEPI.replaceAllUsesWith(NSel);
3499 GEPI.eraseFromParent();
3500 Instruction *NSelI = cast<Instruction>(NSel);
3501 Visited.insert(NSelI);
3502 enqueueUsers(*NSelI);
3503
3504 LLVM_DEBUG(dbgs() << "\n to: " << *NTruedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "\n to: " << *NTrue
<< "\n " << *NFalse << "\n "
<< *NSel << '\n'; } } while (false)
3505 << "\n " << *NFalsedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "\n to: " << *NTrue
<< "\n " << *NFalse << "\n "
<< *NSel << '\n'; } } while (false)
3506 << "\n " << *NSel << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "\n to: " << *NTrue
<< "\n " << *NFalse << "\n "
<< *NSel << '\n'; } } while (false)
;
3507
3508 return true;
3509 }
3510
3511 // Fold gep (phi ptr1, ptr2) => phi gep(ptr1), gep(ptr2)
3512 bool foldGEPPhi(GetElementPtrInst &GEPI) {
3513 if (!GEPI.hasAllConstantIndices())
3514 return false;
3515
3516 PHINode *PHI = cast<PHINode>(GEPI.getPointerOperand());
3517 if (GEPI.getParent() != PHI->getParent() ||
3518 llvm::any_of(PHI->incoming_values(), [](Value *In)
3519 { Instruction *I = dyn_cast<Instruction>(In);
3520 return !I || isa<GetElementPtrInst>(I) || isa<PHINode>(I) ||
3521 succ_empty(I->getParent()) ||
3522 !I->getParent()->isLegalToHoistInto();
3523 }))
3524 return false;
3525
3526 LLVM_DEBUG(dbgs() << " Rewriting gep(phi) -> phi(gep):"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Rewriting gep(phi) -> phi(gep):"
<< "\n original: " << *PHI << "\n "
<< GEPI << "\n to: "; } } while (false)
3527 << "\n original: " << *PHIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Rewriting gep(phi) -> phi(gep):"
<< "\n original: " << *PHI << "\n "
<< GEPI << "\n to: "; } } while (false)
3528 << "\n " << GEPIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Rewriting gep(phi) -> phi(gep):"
<< "\n original: " << *PHI << "\n "
<< GEPI << "\n to: "; } } while (false)
3529 << "\n to: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Rewriting gep(phi) -> phi(gep):"
<< "\n original: " << *PHI << "\n "
<< GEPI << "\n to: "; } } while (false)
;
3530
3531 SmallVector<Value *, 4> Index(GEPI.indices());
3532 bool IsInBounds = GEPI.isInBounds();
3533 IRB.SetInsertPoint(GEPI.getParent()->getFirstNonPHI());
3534 PHINode *NewPN = IRB.CreatePHI(GEPI.getType(), PHI->getNumIncomingValues(),
3535 PHI->getName() + ".sroa.phi");
3536 for (unsigned I = 0, E = PHI->getNumIncomingValues(); I != E; ++I) {
3537 BasicBlock *B = PHI->getIncomingBlock(I);
3538 Value *NewVal = nullptr;
3539 int Idx = NewPN->getBasicBlockIndex(B);
3540 if (Idx >= 0) {
3541 NewVal = NewPN->getIncomingValue(Idx);
3542 } else {
3543 Instruction *In = cast<Instruction>(PHI->getIncomingValue(I));
3544
3545 IRB.SetInsertPoint(In->getParent(), std::next(In->getIterator()));
3546 Type *Ty = GEPI.getSourceElementType();
3547 NewVal = IsInBounds ? IRB.CreateInBoundsGEP(Ty, In, Index,
3548 In->getName() + ".sroa.gep")
3549 : IRB.CreateGEP(Ty, In, Index,
3550 In->getName() + ".sroa.gep");
3551 }
3552 NewPN->addIncoming(NewVal, B);
3553 }
3554
3555 Visited.erase(&GEPI);
3556 GEPI.replaceAllUsesWith(NewPN);
3557 GEPI.eraseFromParent();
3558 Visited.insert(NewPN);
3559 enqueueUsers(*NewPN);
3560
3561 LLVM_DEBUG(for (Value *In : NewPN->incoming_values())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { for (Value *In : NewPN->incoming_values()) dbgs
() << "\n " << *In; dbgs() << "\n "
<< *NewPN << '\n'; } } while (false)
3562 dbgs() << "\n " << *In;do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { for (Value *In : NewPN->incoming_values()) dbgs
() << "\n " << *In; dbgs() << "\n "
<< *NewPN << '\n'; } } while (false)
3563 dbgs() << "\n " << *NewPN << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { for (Value *In : NewPN->incoming_values()) dbgs
() << "\n " << *In; dbgs() << "\n "
<< *NewPN << '\n'; } } while (false)
;
3564
3565 return true;
3566 }
3567
3568 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3569 if (isa<SelectInst>(GEPI.getPointerOperand()) &&
3570 foldGEPSelect(GEPI))
3571 return true;
3572
3573 if (isa<PHINode>(GEPI.getPointerOperand()) &&
3574 foldGEPPhi(GEPI))
3575 return true;
3576
3577 enqueueUsers(GEPI);
3578 return false;
3579 }
3580
3581 bool visitPHINode(PHINode &PN) {
3582 enqueueUsers(PN);
3583 return false;
3584 }
3585
3586 bool visitSelectInst(SelectInst &SI) {
3587 enqueueUsers(SI);
3588 return false;
3589 }
3590};
3591
3592} // end anonymous namespace
3593
3594/// Strip aggregate type wrapping.
3595///
3596/// This removes no-op aggregate types wrapping an underlying type. It will
3597/// strip as many layers of types as it can without changing either the type
3598/// size or the allocated size.
3599static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
3600 if (Ty->isSingleValueType())
3601 return Ty;
3602
3603 uint64_t AllocSize = DL.getTypeAllocSize(Ty).getFixedSize();
3604 uint64_t TypeSize = DL.getTypeSizeInBits(Ty).getFixedSize();
3605
3606 Type *InnerTy;
3607 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3608 InnerTy = ArrTy->getElementType();
3609 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3610 const StructLayout *SL = DL.getStructLayout(STy);
3611 unsigned Index = SL->getElementContainingOffset(0);
3612 InnerTy = STy->getElementType(Index);
3613 } else {
3614 return Ty;
3615 }
3616
3617 if (AllocSize > DL.getTypeAllocSize(InnerTy).getFixedSize() ||
3618 TypeSize > DL.getTypeSizeInBits(InnerTy).getFixedSize())
3619 return Ty;
3620
3621 return stripAggregateTypeWrapping(DL, InnerTy);
3622}
3623
3624/// Try to find a partition of the aggregate type passed in for a given
3625/// offset and size.
3626///
3627/// This recurses through the aggregate type and tries to compute a subtype
3628/// based on the offset and size. When the offset and size span a sub-section
3629/// of an array, it will even compute a new array type for that sub-section,
3630/// and the same for structs.
3631///
3632/// Note that this routine is very strict and tries to find a partition of the
3633/// type which produces the *exact* right offset and size. It is not forgiving
3634/// when the size or offset cause either end of type-based partition to be off.
3635/// Also, this is a best-effort routine. It is reasonable to give up and not
3636/// return a type if necessary.
3637static Type *getTypePartition(const DataLayout &DL, Type *Ty, uint64_t Offset,
3638 uint64_t Size) {
3639 if (Offset == 0 && DL.getTypeAllocSize(Ty).getFixedSize() == Size)
3640 return stripAggregateTypeWrapping(DL, Ty);
3641 if (Offset > DL.getTypeAllocSize(Ty).getFixedSize() ||
3642 (DL.getTypeAllocSize(Ty).getFixedSize() - Offset) < Size)
3643 return nullptr;
3644
3645 if (isa<ArrayType>(Ty) || isa<VectorType>(Ty)) {
3646 Type *ElementTy;
3647 uint64_t TyNumElements;
3648 if (auto *AT = dyn_cast<ArrayType>(Ty)) {
3649 ElementTy = AT->getElementType();
3650 TyNumElements = AT->getNumElements();
3651 } else {
3652 // FIXME: This isn't right for vectors with non-byte-sized or
3653 // non-power-of-two sized elements.
3654 auto *VT = cast<FixedVectorType>(Ty);
3655 ElementTy = VT->getElementType();
3656 TyNumElements = VT->getNumElements();
3657 }
3658 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy).getFixedSize();
3659 uint64_t NumSkippedElements = Offset / ElementSize;
3660 if (NumSkippedElements >= TyNumElements)
3661 return nullptr;
3662 Offset -= NumSkippedElements * ElementSize;
3663
3664 // First check if we need to recurse.
3665 if (Offset > 0 || Size < ElementSize) {
3666 // Bail if the partition ends in a different array element.
3667 if ((Offset + Size) > ElementSize)
3668 return nullptr;
3669 // Recurse through the element type trying to peel off offset bytes.
3670 return getTypePartition(DL, ElementTy, Offset, Size);
3671 }
3672 assert(Offset == 0)(static_cast <bool> (Offset == 0) ? void (0) : __assert_fail
("Offset == 0", "llvm/lib/Transforms/Scalar/SROA.cpp", 3672,
__extension__ __PRETTY_FUNCTION__))
;
3673
3674 if (Size == ElementSize)
3675 return stripAggregateTypeWrapping(DL, ElementTy);
3676 assert(Size > ElementSize)(static_cast <bool> (Size > ElementSize) ? void (0) :
__assert_fail ("Size > ElementSize", "llvm/lib/Transforms/Scalar/SROA.cpp"
, 3676, __extension__ __PRETTY_FUNCTION__))
;
3677 uint64_t NumElements = Size / ElementSize;
3678 if (NumElements * ElementSize != Size)
3679 return nullptr;
3680 return ArrayType::get(ElementTy, NumElements);
3681 }
3682
3683 StructType *STy = dyn_cast<StructType>(Ty);
3684 if (!STy)
3685 return nullptr;
3686
3687 const StructLayout *SL = DL.getStructLayout(STy);
3688 if (Offset >= SL->getSizeInBytes())
3689 return nullptr;
3690 uint64_t EndOffset = Offset + Size;
3691 if (EndOffset > SL->getSizeInBytes())
3692 return nullptr;
3693
3694 unsigned Index = SL->getElementContainingOffset(Offset);
3695 Offset -= SL->getElementOffset(Index);
3696
3697 Type *ElementTy = STy->getElementType(Index);
3698 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy).getFixedSize();
3699 if (Offset >= ElementSize)
3700 return nullptr; // The offset points into alignment padding.
3701
3702 // See if any partition must be contained by the element.
3703 if (Offset > 0 || Size < ElementSize) {
3704 if ((Offset + Size) > ElementSize)
3705 return nullptr;
3706 return getTypePartition(DL, ElementTy, Offset, Size);
3707 }
3708 assert(Offset == 0)(static_cast <bool> (Offset == 0) ? void (0) : __assert_fail
("Offset == 0", "llvm/lib/Transforms/Scalar/SROA.cpp", 3708,
__extension__ __PRETTY_FUNCTION__))
;
3709
3710 if (Size == ElementSize)
3711 return stripAggregateTypeWrapping(DL, ElementTy);
3712
3713 StructType::element_iterator EI = STy->element_begin() + Index,
3714 EE = STy->element_end();
3715 if (EndOffset < SL->getSizeInBytes()) {
3716 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3717 if (Index == EndIndex)
3718 return nullptr; // Within a single element and its padding.
3719
3720 // Don't try to form "natural" types if the elements don't line up with the
3721 // expected size.
3722 // FIXME: We could potentially recurse down through the last element in the
3723 // sub-struct to find a natural end point.
3724 if (SL->getElementOffset(EndIndex) != EndOffset)
3725 return nullptr;
3726
3727 assert(Index < EndIndex)(static_cast <bool> (Index < EndIndex) ? void (0) : __assert_fail
("Index < EndIndex", "llvm/lib/Transforms/Scalar/SROA.cpp"
, 3727, __extension__ __PRETTY_FUNCTION__))
;
3728 EE = STy->element_begin() + EndIndex;
3729 }
3730
3731 // Try to build up a sub-structure.
3732 StructType *SubTy =
3733 StructType::get(STy->getContext(), makeArrayRef(EI, EE), STy->isPacked());
3734 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3735 if (Size != SubSL->getSizeInBytes())
3736 return nullptr; // The sub-struct doesn't have quite the size needed.
3737
3738 return SubTy;
3739}
3740
3741/// Pre-split loads and stores to simplify rewriting.
3742///
3743/// We want to break up the splittable load+store pairs as much as
3744/// possible. This is important to do as a preprocessing step, as once we
3745/// start rewriting the accesses to partitions of the alloca we lose the
3746/// necessary information to correctly split apart paired loads and stores
3747/// which both point into this alloca. The case to consider is something like
3748/// the following:
3749///
3750/// %a = alloca [12 x i8]
3751/// %gep1 = getelementptr [12 x i8]* %a, i32 0, i32 0
3752/// %gep2 = getelementptr [12 x i8]* %a, i32 0, i32 4
3753/// %gep3 = getelementptr [12 x i8]* %a, i32 0, i32 8
3754/// %iptr1 = bitcast i8* %gep1 to i64*
3755/// %iptr2 = bitcast i8* %gep2 to i64*
3756/// %fptr1 = bitcast i8* %gep1 to float*
3757/// %fptr2 = bitcast i8* %gep2 to float*
3758/// %fptr3 = bitcast i8* %gep3 to float*
3759/// store float 0.0, float* %fptr1
3760/// store float 1.0, float* %fptr2
3761/// %v = load i64* %iptr1
3762/// store i64 %v, i64* %iptr2
3763/// %f1 = load float* %fptr2
3764/// %f2 = load float* %fptr3
3765///
3766/// Here we want to form 3 partitions of the alloca, each 4 bytes large, and
3767/// promote everything so we recover the 2 SSA values that should have been
3768/// there all along.
3769///
3770/// \returns true if any changes are made.
3771bool SROAPass::presplitLoadsAndStores(AllocaInst &AI, AllocaSlices &AS) {
3772 LLVM_DEBUG(dbgs() << "Pre-splitting loads and stores\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "Pre-splitting loads and stores\n"
; } } while (false)
;
3773
3774 // Track the loads and stores which are candidates for pre-splitting here, in
3775 // the order they first appear during the partition scan. These give stable
3776 // iteration order and a basis for tracking which loads and stores we
3777 // actually split.
3778 SmallVector<LoadInst *, 4> Loads;
3779 SmallVector<StoreInst *, 4> Stores;
3780
3781 // We need to accumulate the splits required of each load or store where we
3782 // can find them via a direct lookup. This is important to cross-check loads
3783 // and stores against each other. We also track the slice so that we can kill
3784 // all the slices that end up split.
3785 struct SplitOffsets {
3786 Slice *S;
3787 std::vector<uint64_t> Splits;
3788 };
3789 SmallDenseMap<Instruction *, SplitOffsets, 8> SplitOffsetsMap;
3790
3791 // Track loads out of this alloca which cannot, for any reason, be pre-split.
3792 // This is important as we also cannot pre-split stores of those loads!
3793 // FIXME: This is all pretty gross. It means that we can be more aggressive
3794 // in pre-splitting when the load feeding the store happens to come from
3795 // a separate alloca. Put another way, the effectiveness of SROA would be
3796 // decreased by a frontend which just concatenated all of its local allocas
3797 // into one big flat alloca. But defeating such patterns is exactly the job
3798 // SROA is tasked with! Sadly, to not have this discrepancy we would have
3799 // change store pre-splitting to actually force pre-splitting of the load
3800 // that feeds it *and all stores*. That makes pre-splitting much harder, but
3801 // maybe it would make it more principled?
3802 SmallPtrSet<LoadInst *, 8> UnsplittableLoads;
3803
3804 LLVM_DEBUG(dbgs() << " Searching for candidate loads and stores\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Searching for candidate loads and stores\n"
; } } while (false)
;
3805 for (auto &P : AS.partitions()) {
3806 for (Slice &S : P) {
3807 Instruction *I = cast<Instruction>(S.getUse()->getUser());
3808 if (!S.isSplittable() || S.endOffset() <= P.endOffset()) {
3809 // If this is a load we have to track that it can't participate in any
3810 // pre-splitting. If this is a store of a load we have to track that
3811 // that load also can't participate in any pre-splitting.
3812 if (auto *LI = dyn_cast<LoadInst>(I))
3813 UnsplittableLoads.insert(LI);
3814 else if (auto *SI = dyn_cast<StoreInst>(I))
3815 if (auto *LI = dyn_cast<LoadInst>(SI->getValueOperand()))
3816 UnsplittableLoads.insert(LI);
3817 continue;
3818 }
3819 assert(P.endOffset() > S.beginOffset() &&(static_cast <bool> (P.endOffset() > S.beginOffset()
&& "Empty or backwards partition!") ? void (0) : __assert_fail
("P.endOffset() > S.beginOffset() && \"Empty or backwards partition!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3820, __extension__ __PRETTY_FUNCTION__
))
3820 "Empty or backwards partition!")(static_cast <bool> (P.endOffset() > S.beginOffset()
&& "Empty or backwards partition!") ? void (0) : __assert_fail
("P.endOffset() > S.beginOffset() && \"Empty or backwards partition!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3820, __extension__ __PRETTY_FUNCTION__
))
;
3821
3822 // Determine if this is a pre-splittable slice.
3823 if (auto *LI = dyn_cast<LoadInst>(I)) {
3824 assert(!LI->isVolatile() && "Cannot split volatile loads!")(static_cast <bool> (!LI->isVolatile() && "Cannot split volatile loads!"
) ? void (0) : __assert_fail ("!LI->isVolatile() && \"Cannot split volatile loads!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3824, __extension__ __PRETTY_FUNCTION__
))
;
3825
3826 // The load must be used exclusively to store into other pointers for
3827 // us to be able to arbitrarily pre-split it. The stores must also be
3828 // simple to avoid changing semantics.
3829 auto IsLoadSimplyStored = [](LoadInst *LI) {
3830 for (User *LU : LI->users()) {
3831 auto *SI = dyn_cast<StoreInst>(LU);
3832 if (!SI || !SI->isSimple())
3833 return false;
3834 }
3835 return true;
3836 };
3837 if (!IsLoadSimplyStored(LI)) {
3838 UnsplittableLoads.insert(LI);
3839 continue;
3840 }
3841
3842 Loads.push_back(LI);
3843 } else if (auto *SI = dyn_cast<StoreInst>(I)) {
3844 if (S.getUse() != &SI->getOperandUse(SI->getPointerOperandIndex()))
3845 // Skip stores *of* pointers. FIXME: This shouldn't even be possible!
3846 continue;
3847 auto *StoredLoad = dyn_cast<LoadInst>(SI->getValueOperand());
3848 if (!StoredLoad || !StoredLoad->isSimple())
3849 continue;
3850 assert(!SI->isVolatile() && "Cannot split volatile stores!")(static_cast <bool> (!SI->isVolatile() && "Cannot split volatile stores!"
) ? void (0) : __assert_fail ("!SI->isVolatile() && \"Cannot split volatile stores!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3850, __extension__ __PRETTY_FUNCTION__
))
;
3851
3852 Stores.push_back(SI);
3853 } else {
3854 // Other uses cannot be pre-split.
3855 continue;
3856 }
3857
3858 // Record the initial split.
3859 LLVM_DEBUG(dbgs() << " Candidate: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Candidate: " << *I <<
"\n"; } } while (false)
;
3860 auto &Offsets = SplitOffsetsMap[I];
3861 assert(Offsets.Splits.empty() &&(static_cast <bool> (Offsets.Splits.empty() && "Should not have splits the first time we see an instruction!"
) ? void (0) : __assert_fail ("Offsets.Splits.empty() && \"Should not have splits the first time we see an instruction!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3862, __extension__ __PRETTY_FUNCTION__
))
3862 "Should not have splits the first time we see an instruction!")(static_cast <bool> (Offsets.Splits.empty() && "Should not have splits the first time we see an instruction!"
) ? void (0) : __assert_fail ("Offsets.Splits.empty() && \"Should not have splits the first time we see an instruction!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3862, __extension__ __PRETTY_FUNCTION__
))
;
3863 Offsets.S = &S;
3864 Offsets.Splits.push_back(P.endOffset() - S.beginOffset());
3865 }
3866
3867 // Now scan the already split slices, and add a split for any of them which
3868 // we're going to pre-split.
3869 for (Slice *S : P.splitSliceTails()) {
3870 auto SplitOffsetsMapI =
3871 SplitOffsetsMap.find(cast<Instruction>(S->getUse()->getUser()));
3872 if (SplitOffsetsMapI == SplitOffsetsMap.end())
3873 continue;
3874 auto &Offsets = SplitOffsetsMapI->second;
3875
3876 assert(Offsets.S == S && "Found a mismatched slice!")(static_cast <bool> (Offsets.S == S && "Found a mismatched slice!"
) ? void (0) : __assert_fail ("Offsets.S == S && \"Found a mismatched slice!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3876, __extension__ __PRETTY_FUNCTION__
))
;
3877 assert(!Offsets.Splits.empty() &&(static_cast <bool> (!Offsets.Splits.empty() &&
"Cannot have an empty set of splits on the second partition!"
) ? void (0) : __assert_fail ("!Offsets.Splits.empty() && \"Cannot have an empty set of splits on the second partition!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3878, __extension__ __PRETTY_FUNCTION__
))
3878 "Cannot have an empty set of splits on the second partition!")(static_cast <bool> (!Offsets.Splits.empty() &&
"Cannot have an empty set of splits on the second partition!"
) ? void (0) : __assert_fail ("!Offsets.Splits.empty() && \"Cannot have an empty set of splits on the second partition!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3878, __extension__ __PRETTY_FUNCTION__
))
;
3879 assert(Offsets.Splits.back() ==(static_cast <bool> (Offsets.Splits.back() == P.beginOffset
() - Offsets.S->beginOffset() && "Previous split does not end where this one begins!"
) ? void (0) : __assert_fail ("Offsets.Splits.back() == P.beginOffset() - Offsets.S->beginOffset() && \"Previous split does not end where this one begins!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3881, __extension__ __PRETTY_FUNCTION__
))
3880 P.beginOffset() - Offsets.S->beginOffset() &&(static_cast <bool> (Offsets.Splits.back() == P.beginOffset
() - Offsets.S->beginOffset() && "Previous split does not end where this one begins!"
) ? void (0) : __assert_fail ("Offsets.Splits.back() == P.beginOffset() - Offsets.S->beginOffset() && \"Previous split does not end where this one begins!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3881, __extension__ __PRETTY_FUNCTION__
))
3881 "Previous split does not end where this one begins!")(static_cast <bool> (Offsets.Splits.back() == P.beginOffset
() - Offsets.S->beginOffset() && "Previous split does not end where this one begins!"
) ? void (0) : __assert_fail ("Offsets.Splits.back() == P.beginOffset() - Offsets.S->beginOffset() && \"Previous split does not end where this one begins!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3881, __extension__ __PRETTY_FUNCTION__
))
;
3882
3883 // Record each split. The last partition's end isn't needed as the size
3884 // of the slice dictates that.
3885 if (S->endOffset() > P.endOffset())
3886 Offsets.Splits.push_back(P.endOffset() - Offsets.S->beginOffset());
3887 }
3888 }
3889
3890 // We may have split loads where some of their stores are split stores. For
3891 // such loads and stores, we can only pre-split them if their splits exactly
3892 // match relative to their starting offset. We have to verify this prior to
3893 // any rewriting.
3894 llvm::erase_if(Stores, [&UnsplittableLoads, &SplitOffsetsMap](StoreInst *SI) {
3895 // Lookup the load we are storing in our map of split
3896 // offsets.
3897 auto *LI = cast<LoadInst>(SI->getValueOperand());
3898 // If it was completely unsplittable, then we're done,
3899 // and this store can't be pre-split.
3900 if (UnsplittableLoads.count(LI))
3901 return true;
3902
3903 auto LoadOffsetsI = SplitOffsetsMap.find(LI);
3904 if (LoadOffsetsI == SplitOffsetsMap.end())
3905 return false; // Unrelated loads are definitely safe.
3906 auto &LoadOffsets = LoadOffsetsI->second;
3907
3908 // Now lookup the store's offsets.
3909 auto &StoreOffsets = SplitOffsetsMap[SI];
3910
3911 // If the relative offsets of each split in the load and
3912 // store match exactly, then we can split them and we
3913 // don't need to remove them here.
3914 if (LoadOffsets.Splits == StoreOffsets.Splits)
3915 return false;
3916
3917 LLVM_DEBUG(dbgs() << " Mismatched splits for load and store:\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Mismatched splits for load and store:\n"
<< " " << *LI << "\n" << " "
<< *SI << "\n"; } } while (false)
3918 << " " << *LI << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Mismatched splits for load and store:\n"
<< " " << *LI << "\n" << " "
<< *SI << "\n"; } } while (false)
3919 << " " << *SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Mismatched splits for load and store:\n"
<< " " << *LI << "\n" << " "
<< *SI << "\n"; } } while (false)
;
3920
3921 // We've found a store and load that we need to split
3922 // with mismatched relative splits. Just give up on them
3923 // and remove both instructions from our list of
3924 // candidates.
3925 UnsplittableLoads.insert(LI);
3926 return true;
3927 });
3928 // Now we have to go *back* through all the stores, because a later store may
3929 // have caused an earlier store's load to become unsplittable and if it is
3930 // unsplittable for the later store, then we can't rely on it being split in
3931 // the earlier store either.
3932 llvm::erase_if(Stores, [&UnsplittableLoads](StoreInst *SI) {
3933 auto *LI = cast<LoadInst>(SI->getValueOperand());
3934 return UnsplittableLoads.count(LI);
3935 });
3936 // Once we've established all the loads that can't be split for some reason,
3937 // filter any that made it into our list out.
3938 llvm::erase_if(Loads, [&UnsplittableLoads](LoadInst *LI) {
3939 return UnsplittableLoads.count(LI);
3940 });
3941
3942 // If no loads or stores are left, there is no pre-splitting to be done for
3943 // this alloca.
3944 if (Loads.empty() && Stores.empty())
3945 return false;
3946
3947 // From here on, we can't fail and will be building new accesses, so rig up
3948 // an IR builder.
3949 IRBuilderTy IRB(&AI);
3950
3951 // Collect the new slices which we will merge into the alloca slices.
3952 SmallVector<Slice, 4> NewSlices;
3953
3954 // Track any allocas we end up splitting loads and stores for so we iterate
3955 // on them.
3956 SmallPtrSet<AllocaInst *, 4> ResplitPromotableAllocas;
3957
3958 // At this point, we have collected all of the loads and stores we can
3959 // pre-split, and the specific splits needed for them. We actually do the
3960 // splitting in a specific order in order to handle when one of the loads in
3961 // the value operand to one of the stores.
3962 //
3963 // First, we rewrite all of the split loads, and just accumulate each split
3964 // load in a parallel structure. We also build the slices for them and append
3965 // them to the alloca slices.
3966 SmallDenseMap<LoadInst *, std::vector<LoadInst *>, 1> SplitLoadsMap;
3967 std::vector<LoadInst *> SplitLoads;
3968 const DataLayout &DL = AI.getModule()->getDataLayout();
3969 for (LoadInst *LI : Loads) {
3970 SplitLoads.clear();
3971
3972 IntegerType *Ty = cast<IntegerType>(LI->getType());
3973 assert(Ty->getBitWidth() % 8 == 0)(static_cast <bool> (Ty->getBitWidth() % 8 == 0) ? void
(0) : __assert_fail ("Ty->getBitWidth() % 8 == 0", "llvm/lib/Transforms/Scalar/SROA.cpp"
, 3973, __extension__ __PRETTY_FUNCTION__))
;
3974 uint64_t LoadSize = Ty->getBitWidth() / 8;
3975 assert(LoadSize > 0 && "Cannot have a zero-sized integer load!")(static_cast <bool> (LoadSize > 0 && "Cannot have a zero-sized integer load!"
) ? void (0) : __assert_fail ("LoadSize > 0 && \"Cannot have a zero-sized integer load!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3975, __extension__ __PRETTY_FUNCTION__
))
;
3976
3977 auto &Offsets = SplitOffsetsMap[LI];
3978 assert(LoadSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&(static_cast <bool> (LoadSize == Offsets.S->endOffset
() - Offsets.S->beginOffset() && "Slice size should always match load size exactly!"
) ? void (0) : __assert_fail ("LoadSize == Offsets.S->endOffset() - Offsets.S->beginOffset() && \"Slice size should always match load size exactly!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3979, __extension__ __PRETTY_FUNCTION__
))
3979 "Slice size should always match load size exactly!")(static_cast <bool> (LoadSize == Offsets.S->endOffset
() - Offsets.S->beginOffset() && "Slice size should always match load size exactly!"
) ? void (0) : __assert_fail ("LoadSize == Offsets.S->endOffset() - Offsets.S->beginOffset() && \"Slice size should always match load size exactly!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3979, __extension__ __PRETTY_FUNCTION__
))
;
3980 uint64_t BaseOffset = Offsets.S->beginOffset();
3981 assert(BaseOffset + LoadSize > BaseOffset &&(static_cast <bool> (BaseOffset + LoadSize > BaseOffset
&& "Cannot represent alloca access size using 64-bit integers!"
) ? void (0) : __assert_fail ("BaseOffset + LoadSize > BaseOffset && \"Cannot represent alloca access size using 64-bit integers!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3982, __extension__ __PRETTY_FUNCTION__
))
3982 "Cannot represent alloca access size using 64-bit integers!")(static_cast <bool> (BaseOffset + LoadSize > BaseOffset
&& "Cannot represent alloca access size using 64-bit integers!"
) ? void (0) : __assert_fail ("BaseOffset + LoadSize > BaseOffset && \"Cannot represent alloca access size using 64-bit integers!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 3982, __extension__ __PRETTY_FUNCTION__
))
;
3983
3984 Instruction *BasePtr = cast<Instruction>(LI->getPointerOperand());
3985 IRB.SetInsertPoint(LI);
3986
3987 LLVM_DEBUG(dbgs() << " Splitting load: " << *LI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Splitting load: " << *LI
<< "\n"; } } while (false)
;
3988
3989 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3990 int Idx = 0, Size = Offsets.Splits.size();
3991 for (;;) {
3992 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3993 auto AS = LI->getPointerAddressSpace();
3994 auto *PartPtrTy = PartTy->getPointerTo(AS);
3995 LoadInst *PLoad = IRB.CreateAlignedLoad(
3996 PartTy,
3997 getAdjustedPtr(IRB, DL, BasePtr,
3998 APInt(DL.getIndexSizeInBits(AS), PartOffset),
3999 PartPtrTy, BasePtr->getName() + "."),
4000 getAdjustedAlignment(LI, PartOffset),
4001 /*IsVolatile*/ false, LI->getName());
4002 PLoad->copyMetadata(*LI, {LLVMContext::MD_mem_parallel_loop_access,
4003 LLVMContext::MD_access_group});
4004
4005 // Append this load onto the list of split loads so we can find it later
4006 // to rewrite the stores.
4007 SplitLoads.push_back(PLoad);
4008
4009 // Now build a new slice for the alloca.
4010 NewSlices.push_back(
4011 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
4012 &PLoad->getOperandUse(PLoad->getPointerOperandIndex()),
4013 /*IsSplittable*/ false));
4014 LLVM_DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " new slice [" << NewSlices
.back().beginOffset() << ", " << NewSlices.back()
.endOffset() << "): " << *PLoad << "\n"; } }
while (false)
4015 << ", " << NewSlices.back().endOffset()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " new slice [" << NewSlices
.back().beginOffset() << ", " << NewSlices.back()
.endOffset() << "): " << *PLoad << "\n"; } }
while (false)
4016 << "): " << *PLoad << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " new slice [" << NewSlices
.back().beginOffset() << ", " << NewSlices.back()
.endOffset() << "): " << *PLoad << "\n"; } }
while (false)
;
4017
4018 // See if we've handled all the splits.
4019 if (Idx >= Size)
4020 break;
4021
4022 // Setup the next partition.
4023 PartOffset = Offsets.Splits[Idx];
4024 ++Idx;
4025 PartSize = (Idx < Size ? Offsets.Splits[Idx] : LoadSize) - PartOffset;
4026 }
4027
4028 // Now that we have the split loads, do the slow walk over all uses of the
4029 // load and rewrite them as split stores, or save the split loads to use
4030 // below if the store is going to be split there anyways.
4031 bool DeferredStores = false;
4032 for (User *LU : LI->users()) {
4033 StoreInst *SI = cast<StoreInst>(LU);
4034 if (!Stores.empty() && SplitOffsetsMap.count(SI)) {
4035 DeferredStores = true;
4036 LLVM_DEBUG(dbgs() << " Deferred splitting of store: " << *SIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Deferred splitting of store: "
<< *SI << "\n"; } } while (false)
4037 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Deferred splitting of store: "
<< *SI << "\n"; } } while (false)
;
4038 continue;
4039 }
4040
4041 Value *StoreBasePtr = SI->getPointerOperand();
4042 IRB.SetInsertPoint(SI);
4043
4044 LLVM_DEBUG(dbgs() << " Splitting store of load: " << *SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Splitting store of load: " <<
*SI << "\n"; } } while (false)
;
4045
4046 for (int Idx = 0, Size = SplitLoads.size(); Idx < Size; ++Idx) {
4047 LoadInst *PLoad = SplitLoads[Idx];
4048 uint64_t PartOffset = Idx == 0 ? 0 : Offsets.Splits[Idx - 1];
4049 auto *PartPtrTy =
4050 PLoad->getType()->getPointerTo(SI->getPointerAddressSpace());
4051
4052 auto AS = SI->getPointerAddressSpace();
4053 StoreInst *PStore = IRB.CreateAlignedStore(
4054 PLoad,
4055 getAdjustedPtr(IRB, DL, StoreBasePtr,
4056 APInt(DL.getIndexSizeInBits(AS), PartOffset),
4057 PartPtrTy, StoreBasePtr->getName() + "."),
4058 getAdjustedAlignment(SI, PartOffset),
4059 /*IsVolatile*/ false);
4060 PStore->copyMetadata(*SI, {LLVMContext::MD_mem_parallel_loop_access,
4061 LLVMContext::MD_access_group});
4062 LLVM_DEBUG(dbgs() << " +" << PartOffset << ":" << *PStore << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " +" << PartOffset <<
":" << *PStore << "\n"; } } while (false)
;
4063 }
4064
4065 // We want to immediately iterate on any allocas impacted by splitting
4066 // this store, and we have to track any promotable alloca (indicated by
4067 // a direct store) as needing to be resplit because it is no longer
4068 // promotable.
4069 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(StoreBasePtr)) {
4070 ResplitPromotableAllocas.insert(OtherAI);
4071 Worklist.insert(OtherAI);
4072 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
4073 StoreBasePtr->stripInBoundsOffsets())) {
4074 Worklist.insert(OtherAI);
4075 }
4076
4077 // Mark the original store as dead.
4078 DeadInsts.push_back(SI);
4079 }
4080
4081 // Save the split loads if there are deferred stores among the users.
4082 if (DeferredStores)
4083 SplitLoadsMap.insert(std::make_pair(LI, std::move(SplitLoads)));
4084
4085 // Mark the original load as dead and kill the original slice.
4086 DeadInsts.push_back(LI);
4087 Offsets.S->kill();
4088 }
4089
4090 // Second, we rewrite all of the split stores. At this point, we know that
4091 // all loads from this alloca have been split already. For stores of such
4092 // loads, we can simply look up the pre-existing split loads. For stores of
4093 // other loads, we split those loads first and then write split stores of
4094 // them.
4095 for (StoreInst *SI : Stores) {
4096 auto *LI = cast<LoadInst>(SI->getValueOperand());
4097 IntegerType *Ty = cast<IntegerType>(LI->getType());
4098 assert(Ty->getBitWidth() % 8 == 0)(static_cast <bool> (Ty->getBitWidth() % 8 == 0) ? void
(0) : __assert_fail ("Ty->getBitWidth() % 8 == 0", "llvm/lib/Transforms/Scalar/SROA.cpp"
, 4098, __extension__ __PRETTY_FUNCTION__))
;
4099 uint64_t StoreSize = Ty->getBitWidth() / 8;
4100 assert(StoreSize > 0 && "Cannot have a zero-sized integer store!")(static_cast <bool> (StoreSize > 0 && "Cannot have a zero-sized integer store!"
) ? void (0) : __assert_fail ("StoreSize > 0 && \"Cannot have a zero-sized integer store!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 4100, __extension__ __PRETTY_FUNCTION__
))
;
4101
4102 auto &Offsets = SplitOffsetsMap[SI];
4103 assert(StoreSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&(static_cast <bool> (StoreSize == Offsets.S->endOffset
() - Offsets.S->beginOffset() && "Slice size should always match load size exactly!"
) ? void (0) : __assert_fail ("StoreSize == Offsets.S->endOffset() - Offsets.S->beginOffset() && \"Slice size should always match load size exactly!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 4104, __extension__ __PRETTY_FUNCTION__
))
4104 "Slice size should always match load size exactly!")(static_cast <bool> (StoreSize == Offsets.S->endOffset
() - Offsets.S->beginOffset() && "Slice size should always match load size exactly!"
) ? void (0) : __assert_fail ("StoreSize == Offsets.S->endOffset() - Offsets.S->beginOffset() && \"Slice size should always match load size exactly!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 4104, __extension__ __PRETTY_FUNCTION__
))
;
4105 uint64_t BaseOffset = Offsets.S->beginOffset();
4106 assert(BaseOffset + StoreSize > BaseOffset &&(static_cast <bool> (BaseOffset + StoreSize > BaseOffset
&& "Cannot represent alloca access size using 64-bit integers!"
) ? void (0) : __assert_fail ("BaseOffset + StoreSize > BaseOffset && \"Cannot represent alloca access size using 64-bit integers!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 4107, __extension__ __PRETTY_FUNCTION__
))
4107 "Cannot represent alloca access size using 64-bit integers!")(static_cast <bool> (BaseOffset + StoreSize > BaseOffset
&& "Cannot represent alloca access size using 64-bit integers!"
) ? void (0) : __assert_fail ("BaseOffset + StoreSize > BaseOffset && \"Cannot represent alloca access size using 64-bit integers!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 4107, __extension__ __PRETTY_FUNCTION__
))
;
4108
4109 Value *LoadBasePtr = LI->getPointerOperand();
4110 Instruction *StoreBasePtr = cast<Instruction>(SI->getPointerOperand());
4111
4112 LLVM_DEBUG(dbgs() << " Splitting store: " << *SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Splitting store: " << *SI
<< "\n"; } } while (false)
;
4113
4114 // Check whether we have an already split load.
4115 auto SplitLoadsMapI = SplitLoadsMap.find(LI);
4116 std::vector<LoadInst *> *SplitLoads = nullptr;
4117 if (SplitLoadsMapI != SplitLoadsMap.end()) {
4118 SplitLoads = &SplitLoadsMapI->second;
4119 assert(SplitLoads->size() == Offsets.Splits.size() + 1 &&(static_cast <bool> (SplitLoads->size() == Offsets.Splits
.size() + 1 && "Too few split loads for the number of splits in the store!"
) ? void (0) : __assert_fail ("SplitLoads->size() == Offsets.Splits.size() + 1 && \"Too few split loads for the number of splits in the store!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 4120, __extension__ __PRETTY_FUNCTION__
))
4120 "Too few split loads for the number of splits in the store!")(static_cast <bool> (SplitLoads->size() == Offsets.Splits
.size() + 1 && "Too few split loads for the number of splits in the store!"
) ? void (0) : __assert_fail ("SplitLoads->size() == Offsets.Splits.size() + 1 && \"Too few split loads for the number of splits in the store!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 4120, __extension__ __PRETTY_FUNCTION__
))
;
4121 } else {
4122 LLVM_DEBUG(dbgs() << " of load: " << *LI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " of load: " << *LI
<< "\n"; } } while (false)
;
4123 }
4124
4125 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
4126 int Idx = 0, Size = Offsets.Splits.size();
4127 for (;;) {
4128 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
4129 auto *LoadPartPtrTy = PartTy->getPointerTo(LI->getPointerAddressSpace());
4130 auto *StorePartPtrTy = PartTy->getPointerTo(SI->getPointerAddressSpace());
4131
4132 // Either lookup a split load or create one.
4133 LoadInst *PLoad;
4134 if (SplitLoads) {
4135 PLoad = (*SplitLoads)[Idx];
4136 } else {
4137 IRB.SetInsertPoint(LI);
4138 auto AS = LI->getPointerAddressSpace();
4139 PLoad = IRB.CreateAlignedLoad(
4140 PartTy,
4141 getAdjustedPtr(IRB, DL, LoadBasePtr,
4142 APInt(DL.getIndexSizeInBits(AS), PartOffset),
4143 LoadPartPtrTy, LoadBasePtr->getName() + "."),
4144 getAdjustedAlignment(LI, PartOffset),
4145 /*IsVolatile*/ false, LI->getName());
4146 PLoad->copyMetadata(*LI, {LLVMContext::MD_mem_parallel_loop_access,
4147 LLVMContext::MD_access_group});
4148 }
4149
4150 // And store this partition.
4151 IRB.SetInsertPoint(SI);
4152 auto AS = SI->getPointerAddressSpace();
4153 StoreInst *PStore = IRB.CreateAlignedStore(
4154 PLoad,
4155 getAdjustedPtr(IRB, DL, StoreBasePtr,
4156 APInt(DL.getIndexSizeInBits(AS), PartOffset),
4157 StorePartPtrTy, StoreBasePtr->getName() + "."),
4158 getAdjustedAlignment(SI, PartOffset),
4159 /*IsVolatile*/ false);
4160 PStore->copyMetadata(*SI, {LLVMContext::MD_mem_parallel_loop_access,
4161 LLVMContext::MD_access_group});
4162
4163 // Now build a new slice for the alloca.
4164 NewSlices.push_back(
4165 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
4166 &PStore->getOperandUse(PStore->getPointerOperandIndex()),
4167 /*IsSplittable*/ false));
4168 LLVM_DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " new slice [" << NewSlices
.back().beginOffset() << ", " << NewSlices.back()
.endOffset() << "): " << *PStore << "\n"; }
} while (false)
4169 << ", " << NewSlices.back().endOffset()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " new slice [" << NewSlices
.back().beginOffset() << ", " << NewSlices.back()
.endOffset() << "): " << *PStore << "\n"; }
} while (false)
4170 << "): " << *PStore << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " new slice [" << NewSlices
.back().beginOffset() << ", " << NewSlices.back()
.endOffset() << "): " << *PStore << "\n"; }
} while (false)
;
4171 if (!SplitLoads) {
4172 LLVM_DEBUG(dbgs() << " of split load: " << *PLoad << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " of split load: " << *
PLoad << "\n"; } } while (false)
;
4173 }
4174
4175 // See if we've finished all the splits.
4176 if (Idx >= Size)
4177 break;
4178
4179 // Setup the next partition.
4180 PartOffset = Offsets.Splits[Idx];
4181 ++Idx;
4182 PartSize = (Idx < Size ? Offsets.Splits[Idx] : StoreSize) - PartOffset;
4183 }
4184
4185 // We want to immediately iterate on any allocas impacted by splitting
4186 // this load, which is only relevant if it isn't a load of this alloca and
4187 // thus we didn't already split the loads above. We also have to keep track
4188 // of any promotable allocas we split loads on as they can no longer be
4189 // promoted.
4190 if (!SplitLoads) {
4191 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(LoadBasePtr)) {
4192 assert(OtherAI != &AI && "We can't re-split our own alloca!")(static_cast <bool> (OtherAI != &AI && "We can't re-split our own alloca!"
) ? void (0) : __assert_fail ("OtherAI != &AI && \"We can't re-split our own alloca!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 4192, __extension__ __PRETTY_FUNCTION__
))
;
4193 ResplitPromotableAllocas.insert(OtherAI);
4194 Worklist.insert(OtherAI);
4195 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
4196 LoadBasePtr->stripInBoundsOffsets())) {
4197 assert(OtherAI != &AI && "We can't re-split our own alloca!")(static_cast <bool> (OtherAI != &AI && "We can't re-split our own alloca!"
) ? void (0) : __assert_fail ("OtherAI != &AI && \"We can't re-split our own alloca!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 4197, __extension__ __PRETTY_FUNCTION__
))
;
4198 Worklist.insert(OtherAI);
4199 }
4200 }
4201
4202 // Mark the original store as dead now that we've split it up and kill its
4203 // slice. Note that we leave the original load in place unless this store
4204 // was its only use. It may in turn be split up if it is an alloca load
4205 // for some other alloca, but it may be a normal load. This may introduce
4206 // redundant loads, but where those can be merged the rest of the optimizer
4207 // should handle the merging, and this uncovers SSA splits which is more
4208 // important. In practice, the original loads will almost always be fully
4209 // split and removed eventually, and the splits will be merged by any
4210 // trivial CSE, including instcombine.
4211 if (LI->hasOneUse()) {
4212 assert(*LI->user_begin() == SI && "Single use isn't this store!")(static_cast <bool> (*LI->user_begin() == SI &&
"Single use isn't this store!") ? void (0) : __assert_fail (
"*LI->user_begin() == SI && \"Single use isn't this store!\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 4212, __extension__ __PRETTY_FUNCTION__
))
;
4213 DeadInsts.push_back(LI);
4214 }
4215 DeadInsts.push_back(SI);
4216 Offsets.S->kill();
4217 }
4218
4219 // Remove the killed slices that have ben pre-split.
4220 llvm::erase_if(AS, [](const Slice &S) { return S.isDead(); });
4221
4222 // Insert our new slices. This will sort and merge them into the sorted
4223 // sequence.
4224 AS.insert(NewSlices);
4225
4226 LLVM_DEBUG(dbgs() << " Pre-split slices:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Pre-split slices:\n"; } } while
(false)
;
4227#ifndef NDEBUG
4228 for (auto I = AS.begin(), E = AS.end(); I != E; ++I)
4229 LLVM_DEBUG(AS.print(dbgs(), I, " "))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { AS.print(dbgs(), I, " "); } } while (false)
;
4230#endif
4231
4232 // Finally, don't try to promote any allocas that new require re-splitting.
4233 // They have already been added to the worklist above.
4234 llvm::erase_if(PromotableAllocas, [&](AllocaInst *AI) {
4235 return ResplitPromotableAllocas.count(AI);
4236 });
4237
4238 return true;
4239}
4240
4241/// Rewrite an alloca partition's users.
4242///
4243/// This routine drives both of the rewriting goals of the SROA pass. It tries
4244/// to rewrite uses of an alloca partition to be conducive for SSA value
4245/// promotion. If the partition needs a new, more refined alloca, this will
4246/// build that new alloca, preserving as much type information as possible, and
4247/// rewrite the uses of the old alloca to point at the new one and have the
4248/// appropriate new offsets. It also evaluates how successful the rewrite was
4249/// at enabling promotion and if it was successful queues the alloca to be
4250/// promoted.
4251AllocaInst *SROAPass::rewritePartition(AllocaInst &AI, AllocaSlices &AS,
4252 Partition &P) {
4253 // Try to compute a friendly type for this partition of the alloca. This
4254 // won't always succeed, in which case we fall back to a legal integer type
4255 // or an i8 array of an appropriate size.
4256 Type *SliceTy = nullptr;
4257 const DataLayout &DL = AI.getModule()->getDataLayout();
4258 std::pair<Type *, IntegerType *> CommonUseTy =
4259 findCommonType(P.begin(), P.end(), P.endOffset());
4260 // Do all uses operate on the same type?
4261 if (CommonUseTy.first)
4262 if (DL.getTypeAllocSize(CommonUseTy.first).getFixedSize() >= P.size())
4263 SliceTy = CommonUseTy.first;
4264 // If not, can we find an appropriate subtype in the original allocated type?
4265 if (!SliceTy)
4266 if (Type *TypePartitionTy = getTypePartition(DL, AI.getAllocatedType(),
4267 P.beginOffset(), P.size()))
4268 SliceTy = TypePartitionTy;
4269 // If still not, can we use the largest bitwidth integer type used?
4270 if (!SliceTy && CommonUseTy.second)
4271 if (DL.getTypeAllocSize(CommonUseTy.second).getFixedSize() >= P.size())
4272 SliceTy = CommonUseTy.second;
4273 if ((!SliceTy || (SliceTy->isArrayTy() &&
4274 SliceTy->getArrayElementType()->isIntegerTy())) &&
4275 DL.isLegalInteger(P.size() * 8))
4276 SliceTy = Type::getIntNTy(*C, P.size() * 8);
4277 if (!SliceTy)
4278 SliceTy = ArrayType::get(Type::getInt8Ty(*C), P.size());
4279 assert(DL.getTypeAllocSize(SliceTy).getFixedSize() >= P.size())(static_cast <bool> (DL.getTypeAllocSize(SliceTy).getFixedSize
() >= P.size()) ? void (0) : __assert_fail ("DL.getTypeAllocSize(SliceTy).getFixedSize() >= P.size()"
, "llvm/lib/Transforms/Scalar/SROA.cpp", 4279, __extension__ __PRETTY_FUNCTION__
))
;
4280
4281 bool IsIntegerPromotable = isIntegerWideningViable(P, SliceTy, DL);
4282
4283 VectorType *VecTy =
4284 IsIntegerPromotable ? nullptr : isVectorPromotionViable(P, DL);
4285 if (VecTy)
4286 SliceTy = VecTy;
4287
4288 // Check for the case where we're going to rewrite to a new alloca of the
4289 // exact same type as the original, and with the same access offsets. In that
4290 // case, re-use the existing alloca, but still run through the rewriter to
4291 // perform phi and select speculation.
4292 // P.beginOffset() can be non-zero even with the same type in a case with
4293 // out-of-bounds access (e.g. @PR35657 function in SROA/basictest.ll).
4294 AllocaInst *NewAI;
4295 if (SliceTy == AI.getAllocatedType() && P.beginOffset() == 0) {
4296 NewAI = &AI;
4297 // FIXME: We should be able to bail at this point with "nothing changed".
4298 // FIXME: We might want to defer PHI speculation until after here.
4299 // FIXME: return nullptr;
4300 } else {
4301 // Make sure the alignment is compatible with P.beginOffset().
4302 const Align Alignment = commonAlignment(AI.getAlign(), P.beginOffset());
4303 // If we will get at least this much alignment from the type alone, leave
4304 // the alloca's alignment unconstrained.
4305 const bool IsUnconstrained = Alignment <= DL.getABITypeAlign(SliceTy);
4306 NewAI = new AllocaInst(
4307 SliceTy, AI.getType()->getAddressSpace(), nullptr,
4308 IsUnconstrained ? DL.getPrefTypeAlign(SliceTy) : Alignment,
4309 AI.getName() + ".sroa." + Twine(P.begin() - AS.begin()), &AI);
4310 // Copy the old AI debug location over to the new one.
4311 NewAI->setDebugLoc(AI.getDebugLoc());
4312 ++NumNewAllocas;
4313 }
4314
4315 LLVM_DEBUG(dbgs() << "Rewriting alloca partition "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "Rewriting alloca partition " <<
"[" << P.beginOffset() << "," << P.endOffset
() << ") to: " << *NewAI << "\n"; } } while
(false)
4316 << "[" << P.beginOffset() << "," << P.endOffset()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "Rewriting alloca partition " <<
"[" << P.beginOffset() << "," << P.endOffset
() << ") to: " << *NewAI << "\n"; } } while
(false)
4317 << ") to: " << *NewAI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "Rewriting alloca partition " <<
"[" << P.beginOffset() << "," << P.endOffset
() << ") to: " << *NewAI << "\n"; } } while
(false)
;
4318
4319 // Track the high watermark on the worklist as it is only relevant for
4320 // promoted allocas. We will reset it to this point if the alloca is not in
4321 // fact scheduled for promotion.
4322 unsigned PPWOldSize = PostPromotionWorklist.size();
4323 unsigned NumUses = 0;
4324 SmallSetVector<PHINode *, 8> PHIUsers;
4325 SmallSetVector<SelectInst *, 8> SelectUsers;
4326
4327 AllocaSliceRewriter Rewriter(DL, AS, *this, AI, *NewAI, P.beginOffset(),
4328 P.endOffset(), IsIntegerPromotable, VecTy,
4329 PHIUsers, SelectUsers);
4330 bool Promotable = true;
4331 for (Slice *S : P.splitSliceTails()) {
4332 Promotable &= Rewriter.visit(S);
4333 ++NumUses;
4334 }
4335 for (Slice &S : P) {
4336 Promotable &= Rewriter.visit(&S);
4337 ++NumUses;
4338 }
4339
4340 NumAllocaPartitionUses += NumUses;
4341 MaxUsesPerAllocaPartition.updateMax(NumUses);
4342
4343 // Now that we've processed all the slices in the new partition, check if any
4344 // PHIs or Selects would block promotion.
4345 for (PHINode *PHI : PHIUsers)
4346 if (!isSafePHIToSpeculate(*PHI)) {
4347 Promotable = false;
4348 PHIUsers.clear();
4349 SelectUsers.clear();
4350 break;
4351 }
4352
4353 for (SelectInst *Sel : SelectUsers)
4354 if (!isSafeSelectToSpeculate(*Sel)) {
4355 Promotable = false;
4356 PHIUsers.clear();
4357 SelectUsers.clear();
4358 break;
4359 }
4360
4361 if (Promotable) {
4362 for (Use *U : AS.getDeadUsesIfPromotable()) {
4363 auto *OldInst = dyn_cast<Instruction>(U->get());
4364 Value::dropDroppableUse(*U);
4365 if (OldInst)
4366 if (isInstructionTriviallyDead(OldInst))
4367 DeadInsts.push_back(OldInst);
4368 }
4369 if (PHIUsers.empty() && SelectUsers.empty()) {
4370 // Promote the alloca.
4371 PromotableAllocas.push_back(NewAI);
4372 } else {
4373 // If we have either PHIs or Selects to speculate, add them to those
4374 // worklists and re-queue the new alloca so that we promote in on the
4375 // next iteration.
4376 for (PHINode *PHIUser : PHIUsers)
4377 SpeculatablePHIs.insert(PHIUser);
4378 for (SelectInst *SelectUser : SelectUsers)
4379 SpeculatableSelects.insert(SelectUser);
4380 Worklist.insert(NewAI);
4381 }
4382 } else {
4383 // Drop any post-promotion work items if promotion didn't happen.
4384 while (PostPromotionWorklist.size() > PPWOldSize)
4385 PostPromotionWorklist.pop_back();
4386
4387 // We couldn't promote and we didn't create a new partition, nothing
4388 // happened.
4389 if (NewAI == &AI)
4390 return nullptr;
4391
4392 // If we can't promote the alloca, iterate on it to check for new
4393 // refinements exposed by splitting the current alloca. Don't iterate on an
4394 // alloca which didn't actually change and didn't get promoted.
4395 Worklist.insert(NewAI);
4396 }
4397
4398 return NewAI;
4399}
4400
4401/// Walks the slices of an alloca and form partitions based on them,
4402/// rewriting each of their uses.
4403bool SROAPass::splitAlloca(AllocaInst &AI, AllocaSlices &AS) {
4404 if (AS.begin() == AS.end())
4405 return false;
4406
4407 unsigned NumPartitions = 0;
4408 bool Changed = false;
4409 const DataLayout &DL = AI.getModule()->getDataLayout();
4410
4411 // First try to pre-split loads and stores.
4412 Changed |= presplitLoadsAndStores(AI, AS);
4413
4414 // Now that we have identified any pre-splitting opportunities,
4415 // mark loads and stores unsplittable except for the following case.
4416 // We leave a slice splittable if all other slices are disjoint or fully
4417 // included in the slice, such as whole-alloca loads and stores.
4418 // If we fail to split these during pre-splitting, we want to force them
4419 // to be rewritten into a partition.
4420 bool IsSorted = true;
4421
4422 uint64_t AllocaSize =
4423 DL.getTypeAllocSize(AI.getAllocatedType()).getFixedSize();
4424 const uint64_t MaxBitVectorSize = 1024;
4425 if (AllocaSize <= MaxBitVectorSize) {
4426 // If a byte boundary is included in any load or store, a slice starting or
4427 // ending at the boundary is not splittable.
4428 SmallBitVector SplittableOffset(AllocaSize + 1, true);
4429 for (Slice &S : AS)
4430 for (unsigned O = S.beginOffset() + 1;
4431 O < S.endOffset() && O < AllocaSize; O++)
4432 SplittableOffset.reset(O);
4433
4434 for (Slice &S : AS) {
4435 if (!S.isSplittable())
4436 continue;
4437
4438 if ((S.beginOffset() > AllocaSize || SplittableOffset[S.beginOffset()]) &&
4439 (S.endOffset() > AllocaSize || SplittableOffset[S.endOffset()]))
4440 continue;
4441
4442 if (isa<LoadInst>(S.getUse()->getUser()) ||
4443 isa<StoreInst>(S.getUse()->getUser())) {
4444 S.makeUnsplittable();
4445 IsSorted = false;
4446 }
4447 }
4448 }
4449 else {
4450 // We only allow whole-alloca splittable loads and stores
4451 // for a large alloca to avoid creating too large BitVector.
4452 for (Slice &S : AS) {
4453 if (!S.isSplittable())
4454 continue;
4455
4456 if (S.beginOffset() == 0 && S.endOffset() >= AllocaSize)
4457 continue;
4458
4459 if (isa<LoadInst>(S.getUse()->getUser()) ||
4460 isa<StoreInst>(S.getUse()->getUser())) {
4461 S.makeUnsplittable();
4462 IsSorted = false;
4463 }
4464 }
4465 }
4466
4467 if (!IsSorted)
4468 llvm::sort(AS);
4469
4470 /// Describes the allocas introduced by rewritePartition in order to migrate
4471 /// the debug info.
4472 struct Fragment {
4473 AllocaInst *Alloca;
4474 uint64_t Offset;
4475 uint64_t Size;
4476 Fragment(AllocaInst *AI, uint64_t O, uint64_t S)
4477 : Alloca(AI), Offset(O), Size(S) {}
4478 };
4479 SmallVector<Fragment, 4> Fragments;
4480
4481 // Rewrite each partition.
4482 for (auto &P : AS.partitions()) {
4483 if (AllocaInst *NewAI = rewritePartition(AI, AS, P)) {
4484 Changed = true;
4485 if (NewAI != &AI) {
4486 uint64_t SizeOfByte = 8;
4487 uint64_t AllocaSize =
4488 DL.getTypeSizeInBits(NewAI->getAllocatedType()).getFixedSize();
4489 // Don't include any padding.
4490 uint64_t Size = std::min(AllocaSize, P.size() * SizeOfByte);
4491 Fragments.push_back(Fragment(NewAI, P.beginOffset() * SizeOfByte, Size));
4492 }
4493 }
4494 ++NumPartitions;
4495 }
4496
4497 NumAllocaPartitions += NumPartitions;
4498 MaxPartitionsPerAlloca.updateMax(NumPartitions);
4499
4500 // Migrate debug information from the old alloca to the new alloca(s)
4501 // and the individual partitions.
4502 TinyPtrVector<DbgVariableIntrinsic *> DbgDeclares = FindDbgAddrUses(&AI);
4503 for (DbgVariableIntrinsic *DbgDeclare : DbgDeclares) {
4504 auto *Expr = DbgDeclare->getExpression();
4505 DIBuilder DIB(*AI.getModule(), /*AllowUnresolved*/ false);
4506 uint64_t AllocaSize =
4507 DL.getTypeSizeInBits(AI.getAllocatedType()).getFixedSize();
4508 for (auto Fragment : Fragments) {
4509 // Create a fragment expression describing the new partition or reuse AI's
4510 // expression if there is only one partition.
4511 auto *FragmentExpr = Expr;
4512 if (Fragment.Size < AllocaSize || Expr->isFragment()) {
4513 // If this alloca is already a scalar replacement of a larger aggregate,
4514 // Fragment.Offset describes the offset inside the scalar.
4515 auto ExprFragment = Expr->getFragmentInfo();
4516 uint64_t Offset = ExprFragment ? ExprFragment->OffsetInBits : 0;
4517 uint64_t Start = Offset + Fragment.Offset;
4518 uint64_t Size = Fragment.Size;
4519 if (ExprFragment) {
4520 uint64_t AbsEnd =
4521 ExprFragment->OffsetInBits + ExprFragment->SizeInBits;
4522 if (Start >= AbsEnd)
4523 // No need to describe a SROAed padding.
4524 continue;
4525 Size = std::min(Size, AbsEnd - Start);
4526 }
4527 // The new, smaller fragment is stenciled out from the old fragment.
4528 if (auto OrigFragment = FragmentExpr->getFragmentInfo()) {
4529 assert(Start >= OrigFragment->OffsetInBits &&(static_cast <bool> (Start >= OrigFragment->OffsetInBits
&& "new fragment is outside of original fragment") ?
void (0) : __assert_fail ("Start >= OrigFragment->OffsetInBits && \"new fragment is outside of original fragment\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 4530, __extension__ __PRETTY_FUNCTION__
))
4530 "new fragment is outside of original fragment")(static_cast <bool> (Start >= OrigFragment->OffsetInBits
&& "new fragment is outside of original fragment") ?
void (0) : __assert_fail ("Start >= OrigFragment->OffsetInBits && \"new fragment is outside of original fragment\""
, "llvm/lib/Transforms/Scalar/SROA.cpp", 4530, __extension__ __PRETTY_FUNCTION__
))
;
4531 Start -= OrigFragment->OffsetInBits;
4532 }
4533
4534 // The alloca may be larger than the variable.
4535 auto VarSize = DbgDeclare->getVariable()->getSizeInBits();
4536 if (VarSize) {
4537 if (Size > *VarSize)
4538 Size = *VarSize;
4539 if (Size == 0 || Start + Size > *VarSize)
4540 continue;
4541 }
4542
4543 // Avoid creating a fragment expression that covers the entire variable.
4544 if (!VarSize || *VarSize != Size) {
4545 if (auto E =
4546 DIExpression::createFragmentExpression(Expr, Start, Size))
4547 FragmentExpr = *E;
4548 else
4549 continue;
4550 }
4551 }
4552
4553 // Remove any existing intrinsics on the new alloca describing
4554 // the variable fragment.
4555 for (DbgVariableIntrinsic *OldDII : FindDbgAddrUses(Fragment.Alloca)) {
4556 auto SameVariableFragment = [](const DbgVariableIntrinsic *LHS,
4557 const DbgVariableIntrinsic *RHS) {
4558 return LHS->getVariable() == RHS->getVariable() &&
4559 LHS->getDebugLoc()->getInlinedAt() ==
4560 RHS->getDebugLoc()->getInlinedAt();
4561 };
4562 if (SameVariableFragment(OldDII, DbgDeclare))
4563 OldDII->eraseFromParent();
4564 }
4565
4566 DIB.insertDeclare(Fragment.Alloca, DbgDeclare->getVariable(), FragmentExpr,
4567 DbgDeclare->getDebugLoc(), &AI);
4568 }
4569 }
4570 return Changed;
4571}
4572
4573/// Clobber a use with poison, deleting the used value if it becomes dead.
4574void SROAPass::clobberUse(Use &U) {
4575 Value *OldV = U;
4576 // Replace the use with an poison value.
4577 U = PoisonValue::get(OldV->getType());
4578
4579 // Check for this making an instruction dead. We have to garbage collect
4580 // all the dead instructions to ensure the uses of any alloca end up being
4581 // minimal.
4582 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
4583 if (isInstructionTriviallyDead(OldI)) {
4584 DeadInsts.push_back(OldI);
4585 }
4586}
4587
4588/// Analyze an alloca for SROA.
4589///
4590/// This analyzes the alloca to ensure we can reason about it, builds
4591/// the slices of the alloca, and then hands it off to be split and
4592/// rewritten as needed.
4593bool SROAPass::runOnAlloca(AllocaInst &AI) {
4594 LLVM_DEBUG(dbgs() << "SROA alloca: " << AI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "SROA alloca: " << AI <<
"\n"; } } while (false)
;
4595 ++NumAllocasAnalyzed;
4596
4597 // Special case dead allocas, as they're trivial.
4598 if (AI.use_empty()) {
4599 AI.eraseFromParent();
4600 return true;
4601 }
4602 const DataLayout &DL = AI.getModule()->getDataLayout();
4603
4604 // Skip alloca forms that this analysis can't handle.
4605 auto *AT = AI.getAllocatedType();
4606 if (AI.isArrayAllocation() || !AT->isSized() || isa<ScalableVectorType>(AT) ||
4607 DL.getTypeAllocSize(AT).getFixedSize() == 0)
4608 return false;
4609
4610 bool Changed = false;
4611
4612 // First, split any FCA loads and stores touching this alloca to promote
4613 // better splitting and promotion opportunities.
4614 IRBuilderTy IRB(&AI);
4615 AggLoadStoreRewriter AggRewriter(DL, IRB);
4616 Changed |= AggRewriter.rewrite(AI);
4617
4618 // Build the slices using a recursive instruction-visiting builder.
4619 AllocaSlices AS(DL, AI);
4620 LLVM_DEBUG(AS.print(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { AS.print(dbgs()); } } while (false)
;
4621 if (AS.isEscaped())
4622 return Changed;
4623
4624 // Delete all the dead users of this alloca before splitting and rewriting it.
4625 for (Instruction *DeadUser : AS.getDeadUsers()) {
4626 // Free up everything used by this instruction.
4627 for (Use &DeadOp : DeadUser->operands())
4628 clobberUse(DeadOp);
4629
4630 // Now replace the uses of this instruction.
4631 DeadUser->replaceAllUsesWith(PoisonValue::get(DeadUser->getType()));
4632
4633 // And mark it for deletion.
4634 DeadInsts.push_back(DeadUser);
4635 Changed = true;
4636 }
4637 for (Use *DeadOp : AS.getDeadOperands()) {
4638 clobberUse(*DeadOp);
4639 Changed = true;
4640 }
4641
4642 // No slices to split. Leave the dead alloca for a later pass to clean up.
4643 if (AS.begin() == AS.end())
4644 return Changed;
4645
4646 Changed |= splitAlloca(AI, AS);
4647
4648 LLVM_DEBUG(dbgs() << " Speculating PHIs\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Speculating PHIs\n"; } } while
(false)
;
4649 while (!SpeculatablePHIs.empty())
4650 speculatePHINodeLoads(IRB, *SpeculatablePHIs.pop_back_val());
4651
4652 LLVM_DEBUG(dbgs() << " Speculating Selects\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Speculating Selects\n"; } } while
(false)
;
4653 while (!SpeculatableSelects.empty())
4654 speculateSelectInstLoads(IRB, *SpeculatableSelects.pop_back_val());
4655
4656 return Changed;
4657}
4658
4659/// Delete the dead instructions accumulated in this run.
4660///
4661/// Recursively deletes the dead instructions we've accumulated. This is done
4662/// at the very end to maximize locality of the recursive delete and to
4663/// minimize the problems of invalidated instruction pointers as such pointers
4664/// are used heavily in the intermediate stages of the algorithm.
4665///
4666/// We also record the alloca instructions deleted here so that they aren't
4667/// subsequently handed to mem2reg to promote.
4668bool SROAPass::deleteDeadInstructions(
4669 SmallPtrSetImpl<AllocaInst *> &DeletedAllocas) {
4670 bool Changed = false;
4671 while (!DeadInsts.empty()) {
4672 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
4673 if (!I) continue;
4674 LLVM_DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "Deleting dead instruction: " <<
*I << "\n"; } } while (false)
;
4675
4676 // If the instruction is an alloca, find the possible dbg.declare connected
4677 // to it, and remove it too. We must do this before calling RAUW or we will
4678 // not be able to find it.
4679 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
4680 DeletedAllocas.insert(AI);
4681 for (DbgVariableIntrinsic *OldDII : FindDbgAddrUses(AI))
4682 OldDII->eraseFromParent();
4683 }
4684
4685 I->replaceAllUsesWith(UndefValue::get(I->getType()));
4686
4687 for (Use &Operand : I->operands())
4688 if (Instruction *U = dyn_cast<Instruction>(Operand)) {
4689 // Zero out the operand and see if it becomes trivially dead.
4690 Operand = nullptr;
4691 if (isInstructionTriviallyDead(U))
4692 DeadInsts.push_back(U);
4693 }
4694
4695 ++NumDeleted;
4696 I->eraseFromParent();
4697 Changed = true;
4698 }
4699 return Changed;
4700}
4701
4702/// Promote the allocas, using the best available technique.
4703///
4704/// This attempts to promote whatever allocas have been identified as viable in
4705/// the PromotableAllocas list. If that list is empty, there is nothing to do.
4706/// This function returns whether any promotion occurred.
4707bool SROAPass::promoteAllocas(Function &F) {
4708 if (PromotableAllocas.empty())
4709 return false;
4710
4711 NumPromoted += PromotableAllocas.size();
4712
4713 LLVM_DEBUG(dbgs() << "Promoting allocas with mem2reg...\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "Promoting allocas with mem2reg...\n"
; } } while (false)
;
4714 PromoteMemToReg(PromotableAllocas, *DT, AC);
4715 PromotableAllocas.clear();
4716 return true;
4717}
4718
4719PreservedAnalyses SROAPass::runImpl(Function &F, DominatorTree &RunDT,
4720 AssumptionCache &RunAC) {
4721 LLVM_DEBUG(dbgs() << "SROA function: " << F.getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "SROA function: " << F.getName
() << "\n"; } } while (false)
;
4722 C = &F.getContext();
4723 DT = &RunDT;
4724 AC = &RunAC;
4725
4726 BasicBlock &EntryBB = F.getEntryBlock();
4727 for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
4728 I != E; ++I) {
4729 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
4730 if (isa<ScalableVectorType>(AI->getAllocatedType())) {
4731 if (isAllocaPromotable(AI))
4732 PromotableAllocas.push_back(AI);
4733 } else {
4734 Worklist.insert(AI);
4735 }
4736 }
4737 }
4738
4739 bool Changed = false;
4740 // A set of deleted alloca instruction pointers which should be removed from
4741 // the list of promotable allocas.
4742 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
4743
4744 do {
4745 while (!Worklist.empty()) {
4746 Changed |= runOnAlloca(*Worklist.pop_back_val());
4747 Changed |= deleteDeadInstructions(DeletedAllocas);
4748
4749 // Remove the deleted allocas from various lists so that we don't try to
4750 // continue processing them.
4751 if (!DeletedAllocas.empty()) {
4752 auto IsInSet = [&](AllocaInst *AI) { return DeletedAllocas.count(AI); };
4753 Worklist.remove_if(IsInSet);
4754 PostPromotionWorklist.remove_if(IsInSet);
4755 llvm::erase_if(PromotableAllocas, IsInSet);
4756 DeletedAllocas.clear();
4757 }
4758 }
4759
4760 Changed |= promoteAllocas(F);
4761
4762 Worklist = PostPromotionWorklist;
4763 PostPromotionWorklist.clear();
4764 } while (!Worklist.empty());
4765
4766 if (!Changed)
4767 return PreservedAnalyses::all();
4768
4769 PreservedAnalyses PA;
4770 PA.preserveSet<CFGAnalyses>();
4771 return PA;
4772}
4773
4774PreservedAnalyses SROAPass::run(Function &F, FunctionAnalysisManager &AM) {
4775 return runImpl(F, AM.getResult<DominatorTreeAnalysis>(F),
4776 AM.getResult<AssumptionAnalysis>(F));
4777}
4778
4779/// A legacy pass for the legacy pass manager that wraps the \c SROA pass.
4780///
4781/// This is in the llvm namespace purely to allow it to be a friend of the \c
4782/// SROA pass.
4783class llvm::sroa::SROALegacyPass : public FunctionPass {
4784 /// The SROA implementation.
4785 SROAPass Impl;
4786
4787public:
4788 static char ID;
4789
4790 SROALegacyPass() : FunctionPass(ID) {
4791 initializeSROALegacyPassPass(*PassRegistry::getPassRegistry());
4792 }
4793
4794 bool runOnFunction(Function &F) override {
4795 if (skipFunction(F))
4796 return false;
4797
4798 auto PA = Impl.runImpl(
4799 F, getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
4800 getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F));
4801 return !PA.areAllPreserved();
4802 }
4803
4804 void getAnalysisUsage(AnalysisUsage &AU) const override {
4805 AU.addRequired<AssumptionCacheTracker>();
4806 AU.addRequired<DominatorTreeWrapperPass>();
4807 AU.addPreserved<GlobalsAAWrapperPass>();
4808 AU.setPreservesCFG();
4809 }
4810
4811 StringRef getPassName() const override { return "SROA"; }
4812};
4813
4814char SROALegacyPass::ID = 0;
4815
4816FunctionPass *llvm::createSROAPass() { return new SROALegacyPass(); }
4817
4818INITIALIZE_PASS_BEGIN(SROALegacyPass, "sroa",static void *initializeSROALegacyPassPassOnce(PassRegistry &
Registry) {
4819 "Scalar Replacement Of Aggregates", false, false)static void *initializeSROALegacyPassPassOnce(PassRegistry &
Registry) {
4820INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry);
4821INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);
4822INITIALIZE_PASS_END(SROALegacyPass, "sroa", "Scalar Replacement Of Aggregates",PassInfo *PI = new PassInfo( "Scalar Replacement Of Aggregates"
, "sroa", &SROALegacyPass::ID, PassInfo::NormalCtor_t(callDefaultCtor
<SROALegacyPass>), false, false); Registry.registerPass
(*PI, true); return PI; } static llvm::once_flag InitializeSROALegacyPassPassFlag
; void llvm::initializeSROALegacyPassPass(PassRegistry &Registry
) { llvm::call_once(InitializeSROALegacyPassPassFlag, initializeSROALegacyPassPassOnce
, std::ref(Registry)); }
4823 false, false)PassInfo *PI = new PassInfo( "Scalar Replacement Of Aggregates"
, "sroa", &SROALegacyPass::ID, PassInfo::NormalCtor_t(callDefaultCtor
<SROALegacyPass>), false, false); Registry.registerPass
(*PI, true); return PI; } static llvm::once_flag InitializeSROALegacyPassPassFlag
; void llvm::initializeSROALegacyPassPass(PassRegistry &Registry
) { llvm::call_once(InitializeSROALegacyPassPassFlag, initializeSROALegacyPassPassOnce
, std::ref(Registry)); }