Bug Summary

File:llvm/lib/Transforms/Scalar/SROA.cpp
Warning:line 2932, 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 -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 -fno-rounding-math -mconstructor-aliases -munwind-tables -target-cpu x86-64 -tune-cpu generic -debugger-tuning=gdb -ffunction-sections -fdata-sections -fcoverage-compilation-dir=/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/build-llvm/lib/Transforms/Scalar -resource-dir /usr/lib/llvm-13/lib/clang/13.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/build-llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/build-llvm/include -I /build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/include -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-13/lib/clang/13.0.0/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 -O2 -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-13~++20210613111130+5be314f79ba7/build-llvm/lib/Transforms/Scalar -fdebug-prefix-map=/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7=. -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -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-2021-06-13-111025-38230-1 -x c++ /build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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 const 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 undef 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 /// undef 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 undef 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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()"
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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()", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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?\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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.\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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.\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/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)"
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1007, __extension__ __PRETTY_FUNCTION__))
;
1008 if (I.use_empty())
1009 return markAsDead(I);
1010
1011 // TODO: We could use SimplifyInstruction here to fold PHINodes and
1012 // SelectInsts. However, doing so requires to change the current
1013 // dead-operand-tracking mechanism. For instance, suppose neither loading
1014 // from %U nor %other traps. Then "load (select undef, %U, %other)" does not
1015 // trap either. However, if we simply replace %U with undef using the
1016 // current dead-operand-tracking mechanism, "load (select undef, undef,
1017 // %other)" may trap because the select may return the first operand
1018 // "undef".
1019 if (Value *Result = foldPHINodeOrSelectInst(I)) {
1020 if (Result == *U)
1021 // If the result of the constant fold will be the pointer, recurse
1022 // through the PHI/select as if we had RAUW'ed it.
1023 enqueueUsers(I);
1024 else
1025 // Otherwise the operand to the PHI/select is dead, and we can replace
1026 // it with undef.
1027 AS.DeadOperands.push_back(U);
1028
1029 return;
1030 }
1031
1032 if (!IsOffsetKnown)
1033 return PI.setAborted(&I);
1034
1035 // See if we already have computed info on this node.
1036 uint64_t &Size = PHIOrSelectSizes[&I];
1037 if (!Size) {
1038 // This is a new PHI/Select, check for an unsafe use of it.
1039 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&I, Size))
1040 return PI.setAborted(UnsafeI);
1041 }
1042
1043 // For PHI and select operands outside the alloca, we can't nuke the entire
1044 // phi or select -- the other side might still be relevant, so we special
1045 // case them here and use a separate structure to track the operands
1046 // themselves which should be replaced with undef.
1047 // FIXME: This should instead be escaped in the event we're instrumenting
1048 // for address sanitization.
1049 if (Offset.uge(AllocSize)) {
1050 AS.DeadOperands.push_back(U);
1051 return;
1052 }
1053
1054 insertUse(I, Offset, Size);
1055 }
1056
1057 void visitPHINode(PHINode &PN) { visitPHINodeOrSelectInst(PN); }
1058
1059 void visitSelectInst(SelectInst &SI) { visitPHINodeOrSelectInst(SI); }
1060
1061 /// Disable SROA entirely if there are unhandled users of the alloca.
1062 void visitInstruction(Instruction &I) { PI.setAborted(&I); }
1063};
1064
1065AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
1066 :
1067#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1068 AI(AI),
1069#endif
1070 PointerEscapingInstr(nullptr) {
1071 SliceBuilder PB(DL, AI, *this);
1072 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
1073 if (PtrI.isEscaped() || PtrI.isAborted()) {
1074 // FIXME: We should sink the escape vs. abort info into the caller nicely,
1075 // possibly by just storing the PtrInfo in the AllocaSlices.
1076 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
1077 : PtrI.getAbortingInst();
1078 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1078, __extension__ __PRETTY_FUNCTION__))
;
1079 return;
1080 }
1081
1082 llvm::erase_if(Slices, [](const Slice &S) { return S.isDead(); });
1083
1084 // Sort the uses. This arranges for the offsets to be in ascending order,
1085 // and the sizes to be in descending order.
1086 llvm::stable_sort(Slices);
1087}
1088
1089#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1090
1091void AllocaSlices::print(raw_ostream &OS, const_iterator I,
1092 StringRef Indent) const {
1093 printSlice(OS, I, Indent);
1094 OS << "\n";
1095 printUse(OS, I, Indent);
1096}
1097
1098void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
1099 StringRef Indent) const {
1100 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
1101 << " slice #" << (I - begin())
1102 << (I->isSplittable() ? " (splittable)" : "");
1103}
1104
1105void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
1106 StringRef Indent) const {
1107 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
1108}
1109
1110void AllocaSlices::print(raw_ostream &OS) const {
1111 if (PointerEscapingInstr) {
1112 OS << "Can't analyze slices for alloca: " << AI << "\n"
1113 << " A pointer to this alloca escaped by:\n"
1114 << " " << *PointerEscapingInstr << "\n";
1115 return;
1116 }
1117
1118 OS << "Slices of alloca: " << AI << "\n";
1119 for (const_iterator I = begin(), E = end(); I != E; ++I)
1120 print(OS, I);
1121}
1122
1123LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void AllocaSlices::dump(const_iterator I) const {
1124 print(dbgs(), I);
1125}
1126LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void AllocaSlices::dump() const { print(dbgs()); }
1127
1128#endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1129
1130/// Walk the range of a partitioning looking for a common type to cover this
1131/// sequence of slices.
1132static std::pair<Type *, IntegerType *>
1133findCommonType(AllocaSlices::const_iterator B, AllocaSlices::const_iterator E,
1134 uint64_t EndOffset) {
1135 Type *Ty = nullptr;
1136 bool TyIsCommon = true;
1137 IntegerType *ITy = nullptr;
1138
1139 // Note that we need to look at *every* alloca slice's Use to ensure we
1140 // always get consistent results regardless of the order of slices.
1141 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1142 Use *U = I->getUse();
1143 if (isa<IntrinsicInst>(*U->getUser()))
1144 continue;
1145 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1146 continue;
1147
1148 Type *UserTy = nullptr;
1149 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1150 UserTy = LI->getType();
1151 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1152 UserTy = SI->getValueOperand()->getType();
1153 }
1154
1155 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1156 // If the type is larger than the partition, skip it. We only encounter
1157 // this for split integer operations where we want to use the type of the
1158 // entity causing the split. Also skip if the type is not a byte width
1159 // multiple.
1160 if (UserITy->getBitWidth() % 8 != 0 ||
1161 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1162 continue;
1163
1164 // Track the largest bitwidth integer type used in this way in case there
1165 // is no common type.
1166 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1167 ITy = UserITy;
1168 }
1169
1170 // To avoid depending on the order of slices, Ty and TyIsCommon must not
1171 // depend on types skipped above.
1172 if (!UserTy || (Ty && Ty != UserTy))
1173 TyIsCommon = false; // Give up on anything but an iN type.
1174 else
1175 Ty = UserTy;
1176 }
1177
1178 return {TyIsCommon ? Ty : nullptr, ITy};
1179}
1180
1181/// PHI instructions that use an alloca and are subsequently loaded can be
1182/// rewritten to load both input pointers in the pred blocks and then PHI the
1183/// results, allowing the load of the alloca to be promoted.
1184/// From this:
1185/// %P2 = phi [i32* %Alloca, i32* %Other]
1186/// %V = load i32* %P2
1187/// to:
1188/// %V1 = load i32* %Alloca -> will be mem2reg'd
1189/// ...
1190/// %V2 = load i32* %Other
1191/// ...
1192/// %V = phi [i32 %V1, i32 %V2]
1193///
1194/// We can do this to a select if its only uses are loads and if the operands
1195/// to the select can be loaded unconditionally.
1196///
1197/// FIXME: This should be hoisted into a generic utility, likely in
1198/// Transforms/Util/Local.h
1199static bool isSafePHIToSpeculate(PHINode &PN) {
1200 const DataLayout &DL = PN.getModule()->getDataLayout();
1201
1202 // For now, we can only do this promotion if the load is in the same block
1203 // as the PHI, and if there are no stores between the phi and load.
1204 // TODO: Allow recursive phi users.
1205 // TODO: Allow stores.
1206 BasicBlock *BB = PN.getParent();
1207 Align MaxAlign;
1208 uint64_t APWidth = DL.getIndexTypeSizeInBits(PN.getType());
1209 APInt MaxSize(APWidth, 0);
1210 bool HaveLoad = false;
1211 for (User *U : PN.users()) {
1212 LoadInst *LI = dyn_cast<LoadInst>(U);
1213 if (!LI || !LI->isSimple())
1214 return false;
1215
1216 // For now we only allow loads in the same block as the PHI. This is
1217 // a common case that happens when instcombine merges two loads through
1218 // a PHI.
1219 if (LI->getParent() != BB)
1220 return false;
1221
1222 // Ensure that there are no instructions between the PHI and the load that
1223 // could store.
1224 for (BasicBlock::iterator BBI(PN); &*BBI != LI; ++BBI)
1225 if (BBI->mayWriteToMemory())
1226 return false;
1227
1228 uint64_t Size = DL.getTypeStoreSize(LI->getType()).getFixedSize();
1229 MaxAlign = std::max(MaxAlign, LI->getAlign());
1230 MaxSize = MaxSize.ult(Size) ? APInt(APWidth, Size) : MaxSize;
1231 HaveLoad = true;
1232 }
1233
1234 if (!HaveLoad)
1235 return false;
1236
1237 // We can only transform this if it is safe to push the loads into the
1238 // predecessor blocks. The only thing to watch out for is that we can't put
1239 // a possibly trapping load in the predecessor if it is a critical edge.
1240 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1241 Instruction *TI = PN.getIncomingBlock(Idx)->getTerminator();
1242 Value *InVal = PN.getIncomingValue(Idx);
1243
1244 // If the value is produced by the terminator of the predecessor (an
1245 // invoke) or it has side-effects, there is no valid place to put a load
1246 // in the predecessor.
1247 if (TI == InVal || TI->mayHaveSideEffects())
1248 return false;
1249
1250 // If the predecessor has a single successor, then the edge isn't
1251 // critical.
1252 if (TI->getNumSuccessors() == 1)
1253 continue;
1254
1255 // If this pointer is always safe to load, or if we can prove that there
1256 // is already a load in the block, then we can move the load to the pred
1257 // block.
1258 if (isSafeToLoadUnconditionally(InVal, MaxAlign, MaxSize, DL, TI))
1259 continue;
1260
1261 return false;
1262 }
1263
1264 return true;
1265}
1266
1267static void speculatePHINodeLoads(PHINode &PN) {
1268 LLVM_DEBUG(dbgs() << " original: " << PN << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << PN <<
"\n"; } } while (false)
;
1269
1270 LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1271 Type *LoadTy = SomeLoad->getType();
1272 IRBuilderTy PHIBuilder(&PN);
1273 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1274 PN.getName() + ".sroa.speculated");
1275
1276 // Get the AA tags and alignment to use from one of the loads. It does not
1277 // matter which one we get and if any differ.
1278 AAMDNodes AATags;
1279 SomeLoad->getAAMetadata(AATags);
1280 Align Alignment = SomeLoad->getAlign();
1281
1282 // Rewrite all loads of the PN to use the new PHI.
1283 while (!PN.use_empty()) {
1284 LoadInst *LI = cast<LoadInst>(PN.user_back());
1285 LI->replaceAllUsesWith(NewPN);
1286 LI->eraseFromParent();
1287 }
1288
1289 // Inject loads into all of the pred blocks.
1290 DenseMap<BasicBlock*, Value*> InjectedLoads;
1291 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1292 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1293 Value *InVal = PN.getIncomingValue(Idx);
1294
1295 // A PHI node is allowed to have multiple (duplicated) entries for the same
1296 // basic block, as long as the value is the same. So if we already injected
1297 // a load in the predecessor, then we should reuse the same load for all
1298 // duplicated entries.
1299 if (Value* V = InjectedLoads.lookup(Pred)) {
1300 NewPN->addIncoming(V, Pred);
1301 continue;
1302 }
1303
1304 Instruction *TI = Pred->getTerminator();
1305 IRBuilderTy PredBuilder(TI);
1306
1307 LoadInst *Load = PredBuilder.CreateAlignedLoad(
1308 LoadTy, InVal, Alignment,
1309 (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1310 ++NumLoadsSpeculated;
1311 if (AATags)
1312 Load->setAAMetadata(AATags);
1313 NewPN->addIncoming(Load, Pred);
1314 InjectedLoads[Pred] = Load;
1315 }
1316
1317 LLVM_DEBUG(dbgs() << " speculated to: " << *NewPN << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " speculated to: " <<
*NewPN << "\n"; } } while (false)
;
1318 PN.eraseFromParent();
1319}
1320
1321/// Select instructions that use an alloca and are subsequently loaded can be
1322/// rewritten to load both input pointers and then select between the result,
1323/// allowing the load of the alloca to be promoted.
1324/// From this:
1325/// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1326/// %V = load i32* %P2
1327/// to:
1328/// %V1 = load i32* %Alloca -> will be mem2reg'd
1329/// %V2 = load i32* %Other
1330/// %V = select i1 %cond, i32 %V1, i32 %V2
1331///
1332/// We can do this to a select if its only uses are loads and if the operand
1333/// to the select can be loaded unconditionally.
1334static bool isSafeSelectToSpeculate(SelectInst &SI) {
1335 Value *TValue = SI.getTrueValue();
1336 Value *FValue = SI.getFalseValue();
1337 const DataLayout &DL = SI.getModule()->getDataLayout();
1338
1339 for (User *U : SI.users()) {
1340 LoadInst *LI = dyn_cast<LoadInst>(U);
1341 if (!LI || !LI->isSimple())
1342 return false;
1343
1344 // Both operands to the select need to be dereferenceable, either
1345 // absolutely (e.g. allocas) or at this point because we can see other
1346 // accesses to it.
1347 if (!isSafeToLoadUnconditionally(TValue, LI->getType(),
1348 LI->getAlign(), DL, LI))
1349 return false;
1350 if (!isSafeToLoadUnconditionally(FValue, LI->getType(),
1351 LI->getAlign(), DL, LI))
1352 return false;
1353 }
1354
1355 return true;
1356}
1357
1358static void speculateSelectInstLoads(SelectInst &SI) {
1359 LLVM_DEBUG(dbgs() << " original: " << SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << SI <<
"\n"; } } while (false)
;
1360
1361 IRBuilderTy IRB(&SI);
1362 Value *TV = SI.getTrueValue();
1363 Value *FV = SI.getFalseValue();
1364 // Replace the loads of the select with a select of two loads.
1365 while (!SI.use_empty()) {
1366 LoadInst *LI = cast<LoadInst>(SI.user_back());
1367 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1367, __extension__ __PRETTY_FUNCTION__))
;
1368
1369 IRB.SetInsertPoint(LI);
1370 LoadInst *TL = IRB.CreateLoad(LI->getType(), TV,
1371 LI->getName() + ".sroa.speculate.load.true");
1372 LoadInst *FL = IRB.CreateLoad(LI->getType(), FV,
1373 LI->getName() + ".sroa.speculate.load.false");
1374 NumLoadsSpeculated += 2;
1375
1376 // Transfer alignment and AA info if present.
1377 TL->setAlignment(LI->getAlign());
1378 FL->setAlignment(LI->getAlign());
1379
1380 AAMDNodes Tags;
1381 LI->getAAMetadata(Tags);
1382 if (Tags) {
1383 TL->setAAMetadata(Tags);
1384 FL->setAAMetadata(Tags);
1385 }
1386
1387 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1388 LI->getName() + ".sroa.speculated");
1389
1390 LLVM_DEBUG(dbgs() << " speculated to: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " speculated to: " <<
*V << "\n"; } } while (false)
;
1391 LI->replaceAllUsesWith(V);
1392 LI->eraseFromParent();
1393 }
1394 SI.eraseFromParent();
1395}
1396
1397/// Build a GEP out of a base pointer and indices.
1398///
1399/// This will return the BasePtr if that is valid, or build a new GEP
1400/// instruction using the IRBuilder if GEP-ing is needed.
1401static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1402 SmallVectorImpl<Value *> &Indices,
1403 const Twine &NamePrefix) {
1404 if (Indices.empty())
1405 return BasePtr;
1406
1407 // A single zero index is a no-op, so check for this and avoid building a GEP
1408 // in that case.
1409 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1410 return BasePtr;
1411
1412 return IRB.CreateInBoundsGEP(BasePtr->getType()->getPointerElementType(),
1413 BasePtr, Indices, NamePrefix + "sroa_idx");
1414}
1415
1416/// Get a natural GEP off of the BasePtr walking through Ty toward
1417/// TargetTy without changing the offset of the pointer.
1418///
1419/// This routine assumes we've already established a properly offset GEP with
1420/// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1421/// zero-indices down through type layers until we find one the same as
1422/// TargetTy. If we can't find one with the same type, we at least try to use
1423/// one with the same size. If none of that works, we just produce the GEP as
1424/// indicated by Indices to have the correct offset.
1425static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1426 Value *BasePtr, Type *Ty, Type *TargetTy,
1427 SmallVectorImpl<Value *> &Indices,
1428 const Twine &NamePrefix) {
1429 if (Ty == TargetTy)
1430 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1431
1432 // Offset size to use for the indices.
1433 unsigned OffsetSize = DL.getIndexTypeSizeInBits(BasePtr->getType());
1434
1435 // See if we can descend into a struct and locate a field with the correct
1436 // type.
1437 unsigned NumLayers = 0;
1438 Type *ElementTy = Ty;
1439 do {
1440 if (ElementTy->isPointerTy())
1441 break;
1442
1443 if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1444 ElementTy = ArrayTy->getElementType();
1445 Indices.push_back(IRB.getIntN(OffsetSize, 0));
1446 } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1447 ElementTy = VectorTy->getElementType();
1448 Indices.push_back(IRB.getInt32(0));
1449 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1450 if (STy->element_begin() == STy->element_end())
1451 break; // Nothing left to descend into.
1452 ElementTy = *STy->element_begin();
1453 Indices.push_back(IRB.getInt32(0));
1454 } else {
1455 break;
1456 }
1457 ++NumLayers;
1458 } while (ElementTy != TargetTy);
1459 if (ElementTy != TargetTy)
1460 Indices.erase(Indices.end() - NumLayers, Indices.end());
1461
1462 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1463}
1464
1465/// Recursively compute indices for a natural GEP.
1466///
1467/// This is the recursive step for getNaturalGEPWithOffset that walks down the
1468/// element types adding appropriate indices for the GEP.
1469static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1470 Value *Ptr, Type *Ty, APInt &Offset,
1471 Type *TargetTy,
1472 SmallVectorImpl<Value *> &Indices,
1473 const Twine &NamePrefix) {
1474 if (Offset == 0)
1475 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices,
1476 NamePrefix);
1477
1478 // We can't recurse through pointer types.
1479 if (Ty->isPointerTy())
1480 return nullptr;
1481
1482 // We try to analyze GEPs over vectors here, but note that these GEPs are
1483 // extremely poorly defined currently. The long-term goal is to remove GEPing
1484 // over a vector from the IR completely.
1485 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1486 unsigned ElementSizeInBits =
1487 DL.getTypeSizeInBits(VecTy->getScalarType()).getFixedSize();
1488 if (ElementSizeInBits % 8 != 0) {
1489 // GEPs over non-multiple of 8 size vector elements are invalid.
1490 return nullptr;
1491 }
1492 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1493 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1494 if (NumSkippedElements.ugt(cast<FixedVectorType>(VecTy)->getNumElements()))
1495 return nullptr;
1496 Offset -= NumSkippedElements * ElementSize;
1497 Indices.push_back(IRB.getInt(NumSkippedElements));
1498 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1499 Offset, TargetTy, Indices, NamePrefix);
1500 }
1501
1502 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1503 Type *ElementTy = ArrTy->getElementType();
1504 APInt ElementSize(Offset.getBitWidth(),
1505 DL.getTypeAllocSize(ElementTy).getFixedSize());
1506 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1507 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1508 return nullptr;
1509
1510 Offset -= NumSkippedElements * ElementSize;
1511 Indices.push_back(IRB.getInt(NumSkippedElements));
1512 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1513 Indices, NamePrefix);
1514 }
1515
1516 StructType *STy = dyn_cast<StructType>(Ty);
1517 if (!STy)
1518 return nullptr;
1519
1520 const StructLayout *SL = DL.getStructLayout(STy);
1521 uint64_t StructOffset = Offset.getZExtValue();
1522 if (StructOffset >= SL->getSizeInBytes())
1523 return nullptr;
1524 unsigned Index = SL->getElementContainingOffset(StructOffset);
1525 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1526 Type *ElementTy = STy->getElementType(Index);
1527 if (Offset.uge(DL.getTypeAllocSize(ElementTy).getFixedSize()))
1528 return nullptr; // The offset points into alignment padding.
1529
1530 Indices.push_back(IRB.getInt32(Index));
1531 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1532 Indices, NamePrefix);
1533}
1534
1535/// Get a natural GEP from a base pointer to a particular offset and
1536/// resulting in a particular type.
1537///
1538/// The goal is to produce a "natural" looking GEP that works with the existing
1539/// composite types to arrive at the appropriate offset and element type for
1540/// a pointer. TargetTy is the element type the returned GEP should point-to if
1541/// possible. We recurse by decreasing Offset, adding the appropriate index to
1542/// Indices, and setting Ty to the result subtype.
1543///
1544/// If no natural GEP can be constructed, this function returns null.
1545static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1546 Value *Ptr, APInt Offset, Type *TargetTy,
1547 SmallVectorImpl<Value *> &Indices,
1548 const Twine &NamePrefix) {
1549 PointerType *Ty = cast<PointerType>(Ptr->getType());
1550
1551 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1552 // an i8.
1553 if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1554 return nullptr;
1555
1556 Type *ElementTy = Ty->getElementType();
1557 if (!ElementTy->isSized())
1558 return nullptr; // We can't GEP through an unsized element.
1559 if (isa<ScalableVectorType>(ElementTy))
1560 return nullptr;
1561 APInt ElementSize(Offset.getBitWidth(),
1562 DL.getTypeAllocSize(ElementTy).getFixedSize());
1563 if (ElementSize == 0)
1564 return nullptr; // Zero-length arrays can't help us build a natural GEP.
1565 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1566
1567 Offset -= NumSkippedElements * ElementSize;
1568 Indices.push_back(IRB.getInt(NumSkippedElements));
1569 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1570 Indices, NamePrefix);
1571}
1572
1573/// Compute an adjusted pointer from Ptr by Offset bytes where the
1574/// resulting pointer has PointerTy.
1575///
1576/// This tries very hard to compute a "natural" GEP which arrives at the offset
1577/// and produces the pointer type desired. Where it cannot, it will try to use
1578/// the natural GEP to arrive at the offset and bitcast to the type. Where that
1579/// fails, it will try to use an existing i8* and GEP to the byte offset and
1580/// bitcast to the type.
1581///
1582/// The strategy for finding the more natural GEPs is to peel off layers of the
1583/// pointer, walking back through bit casts and GEPs, searching for a base
1584/// pointer from which we can compute a natural GEP with the desired
1585/// properties. The algorithm tries to fold as many constant indices into
1586/// a single GEP as possible, thus making each GEP more independent of the
1587/// surrounding code.
1588static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1589 APInt Offset, Type *PointerTy,
1590 const Twine &NamePrefix) {
1591 // Even though we don't look through PHI nodes, we could be called on an
1592 // instruction in an unreachable block, which may be on a cycle.
1593 SmallPtrSet<Value *, 4> Visited;
1594 Visited.insert(Ptr);
1595 SmallVector<Value *, 4> Indices;
1596
1597 // We may end up computing an offset pointer that has the wrong type. If we
1598 // never are able to compute one directly that has the correct type, we'll
1599 // fall back to it, so keep it and the base it was computed from around here.
1600 Value *OffsetPtr = nullptr;
1601 Value *OffsetBasePtr;
1602
1603 // Remember any i8 pointer we come across to re-use if we need to do a raw
1604 // byte offset.
1605 Value *Int8Ptr = nullptr;
1606 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1607
1608 PointerType *TargetPtrTy = cast<PointerType>(PointerTy);
1609 Type *TargetTy = TargetPtrTy->getElementType();
1610
1611 // As `addrspacecast` is , `Ptr` (the storage pointer) may have different
1612 // address space from the expected `PointerTy` (the pointer to be used).
1613 // Adjust the pointer type based the original storage pointer.
1614 auto AS = cast<PointerType>(Ptr->getType())->getAddressSpace();
1615 PointerTy = TargetTy->getPointerTo(AS);
1616
1617 do {
1618 // First fold any existing GEPs into the offset.
1619 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1620 APInt GEPOffset(Offset.getBitWidth(), 0);
1621 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1622 break;
1623 Offset += GEPOffset;
1624 Ptr = GEP->getPointerOperand();
1625 if (!Visited.insert(Ptr).second)
1626 break;
1627 }
1628
1629 // See if we can perform a natural GEP here.
1630 Indices.clear();
1631 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1632 Indices, NamePrefix)) {
1633 // If we have a new natural pointer at the offset, clear out any old
1634 // offset pointer we computed. Unless it is the base pointer or
1635 // a non-instruction, we built a GEP we don't need. Zap it.
1636 if (OffsetPtr && OffsetPtr != OffsetBasePtr)
1637 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr)) {
1638 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1638, __extension__ __PRETTY_FUNCTION__))
;
1639 I->eraseFromParent();
1640 }
1641 OffsetPtr = P;
1642 OffsetBasePtr = Ptr;
1643 // If we also found a pointer of the right type, we're done.
1644 if (P->getType() == PointerTy)
1645 break;
1646 }
1647
1648 // Stash this pointer if we've found an i8*.
1649 if (Ptr->getType()->isIntegerTy(8)) {
1650 Int8Ptr = Ptr;
1651 Int8PtrOffset = Offset;
1652 }
1653
1654 // Peel off a layer of the pointer and update the offset appropriately.
1655 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1656 Ptr = cast<Operator>(Ptr)->getOperand(0);
1657 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1658 if (GA->isInterposable())
1659 break;
1660 Ptr = GA->getAliasee();
1661 } else {
1662 break;
1663 }
1664 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1664, __extension__ __PRETTY_FUNCTION__))
;
1665 } while (Visited.insert(Ptr).second);
1666
1667 if (!OffsetPtr) {
1668 if (!Int8Ptr) {
1669 Int8Ptr = IRB.CreateBitCast(
1670 Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1671 NamePrefix + "sroa_raw_cast");
1672 Int8PtrOffset = Offset;
1673 }
1674
1675 OffsetPtr = Int8PtrOffset == 0
1676 ? Int8Ptr
1677 : IRB.CreateInBoundsGEP(IRB.getInt8Ty(), Int8Ptr,
1678 IRB.getInt(Int8PtrOffset),
1679 NamePrefix + "sroa_raw_idx");
1680 }
1681 Ptr = OffsetPtr;
1682
1683 // On the off chance we were targeting i8*, guard the bitcast here.
1684 if (cast<PointerType>(Ptr->getType()) != TargetPtrTy) {
1685 Ptr = IRB.CreatePointerBitCastOrAddrSpaceCast(Ptr,
1686 TargetPtrTy,
1687 NamePrefix + "sroa_cast");
1688 }
1689
1690 return Ptr;
1691}
1692
1693/// Compute the adjusted alignment for a load or store from an offset.
1694static Align getAdjustedAlignment(Instruction *I, uint64_t Offset) {
1695 return commonAlignment(getLoadStoreAlignment(I), Offset);
1696}
1697
1698/// Test whether we can convert a value from the old to the new type.
1699///
1700/// This predicate should be used to guard calls to convertValue in order to
1701/// ensure that we only try to convert viable values. The strategy is that we
1702/// will peel off single element struct and array wrappings to get to an
1703/// underlying value, and convert that value.
1704static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1705 if (OldTy == NewTy)
1706 return true;
1707
1708 // For integer types, we can't handle any bit-width differences. This would
1709 // break both vector conversions with extension and introduce endianness
1710 // issues when in conjunction with loads and stores.
1711 if (isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) {
1712 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1714, __extension__ __PRETTY_FUNCTION__))
1713 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1714, __extension__ __PRETTY_FUNCTION__))
1714 "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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1714, __extension__ __PRETTY_FUNCTION__))
;
1715 return false;
1716 }
1717
1718 if (DL.getTypeSizeInBits(NewTy).getFixedSize() !=
1719 DL.getTypeSizeInBits(OldTy).getFixedSize())
1720 return false;
1721 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1722 return false;
1723
1724 // We can convert pointers to integers and vice-versa. Same for vectors
1725 // of pointers and integers.
1726 OldTy = OldTy->getScalarType();
1727 NewTy = NewTy->getScalarType();
1728 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1729 if (NewTy->isPointerTy() && OldTy->isPointerTy()) {
1730 unsigned OldAS = OldTy->getPointerAddressSpace();
1731 unsigned NewAS = NewTy->getPointerAddressSpace();
1732 // Convert pointers if they are pointers from the same address space or
1733 // different integral (not non-integral) address spaces with the same
1734 // pointer size.
1735 return OldAS == NewAS ||
1736 (!DL.isNonIntegralAddressSpace(OldAS) &&
1737 !DL.isNonIntegralAddressSpace(NewAS) &&
1738 DL.getPointerSize(OldAS) == DL.getPointerSize(NewAS));
1739 }
1740
1741 // We can convert integers to integral pointers, but not to non-integral
1742 // pointers.
1743 if (OldTy->isIntegerTy())
1744 return !DL.isNonIntegralPointerType(NewTy);
1745
1746 // We can convert integral pointers to integers, but non-integral pointers
1747 // need to remain pointers.
1748 if (!DL.isNonIntegralPointerType(OldTy))
1749 return NewTy->isIntegerTy();
1750
1751 return false;
1752 }
1753
1754 return true;
1755}
1756
1757/// Generic routine to convert an SSA value to a value of a different
1758/// type.
1759///
1760/// This will try various different casting techniques, such as bitcasts,
1761/// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1762/// two types for viability with this routine.
1763static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1764 Type *NewTy) {
1765 Type *OldTy = V->getType();
1766 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1766, __extension__ __PRETTY_FUNCTION__))
;
1767
1768 if (OldTy == NewTy)
1769 return V;
1770
1771 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.\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1772, __extension__ __PRETTY_FUNCTION__))
1772 "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.\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1772, __extension__ __PRETTY_FUNCTION__))
;
1773
1774 // See if we need inttoptr for this type pair. May require additional bitcast.
1775 if (OldTy->isIntOrIntVectorTy() && NewTy->isPtrOrPtrVectorTy()) {
1776 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1777 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1778 // Expand <4 x i32> to <2 x i8*> --> <4 x i32> to <2 x i64> to <2 x i8*>
1779 // Directly handle i64 to i8*
1780 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1781 NewTy);
1782 }
1783
1784 // See if we need ptrtoint for this type pair. May require additional bitcast.
1785 if (OldTy->isPtrOrPtrVectorTy() && NewTy->isIntOrIntVectorTy()) {
1786 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1787 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1788 // Expand <2 x i8*> to <4 x i32> --> <2 x i8*> to <2 x i64> to <4 x i32>
1789 // Expand i8* to i64 --> i8* to i64 to i64
1790 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1791 NewTy);
1792 }
1793
1794 if (OldTy->isPtrOrPtrVectorTy() && NewTy->isPtrOrPtrVectorTy()) {
1795 unsigned OldAS = OldTy->getPointerAddressSpace();
1796 unsigned NewAS = NewTy->getPointerAddressSpace();
1797 // To convert pointers with different address spaces (they are already
1798 // checked convertible, i.e. they have the same pointer size), so far we
1799 // cannot use `bitcast` (which has restrict on the same address space) or
1800 // `addrspacecast` (which is not always no-op casting). Instead, use a pair
1801 // of no-op `ptrtoint`/`inttoptr` casts through an integer with the same bit
1802 // size.
1803 if (OldAS != NewAS) {
1804 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)"
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1804, __extension__ __PRETTY_FUNCTION__))
;
1805 return IRB.CreateIntToPtr(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1806 NewTy);
1807 }
1808 }
1809
1810 return IRB.CreateBitCast(V, NewTy);
1811}
1812
1813/// Test whether the given slice use can be promoted to a vector.
1814///
1815/// This function is called to test each entry in a partition which is slated
1816/// for a single slice.
1817static bool isVectorPromotionViableForSlice(Partition &P, const Slice &S,
1818 VectorType *Ty,
1819 uint64_t ElementSize,
1820 const DataLayout &DL) {
1821 // First validate the slice offsets.
1822 uint64_t BeginOffset =
1823 std::max(S.beginOffset(), P.beginOffset()) - P.beginOffset();
1824 uint64_t BeginIndex = BeginOffset / ElementSize;
1825 if (BeginIndex * ElementSize != BeginOffset ||
1826 BeginIndex >= cast<FixedVectorType>(Ty)->getNumElements())
1827 return false;
1828 uint64_t EndOffset =
1829 std::min(S.endOffset(), P.endOffset()) - P.beginOffset();
1830 uint64_t EndIndex = EndOffset / ElementSize;
1831 if (EndIndex * ElementSize != EndOffset ||
1832 EndIndex > cast<FixedVectorType>(Ty)->getNumElements())
1833 return false;
1834
1835 assert(EndIndex > BeginIndex && "Empty vector!")(static_cast <bool> (EndIndex > BeginIndex &&
"Empty vector!") ? void (0) : __assert_fail ("EndIndex > BeginIndex && \"Empty vector!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1835, __extension__ __PRETTY_FUNCTION__))
;
1836 uint64_t NumElements = EndIndex - BeginIndex;
1837 Type *SliceTy = (NumElements == 1)
1838 ? Ty->getElementType()
1839 : FixedVectorType::get(Ty->getElementType(), NumElements);
1840
1841 Type *SplitIntTy =
1842 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1843
1844 Use *U = S.getUse();
1845
1846 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1847 if (MI->isVolatile())
1848 return false;
1849 if (!S.isSplittable())
1850 return false; // Skip any unsplittable intrinsics.
1851 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1852 if (!II->isLifetimeStartOrEnd() && !II->isDroppable())
1853 return false;
1854 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1855 // Disable vector promotion when there are loads or stores of an FCA.
1856 return false;
1857 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1858 if (LI->isVolatile())
1859 return false;
1860 Type *LTy = LI->getType();
1861 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1862 assert(LTy->isIntegerTy())(static_cast <bool> (LTy->isIntegerTy()) ? void (0) :
__assert_fail ("LTy->isIntegerTy()", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1862, __extension__ __PRETTY_FUNCTION__))
;
1863 LTy = SplitIntTy;
1864 }
1865 if (!canConvertValue(DL, SliceTy, LTy))
1866 return false;
1867 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1868 if (SI->isVolatile())
1869 return false;
1870 Type *STy = SI->getValueOperand()->getType();
1871 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1872 assert(STy->isIntegerTy())(static_cast <bool> (STy->isIntegerTy()) ? void (0) :
__assert_fail ("STy->isIntegerTy()", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1872, __extension__ __PRETTY_FUNCTION__))
;
1873 STy = SplitIntTy;
1874 }
1875 if (!canConvertValue(DL, STy, SliceTy))
1876 return false;
1877 } else {
1878 return false;
1879 }
1880
1881 return true;
1882}
1883
1884/// Test whether the given alloca partitioning and range of slices can be
1885/// promoted to a vector.
1886///
1887/// This is a quick test to check whether we can rewrite a particular alloca
1888/// partition (and its newly formed alloca) into a vector alloca with only
1889/// whole-vector loads and stores such that it could be promoted to a vector
1890/// SSA value. We only can ensure this for a limited set of operations, and we
1891/// don't want to do the rewrites unless we are confident that the result will
1892/// be promotable, so we have an early test here.
1893static VectorType *isVectorPromotionViable(Partition &P, const DataLayout &DL) {
1894 // Collect the candidate types for vector-based promotion. Also track whether
1895 // we have different element types.
1896 SmallVector<VectorType *, 4> CandidateTys;
1897 Type *CommonEltTy = nullptr;
1898 bool HaveCommonEltTy = true;
1899 auto CheckCandidateType = [&](Type *Ty) {
1900 if (auto *VTy = dyn_cast<VectorType>(Ty)) {
1901 // Return if bitcast to vectors is different for total size in bits.
1902 if (!CandidateTys.empty()) {
1903 VectorType *V = CandidateTys[0];
1904 if (DL.getTypeSizeInBits(VTy).getFixedSize() !=
1905 DL.getTypeSizeInBits(V).getFixedSize()) {
1906 CandidateTys.clear();
1907 return;
1908 }
1909 }
1910 CandidateTys.push_back(VTy);
1911 if (!CommonEltTy)
1912 CommonEltTy = VTy->getElementType();
1913 else if (CommonEltTy != VTy->getElementType())
1914 HaveCommonEltTy = false;
1915 }
1916 };
1917 // Consider any loads or stores that are the exact size of the slice.
1918 for (const Slice &S : P)
1919 if (S.beginOffset() == P.beginOffset() &&
1920 S.endOffset() == P.endOffset()) {
1921 if (auto *LI = dyn_cast<LoadInst>(S.getUse()->getUser()))
1922 CheckCandidateType(LI->getType());
1923 else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser()))
1924 CheckCandidateType(SI->getValueOperand()->getType());
1925 }
1926
1927 // If we didn't find a vector type, nothing to do here.
1928 if (CandidateTys.empty())
1929 return nullptr;
1930
1931 // Remove non-integer vector types if we had multiple common element types.
1932 // FIXME: It'd be nice to replace them with integer vector types, but we can't
1933 // do that until all the backends are known to produce good code for all
1934 // integer vector types.
1935 if (!HaveCommonEltTy) {
1936 llvm::erase_if(CandidateTys, [](VectorType *VTy) {
1937 return !VTy->getElementType()->isIntegerTy();
1938 });
1939
1940 // If there were no integer vector types, give up.
1941 if (CandidateTys.empty())
1942 return nullptr;
1943
1944 // Rank the remaining candidate vector types. This is easy because we know
1945 // they're all integer vectors. We sort by ascending number of elements.
1946 auto RankVectorTypes = [&DL](VectorType *RHSTy, VectorType *LHSTy) {
1947 (void)DL;
1948 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1950, __extension__ __PRETTY_FUNCTION__))
1949 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1950, __extension__ __PRETTY_FUNCTION__))
1950 "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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1950, __extension__ __PRETTY_FUNCTION__))
;
1951 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1952, __extension__ __PRETTY_FUNCTION__))
1952 "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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1952, __extension__ __PRETTY_FUNCTION__))
;
1953 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1954, __extension__ __PRETTY_FUNCTION__))
1954 "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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1954, __extension__ __PRETTY_FUNCTION__))
;
1955 return cast<FixedVectorType>(RHSTy)->getNumElements() <
1956 cast<FixedVectorType>(LHSTy)->getNumElements();
1957 };
1958 llvm::sort(CandidateTys, RankVectorTypes);
1959 CandidateTys.erase(
1960 std::unique(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes),
1961 CandidateTys.end());
1962 } else {
1963// The only way to have the same element type in every vector type is to
1964// have the same vector type. Check that and remove all but one.
1965#ifndef NDEBUG
1966 for (VectorType *VTy : CandidateTys) {
1967 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1968, __extension__ __PRETTY_FUNCTION__))
1968 "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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1968, __extension__ __PRETTY_FUNCTION__))
;
1969 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1970, __extension__ __PRETTY_FUNCTION__))
1970 "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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1970, __extension__ __PRETTY_FUNCTION__))
;
1971 }
1972#endif
1973 CandidateTys.resize(1);
1974 }
1975
1976 // Try each vector type, and return the one which works.
1977 auto CheckVectorTypeForPromotion = [&](VectorType *VTy) {
1978 uint64_t ElementSize =
1979 DL.getTypeSizeInBits(VTy->getElementType()).getFixedSize();
1980
1981 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1982 // that aren't byte sized.
1983 if (ElementSize % 8)
1984 return false;
1985 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?\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1986, __extension__ __PRETTY_FUNCTION__))
1986 "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?\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1986, __extension__ __PRETTY_FUNCTION__))
;
1987 ElementSize /= 8;
1988
1989 for (const Slice &S : P)
1990 if (!isVectorPromotionViableForSlice(P, S, VTy, ElementSize, DL))
1991 return false;
1992
1993 for (const Slice *S : P.splitSliceTails())
1994 if (!isVectorPromotionViableForSlice(P, *S, VTy, ElementSize, DL))
1995 return false;
1996
1997 return true;
1998 };
1999 for (VectorType *VTy : CandidateTys)
2000 if (CheckVectorTypeForPromotion(VTy))
2001 return VTy;
2002
2003 return nullptr;
2004}
2005
2006/// Test whether a slice of an alloca is valid for integer widening.
2007///
2008/// This implements the necessary checking for the \c isIntegerWideningViable
2009/// test below on a single slice of the alloca.
2010static bool isIntegerWideningViableForSlice(const Slice &S,
2011 uint64_t AllocBeginOffset,
2012 Type *AllocaTy,
2013 const DataLayout &DL,
2014 bool &WholeAllocaOp) {
2015 uint64_t Size = DL.getTypeStoreSize(AllocaTy).getFixedSize();
2016
2017 uint64_t RelBegin = S.beginOffset() - AllocBeginOffset;
2018 uint64_t RelEnd = S.endOffset() - AllocBeginOffset;
2019
2020 // We can't reasonably handle cases where the load or store extends past
2021 // the end of the alloca's type and into its padding.
2022 if (RelEnd > Size)
2023 return false;
2024
2025 Use *U = S.getUse();
2026
2027 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
2028 if (LI->isVolatile())
2029 return false;
2030 // We can't handle loads that extend past the allocated memory.
2031 if (DL.getTypeStoreSize(LI->getType()).getFixedSize() > Size)
2032 return false;
2033 // So far, AllocaSliceRewriter does not support widening split slice tails
2034 // in rewriteIntegerLoad.
2035 if (S.beginOffset() < AllocBeginOffset)
2036 return false;
2037 // Note that we don't count vector loads or stores as whole-alloca
2038 // operations which enable integer widening because we would prefer to use
2039 // vector widening instead.
2040 if (!isa<VectorType>(LI->getType()) && RelBegin == 0 && RelEnd == Size)
2041 WholeAllocaOp = true;
2042 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
2043 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy).getFixedSize())
2044 return false;
2045 } else if (RelBegin != 0 || RelEnd != Size ||
2046 !canConvertValue(DL, AllocaTy, LI->getType())) {
2047 // Non-integer loads need to be convertible from the alloca type so that
2048 // they are promotable.
2049 return false;
2050 }
2051 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
2052 Type *ValueTy = SI->getValueOperand()->getType();
2053 if (SI->isVolatile())
2054 return false;
2055 // We can't handle stores that extend past the allocated memory.
2056 if (DL.getTypeStoreSize(ValueTy).getFixedSize() > Size)
2057 return false;
2058 // So far, AllocaSliceRewriter does not support widening split slice tails
2059 // in rewriteIntegerStore.
2060 if (S.beginOffset() < AllocBeginOffset)
2061 return false;
2062 // Note that we don't count vector loads or stores as whole-alloca
2063 // operations which enable integer widening because we would prefer to use
2064 // vector widening instead.
2065 if (!isa<VectorType>(ValueTy) && RelBegin == 0 && RelEnd == Size)
2066 WholeAllocaOp = true;
2067 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
2068 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy).getFixedSize())
2069 return false;
2070 } else if (RelBegin != 0 || RelEnd != Size ||
2071 !canConvertValue(DL, ValueTy, AllocaTy)) {
2072 // Non-integer stores need to be convertible to the alloca type so that
2073 // they are promotable.
2074 return false;
2075 }
2076 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
2077 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
2078 return false;
2079 if (!S.isSplittable())
2080 return false; // Skip any unsplittable intrinsics.
2081 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
2082 if (!II->isLifetimeStartOrEnd() && !II->isDroppable())
2083 return false;
2084 } else {
2085 return false;
2086 }
2087
2088 return true;
2089}
2090
2091/// Test whether the given alloca partition's integer operations can be
2092/// widened to promotable ones.
2093///
2094/// This is a quick test to check whether we can rewrite the integer loads and
2095/// stores to a particular alloca into wider loads and stores and be able to
2096/// promote the resulting alloca.
2097static bool isIntegerWideningViable(Partition &P, Type *AllocaTy,
2098 const DataLayout &DL) {
2099 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy).getFixedSize();
2100 // Don't create integer types larger than the maximum bitwidth.
2101 if (SizeInBits > IntegerType::MAX_INT_BITS)
2102 return false;
2103
2104 // Don't try to handle allocas with bit-padding.
2105 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy).getFixedSize())
2106 return false;
2107
2108 // We need to ensure that an integer type with the appropriate bitwidth can
2109 // be converted to the alloca type, whatever that is. We don't want to force
2110 // the alloca itself to have an integer type if there is a more suitable one.
2111 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
2112 if (!canConvertValue(DL, AllocaTy, IntTy) ||
2113 !canConvertValue(DL, IntTy, AllocaTy))
2114 return false;
2115
2116 // While examining uses, we ensure that the alloca has a covering load or
2117 // store. We don't want to widen the integer operations only to fail to
2118 // promote due to some other unsplittable entry (which we may make splittable
2119 // later). However, if there are only splittable uses, go ahead and assume
2120 // that we cover the alloca.
2121 // FIXME: We shouldn't consider split slices that happen to start in the
2122 // partition here...
2123 bool WholeAllocaOp = P.empty() && DL.isLegalInteger(SizeInBits);
2124
2125 for (const Slice &S : P)
2126 if (!isIntegerWideningViableForSlice(S, P.beginOffset(), AllocaTy, DL,
2127 WholeAllocaOp))
2128 return false;
2129
2130 for (const Slice *S : P.splitSliceTails())
2131 if (!isIntegerWideningViableForSlice(*S, P.beginOffset(), AllocaTy, DL,
2132 WholeAllocaOp))
2133 return false;
2134
2135 return WholeAllocaOp;
2136}
2137
2138static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
2139 IntegerType *Ty, uint64_t Offset,
2140 const Twine &Name) {
2141 LLVM_DEBUG(dbgs() << " start: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " start: " << *V <<
"\n"; } } while (false)
;
2142 IntegerType *IntTy = cast<IntegerType>(V->getType());
2143 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2145, __extension__ __PRETTY_FUNCTION__))
2144 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2145, __extension__ __PRETTY_FUNCTION__))
2145 "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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2145, __extension__ __PRETTY_FUNCTION__))
;
2146 uint64_t ShAmt = 8 * Offset;
2147 if (DL.isBigEndian())
2148 ShAmt = 8 * (DL.getTypeStoreSize(IntTy).getFixedSize() -
2149 DL.getTypeStoreSize(Ty).getFixedSize() - Offset);
2150 if (ShAmt) {
2151 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2152 LLVM_DEBUG(dbgs() << " shifted: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " shifted: " << *V <<
"\n"; } } while (false)
;
2153 }
2154 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2155, __extension__ __PRETTY_FUNCTION__))
2155 "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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2155, __extension__ __PRETTY_FUNCTION__))
;
2156 if (Ty != IntTy) {
2157 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2158 LLVM_DEBUG(dbgs() << " trunced: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " trunced: " << *V <<
"\n"; } } while (false)
;
2159 }
2160 return V;
2161}
2162
2163static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
2164 Value *V, uint64_t Offset, const Twine &Name) {
2165 IntegerType *IntTy = cast<IntegerType>(Old->getType());
2166 IntegerType *Ty = cast<IntegerType>(V->getType());
2167 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2168, __extension__ __PRETTY_FUNCTION__))
2168 "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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2168, __extension__ __PRETTY_FUNCTION__))
;
2169 LLVM_DEBUG(dbgs() << " start: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " start: " << *V <<
"\n"; } } while (false)
;
2170 if (Ty != IntTy) {
2171 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2172 LLVM_DEBUG(dbgs() << " extended: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " extended: " << *V <<
"\n"; } } while (false)
;
2173 }
2174 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2176, __extension__ __PRETTY_FUNCTION__))
2175 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2176, __extension__ __PRETTY_FUNCTION__))
2176 "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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2176, __extension__ __PRETTY_FUNCTION__))
;
2177 uint64_t ShAmt = 8 * Offset;
2178 if (DL.isBigEndian())
2179 ShAmt = 8 * (DL.getTypeStoreSize(IntTy).getFixedSize() -
2180 DL.getTypeStoreSize(Ty).getFixedSize() - Offset);
2181 if (ShAmt) {
2182 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2183 LLVM_DEBUG(dbgs() << " shifted: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " shifted: " << *V <<
"\n"; } } while (false)
;
2184 }
2185
2186 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2187 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2188 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2189 LLVM_DEBUG(dbgs() << " masked: " << *Old << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " masked: " << *Old <<
"\n"; } } while (false)
;
2190 V = IRB.CreateOr(Old, V, Name + ".insert");
2191 LLVM_DEBUG(dbgs() << " inserted: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " inserted: " << *V <<
"\n"; } } while (false)
;
2192 }
2193 return V;
2194}
2195
2196static Value *extractVector(IRBuilderTy &IRB, Value *V, unsigned BeginIndex,
2197 unsigned EndIndex, const Twine &Name) {
2198 auto *VecTy = cast<FixedVectorType>(V->getType());
2199 unsigned NumElements = EndIndex - BeginIndex;
2200 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2200, __extension__ __PRETTY_FUNCTION__))
;
2201
2202 if (NumElements == VecTy->getNumElements())
2203 return V;
2204
2205 if (NumElements == 1) {
2206 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2207 Name + ".extract");
2208 LLVM_DEBUG(dbgs() << " extract: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " extract: " << *V <<
"\n"; } } while (false)
;
2209 return V;
2210 }
2211
2212 SmallVector<int, 8> Mask;
2213 Mask.reserve(NumElements);
2214 for (unsigned i = BeginIndex; i != EndIndex; ++i)
2215 Mask.push_back(i);
2216 V = IRB.CreateShuffleVector(V, Mask, Name + ".extract");
2217 LLVM_DEBUG(dbgs() << " shuffle: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " shuffle: " << *V <<
"\n"; } } while (false)
;
2218 return V;
2219}
2220
2221static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
2222 unsigned BeginIndex, const Twine &Name) {
2223 VectorType *VecTy = cast<VectorType>(Old->getType());
2224 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2224, __extension__ __PRETTY_FUNCTION__))
;
2225
2226 VectorType *Ty = dyn_cast<VectorType>(V->getType());
2227 if (!Ty) {
2228 // Single element to insert.
2229 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2230 Name + ".insert");
2231 LLVM_DEBUG(dbgs() << " insert: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " insert: " << *V <<
"\n"; } } while (false)
;
2232 return V;
2233 }
2234
2235 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2237, __extension__ __PRETTY_FUNCTION__))
2236 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2237, __extension__ __PRETTY_FUNCTION__))
2237 "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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2237, __extension__ __PRETTY_FUNCTION__))
;
2238 if (cast<FixedVectorType>(Ty)->getNumElements() ==
2239 cast<FixedVectorType>(VecTy)->getNumElements()) {
2240 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2240, __extension__ __PRETTY_FUNCTION__))
;
2241 return V;
2242 }
2243 unsigned EndIndex = BeginIndex + cast<FixedVectorType>(Ty)->getNumElements();
2244
2245 // When inserting a smaller vector into the larger to store, we first
2246 // use a shuffle vector to widen it with undef elements, and then
2247 // a second shuffle vector to select between the loaded vector and the
2248 // incoming vector.
2249 SmallVector<int, 8> Mask;
2250 Mask.reserve(cast<FixedVectorType>(VecTy)->getNumElements());
2251 for (unsigned i = 0; i != cast<FixedVectorType>(VecTy)->getNumElements(); ++i)
2252 if (i >= BeginIndex && i < EndIndex)
2253 Mask.push_back(i - BeginIndex);
2254 else
2255 Mask.push_back(-1);
2256 V = IRB.CreateShuffleVector(V, Mask, Name + ".expand");
2257 LLVM_DEBUG(dbgs() << " shuffle: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " shuffle: " << *V <<
"\n"; } } while (false)
;
2258
2259 SmallVector<Constant *, 8> Mask2;
2260 Mask2.reserve(cast<FixedVectorType>(VecTy)->getNumElements());
2261 for (unsigned i = 0; i != cast<FixedVectorType>(VecTy)->getNumElements(); ++i)
2262 Mask2.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
2263
2264 V = IRB.CreateSelect(ConstantVector::get(Mask2), V, Old, Name + "blend");
2265
2266 LLVM_DEBUG(dbgs() << " blend: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " blend: " << *V <<
"\n"; } } while (false)
;
2267 return V;
2268}
2269
2270/// Visitor to rewrite instructions using p particular slice of an alloca
2271/// to use a new alloca.
2272///
2273/// Also implements the rewriting to vector-based accesses when the partition
2274/// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2275/// lives here.
2276class llvm::sroa::AllocaSliceRewriter
2277 : public InstVisitor<AllocaSliceRewriter, bool> {
2278 // Befriend the base class so it can delegate to private visit methods.
2279 friend class InstVisitor<AllocaSliceRewriter, bool>;
2280
2281 using Base = InstVisitor<AllocaSliceRewriter, bool>;
2282
2283 const DataLayout &DL;
2284 AllocaSlices &AS;
2285 SROA &Pass;
2286 AllocaInst &OldAI, &NewAI;
2287 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2288 Type *NewAllocaTy;
2289
2290 // This is a convenience and flag variable that will be null unless the new
2291 // alloca's integer operations should be widened to this integer type due to
2292 // passing isIntegerWideningViable above. If it is non-null, the desired
2293 // integer type will be stored here for easy access during rewriting.
2294 IntegerType *IntTy;
2295
2296 // If we are rewriting an alloca partition which can be written as pure
2297 // vector operations, we stash extra information here. When VecTy is
2298 // non-null, we have some strict guarantees about the rewritten alloca:
2299 // - The new alloca is exactly the size of the vector type here.
2300 // - The accesses all either map to the entire vector or to a single
2301 // element.
2302 // - The set of accessing instructions is only one of those handled above
2303 // in isVectorPromotionViable. Generally these are the same access kinds
2304 // which are promotable via mem2reg.
2305 VectorType *VecTy;
2306 Type *ElementTy;
2307 uint64_t ElementSize;
2308
2309 // The original offset of the slice currently being rewritten relative to
2310 // the original alloca.
2311 uint64_t BeginOffset = 0;
2312 uint64_t EndOffset = 0;
2313
2314 // The new offsets of the slice currently being rewritten relative to the
2315 // original alloca.
2316 uint64_t NewBeginOffset = 0, NewEndOffset = 0;
2317
2318 uint64_t SliceSize = 0;
2319 bool IsSplittable = false;
2320 bool IsSplit = false;
2321 Use *OldUse = nullptr;
2322 Instruction *OldPtr = nullptr;
2323
2324 // Track post-rewrite users which are PHI nodes and Selects.
2325 SmallSetVector<PHINode *, 8> &PHIUsers;
2326 SmallSetVector<SelectInst *, 8> &SelectUsers;
2327
2328 // Utility IR builder, whose name prefix is setup for each visited use, and
2329 // the insertion point is set to point to the user.
2330 IRBuilderTy IRB;
2331
2332public:
2333 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &AS, SROA &Pass,
2334 AllocaInst &OldAI, AllocaInst &NewAI,
2335 uint64_t NewAllocaBeginOffset,
2336 uint64_t NewAllocaEndOffset, bool IsIntegerPromotable,
2337 VectorType *PromotableVecTy,
2338 SmallSetVector<PHINode *, 8> &PHIUsers,
2339 SmallSetVector<SelectInst *, 8> &SelectUsers)
2340 : DL(DL), AS(AS), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2341 NewAllocaBeginOffset(NewAllocaBeginOffset),
2342 NewAllocaEndOffset(NewAllocaEndOffset),
2343 NewAllocaTy(NewAI.getAllocatedType()),
2344 IntTy(
2345 IsIntegerPromotable
2346 ? Type::getIntNTy(NewAI.getContext(),
2347 DL.getTypeSizeInBits(NewAI.getAllocatedType())
2348 .getFixedSize())
2349 : nullptr),
2350 VecTy(PromotableVecTy),
2351 ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2352 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy).getFixedSize() / 8
2353 : 0),
2354 PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2355 IRB(NewAI.getContext(), ConstantFolder()) {
2356 if (VecTy) {
2357 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2358, __extension__ __PRETTY_FUNCTION__))
2358 "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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2358, __extension__ __PRETTY_FUNCTION__))
;
2359 ++NumVectorized;
2360 }
2361 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)"
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2361, __extension__ __PRETTY_FUNCTION__))
;
2362 }
2363
2364 bool visit(AllocaSlices::const_iterator I) {
2365 bool CanSROA = true;
2366 BeginOffset = I->beginOffset();
2367 EndOffset = I->endOffset();
2368 IsSplittable = I->isSplittable();
2369 IsSplit =
2370 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2371 LLVM_DEBUG(dbgs() << " rewriting " << (IsSplit ? "split " : ""))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " rewriting " << (IsSplit ?
"split " : ""); } } while (false)
;
2372 LLVM_DEBUG(AS.printSlice(dbgs(), I, ""))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { AS.printSlice(dbgs(), I, ""); } } while (false)
;
2373 LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "\n"; } } while (false)
;
2374
2375 // Compute the intersecting offset range.
2376 assert(BeginOffset < NewAllocaEndOffset)(static_cast <bool> (BeginOffset < NewAllocaEndOffset
) ? void (0) : __assert_fail ("BeginOffset < NewAllocaEndOffset"
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2376, __extension__ __PRETTY_FUNCTION__))
;
2377 assert(EndOffset > NewAllocaBeginOffset)(static_cast <bool> (EndOffset > NewAllocaBeginOffset
) ? void (0) : __assert_fail ("EndOffset > NewAllocaBeginOffset"
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2377, __extension__ __PRETTY_FUNCTION__))
;
2378 NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2379 NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2380
2381 SliceSize = NewEndOffset - NewBeginOffset;
2382
2383 OldUse = I->getUse();
2384 OldPtr = cast<Instruction>(OldUse->get());
2385
2386 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2387 IRB.SetInsertPoint(OldUserI);
2388 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2389 IRB.getInserter().SetNamePrefix(
2390 Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2391
2392 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2393 if (VecTy || IntTy)
2394 assert(CanSROA)(static_cast <bool> (CanSROA) ? void (0) : __assert_fail
("CanSROA", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2394, __extension__ __PRETTY_FUNCTION__))
;
2395 return CanSROA;
2396 }
2397
2398private:
2399 // Make sure the other visit overloads are visible.
2400 using Base::visit;
2401
2402 // Every instruction which can end up as a user must have a rewrite rule.
2403 bool visitInstruction(Instruction &I) {
2404 LLVM_DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " !!!! Cannot rewrite: " <<
I << "\n"; } } while (false)
;
2405 llvm_unreachable("No rewrite rule for this instruction!")::llvm::llvm_unreachable_internal("No rewrite rule for this instruction!"
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2405)
;
2406 }
2407
2408 Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2409 // Note that the offset computation can use BeginOffset or NewBeginOffset
2410 // interchangeably for unsplit slices.
2411 assert(IsSplit || BeginOffset == NewBeginOffset)(static_cast <bool> (IsSplit || BeginOffset == NewBeginOffset
) ? void (0) : __assert_fail ("IsSplit || BeginOffset == NewBeginOffset"
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2411, __extension__ __PRETTY_FUNCTION__))
;
2412 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2413
2414#ifndef NDEBUG
2415 StringRef OldName = OldPtr->getName();
2416 // Skip through the last '.sroa.' component of the name.
2417 size_t LastSROAPrefix = OldName.rfind(".sroa.");
2418 if (LastSROAPrefix != StringRef::npos) {
2419 OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2420 // Look for an SROA slice index.
2421 size_t IndexEnd = OldName.find_first_not_of("0123456789");
2422 if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2423 // Strip the index and look for the offset.
2424 OldName = OldName.substr(IndexEnd + 1);
2425 size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2426 if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2427 // Strip the offset.
2428 OldName = OldName.substr(OffsetEnd + 1);
2429 }
2430 }
2431 // Strip any SROA suffixes as well.
2432 OldName = OldName.substr(0, OldName.find(".sroa_"));
2433#endif
2434
2435 return getAdjustedPtr(IRB, DL, &NewAI,
2436 APInt(DL.getIndexTypeSizeInBits(PointerTy), Offset),
2437 PointerTy,
2438#ifndef NDEBUG
2439 Twine(OldName) + "."
2440#else
2441 Twine()
2442#endif
2443 );
2444 }
2445
2446 /// Compute suitable alignment to access this slice of the *new*
2447 /// alloca.
2448 ///
2449 /// You can optionally pass a type to this routine and if that type's ABI
2450 /// alignment is itself suitable, this will return zero.
2451 Align getSliceAlign() {
2452 return commonAlignment(NewAI.getAlign(),
2453 NewBeginOffset - NewAllocaBeginOffset);
2454 }
2455
2456 unsigned getIndex(uint64_t Offset) {
2457 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2457, __extension__ __PRETTY_FUNCTION__))
;
2458 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2459 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2459, __extension__ __PRETTY_FUNCTION__))
;
2460 uint32_t Index = RelOffset / ElementSize;
2461 assert(Index * ElementSize == RelOffset)(static_cast <bool> (Index * ElementSize == RelOffset) ?
void (0) : __assert_fail ("Index * ElementSize == RelOffset"
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2461, __extension__ __PRETTY_FUNCTION__))
;
2462 return Index;
2463 }
2464
2465 void deleteIfTriviallyDead(Value *V) {
2466 Instruction *I = cast<Instruction>(V);
2467 if (isInstructionTriviallyDead(I))
2468 Pass.DeadInsts.push_back(I);
2469 }
2470
2471 Value *rewriteVectorizedLoadInst(LoadInst &LI) {
2472 unsigned BeginIndex = getIndex(NewBeginOffset);
2473 unsigned EndIndex = getIndex(NewEndOffset);
2474 assert(EndIndex > BeginIndex && "Empty vector!")(static_cast <bool> (EndIndex > BeginIndex &&
"Empty vector!") ? void (0) : __assert_fail ("EndIndex > BeginIndex && \"Empty vector!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2474, __extension__ __PRETTY_FUNCTION__))
;
2475
2476 LoadInst *Load = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2477 NewAI.getAlign(), "load");
2478
2479 Load->copyMetadata(LI, {LLVMContext::MD_mem_parallel_loop_access,
2480 LLVMContext::MD_access_group});
2481 return extractVector(IRB, Load, BeginIndex, EndIndex, "vec");
2482 }
2483
2484 Value *rewriteIntegerLoad(LoadInst &LI) {
2485 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2485, __extension__ __PRETTY_FUNCTION__))
;
2486 assert(!LI.isVolatile())(static_cast <bool> (!LI.isVolatile()) ? void (0) : __assert_fail
("!LI.isVolatile()", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2486, __extension__ __PRETTY_FUNCTION__))
;
2487 Value *V = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2488 NewAI.getAlign(), "load");
2489 V = convertValue(DL, IRB, V, IntTy);
2490 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2490, __extension__ __PRETTY_FUNCTION__))
;
2491 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2492 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset) {
2493 IntegerType *ExtractTy = Type::getIntNTy(LI.getContext(), SliceSize * 8);
2494 V = extractInteger(DL, IRB, V, ExtractTy, Offset, "extract");
2495 }
2496 // It is possible that the extracted type is not the load type. This
2497 // happens if there is a load past the end of the alloca, and as
2498 // a consequence the slice is narrower but still a candidate for integer
2499 // lowering. To handle this case, we just zero extend the extracted
2500 // integer.
2501 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2502, __extension__ __PRETTY_FUNCTION__))
2502 "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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2502, __extension__ __PRETTY_FUNCTION__))
;
2503 if (cast<IntegerType>(LI.getType())->getBitWidth() > SliceSize * 8)
2504 V = IRB.CreateZExt(V, LI.getType());
2505 return V;
2506 }
2507
2508 bool visitLoadInst(LoadInst &LI) {
2509 LLVM_DEBUG(dbgs() << " original: " << LI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << LI <<
"\n"; } } while (false)
;
2510 Value *OldOp = LI.getOperand(0);
2511 assert(OldOp == OldPtr)(static_cast <bool> (OldOp == OldPtr) ? void (0) : __assert_fail
("OldOp == OldPtr", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2511, __extension__ __PRETTY_FUNCTION__))
;
2512
2513 AAMDNodes AATags;
2514 LI.getAAMetadata(AATags);
2515
2516 unsigned AS = LI.getPointerAddressSpace();
2517
2518 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2519 : LI.getType();
2520 const bool IsLoadPastEnd =
2521 DL.getTypeStoreSize(TargetTy).getFixedSize() > SliceSize;
2522 bool IsPtrAdjusted = false;
2523 Value *V;
2524 if (VecTy) {
2525 V = rewriteVectorizedLoadInst(LI);
2526 } else if (IntTy && LI.getType()->isIntegerTy()) {
2527 V = rewriteIntegerLoad(LI);
2528 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2529 NewEndOffset == NewAllocaEndOffset &&
2530 (canConvertValue(DL, NewAllocaTy, TargetTy) ||
2531 (IsLoadPastEnd && NewAllocaTy->isIntegerTy() &&
2532 TargetTy->isIntegerTy()))) {
2533 LoadInst *NewLI = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2534 NewAI.getAlign(), LI.isVolatile(),
2535 LI.getName());
2536 if (AATags)
2537 NewLI->setAAMetadata(AATags.shift(NewBeginOffset - BeginOffset));
2538 if (LI.isVolatile())
2539 NewLI->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
2540 if (NewLI->isAtomic())
2541 NewLI->setAlignment(LI.getAlign());
2542
2543 // Any !nonnull metadata or !range metadata on the old load is also valid
2544 // on the new load. This is even true in some cases even when the loads
2545 // are different types, for example by mapping !nonnull metadata to
2546 // !range metadata by modeling the null pointer constant converted to the
2547 // integer type.
2548 // FIXME: Add support for range metadata here. Currently the utilities
2549 // for this don't propagate range metadata in trivial cases from one
2550 // integer load to another, don't handle non-addrspace-0 null pointers
2551 // correctly, and don't have any support for mapping ranges as the
2552 // integer type becomes winder or narrower.
2553 if (MDNode *N = LI.getMetadata(LLVMContext::MD_nonnull))
2554 copyNonnullMetadata(LI, N, *NewLI);
2555
2556 // Try to preserve nonnull metadata
2557 V = NewLI;
2558
2559 // If this is an integer load past the end of the slice (which means the
2560 // bytes outside the slice are undef or this load is dead) just forcibly
2561 // fix the integer size with correct handling of endianness.
2562 if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2563 if (auto *TITy = dyn_cast<IntegerType>(TargetTy))
2564 if (AITy->getBitWidth() < TITy->getBitWidth()) {
2565 V = IRB.CreateZExt(V, TITy, "load.ext");
2566 if (DL.isBigEndian())
2567 V = IRB.CreateShl(V, TITy->getBitWidth() - AITy->getBitWidth(),
2568 "endian_shift");
2569 }
2570 } else {
2571 Type *LTy = TargetTy->getPointerTo(AS);
2572 LoadInst *NewLI =
2573 IRB.CreateAlignedLoad(TargetTy, getNewAllocaSlicePtr(IRB, LTy),
2574 getSliceAlign(), LI.isVolatile(), LI.getName());
2575 if (AATags)
2576 NewLI->setAAMetadata(AATags.shift(NewBeginOffset - BeginOffset));
2577 if (LI.isVolatile())
2578 NewLI->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
2579 NewLI->copyMetadata(LI, {LLVMContext::MD_mem_parallel_loop_access,
2580 LLVMContext::MD_access_group});
2581
2582 V = NewLI;
2583 IsPtrAdjusted = true;
2584 }
2585 V = convertValue(DL, IRB, V, TargetTy);
2586
2587 if (IsSplit) {
2588 assert(!LI.isVolatile())(static_cast <bool> (!LI.isVolatile()) ? void (0) : __assert_fail
("!LI.isVolatile()", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2588, __extension__ __PRETTY_FUNCTION__))
;
2589 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2590, __extension__ __PRETTY_FUNCTION__))
2590 "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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2590, __extension__ __PRETTY_FUNCTION__))
;
2591 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2592, __extension__ __PRETTY_FUNCTION__))
2592 "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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2592, __extension__ __PRETTY_FUNCTION__))
;
2593 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2594, __extension__ __PRETTY_FUNCTION__))
2594 "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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2594, __extension__ __PRETTY_FUNCTION__))
;
2595 // Move the insertion point just past the load so that we can refer to it.
2596 IRB.SetInsertPoint(&*std::next(BasicBlock::iterator(&LI)));
2597 // Create a placeholder value with the same type as LI to use as the
2598 // basis for the new value. This allows us to replace the uses of LI with
2599 // the computed value, and then replace the placeholder with LI, leaving
2600 // LI only used for this computation.
2601 Value *Placeholder = new LoadInst(
2602 LI.getType(), UndefValue::get(LI.getType()->getPointerTo(AS)), "",
2603 false, Align(1));
2604 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset - BeginOffset,
2605 "insert");
2606 LI.replaceAllUsesWith(V);
2607 Placeholder->replaceAllUsesWith(&LI);
2608 Placeholder->deleteValue();
2609 } else {
2610 LI.replaceAllUsesWith(V);
2611 }
2612
2613 Pass.DeadInsts.push_back(&LI);
2614 deleteIfTriviallyDead(OldOp);
2615 LLVM_DEBUG(dbgs() << " to: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *V <<
"\n"; } } while (false)
;
2616 return !LI.isVolatile() && !IsPtrAdjusted;
2617 }
2618
2619 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp,
2620 AAMDNodes AATags) {
2621 if (V->getType() != VecTy) {
2622 unsigned BeginIndex = getIndex(NewBeginOffset);
2623 unsigned EndIndex = getIndex(NewEndOffset);
2624 assert(EndIndex > BeginIndex && "Empty vector!")(static_cast <bool> (EndIndex > BeginIndex &&
"Empty vector!") ? void (0) : __assert_fail ("EndIndex > BeginIndex && \"Empty vector!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2624, __extension__ __PRETTY_FUNCTION__))
;
2625 unsigned NumElements = EndIndex - BeginIndex;
2626 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2627, __extension__ __PRETTY_FUNCTION__))
2627 "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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2627, __extension__ __PRETTY_FUNCTION__))
;
2628 Type *SliceTy = (NumElements == 1)
2629 ? ElementTy
2630 : FixedVectorType::get(ElementTy, NumElements);
2631 if (V->getType() != SliceTy)
2632 V = convertValue(DL, IRB, V, SliceTy);
2633
2634 // Mix in the existing elements.
2635 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2636 NewAI.getAlign(), "load");
2637 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2638 }
2639 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlign());
2640 Store->copyMetadata(SI, {LLVMContext::MD_mem_parallel_loop_access,
2641 LLVMContext::MD_access_group});
2642 if (AATags)
2643 Store->setAAMetadata(AATags.shift(NewBeginOffset - BeginOffset));
2644 Pass.DeadInsts.push_back(&SI);
2645
2646 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *Store <<
"\n"; } } while (false)
;
2647 return true;
2648 }
2649
2650 bool rewriteIntegerStore(Value *V, StoreInst &SI, AAMDNodes AATags) {
2651 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2651, __extension__ __PRETTY_FUNCTION__))
;
2652 assert(!SI.isVolatile())(static_cast <bool> (!SI.isVolatile()) ? void (0) : __assert_fail
("!SI.isVolatile()", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2652, __extension__ __PRETTY_FUNCTION__))
;
2653 if (DL.getTypeSizeInBits(V->getType()).getFixedSize() !=
2654 IntTy->getBitWidth()) {
2655 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2656 NewAI.getAlign(), "oldload");
2657 Old = convertValue(DL, IRB, Old, IntTy);
2658 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2658, __extension__ __PRETTY_FUNCTION__))
;
2659 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2660 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset, "insert");
2661 }
2662 V = convertValue(DL, IRB, V, NewAllocaTy);
2663 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlign());
2664 Store->copyMetadata(SI, {LLVMContext::MD_mem_parallel_loop_access,
2665 LLVMContext::MD_access_group});
2666 if (AATags)
2667 Store->setAAMetadata(AATags.shift(NewBeginOffset - BeginOffset));
2668 Pass.DeadInsts.push_back(&SI);
2669 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *Store <<
"\n"; } } while (false)
;
2670 return true;
2671 }
2672
2673 bool visitStoreInst(StoreInst &SI) {
2674 LLVM_DEBUG(dbgs() << " original: " << SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << SI <<
"\n"; } } while (false)
;
2675 Value *OldOp = SI.getOperand(1);
2676 assert(OldOp == OldPtr)(static_cast <bool> (OldOp == OldPtr) ? void (0) : __assert_fail
("OldOp == OldPtr", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2676, __extension__ __PRETTY_FUNCTION__))
;
2677
2678 AAMDNodes AATags;
2679 SI.getAAMetadata(AATags);
2680
2681 Value *V = SI.getValueOperand();
2682
2683 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2684 // alloca that should be re-examined after promoting this alloca.
2685 if (V->getType()->isPointerTy())
2686 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2687 Pass.PostPromotionWorklist.insert(AI);
2688
2689 if (SliceSize < DL.getTypeStoreSize(V->getType()).getFixedSize()) {
2690 assert(!SI.isVolatile())(static_cast <bool> (!SI.isVolatile()) ? void (0) : __assert_fail
("!SI.isVolatile()", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2690, __extension__ __PRETTY_FUNCTION__))
;
2691 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2692, __extension__ __PRETTY_FUNCTION__))
2692 "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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2692, __extension__ __PRETTY_FUNCTION__))
;
2693 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2694, __extension__ __PRETTY_FUNCTION__))
2694 "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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2694, __extension__ __PRETTY_FUNCTION__))
;
2695 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2696 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset - BeginOffset,
2697 "extract");
2698 }
2699
2700 if (VecTy)
2701 return rewriteVectorizedStoreInst(V, SI, OldOp, AATags);
2702 if (IntTy && V->getType()->isIntegerTy())
2703 return rewriteIntegerStore(V, SI, AATags);
2704
2705 const bool IsStorePastEnd =
2706 DL.getTypeStoreSize(V->getType()).getFixedSize() > SliceSize;
2707 StoreInst *NewSI;
2708 if (NewBeginOffset == NewAllocaBeginOffset &&
2709 NewEndOffset == NewAllocaEndOffset &&
2710 (canConvertValue(DL, V->getType(), NewAllocaTy) ||
2711 (IsStorePastEnd && NewAllocaTy->isIntegerTy() &&
2712 V->getType()->isIntegerTy()))) {
2713 // If this is an integer store past the end of slice (and thus the bytes
2714 // past that point are irrelevant or this is unreachable), truncate the
2715 // value prior to storing.
2716 if (auto *VITy = dyn_cast<IntegerType>(V->getType()))
2717 if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2718 if (VITy->getBitWidth() > AITy->getBitWidth()) {
2719 if (DL.isBigEndian())
2720 V = IRB.CreateLShr(V, VITy->getBitWidth() - AITy->getBitWidth(),
2721 "endian_shift");
2722 V = IRB.CreateTrunc(V, AITy, "load.trunc");
2723 }
2724
2725 V = convertValue(DL, IRB, V, NewAllocaTy);
2726 NewSI =
2727 IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlign(), SI.isVolatile());
2728 } else {
2729 unsigned AS = SI.getPointerAddressSpace();
2730 Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo(AS));
2731 NewSI =
2732 IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(), SI.isVolatile());
2733 }
2734 NewSI->copyMetadata(SI, {LLVMContext::MD_mem_parallel_loop_access,
2735 LLVMContext::MD_access_group});
2736 if (AATags)
2737 NewSI->setAAMetadata(AATags.shift(NewBeginOffset - BeginOffset));
2738 if (SI.isVolatile())
2739 NewSI->setAtomic(SI.getOrdering(), SI.getSyncScopeID());
2740 if (NewSI->isAtomic())
2741 NewSI->setAlignment(SI.getAlign());
2742 Pass.DeadInsts.push_back(&SI);
2743 deleteIfTriviallyDead(OldOp);
2744
2745 LLVM_DEBUG(dbgs() << " to: " << *NewSI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *NewSI <<
"\n"; } } while (false)
;
2746 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2747 }
2748
2749 /// Compute an integer value from splatting an i8 across the given
2750 /// number of bytes.
2751 ///
2752 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2753 /// call this routine.
2754 /// FIXME: Heed the advice above.
2755 ///
2756 /// \param V The i8 value to splat.
2757 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2758 Value *getIntegerSplat(Value *V, unsigned Size) {
2759 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.\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2759, __extension__ __PRETTY_FUNCTION__))
;
2760 IntegerType *VTy = cast<IntegerType>(V->getType());
2761 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2761, __extension__ __PRETTY_FUNCTION__))
;
2762 if (Size == 1)
2763 return V;
2764
2765 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size * 8);
2766 V = IRB.CreateMul(
2767 IRB.CreateZExt(V, SplatIntTy, "zext"),
2768 ConstantExpr::getUDiv(
2769 Constant::getAllOnesValue(SplatIntTy),
2770 ConstantExpr::getZExt(Constant::getAllOnesValue(V->getType()),
2771 SplatIntTy)),
2772 "isplat");
2773 return V;
2774 }
2775
2776 /// Compute a vector splat for a given element value.
2777 Value *getVectorSplat(Value *V, unsigned NumElements) {
2778 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2779 LLVM_DEBUG(dbgs() << " splat: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " splat: " << *V <<
"\n"; } } while (false)
;
2780 return V;
2781 }
2782
2783 bool visitMemSetInst(MemSetInst &II) {
2784 LLVM_DEBUG(dbgs() << " original: " << II << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << II <<
"\n"; } } while (false)
;
2785 assert(II.getRawDest() == OldPtr)(static_cast <bool> (II.getRawDest() == OldPtr) ? void (
0) : __assert_fail ("II.getRawDest() == OldPtr", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2785, __extension__ __PRETTY_FUNCTION__))
;
2786
2787 AAMDNodes AATags;
2788 II.getAAMetadata(AATags);
2789
2790 // If the memset has a variable size, it cannot be split, just adjust the
2791 // pointer to the new alloca.
2792 if (!isa<Constant>(II.getLength())) {
2793 assert(!IsSplit)(static_cast <bool> (!IsSplit) ? void (0) : __assert_fail
("!IsSplit", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2793, __extension__ __PRETTY_FUNCTION__))
;
2794 assert(NewBeginOffset == BeginOffset)(static_cast <bool> (NewBeginOffset == BeginOffset) ? void
(0) : __assert_fail ("NewBeginOffset == BeginOffset", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2794, __extension__ __PRETTY_FUNCTION__))
;
2795 II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2796 II.setDestAlignment(getSliceAlign());
2797
2798 deleteIfTriviallyDead(OldPtr);
2799 return false;
2800 }
2801
2802 // Record this instruction for deletion.
2803 Pass.DeadInsts.push_back(&II);
2804
2805 Type *AllocaTy = NewAI.getAllocatedType();
2806 Type *ScalarTy = AllocaTy->getScalarType();
2807
2808 const bool CanContinue = [&]() {
2809 if (VecTy || IntTy)
2810 return true;
2811 if (BeginOffset > NewAllocaBeginOffset ||
2812 EndOffset < NewAllocaEndOffset)
2813 return false;
2814 auto *C = cast<ConstantInt>(II.getLength());
2815 if (C->getBitWidth() > 64)
2816 return false;
2817 const auto Len = C->getZExtValue();
2818 auto *Int8Ty = IntegerType::getInt8Ty(NewAI.getContext());
2819 auto *SrcTy = FixedVectorType::get(Int8Ty, Len);
2820 return canConvertValue(DL, SrcTy, AllocaTy) &&
2821 DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy).getFixedSize());
2822 }();
2823
2824 // If this doesn't map cleanly onto the alloca type, and that type isn't
2825 // a single value type, just emit a memset.
2826 if (!CanContinue) {
2827 Type *SizeTy = II.getLength()->getType();
2828 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2829 CallInst *New = IRB.CreateMemSet(
2830 getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2831 MaybeAlign(getSliceAlign()), II.isVolatile());
2832 if (AATags)
2833 New->setAAMetadata(AATags.shift(NewBeginOffset - BeginOffset));
2834 LLVM_DEBUG(dbgs() << " to: " << *New << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *New <<
"\n"; } } while (false)
;
2835 return false;
2836 }
2837
2838 // If we can represent this as a simple value, we have to build the actual
2839 // value to store, which requires expanding the byte present in memset to
2840 // a sensible representation for the alloca type. This is essentially
2841 // splatting the byte to a sufficiently wide integer, splatting it across
2842 // any desired vector width, and bitcasting to the final type.
2843 Value *V;
2844
2845 if (VecTy) {
2846 // If this is a memset of a vectorized alloca, insert it.
2847 assert(ElementTy == ScalarTy)(static_cast <bool> (ElementTy == ScalarTy) ? void (0) :
__assert_fail ("ElementTy == ScalarTy", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2847, __extension__ __PRETTY_FUNCTION__))
;
2848
2849 unsigned BeginIndex = getIndex(NewBeginOffset);
2850 unsigned EndIndex = getIndex(NewEndOffset);
2851 assert(EndIndex > BeginIndex && "Empty vector!")(static_cast <bool> (EndIndex > BeginIndex &&
"Empty vector!") ? void (0) : __assert_fail ("EndIndex > BeginIndex && \"Empty vector!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2851, __extension__ __PRETTY_FUNCTION__))
;
2852 unsigned NumElements = EndIndex - BeginIndex;
2853 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2854, __extension__ __PRETTY_FUNCTION__))
2854 "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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2854, __extension__ __PRETTY_FUNCTION__))
;
2855
2856 Value *Splat = getIntegerSplat(
2857 II.getValue(), DL.getTypeSizeInBits(ElementTy).getFixedSize() / 8);
2858 Splat = convertValue(DL, IRB, Splat, ElementTy);
2859 if (NumElements > 1)
2860 Splat = getVectorSplat(Splat, NumElements);
2861
2862 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2863 NewAI.getAlign(), "oldload");
2864 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2865 } else if (IntTy) {
2866 // If this is a memset on an alloca where we can widen stores, insert the
2867 // set integer.
2868 assert(!II.isVolatile())(static_cast <bool> (!II.isVolatile()) ? void (0) : __assert_fail
("!II.isVolatile()", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2868, __extension__ __PRETTY_FUNCTION__))
;
2869
2870 uint64_t Size = NewEndOffset - NewBeginOffset;
2871 V = getIntegerSplat(II.getValue(), Size);
2872
2873 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2874 EndOffset != NewAllocaBeginOffset)) {
2875 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2876 NewAI.getAlign(), "oldload");
2877 Old = convertValue(DL, IRB, Old, IntTy);
2878 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2879 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2880 } else {
2881 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2882, __extension__ __PRETTY_FUNCTION__))
2882 "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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2882, __extension__ __PRETTY_FUNCTION__))
;
2883 }
2884 V = convertValue(DL, IRB, V, AllocaTy);
2885 } else {
2886 // Established these invariants above.
2887 assert(NewBeginOffset == NewAllocaBeginOffset)(static_cast <bool> (NewBeginOffset == NewAllocaBeginOffset
) ? void (0) : __assert_fail ("NewBeginOffset == NewAllocaBeginOffset"
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2887, __extension__ __PRETTY_FUNCTION__))
;
2888 assert(NewEndOffset == NewAllocaEndOffset)(static_cast <bool> (NewEndOffset == NewAllocaEndOffset
) ? void (0) : __assert_fail ("NewEndOffset == NewAllocaEndOffset"
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2888, __extension__ __PRETTY_FUNCTION__))
;
2889
2890 V = getIntegerSplat(II.getValue(),
2891 DL.getTypeSizeInBits(ScalarTy).getFixedSize() / 8);
2892 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2893 V = getVectorSplat(
2894 V, cast<FixedVectorType>(AllocaVecTy)->getNumElements());
2895
2896 V = convertValue(DL, IRB, V, AllocaTy);
2897 }
2898
2899 StoreInst *New =
2900 IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlign(), II.isVolatile());
2901 New->copyMetadata(II, {LLVMContext::MD_mem_parallel_loop_access,
2902 LLVMContext::MD_access_group});
2903 if (AATags)
2904 New->setAAMetadata(AATags.shift(NewBeginOffset - BeginOffset));
2905 LLVM_DEBUG(dbgs() << " to: " << *New << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *New <<
"\n"; } } while (false)
;
2906 return !II.isVolatile();
2907 }
2908
2909 bool visitMemTransferInst(MemTransferInst &II) {
2910 // Rewriting of memory transfer instructions can be a bit tricky. We break
2911 // them into two categories: split intrinsics and unsplit intrinsics.
2912
2913 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
2914
2915 AAMDNodes AATags;
2916 II.getAAMetadata(AATags);
2917
2918 bool IsDest = &II.getRawDestUse() == OldUse;
2919 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)"
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2920, __extension__ __PRETTY_FUNCTION__))
3
Assuming 'IsDest' is false
4
Assuming pointer value is null
5
'?' condition is true
2920 (!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)"
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2920, __extension__ __PRETTY_FUNCTION__))
;
2921
2922 MaybeAlign SliceAlign = getSliceAlign();
2923
2924 // For unsplit intrinsics, we simply modify the source and destination
2925 // pointers in place. This isn't just an optimization, it is a matter of
2926 // correctness. With unsplit intrinsics we may be dealing with transfers
2927 // within a single alloca before SROA ran, or with transfers that have
2928 // a variable length. We may also be dealing with memmove instead of
2929 // memcpy, and so simply updating the pointers is the necessary for us to
2930 // update both source and dest of a single call.
2931 if (!IsSplittable) {
6
Assuming field 'IsSplittable' is false
7
Taking true branch
2932 Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
8
Called C++ object pointer is null
2933 if (IsDest) {
2934 II.setDest(AdjustedPtr);
2935 II.setDestAlignment(SliceAlign);
2936 }
2937 else {
2938 II.setSource(AdjustedPtr);
2939 II.setSourceAlignment(SliceAlign);
2940 }
2941
2942 LLVM_DEBUG(dbgs() << " to: " << II << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << II <<
"\n"; } } while (false)
;
2943 deleteIfTriviallyDead(OldPtr);
2944 return false;
2945 }
2946 // For split transfer intrinsics we have an incredibly useful assurance:
2947 // the source and destination do not reside within the same alloca, and at
2948 // least one of them does not escape. This means that we can replace
2949 // memmove with memcpy, and we don't need to worry about all manner of
2950 // downsides to splitting and transforming the operations.
2951
2952 // If this doesn't map cleanly onto the alloca type, and that type isn't
2953 // a single value type, just emit a memcpy.
2954 bool EmitMemCpy =
2955 !VecTy && !IntTy &&
2956 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2957 SliceSize !=
2958 DL.getTypeStoreSize(NewAI.getAllocatedType()).getFixedSize() ||
2959 !NewAI.getAllocatedType()->isSingleValueType());
2960
2961 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2962 // size hasn't been shrunk based on analysis of the viable range, this is
2963 // a no-op.
2964 if (EmitMemCpy && &OldAI == &NewAI) {
2965 // Ensure the start lines up.
2966 assert(NewBeginOffset == BeginOffset)(static_cast <bool> (NewBeginOffset == BeginOffset) ? void
(0) : __assert_fail ("NewBeginOffset == BeginOffset", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2966, __extension__ __PRETTY_FUNCTION__))
;
2967
2968 // Rewrite the size as needed.
2969 if (NewEndOffset != EndOffset)
2970 II.setLength(ConstantInt::get(II.getLength()->getType(),
2971 NewEndOffset - NewBeginOffset));
2972 return false;
2973 }
2974 // Record this instruction for deletion.
2975 Pass.DeadInsts.push_back(&II);
2976
2977 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2978 // alloca that should be re-examined after rewriting this instruction.
2979 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2980 if (AllocaInst *AI =
2981 dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2982 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.\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2983, __extension__ __PRETTY_FUNCTION__))
2983 "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.\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2983, __extension__ __PRETTY_FUNCTION__))
;
2984 Pass.Worklist.insert(AI);
2985 }
2986
2987 Type *OtherPtrTy = OtherPtr->getType();
2988 unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2989
2990 // Compute the relative offset for the other pointer within the transfer.
2991 unsigned OffsetWidth = DL.getIndexSizeInBits(OtherAS);
2992 APInt OtherOffset(OffsetWidth, NewBeginOffset - BeginOffset);
2993 Align OtherAlign =
2994 (IsDest ? II.getSourceAlign() : II.getDestAlign()).valueOrOne();
2995 OtherAlign =
2996 commonAlignment(OtherAlign, OtherOffset.zextOrTrunc(64).getZExtValue());
2997
2998 if (EmitMemCpy) {
2999 // Compute the other pointer, folding as much as possible to produce
3000 // a single, simple GEP in most cases.
3001 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
3002 OtherPtr->getName() + ".");
3003
3004 Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3005 Type *SizeTy = II.getLength()->getType();
3006 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
3007
3008 Value *DestPtr, *SrcPtr;
3009 MaybeAlign DestAlign, SrcAlign;
3010 // Note: IsDest is true iff we're copying into the new alloca slice
3011 if (IsDest) {
3012 DestPtr = OurPtr;
3013 DestAlign = SliceAlign;
3014 SrcPtr = OtherPtr;
3015 SrcAlign = OtherAlign;
3016 } else {
3017 DestPtr = OtherPtr;
3018 DestAlign = OtherAlign;
3019 SrcPtr = OurPtr;
3020 SrcAlign = SliceAlign;
3021 }
3022 CallInst *New = IRB.CreateMemCpy(DestPtr, DestAlign, SrcPtr, SrcAlign,
3023 Size, II.isVolatile());
3024 if (AATags)
3025 New->setAAMetadata(AATags.shift(NewBeginOffset - BeginOffset));
3026 LLVM_DEBUG(dbgs() << " to: " << *New << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *New <<
"\n"; } } while (false)
;
3027 return false;
3028 }
3029
3030 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
3031 NewEndOffset == NewAllocaEndOffset;
3032 uint64_t Size = NewEndOffset - NewBeginOffset;
3033 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
3034 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
3035 unsigned NumElements = EndIndex - BeginIndex;
3036 IntegerType *SubIntTy =
3037 IntTy ? Type::getIntNTy(IntTy->getContext(), Size * 8) : nullptr;
3038
3039 // Reset the other pointer type to match the register type we're going to
3040 // use, but using the address space of the original other pointer.
3041 Type *OtherTy;
3042 if (VecTy && !IsWholeAlloca) {
3043 if (NumElements == 1)
3044 OtherTy = VecTy->getElementType();
3045 else
3046 OtherTy = FixedVectorType::get(VecTy->getElementType(), NumElements);
3047 } else if (IntTy && !IsWholeAlloca) {
3048 OtherTy = SubIntTy;
3049 } else {
3050 OtherTy = NewAllocaTy;
3051 }
3052 OtherPtrTy = OtherTy->getPointerTo(OtherAS);
3053
3054 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
3055 OtherPtr->getName() + ".");
3056 MaybeAlign SrcAlign = OtherAlign;
3057 Value *DstPtr = &NewAI;
3058 MaybeAlign DstAlign = SliceAlign;
3059 if (!IsDest) {
3060 std::swap(SrcPtr, DstPtr);
3061 std::swap(SrcAlign, DstAlign);
3062 }
3063
3064 Value *Src;
3065 if (VecTy && !IsWholeAlloca && !IsDest) {
3066 Src = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
3067 NewAI.getAlign(), "load");
3068 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
3069 } else if (IntTy && !IsWholeAlloca && !IsDest) {
3070 Src = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
3071 NewAI.getAlign(), "load");
3072 Src = convertValue(DL, IRB, Src, IntTy);
3073 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
3074 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
3075 } else {
3076 LoadInst *Load = IRB.CreateAlignedLoad(OtherTy, SrcPtr, SrcAlign,
3077 II.isVolatile(), "copyload");
3078 Load->copyMetadata(II, {LLVMContext::MD_mem_parallel_loop_access,
3079 LLVMContext::MD_access_group});
3080 if (AATags)
3081 Load->setAAMetadata(AATags.shift(NewBeginOffset - BeginOffset));
3082 Src = Load;
3083 }
3084
3085 if (VecTy && !IsWholeAlloca && IsDest) {
3086 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
3087 NewAI.getAlign(), "oldload");
3088 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
3089 } else if (IntTy && !IsWholeAlloca && IsDest) {
3090 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
3091 NewAI.getAlign(), "oldload");
3092 Old = convertValue(DL, IRB, Old, IntTy);
3093 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
3094 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
3095 Src = convertValue(DL, IRB, Src, NewAllocaTy);
3096 }
3097
3098 StoreInst *Store = cast<StoreInst>(
3099 IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
3100 Store->copyMetadata(II, {LLVMContext::MD_mem_parallel_loop_access,
3101 LLVMContext::MD_access_group});
3102 if (AATags)
3103 Store->setAAMetadata(AATags.shift(NewBeginOffset - BeginOffset));
3104 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *Store <<
"\n"; } } while (false)
;
3105 return !II.isVolatile();
3106 }
3107
3108 bool visitIntrinsicInst(IntrinsicInst &II) {
3109 assert((II.isLifetimeStartOrEnd() || II.isDroppable()) &&(static_cast <bool> ((II.isLifetimeStartOrEnd() || II.isDroppable
()) && "Unexpected intrinsic!") ? void (0) : __assert_fail
("(II.isLifetimeStartOrEnd() || II.isDroppable()) && \"Unexpected intrinsic!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3110, __extension__ __PRETTY_FUNCTION__))
3110 "Unexpected intrinsic!")(static_cast <bool> ((II.isLifetimeStartOrEnd() || II.isDroppable
()) && "Unexpected intrinsic!") ? void (0) : __assert_fail
("(II.isLifetimeStartOrEnd() || II.isDroppable()) && \"Unexpected intrinsic!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3110, __extension__ __PRETTY_FUNCTION__))
;
3111 LLVM_DEBUG(dbgs() << " original: " << II << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << II <<
"\n"; } } while (false)
;
3112
3113 // Record this instruction for deletion.
3114 Pass.DeadInsts.push_back(&II);
3115
3116 if (II.isDroppable()) {
3117 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3117, __extension__ __PRETTY_FUNCTION__))
;
3118 // TODO For now we forget assumed information, this can be improved.
3119 OldPtr->dropDroppableUsesIn(II);
3120 return true;
3121 }
3122
3123 assert(II.getArgOperand(1) == OldPtr)(static_cast <bool> (II.getArgOperand(1) == OldPtr) ? void
(0) : __assert_fail ("II.getArgOperand(1) == OldPtr", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3123, __extension__ __PRETTY_FUNCTION__))
;
3124 // Lifetime intrinsics are only promotable if they cover the whole alloca.
3125 // Therefore, we drop lifetime intrinsics which don't cover the whole
3126 // alloca.
3127 // (In theory, intrinsics which partially cover an alloca could be
3128 // promoted, but PromoteMemToReg doesn't handle that case.)
3129 // FIXME: Check whether the alloca is promotable before dropping the
3130 // lifetime intrinsics?
3131 if (NewBeginOffset != NewAllocaBeginOffset ||
3132 NewEndOffset != NewAllocaEndOffset)
3133 return true;
3134
3135 ConstantInt *Size =
3136 ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
3137 NewEndOffset - NewBeginOffset);
3138 // Lifetime intrinsics always expect an i8* so directly get such a pointer
3139 // for the new alloca slice.
3140 Type *PointerTy = IRB.getInt8PtrTy(OldPtr->getType()->getPointerAddressSpace());
3141 Value *Ptr = getNewAllocaSlicePtr(IRB, PointerTy);
3142 Value *New;
3143 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
3144 New = IRB.CreateLifetimeStart(Ptr, Size);
3145 else
3146 New = IRB.CreateLifetimeEnd(Ptr, Size);
3147
3148 (void)New;
3149 LLVM_DEBUG(dbgs() << " to: " << *New << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *New <<
"\n"; } } while (false)
;
3150
3151 return true;
3152 }
3153
3154 void fixLoadStoreAlign(Instruction &Root) {
3155 // This algorithm implements the same visitor loop as
3156 // hasUnsafePHIOrSelectUse, and fixes the alignment of each load
3157 // or store found.
3158 SmallPtrSet<Instruction *, 4> Visited;
3159 SmallVector<Instruction *, 4> Uses;
3160 Visited.insert(&Root);
3161 Uses.push_back(&Root);
3162 do {
3163 Instruction *I = Uses.pop_back_val();
3164
3165 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3166 LI->setAlignment(std::min(LI->getAlign(), getSliceAlign()));
3167 continue;
3168 }
3169 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3170 SI->setAlignment(std::min(SI->getAlign(), getSliceAlign()));
3171 continue;
3172 }
3173
3174 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)"
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3176, __extension__ __PRETTY_FUNCTION__))
3175 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)"
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3176, __extension__ __PRETTY_FUNCTION__))
3176 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)"
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3176, __extension__ __PRETTY_FUNCTION__))
;
3177 for (User *U : I->users())
3178 if (Visited.insert(cast<Instruction>(U)).second)
3179 Uses.push_back(cast<Instruction>(U));
3180 } while (!Uses.empty());
3181 }
3182
3183 bool visitPHINode(PHINode &PN) {
3184 LLVM_DEBUG(dbgs() << " original: " << PN << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << PN <<
"\n"; } } while (false)
;
3185 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable")(static_cast <bool> (BeginOffset >= NewAllocaBeginOffset
&& "PHIs are unsplittable") ? void (0) : __assert_fail
("BeginOffset >= NewAllocaBeginOffset && \"PHIs are unsplittable\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3185, __extension__ __PRETTY_FUNCTION__))
;
3186 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable")(static_cast <bool> (EndOffset <= NewAllocaEndOffset
&& "PHIs are unsplittable") ? void (0) : __assert_fail
("EndOffset <= NewAllocaEndOffset && \"PHIs are unsplittable\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3186, __extension__ __PRETTY_FUNCTION__))
;
3187
3188 // We would like to compute a new pointer in only one place, but have it be
3189 // as local as possible to the PHI. To do that, we re-use the location of
3190 // the old pointer, which necessarily must be in the right position to
3191 // dominate the PHI.
3192 IRBuilderBase::InsertPointGuard Guard(IRB);
3193 if (isa<PHINode>(OldPtr))
3194 IRB.SetInsertPoint(&*OldPtr->getParent()->getFirstInsertionPt());
3195 else
3196 IRB.SetInsertPoint(OldPtr);
3197 IRB.SetCurrentDebugLocation(OldPtr->getDebugLoc());
3198
3199 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3200 // Replace the operands which were using the old pointer.
3201 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
3202
3203 LLVM_DEBUG(dbgs() << " to: " << PN << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << PN <<
"\n"; } } while (false)
;
3204 deleteIfTriviallyDead(OldPtr);
3205
3206 // Fix the alignment of any loads or stores using this PHI node.
3207 fixLoadStoreAlign(PN);
3208
3209 // PHIs can't be promoted on their own, but often can be speculated. We
3210 // check the speculation outside of the rewriter so that we see the
3211 // fully-rewritten alloca.
3212 PHIUsers.insert(&PN);
3213 return true;
3214 }
3215
3216 bool visitSelectInst(SelectInst &SI) {
3217 LLVM_DEBUG(dbgs() << " original: " << SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << SI <<
"\n"; } } while (false)
;
3218 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3219, __extension__ __PRETTY_FUNCTION__))
3219 "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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3219, __extension__ __PRETTY_FUNCTION__))
;
3220 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable")(static_cast <bool> (BeginOffset >= NewAllocaBeginOffset
&& "Selects are unsplittable") ? void (0) : __assert_fail
("BeginOffset >= NewAllocaBeginOffset && \"Selects are unsplittable\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3220, __extension__ __PRETTY_FUNCTION__))
;
3221 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable")(static_cast <bool> (EndOffset <= NewAllocaEndOffset
&& "Selects are unsplittable") ? void (0) : __assert_fail
("EndOffset <= NewAllocaEndOffset && \"Selects are unsplittable\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3221, __extension__ __PRETTY_FUNCTION__))
;
3222
3223 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3224 // Replace the operands which were using the old pointer.
3225 if (SI.getOperand(1) == OldPtr)
3226 SI.setOperand(1, NewPtr);
3227 if (SI.getOperand(2) == OldPtr)
3228 SI.setOperand(2, NewPtr);
3229
3230 LLVM_DEBUG(dbgs() << " to: " << SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << SI <<
"\n"; } } while (false)
;
3231 deleteIfTriviallyDead(OldPtr);
3232
3233 // Fix the alignment of any loads or stores using this select.
3234 fixLoadStoreAlign(SI);
3235
3236 // Selects can't be promoted on their own, but often can be speculated. We
3237 // check the speculation outside of the rewriter so that we see the
3238 // fully-rewritten alloca.
3239 SelectUsers.insert(&SI);
3240 return true;
3241 }
3242};
3243
3244namespace {
3245
3246/// Visitor to rewrite aggregate loads and stores as scalar.
3247///
3248/// This pass aggressively rewrites all aggregate loads and stores on
3249/// a particular pointer (or any pointer derived from it which we can identify)
3250/// with scalar loads and stores.
3251class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
3252 // Befriend the base class so it can delegate to private visit methods.
3253 friend class InstVisitor<AggLoadStoreRewriter, bool>;
3254
3255 /// Queue of pointer uses to analyze and potentially rewrite.
3256 SmallVector<Use *, 8> Queue;
3257
3258 /// Set to prevent us from cycling with phi nodes and loops.
3259 SmallPtrSet<User *, 8> Visited;
3260
3261 /// The current pointer use being rewritten. This is used to dig up the used
3262 /// value (as opposed to the user).
3263 Use *U = nullptr;
3264
3265 /// Used to calculate offsets, and hence alignment, of subobjects.
3266 const DataLayout &DL;
3267
3268public:
3269 AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
3270
3271 /// Rewrite loads and stores through a pointer and all pointers derived from
3272 /// it.
3273 bool rewrite(Instruction &I) {
3274 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)
;
3275 enqueueUsers(I);
3276 bool Changed = false;
3277 while (!Queue.empty()) {
3278 U = Queue.pop_back_val();
3279 Changed |= visit(cast<Instruction>(U->getUser()));
3280 }
3281 return Changed;
3282 }
3283
3284private:
3285 /// Enqueue all the users of the given instruction for further processing.
3286 /// This uses a set to de-duplicate users.
3287 void enqueueUsers(Instruction &I) {
3288 for (Use &U : I.uses())
3289 if (Visited.insert(U.getUser()).second)
3290 Queue.push_back(&U);
3291 }
3292
3293 // Conservative default is to not rewrite anything.
3294 bool visitInstruction(Instruction &I) { return false; }
3295
3296 /// Generic recursive split emission class.
3297 template <typename Derived> class OpSplitter {
3298 protected:
3299 /// The builder used to form new instructions.
3300 IRBuilderTy IRB;
3301
3302 /// The indices which to be used with insert- or extractvalue to select the
3303 /// appropriate value within the aggregate.
3304 SmallVector<unsigned, 4> Indices;
3305
3306 /// The indices to a GEP instruction which will move Ptr to the correct slot
3307 /// within the aggregate.
3308 SmallVector<Value *, 4> GEPIndices;
3309
3310 /// The base pointer of the original op, used as a base for GEPing the
3311 /// split operations.
3312 Value *Ptr;
3313
3314 /// The base pointee type being GEPed into.
3315 Type *BaseTy;
3316
3317 /// Known alignment of the base pointer.
3318 Align BaseAlign;
3319
3320 /// To calculate offset of each component so we can correctly deduce
3321 /// alignments.
3322 const DataLayout &DL;
3323
3324 /// Initialize the splitter with an insertion point, Ptr and start with a
3325 /// single zero GEP index.
3326 OpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
3327 Align BaseAlign, const DataLayout &DL)
3328 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr),
3329 BaseTy(BaseTy), BaseAlign(BaseAlign), DL(DL) {}
3330
3331 public:
3332 /// Generic recursive split emission routine.
3333 ///
3334 /// This method recursively splits an aggregate op (load or store) into
3335 /// scalar or vector ops. It splits recursively until it hits a single value
3336 /// and emits that single value operation via the template argument.
3337 ///
3338 /// The logic of this routine relies on GEPs and insertvalue and
3339 /// extractvalue all operating with the same fundamental index list, merely
3340 /// formatted differently (GEPs need actual values).
3341 ///
3342 /// \param Ty The type being split recursively into smaller ops.
3343 /// \param Agg The aggregate value being built up or stored, depending on
3344 /// whether this is splitting a load or a store respectively.
3345 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3346 if (Ty->isSingleValueType()) {
3347 unsigned Offset = DL.getIndexedOffsetInType(BaseTy, GEPIndices);
3348 return static_cast<Derived *>(this)->emitFunc(
3349 Ty, Agg, commonAlignment(BaseAlign, Offset), Name);
3350 }
3351
3352 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3353 unsigned OldSize = Indices.size();
3354 (void)OldSize;
3355 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3356 ++Idx) {
3357 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3357, __extension__ __PRETTY_FUNCTION__))
;
3358 Indices.push_back(Idx);
3359 GEPIndices.push_back(IRB.getInt32(Idx));
3360 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3361 GEPIndices.pop_back();
3362 Indices.pop_back();
3363 }
3364 return;
3365 }
3366
3367 if (StructType *STy = dyn_cast<StructType>(Ty)) {
3368 unsigned OldSize = Indices.size();
3369 (void)OldSize;
3370 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3371 ++Idx) {
3372 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3372, __extension__ __PRETTY_FUNCTION__))
;
3373 Indices.push_back(Idx);
3374 GEPIndices.push_back(IRB.getInt32(Idx));
3375 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3376 GEPIndices.pop_back();
3377 Indices.pop_back();
3378 }
3379 return;
3380 }
3381
3382 llvm_unreachable("Only arrays and structs are aggregate loadable types")::llvm::llvm_unreachable_internal("Only arrays and structs are aggregate loadable types"
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3382)
;
3383 }
3384 };
3385
3386 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3387 AAMDNodes AATags;
3388
3389 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
3390 AAMDNodes AATags, Align BaseAlign, const DataLayout &DL)
3391 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr, BaseTy, BaseAlign,
3392 DL),
3393 AATags(AATags) {}
3394
3395 /// Emit a leaf load of a single value. This is called at the leaves of the
3396 /// recursive emission to actually load values.
3397 void emitFunc(Type *Ty, Value *&Agg, Align Alignment, const Twine &Name) {
3398 assert(Ty->isSingleValueType())(static_cast <bool> (Ty->isSingleValueType()) ? void
(0) : __assert_fail ("Ty->isSingleValueType()", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3398, __extension__ __PRETTY_FUNCTION__))
;
3399 // Load the single value and insert it using the indices.
3400 Value *GEP =
3401 IRB.CreateInBoundsGEP(BaseTy, Ptr, GEPIndices, Name + ".gep");
3402 LoadInst *Load =
3403 IRB.CreateAlignedLoad(Ty, GEP, Alignment, Name + ".load");
3404
3405 APInt Offset(
3406 DL.getIndexSizeInBits(Ptr->getType()->getPointerAddressSpace()), 0);
3407 if (AATags &&
3408 GEPOperator::accumulateConstantOffset(BaseTy, GEPIndices, DL, Offset))
3409 Load->setAAMetadata(AATags.shift(Offset.getZExtValue()));
3410
3411 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3412 LLVM_DEBUG(dbgs() << " to: " << *Load << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *Load <<
"\n"; } } while (false)
;
3413 }
3414 };
3415
3416 bool visitLoadInst(LoadInst &LI) {
3417 assert(LI.getPointerOperand() == *U)(static_cast <bool> (LI.getPointerOperand() == *U) ? void
(0) : __assert_fail ("LI.getPointerOperand() == *U", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3417, __extension__ __PRETTY_FUNCTION__))
;
3418 if (!LI.isSimple() || LI.getType()->isSingleValueType())
3419 return false;
3420
3421 // We have an aggregate being loaded, split it apart.
3422 LLVM_DEBUG(dbgs() << " original: " << LI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << LI <<
"\n"; } } while (false)
;
3423 AAMDNodes AATags;
3424 LI.getAAMetadata(AATags);
3425 LoadOpSplitter Splitter(&LI, *U, LI.getType(), AATags,
3426 getAdjustedAlignment(&LI, 0), DL);
3427 Value *V = UndefValue::get(LI.getType());
3428 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3429 Visited.erase(&LI);
3430 LI.replaceAllUsesWith(V);
3431 LI.eraseFromParent();
3432 return true;
3433 }
3434
3435 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3436 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
3437 AAMDNodes AATags, Align BaseAlign, const DataLayout &DL)
3438 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr, BaseTy, BaseAlign,
3439 DL),
3440 AATags(AATags) {}
3441 AAMDNodes AATags;
3442 /// Emit a leaf store of a single value. This is called at the leaves of the
3443 /// recursive emission to actually produce stores.
3444 void emitFunc(Type *Ty, Value *&Agg, Align Alignment, const Twine &Name) {
3445 assert(Ty->isSingleValueType())(static_cast <bool> (Ty->isSingleValueType()) ? void
(0) : __assert_fail ("Ty->isSingleValueType()", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3445, __extension__ __PRETTY_FUNCTION__))
;
3446 // Extract the single value and store it using the indices.
3447 //
3448 // The gep and extractvalue values are factored out of the CreateStore
3449 // call to make the output independent of the argument evaluation order.
3450 Value *ExtractValue =
3451 IRB.CreateExtractValue(Agg, Indices, Name + ".extract");
3452 Value *InBoundsGEP =
3453 IRB.CreateInBoundsGEP(BaseTy, Ptr, GEPIndices, Name + ".gep");
3454 StoreInst *Store =
3455 IRB.CreateAlignedStore(ExtractValue, InBoundsGEP, Alignment);
3456
3457 APInt Offset(
3458 DL.getIndexSizeInBits(Ptr->getType()->getPointerAddressSpace()), 0);
3459 if (AATags &&
3460 GEPOperator::accumulateConstantOffset(BaseTy, GEPIndices, DL, Offset))
3461 Store->setAAMetadata(AATags.shift(Offset.getZExtValue()));
3462
3463 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *Store <<
"\n"; } } while (false)
;
3464 }
3465 };
3466
3467 bool visitStoreInst(StoreInst &SI) {
3468 if (!SI.isSimple() || SI.getPointerOperand() != *U)
3469 return false;
3470 Value *V = SI.getValueOperand();
3471 if (V->getType()->isSingleValueType())
3472 return false;
3473
3474 // We have an aggregate being stored, split it apart.
3475 LLVM_DEBUG(dbgs() << " original: " << SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << SI <<
"\n"; } } while (false)
;
3476 AAMDNodes AATags;
3477 SI.getAAMetadata(AATags);
3478 StoreOpSplitter Splitter(&SI, *U, V->getType(), AATags,
3479 getAdjustedAlignment(&SI, 0), DL);
3480 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3481 Visited.erase(&SI);
3482 SI.eraseFromParent();
3483 return true;
3484 }
3485
3486 bool visitBitCastInst(BitCastInst &BC) {
3487 enqueueUsers(BC);
3488 return false;
3489 }
3490
3491 bool visitAddrSpaceCastInst(AddrSpaceCastInst &ASC) {
3492 enqueueUsers(ASC);
3493 return false;
3494 }
3495
3496 // Fold gep (select cond, ptr1, ptr2) => select cond, gep(ptr1), gep(ptr2)
3497 bool foldGEPSelect(GetElementPtrInst &GEPI) {
3498 if (!GEPI.hasAllConstantIndices())
3499 return false;
3500
3501 SelectInst *Sel = cast<SelectInst>(GEPI.getPointerOperand());
3502
3503 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)
3504 << "\n original: " << *Seldo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Rewriting gep(select) -> select(gep):"
<< "\n original: " << *Sel << "\n "
<< GEPI; } } while (false)
3505 << "\n " << GEPI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Rewriting gep(select) -> select(gep):"
<< "\n original: " << *Sel << "\n "
<< GEPI; } } while (false)
;
3506
3507 IRBuilderTy Builder(&GEPI);
3508 SmallVector<Value *, 4> Index(GEPI.indices());
3509 bool IsInBounds = GEPI.isInBounds();
3510
3511 Type *Ty = GEPI.getSourceElementType();
3512 Value *True = Sel->getTrueValue();
3513 Value *NTrue =
3514 IsInBounds
3515 ? Builder.CreateInBoundsGEP(Ty, True, Index,
3516 True->getName() + ".sroa.gep")
3517 : Builder.CreateGEP(Ty, True, Index, True->getName() + ".sroa.gep");
3518
3519 Value *False = Sel->getFalseValue();
3520
3521 Value *NFalse =
3522 IsInBounds
3523 ? Builder.CreateInBoundsGEP(Ty, False, Index,
3524 False->getName() + ".sroa.gep")
3525 : Builder.CreateGEP(Ty, False, Index,
3526 False->getName() + ".sroa.gep");
3527
3528 Value *NSel = Builder.CreateSelect(Sel->getCondition(), NTrue, NFalse,
3529 Sel->getName() + ".sroa.sel");
3530 Visited.erase(&GEPI);
3531 GEPI.replaceAllUsesWith(NSel);
3532 GEPI.eraseFromParent();
3533 Instruction *NSelI = cast<Instruction>(NSel);
3534 Visited.insert(NSelI);
3535 enqueueUsers(*NSelI);
3536
3537 LLVM_DEBUG(dbgs() << "\n to: " << *NTruedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "\n to: " << *NTrue
<< "\n " << *NFalse << "\n "
<< *NSel << '\n'; } } while (false)
3538 << "\n " << *NFalsedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "\n to: " << *NTrue
<< "\n " << *NFalse << "\n "
<< *NSel << '\n'; } } while (false)
3539 << "\n " << *NSel << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "\n to: " << *NTrue
<< "\n " << *NFalse << "\n "
<< *NSel << '\n'; } } while (false)
;
3540
3541 return true;
3542 }
3543
3544 // Fold gep (phi ptr1, ptr2) => phi gep(ptr1), gep(ptr2)
3545 bool foldGEPPhi(GetElementPtrInst &GEPI) {
3546 if (!GEPI.hasAllConstantIndices())
3547 return false;
3548
3549 PHINode *PHI = cast<PHINode>(GEPI.getPointerOperand());
3550 if (GEPI.getParent() != PHI->getParent() ||
3551 llvm::any_of(PHI->incoming_values(), [](Value *In)
3552 { Instruction *I = dyn_cast<Instruction>(In);
3553 return !I || isa<GetElementPtrInst>(I) || isa<PHINode>(I) ||
3554 succ_empty(I->getParent()) ||
3555 !I->getParent()->isLegalToHoistInto();
3556 }))
3557 return false;
3558
3559 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)
3560 << "\n original: " << *PHIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Rewriting gep(phi) -> phi(gep):"
<< "\n original: " << *PHI << "\n "
<< GEPI << "\n to: "; } } while (false)
3561 << "\n " << GEPIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Rewriting gep(phi) -> phi(gep):"
<< "\n original: " << *PHI << "\n "
<< GEPI << "\n to: "; } } while (false)
3562 << "\n to: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Rewriting gep(phi) -> phi(gep):"
<< "\n original: " << *PHI << "\n "
<< GEPI << "\n to: "; } } while (false)
;
3563
3564 SmallVector<Value *, 4> Index(GEPI.indices());
3565 bool IsInBounds = GEPI.isInBounds();
3566 IRBuilderTy PHIBuilder(GEPI.getParent()->getFirstNonPHI());
3567 PHINode *NewPN = PHIBuilder.CreatePHI(GEPI.getType(),
3568 PHI->getNumIncomingValues(),
3569 PHI->getName() + ".sroa.phi");
3570 for (unsigned I = 0, E = PHI->getNumIncomingValues(); I != E; ++I) {
3571 BasicBlock *B = PHI->getIncomingBlock(I);
3572 Value *NewVal = nullptr;
3573 int Idx = NewPN->getBasicBlockIndex(B);
3574 if (Idx >= 0) {
3575 NewVal = NewPN->getIncomingValue(Idx);
3576 } else {
3577 Instruction *In = cast<Instruction>(PHI->getIncomingValue(I));
3578
3579 IRBuilderTy B(In->getParent(), std::next(In->getIterator()));
3580 Type *Ty = GEPI.getSourceElementType();
3581 NewVal = IsInBounds
3582 ? B.CreateInBoundsGEP(Ty, In, Index, In->getName() + ".sroa.gep")
3583 : B.CreateGEP(Ty, In, Index, In->getName() + ".sroa.gep");
3584 }
3585 NewPN->addIncoming(NewVal, B);
3586 }
3587
3588 Visited.erase(&GEPI);
3589 GEPI.replaceAllUsesWith(NewPN);
3590 GEPI.eraseFromParent();
3591 Visited.insert(NewPN);
3592 enqueueUsers(*NewPN);
3593
3594 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)
3595 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)
3596 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)
;
3597
3598 return true;
3599 }
3600
3601 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3602 if (isa<SelectInst>(GEPI.getPointerOperand()) &&
3603 foldGEPSelect(GEPI))
3604 return true;
3605
3606 if (isa<PHINode>(GEPI.getPointerOperand()) &&
3607 foldGEPPhi(GEPI))
3608 return true;
3609
3610 enqueueUsers(GEPI);
3611 return false;
3612 }
3613
3614 bool visitPHINode(PHINode &PN) {
3615 enqueueUsers(PN);
3616 return false;
3617 }
3618
3619 bool visitSelectInst(SelectInst &SI) {
3620 enqueueUsers(SI);
3621 return false;
3622 }
3623};
3624
3625} // end anonymous namespace
3626
3627/// Strip aggregate type wrapping.
3628///
3629/// This removes no-op aggregate types wrapping an underlying type. It will
3630/// strip as many layers of types as it can without changing either the type
3631/// size or the allocated size.
3632static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
3633 if (Ty->isSingleValueType())
3634 return Ty;
3635
3636 uint64_t AllocSize = DL.getTypeAllocSize(Ty).getFixedSize();
3637 uint64_t TypeSize = DL.getTypeSizeInBits(Ty).getFixedSize();
3638
3639 Type *InnerTy;
3640 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3641 InnerTy = ArrTy->getElementType();
3642 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3643 const StructLayout *SL = DL.getStructLayout(STy);
3644 unsigned Index = SL->getElementContainingOffset(0);
3645 InnerTy = STy->getElementType(Index);
3646 } else {
3647 return Ty;
3648 }
3649
3650 if (AllocSize > DL.getTypeAllocSize(InnerTy).getFixedSize() ||
3651 TypeSize > DL.getTypeSizeInBits(InnerTy).getFixedSize())
3652 return Ty;
3653
3654 return stripAggregateTypeWrapping(DL, InnerTy);
3655}
3656
3657/// Try to find a partition of the aggregate type passed in for a given
3658/// offset and size.
3659///
3660/// This recurses through the aggregate type and tries to compute a subtype
3661/// based on the offset and size. When the offset and size span a sub-section
3662/// of an array, it will even compute a new array type for that sub-section,
3663/// and the same for structs.
3664///
3665/// Note that this routine is very strict and tries to find a partition of the
3666/// type which produces the *exact* right offset and size. It is not forgiving
3667/// when the size or offset cause either end of type-based partition to be off.
3668/// Also, this is a best-effort routine. It is reasonable to give up and not
3669/// return a type if necessary.
3670static Type *getTypePartition(const DataLayout &DL, Type *Ty, uint64_t Offset,
3671 uint64_t Size) {
3672 if (Offset == 0 && DL.getTypeAllocSize(Ty).getFixedSize() == Size)
3673 return stripAggregateTypeWrapping(DL, Ty);
3674 if (Offset > DL.getTypeAllocSize(Ty).getFixedSize() ||
3675 (DL.getTypeAllocSize(Ty).getFixedSize() - Offset) < Size)
3676 return nullptr;
3677
3678 if (isa<ArrayType>(Ty) || isa<VectorType>(Ty)) {
3679 Type *ElementTy;
3680 uint64_t TyNumElements;
3681 if (auto *AT = dyn_cast<ArrayType>(Ty)) {
3682 ElementTy = AT->getElementType();
3683 TyNumElements = AT->getNumElements();
3684 } else {
3685 // FIXME: This isn't right for vectors with non-byte-sized or
3686 // non-power-of-two sized elements.
3687 auto *VT = cast<FixedVectorType>(Ty);
3688 ElementTy = VT->getElementType();
3689 TyNumElements = VT->getNumElements();
3690 }
3691 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy).getFixedSize();
3692 uint64_t NumSkippedElements = Offset / ElementSize;
3693 if (NumSkippedElements >= TyNumElements)
3694 return nullptr;
3695 Offset -= NumSkippedElements * ElementSize;
3696
3697 // First check if we need to recurse.
3698 if (Offset > 0 || Size < ElementSize) {
3699 // Bail if the partition ends in a different array element.
3700 if ((Offset + Size) > ElementSize)
3701 return nullptr;
3702 // Recurse through the element type trying to peel off offset bytes.
3703 return getTypePartition(DL, ElementTy, Offset, Size);
3704 }
3705 assert(Offset == 0)(static_cast <bool> (Offset == 0) ? void (0) : __assert_fail
("Offset == 0", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3705, __extension__ __PRETTY_FUNCTION__))
;
3706
3707 if (Size == ElementSize)
3708 return stripAggregateTypeWrapping(DL, ElementTy);
3709 assert(Size > ElementSize)(static_cast <bool> (Size > ElementSize) ? void (0) :
__assert_fail ("Size > ElementSize", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3709, __extension__ __PRETTY_FUNCTION__))
;
3710 uint64_t NumElements = Size / ElementSize;
3711 if (NumElements * ElementSize != Size)
3712 return nullptr;
3713 return ArrayType::get(ElementTy, NumElements);
3714 }
3715
3716 StructType *STy = dyn_cast<StructType>(Ty);
3717 if (!STy)
3718 return nullptr;
3719
3720 const StructLayout *SL = DL.getStructLayout(STy);
3721 if (Offset >= SL->getSizeInBytes())
3722 return nullptr;
3723 uint64_t EndOffset = Offset + Size;
3724 if (EndOffset > SL->getSizeInBytes())
3725 return nullptr;
3726
3727 unsigned Index = SL->getElementContainingOffset(Offset);
3728 Offset -= SL->getElementOffset(Index);
3729
3730 Type *ElementTy = STy->getElementType(Index);
3731 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy).getFixedSize();
3732 if (Offset >= ElementSize)
3733 return nullptr; // The offset points into alignment padding.
3734
3735 // See if any partition must be contained by the element.
3736 if (Offset > 0 || Size < ElementSize) {
3737 if ((Offset + Size) > ElementSize)
3738 return nullptr;
3739 return getTypePartition(DL, ElementTy, Offset, Size);
3740 }
3741 assert(Offset == 0)(static_cast <bool> (Offset == 0) ? void (0) : __assert_fail
("Offset == 0", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3741, __extension__ __PRETTY_FUNCTION__))
;
3742
3743 if (Size == ElementSize)
3744 return stripAggregateTypeWrapping(DL, ElementTy);
3745
3746 StructType::element_iterator EI = STy->element_begin() + Index,
3747 EE = STy->element_end();
3748 if (EndOffset < SL->getSizeInBytes()) {
3749 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3750 if (Index == EndIndex)
3751 return nullptr; // Within a single element and its padding.
3752
3753 // Don't try to form "natural" types if the elements don't line up with the
3754 // expected size.
3755 // FIXME: We could potentially recurse down through the last element in the
3756 // sub-struct to find a natural end point.
3757 if (SL->getElementOffset(EndIndex) != EndOffset)
3758 return nullptr;
3759
3760 assert(Index < EndIndex)(static_cast <bool> (Index < EndIndex) ? void (0) : __assert_fail
("Index < EndIndex", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3760, __extension__ __PRETTY_FUNCTION__))
;
3761 EE = STy->element_begin() + EndIndex;
3762 }
3763
3764 // Try to build up a sub-structure.
3765 StructType *SubTy =
3766 StructType::get(STy->getContext(), makeArrayRef(EI, EE), STy->isPacked());
3767 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3768 if (Size != SubSL->getSizeInBytes())
3769 return nullptr; // The sub-struct doesn't have quite the size needed.
3770
3771 return SubTy;
3772}
3773
3774/// Pre-split loads and stores to simplify rewriting.
3775///
3776/// We want to break up the splittable load+store pairs as much as
3777/// possible. This is important to do as a preprocessing step, as once we
3778/// start rewriting the accesses to partitions of the alloca we lose the
3779/// necessary information to correctly split apart paired loads and stores
3780/// which both point into this alloca. The case to consider is something like
3781/// the following:
3782///
3783/// %a = alloca [12 x i8]
3784/// %gep1 = getelementptr [12 x i8]* %a, i32 0, i32 0
3785/// %gep2 = getelementptr [12 x i8]* %a, i32 0, i32 4
3786/// %gep3 = getelementptr [12 x i8]* %a, i32 0, i32 8
3787/// %iptr1 = bitcast i8* %gep1 to i64*
3788/// %iptr2 = bitcast i8* %gep2 to i64*
3789/// %fptr1 = bitcast i8* %gep1 to float*
3790/// %fptr2 = bitcast i8* %gep2 to float*
3791/// %fptr3 = bitcast i8* %gep3 to float*
3792/// store float 0.0, float* %fptr1
3793/// store float 1.0, float* %fptr2
3794/// %v = load i64* %iptr1
3795/// store i64 %v, i64* %iptr2
3796/// %f1 = load float* %fptr2
3797/// %f2 = load float* %fptr3
3798///
3799/// Here we want to form 3 partitions of the alloca, each 4 bytes large, and
3800/// promote everything so we recover the 2 SSA values that should have been
3801/// there all along.
3802///
3803/// \returns true if any changes are made.
3804bool SROA::presplitLoadsAndStores(AllocaInst &AI, AllocaSlices &AS) {
3805 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)
;
3806
3807 // Track the loads and stores which are candidates for pre-splitting here, in
3808 // the order they first appear during the partition scan. These give stable
3809 // iteration order and a basis for tracking which loads and stores we
3810 // actually split.
3811 SmallVector<LoadInst *, 4> Loads;
3812 SmallVector<StoreInst *, 4> Stores;
3813
3814 // We need to accumulate the splits required of each load or store where we
3815 // can find them via a direct lookup. This is important to cross-check loads
3816 // and stores against each other. We also track the slice so that we can kill
3817 // all the slices that end up split.
3818 struct SplitOffsets {
3819 Slice *S;
3820 std::vector<uint64_t> Splits;
3821 };
3822 SmallDenseMap<Instruction *, SplitOffsets, 8> SplitOffsetsMap;
3823
3824 // Track loads out of this alloca which cannot, for any reason, be pre-split.
3825 // This is important as we also cannot pre-split stores of those loads!
3826 // FIXME: This is all pretty gross. It means that we can be more aggressive
3827 // in pre-splitting when the load feeding the store happens to come from
3828 // a separate alloca. Put another way, the effectiveness of SROA would be
3829 // decreased by a frontend which just concatenated all of its local allocas
3830 // into one big flat alloca. But defeating such patterns is exactly the job
3831 // SROA is tasked with! Sadly, to not have this discrepancy we would have
3832 // change store pre-splitting to actually force pre-splitting of the load
3833 // that feeds it *and all stores*. That makes pre-splitting much harder, but
3834 // maybe it would make it more principled?
3835 SmallPtrSet<LoadInst *, 8> UnsplittableLoads;
3836
3837 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)
;
3838 for (auto &P : AS.partitions()) {
3839 for (Slice &S : P) {
3840 Instruction *I = cast<Instruction>(S.getUse()->getUser());
3841 if (!S.isSplittable() || S.endOffset() <= P.endOffset()) {
3842 // If this is a load we have to track that it can't participate in any
3843 // pre-splitting. If this is a store of a load we have to track that
3844 // that load also can't participate in any pre-splitting.
3845 if (auto *LI = dyn_cast<LoadInst>(I))
3846 UnsplittableLoads.insert(LI);
3847 else if (auto *SI = dyn_cast<StoreInst>(I))
3848 if (auto *LI = dyn_cast<LoadInst>(SI->getValueOperand()))
3849 UnsplittableLoads.insert(LI);
3850 continue;
3851 }
3852 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3853, __extension__ __PRETTY_FUNCTION__))
3853 "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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3853, __extension__ __PRETTY_FUNCTION__))
;
3854
3855 // Determine if this is a pre-splittable slice.
3856 if (auto *LI = dyn_cast<LoadInst>(I)) {
3857 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3857, __extension__ __PRETTY_FUNCTION__))
;
3858
3859 // The load must be used exclusively to store into other pointers for
3860 // us to be able to arbitrarily pre-split it. The stores must also be
3861 // simple to avoid changing semantics.
3862 auto IsLoadSimplyStored = [](LoadInst *LI) {
3863 for (User *LU : LI->users()) {
3864 auto *SI = dyn_cast<StoreInst>(LU);
3865 if (!SI || !SI->isSimple())
3866 return false;
3867 }
3868 return true;
3869 };
3870 if (!IsLoadSimplyStored(LI)) {
3871 UnsplittableLoads.insert(LI);
3872 continue;
3873 }
3874
3875 Loads.push_back(LI);
3876 } else if (auto *SI = dyn_cast<StoreInst>(I)) {
3877 if (S.getUse() != &SI->getOperandUse(SI->getPointerOperandIndex()))
3878 // Skip stores *of* pointers. FIXME: This shouldn't even be possible!
3879 continue;
3880 auto *StoredLoad = dyn_cast<LoadInst>(SI->getValueOperand());
3881 if (!StoredLoad || !StoredLoad->isSimple())
3882 continue;
3883 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3883, __extension__ __PRETTY_FUNCTION__))
;
3884
3885 Stores.push_back(SI);
3886 } else {
3887 // Other uses cannot be pre-split.
3888 continue;
3889 }
3890
3891 // Record the initial split.
3892 LLVM_DEBUG(dbgs() << " Candidate: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Candidate: " << *I <<
"\n"; } } while (false)
;
3893 auto &Offsets = SplitOffsetsMap[I];
3894 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3895, __extension__ __PRETTY_FUNCTION__))
3895 "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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3895, __extension__ __PRETTY_FUNCTION__))
;
3896 Offsets.S = &S;
3897 Offsets.Splits.push_back(P.endOffset() - S.beginOffset());
3898 }
3899
3900 // Now scan the already split slices, and add a split for any of them which
3901 // we're going to pre-split.
3902 for (Slice *S : P.splitSliceTails()) {
3903 auto SplitOffsetsMapI =
3904 SplitOffsetsMap.find(cast<Instruction>(S->getUse()->getUser()));
3905 if (SplitOffsetsMapI == SplitOffsetsMap.end())
3906 continue;
3907 auto &Offsets = SplitOffsetsMapI->second;
3908
3909 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3909, __extension__ __PRETTY_FUNCTION__))
;
3910 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3911, __extension__ __PRETTY_FUNCTION__))
3911 "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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3911, __extension__ __PRETTY_FUNCTION__))
;
3912 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3914, __extension__ __PRETTY_FUNCTION__))
3913 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3914, __extension__ __PRETTY_FUNCTION__))
3914 "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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3914, __extension__ __PRETTY_FUNCTION__))
;
3915
3916 // Record each split. The last partition's end isn't needed as the size
3917 // of the slice dictates that.
3918 if (S->endOffset() > P.endOffset())
3919 Offsets.Splits.push_back(P.endOffset() - Offsets.S->beginOffset());
3920 }
3921 }
3922
3923 // We may have split loads where some of their stores are split stores. For
3924 // such loads and stores, we can only pre-split them if their splits exactly
3925 // match relative to their starting offset. We have to verify this prior to
3926 // any rewriting.
3927 llvm::erase_if(Stores, [&UnsplittableLoads, &SplitOffsetsMap](StoreInst *SI) {
3928 // Lookup the load we are storing in our map of split
3929 // offsets.
3930 auto *LI = cast<LoadInst>(SI->getValueOperand());
3931 // If it was completely unsplittable, then we're done,
3932 // and this store can't be pre-split.
3933 if (UnsplittableLoads.count(LI))
3934 return true;
3935
3936 auto LoadOffsetsI = SplitOffsetsMap.find(LI);
3937 if (LoadOffsetsI == SplitOffsetsMap.end())
3938 return false; // Unrelated loads are definitely safe.
3939 auto &LoadOffsets = LoadOffsetsI->second;
3940
3941 // Now lookup the store's offsets.
3942 auto &StoreOffsets = SplitOffsetsMap[SI];
3943
3944 // If the relative offsets of each split in the load and
3945 // store match exactly, then we can split them and we
3946 // don't need to remove them here.
3947 if (LoadOffsets.Splits == StoreOffsets.Splits)
3948 return false;
3949
3950 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)
3951 << " " << *LI << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Mismatched splits for load and store:\n"
<< " " << *LI << "\n" << " "
<< *SI << "\n"; } } while (false)
3952 << " " << *SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Mismatched splits for load and store:\n"
<< " " << *LI << "\n" << " "
<< *SI << "\n"; } } while (false)
;
3953
3954 // We've found a store and load that we need to split
3955 // with mismatched relative splits. Just give up on them
3956 // and remove both instructions from our list of
3957 // candidates.
3958 UnsplittableLoads.insert(LI);
3959 return true;
3960 });
3961 // Now we have to go *back* through all the stores, because a later store may
3962 // have caused an earlier store's load to become unsplittable and if it is
3963 // unsplittable for the later store, then we can't rely on it being split in
3964 // the earlier store either.
3965 llvm::erase_if(Stores, [&UnsplittableLoads](StoreInst *SI) {
3966 auto *LI = cast<LoadInst>(SI->getValueOperand());
3967 return UnsplittableLoads.count(LI);
3968 });
3969 // Once we've established all the loads that can't be split for some reason,
3970 // filter any that made it into our list out.
3971 llvm::erase_if(Loads, [&UnsplittableLoads](LoadInst *LI) {
3972 return UnsplittableLoads.count(LI);
3973 });
3974
3975 // If no loads or stores are left, there is no pre-splitting to be done for
3976 // this alloca.
3977 if (Loads.empty() && Stores.empty())
3978 return false;
3979
3980 // From here on, we can't fail and will be building new accesses, so rig up
3981 // an IR builder.
3982 IRBuilderTy IRB(&AI);
3983
3984 // Collect the new slices which we will merge into the alloca slices.
3985 SmallVector<Slice, 4> NewSlices;
3986
3987 // Track any allocas we end up splitting loads and stores for so we iterate
3988 // on them.
3989 SmallPtrSet<AllocaInst *, 4> ResplitPromotableAllocas;
3990
3991 // At this point, we have collected all of the loads and stores we can
3992 // pre-split, and the specific splits needed for them. We actually do the
3993 // splitting in a specific order in order to handle when one of the loads in
3994 // the value operand to one of the stores.
3995 //
3996 // First, we rewrite all of the split loads, and just accumulate each split
3997 // load in a parallel structure. We also build the slices for them and append
3998 // them to the alloca slices.
3999 SmallDenseMap<LoadInst *, std::vector<LoadInst *>, 1> SplitLoadsMap;
4000 std::vector<LoadInst *> SplitLoads;
4001 const DataLayout &DL = AI.getModule()->getDataLayout();
4002 for (LoadInst *LI : Loads) {
4003 SplitLoads.clear();
4004
4005 IntegerType *Ty = cast<IntegerType>(LI->getType());
4006 assert(Ty->getBitWidth() % 8 == 0)(static_cast <bool> (Ty->getBitWidth() % 8 == 0) ? void
(0) : __assert_fail ("Ty->getBitWidth() % 8 == 0", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4006, __extension__ __PRETTY_FUNCTION__))
;
4007 uint64_t LoadSize = Ty->getBitWidth() / 8;
4008 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4008, __extension__ __PRETTY_FUNCTION__))
;
4009
4010 auto &Offsets = SplitOffsetsMap[LI];
4011 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4012, __extension__ __PRETTY_FUNCTION__))
4012 "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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4012, __extension__ __PRETTY_FUNCTION__))
;
4013 uint64_t BaseOffset = Offsets.S->beginOffset();
4014 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4015, __extension__ __PRETTY_FUNCTION__))
4015 "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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4015, __extension__ __PRETTY_FUNCTION__))
;
4016
4017 Instruction *BasePtr = cast<Instruction>(LI->getPointerOperand());
4018 IRB.SetInsertPoint(LI);
4019
4020 LLVM_DEBUG(dbgs() << " Splitting load: " << *LI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Splitting load: " << *LI
<< "\n"; } } while (false)
;
4021
4022 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
4023 int Idx = 0, Size = Offsets.Splits.size();
4024 for (;;) {
4025 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
4026 auto AS = LI->getPointerAddressSpace();
4027 auto *PartPtrTy = PartTy->getPointerTo(AS);
4028 LoadInst *PLoad = IRB.CreateAlignedLoad(
4029 PartTy,
4030 getAdjustedPtr(IRB, DL, BasePtr,
4031 APInt(DL.getIndexSizeInBits(AS), PartOffset),
4032 PartPtrTy, BasePtr->getName() + "."),
4033 getAdjustedAlignment(LI, PartOffset),
4034 /*IsVolatile*/ false, LI->getName());
4035 PLoad->copyMetadata(*LI, {LLVMContext::MD_mem_parallel_loop_access,
4036 LLVMContext::MD_access_group});
4037
4038 // Append this load onto the list of split loads so we can find it later
4039 // to rewrite the stores.
4040 SplitLoads.push_back(PLoad);
4041
4042 // Now build a new slice for the alloca.
4043 NewSlices.push_back(
4044 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
4045 &PLoad->getOperandUse(PLoad->getPointerOperandIndex()),
4046 /*IsSplittable*/ false));
4047 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)
4048 << ", " << NewSlices.back().endOffset()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " new slice [" << NewSlices
.back().beginOffset() << ", " << NewSlices.back()
.endOffset() << "): " << *PLoad << "\n"; } }
while (false)
4049 << "): " << *PLoad << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " new slice [" << NewSlices
.back().beginOffset() << ", " << NewSlices.back()
.endOffset() << "): " << *PLoad << "\n"; } }
while (false)
;
4050
4051 // See if we've handled all the splits.
4052 if (Idx >= Size)
4053 break;
4054
4055 // Setup the next partition.
4056 PartOffset = Offsets.Splits[Idx];
4057 ++Idx;
4058 PartSize = (Idx < Size ? Offsets.Splits[Idx] : LoadSize) - PartOffset;
4059 }
4060
4061 // Now that we have the split loads, do the slow walk over all uses of the
4062 // load and rewrite them as split stores, or save the split loads to use
4063 // below if the store is going to be split there anyways.
4064 bool DeferredStores = false;
4065 for (User *LU : LI->users()) {
4066 StoreInst *SI = cast<StoreInst>(LU);
4067 if (!Stores.empty() && SplitOffsetsMap.count(SI)) {
4068 DeferredStores = true;
4069 LLVM_DEBUG(dbgs() << " Deferred splitting of store: " << *SIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Deferred splitting of store: "
<< *SI << "\n"; } } while (false)
4070 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Deferred splitting of store: "
<< *SI << "\n"; } } while (false)
;
4071 continue;
4072 }
4073
4074 Value *StoreBasePtr = SI->getPointerOperand();
4075 IRB.SetInsertPoint(SI);
4076
4077 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)
;
4078
4079 for (int Idx = 0, Size = SplitLoads.size(); Idx < Size; ++Idx) {
4080 LoadInst *PLoad = SplitLoads[Idx];
4081 uint64_t PartOffset = Idx == 0 ? 0 : Offsets.Splits[Idx - 1];
4082 auto *PartPtrTy =
4083 PLoad->getType()->getPointerTo(SI->getPointerAddressSpace());
4084
4085 auto AS = SI->getPointerAddressSpace();
4086 StoreInst *PStore = IRB.CreateAlignedStore(
4087 PLoad,
4088 getAdjustedPtr(IRB, DL, StoreBasePtr,
4089 APInt(DL.getIndexSizeInBits(AS), PartOffset),
4090 PartPtrTy, StoreBasePtr->getName() + "."),
4091 getAdjustedAlignment(SI, PartOffset),
4092 /*IsVolatile*/ false);
4093 PStore->copyMetadata(*SI, {LLVMContext::MD_mem_parallel_loop_access,
4094 LLVMContext::MD_access_group});
4095 LLVM_DEBUG(dbgs() << " +" << PartOffset << ":" << *PStore << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " +" << PartOffset <<
":" << *PStore << "\n"; } } while (false)
;
4096 }
4097
4098 // We want to immediately iterate on any allocas impacted by splitting
4099 // this store, and we have to track any promotable alloca (indicated by
4100 // a direct store) as needing to be resplit because it is no longer
4101 // promotable.
4102 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(StoreBasePtr)) {
4103 ResplitPromotableAllocas.insert(OtherAI);
4104 Worklist.insert(OtherAI);
4105 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
4106 StoreBasePtr->stripInBoundsOffsets())) {
4107 Worklist.insert(OtherAI);
4108 }
4109
4110 // Mark the original store as dead.
4111 DeadInsts.push_back(SI);
4112 }
4113
4114 // Save the split loads if there are deferred stores among the users.
4115 if (DeferredStores)
4116 SplitLoadsMap.insert(std::make_pair(LI, std::move(SplitLoads)));
4117
4118 // Mark the original load as dead and kill the original slice.
4119 DeadInsts.push_back(LI);
4120 Offsets.S->kill();
4121 }
4122
4123 // Second, we rewrite all of the split stores. At this point, we know that
4124 // all loads from this alloca have been split already. For stores of such
4125 // loads, we can simply look up the pre-existing split loads. For stores of
4126 // other loads, we split those loads first and then write split stores of
4127 // them.
4128 for (StoreInst *SI : Stores) {
4129 auto *LI = cast<LoadInst>(SI->getValueOperand());
4130 IntegerType *Ty = cast<IntegerType>(LI->getType());
4131 assert(Ty->getBitWidth() % 8 == 0)(static_cast <bool> (Ty->getBitWidth() % 8 == 0) ? void
(0) : __assert_fail ("Ty->getBitWidth() % 8 == 0", "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4131, __extension__ __PRETTY_FUNCTION__))
;
4132 uint64_t StoreSize = Ty->getBitWidth() / 8;
4133 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4133, __extension__ __PRETTY_FUNCTION__))
;
4134
4135 auto &Offsets = SplitOffsetsMap[SI];
4136 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4137, __extension__ __PRETTY_FUNCTION__))
4137 "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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4137, __extension__ __PRETTY_FUNCTION__))
;
4138 uint64_t BaseOffset = Offsets.S->beginOffset();
4139 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4140, __extension__ __PRETTY_FUNCTION__))
4140 "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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4140, __extension__ __PRETTY_FUNCTION__))
;
4141
4142 Value *LoadBasePtr = LI->getPointerOperand();
4143 Instruction *StoreBasePtr = cast<Instruction>(SI->getPointerOperand());
4144
4145 LLVM_DEBUG(dbgs() << " Splitting store: " << *SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Splitting store: " << *SI
<< "\n"; } } while (false)
;
4146
4147 // Check whether we have an already split load.
4148 auto SplitLoadsMapI = SplitLoadsMap.find(LI);
4149 std::vector<LoadInst *> *SplitLoads = nullptr;
4150 if (SplitLoadsMapI != SplitLoadsMap.end()) {
4151 SplitLoads = &SplitLoadsMapI->second;
4152 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4153, __extension__ __PRETTY_FUNCTION__))
4153 "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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4153, __extension__ __PRETTY_FUNCTION__))
;
4154 } else {
4155 LLVM_DEBUG(dbgs() << " of load: " << *LI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " of load: " << *LI
<< "\n"; } } while (false)
;
4156 }
4157
4158 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
4159 int Idx = 0, Size = Offsets.Splits.size();
4160 for (;;) {
4161 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
4162 auto *LoadPartPtrTy = PartTy->getPointerTo(LI->getPointerAddressSpace());
4163 auto *StorePartPtrTy = PartTy->getPointerTo(SI->getPointerAddressSpace());
4164
4165 // Either lookup a split load or create one.
4166 LoadInst *PLoad;
4167 if (SplitLoads) {
4168 PLoad = (*SplitLoads)[Idx];
4169 } else {
4170 IRB.SetInsertPoint(LI);
4171 auto AS = LI->getPointerAddressSpace();
4172 PLoad = IRB.CreateAlignedLoad(
4173 PartTy,
4174 getAdjustedPtr(IRB, DL, LoadBasePtr,
4175 APInt(DL.getIndexSizeInBits(AS), PartOffset),
4176 LoadPartPtrTy, LoadBasePtr->getName() + "."),
4177 getAdjustedAlignment(LI, PartOffset),
4178 /*IsVolatile*/ false, LI->getName());
4179 PLoad->copyMetadata(*LI, {LLVMContext::MD_mem_parallel_loop_access,
4180 LLVMContext::MD_access_group});
4181 }
4182
4183 // And store this partition.
4184 IRB.SetInsertPoint(SI);
4185 auto AS = SI->getPointerAddressSpace();
4186 StoreInst *PStore = IRB.CreateAlignedStore(
4187 PLoad,
4188 getAdjustedPtr(IRB, DL, StoreBasePtr,
4189 APInt(DL.getIndexSizeInBits(AS), PartOffset),
4190 StorePartPtrTy, StoreBasePtr->getName() + "."),
4191 getAdjustedAlignment(SI, PartOffset),
4192 /*IsVolatile*/ false);
4193 PStore->copyMetadata(*SI, {LLVMContext::MD_mem_parallel_loop_access,
4194 LLVMContext::MD_access_group});
4195
4196 // Now build a new slice for the alloca.
4197 NewSlices.push_back(
4198 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
4199 &PStore->getOperandUse(PStore->getPointerOperandIndex()),
4200 /*IsSplittable*/ false));
4201 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)
4202 << ", " << NewSlices.back().endOffset()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " new slice [" << NewSlices
.back().beginOffset() << ", " << NewSlices.back()
.endOffset() << "): " << *PStore << "\n"; }
} while (false)
4203 << "): " << *PStore << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " new slice [" << NewSlices
.back().beginOffset() << ", " << NewSlices.back()
.endOffset() << "): " << *PStore << "\n"; }
} while (false)
;
4204 if (!SplitLoads) {
4205 LLVM_DEBUG(dbgs() << " of split load: " << *PLoad << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " of split load: " << *
PLoad << "\n"; } } while (false)
;
4206 }
4207
4208 // See if we've finished all the splits.
4209 if (Idx >= Size)
4210 break;
4211
4212 // Setup the next partition.
4213 PartOffset = Offsets.Splits[Idx];
4214 ++Idx;
4215 PartSize = (Idx < Size ? Offsets.Splits[Idx] : StoreSize) - PartOffset;
4216 }
4217
4218 // We want to immediately iterate on any allocas impacted by splitting
4219 // this load, which is only relevant if it isn't a load of this alloca and
4220 // thus we didn't already split the loads above. We also have to keep track
4221 // of any promotable allocas we split loads on as they can no longer be
4222 // promoted.
4223 if (!SplitLoads) {
4224 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(LoadBasePtr)) {
4225 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4225, __extension__ __PRETTY_FUNCTION__))
;
4226 ResplitPromotableAllocas.insert(OtherAI);
4227 Worklist.insert(OtherAI);
4228 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
4229 LoadBasePtr->stripInBoundsOffsets())) {
4230 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4230, __extension__ __PRETTY_FUNCTION__))
;
4231 Worklist.insert(OtherAI);
4232 }
4233 }
4234
4235 // Mark the original store as dead now that we've split it up and kill its
4236 // slice. Note that we leave the original load in place unless this store
4237 // was its only use. It may in turn be split up if it is an alloca load
4238 // for some other alloca, but it may be a normal load. This may introduce
4239 // redundant loads, but where those can be merged the rest of the optimizer
4240 // should handle the merging, and this uncovers SSA splits which is more
4241 // important. In practice, the original loads will almost always be fully
4242 // split and removed eventually, and the splits will be merged by any
4243 // trivial CSE, including instcombine.
4244 if (LI->hasOneUse()) {
4245 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!\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4245, __extension__ __PRETTY_FUNCTION__))
;
4246 DeadInsts.push_back(LI);
4247 }
4248 DeadInsts.push_back(SI);
4249 Offsets.S->kill();
4250 }
4251
4252 // Remove the killed slices that have ben pre-split.
4253 llvm::erase_if(AS, [](const Slice &S) { return S.isDead(); });
4254
4255 // Insert our new slices. This will sort and merge them into the sorted
4256 // sequence.
4257 AS.insert(NewSlices);
4258
4259 LLVM_DEBUG(dbgs() << " Pre-split slices:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Pre-split slices:\n"; } } while
(false)
;
4260#ifndef NDEBUG
4261 for (auto I = AS.begin(), E = AS.end(); I != E; ++I)
4262 LLVM_DEBUG(AS.print(dbgs(), I, " "))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { AS.print(dbgs(), I, " "); } } while (false)
;
4263#endif
4264
4265 // Finally, don't try to promote any allocas that new require re-splitting.
4266 // They have already been added to the worklist above.
4267 llvm::erase_if(PromotableAllocas, [&](AllocaInst *AI) {
4268 return ResplitPromotableAllocas.count(AI);
4269 });
4270
4271 return true;
4272}
4273
4274/// Rewrite an alloca partition's users.
4275///
4276/// This routine drives both of the rewriting goals of the SROA pass. It tries
4277/// to rewrite uses of an alloca partition to be conducive for SSA value
4278/// promotion. If the partition needs a new, more refined alloca, this will
4279/// build that new alloca, preserving as much type information as possible, and
4280/// rewrite the uses of the old alloca to point at the new one and have the
4281/// appropriate new offsets. It also evaluates how successful the rewrite was
4282/// at enabling promotion and if it was successful queues the alloca to be
4283/// promoted.
4284AllocaInst *SROA::rewritePartition(AllocaInst &AI, AllocaSlices &AS,
4285 Partition &P) {
4286 // Try to compute a friendly type for this partition of the alloca. This
4287 // won't always succeed, in which case we fall back to a legal integer type
4288 // or an i8 array of an appropriate size.
4289 Type *SliceTy = nullptr;
4290 const DataLayout &DL = AI.getModule()->getDataLayout();
4291 std::pair<Type *, IntegerType *> CommonUseTy =
4292 findCommonType(P.begin(), P.end(), P.endOffset());
4293 // Do all uses operate on the same type?
4294 if (CommonUseTy.first)
4295 if (DL.getTypeAllocSize(CommonUseTy.first).getFixedSize() >= P.size())
4296 SliceTy = CommonUseTy.first;
4297 // If not, can we find an appropriate subtype in the original allocated type?
4298 if (!SliceTy)
4299 if (Type *TypePartitionTy = getTypePartition(DL, AI.getAllocatedType(),
4300 P.beginOffset(), P.size()))
4301 SliceTy = TypePartitionTy;
4302 // If still not, can we use the largest bitwidth integer type used?
4303 if (!SliceTy && CommonUseTy.second)
4304 if (DL.getTypeAllocSize(CommonUseTy.second).getFixedSize() >= P.size())
4305 SliceTy = CommonUseTy.second;
4306 if ((!SliceTy || (SliceTy->isArrayTy() &&
4307 SliceTy->getArrayElementType()->isIntegerTy())) &&
4308 DL.isLegalInteger(P.size() * 8))
4309 SliceTy = Type::getIntNTy(*C, P.size() * 8);
4310 if (!SliceTy)
4311 SliceTy = ArrayType::get(Type::getInt8Ty(*C), P.size());
4312 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()"
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4312, __extension__ __PRETTY_FUNCTION__))
;
4313
4314 bool IsIntegerPromotable = isIntegerWideningViable(P, SliceTy, DL);
4315
4316 VectorType *VecTy =
4317 IsIntegerPromotable ? nullptr : isVectorPromotionViable(P, DL);
4318 if (VecTy)
4319 SliceTy = VecTy;
4320
4321 // Check for the case where we're going to rewrite to a new alloca of the
4322 // exact same type as the original, and with the same access offsets. In that
4323 // case, re-use the existing alloca, but still run through the rewriter to
4324 // perform phi and select speculation.
4325 // P.beginOffset() can be non-zero even with the same type in a case with
4326 // out-of-bounds access (e.g. @PR35657 function in SROA/basictest.ll).
4327 AllocaInst *NewAI;
4328 if (SliceTy == AI.getAllocatedType() && P.beginOffset() == 0) {
4329 NewAI = &AI;
4330 // FIXME: We should be able to bail at this point with "nothing changed".
4331 // FIXME: We might want to defer PHI speculation until after here.
4332 // FIXME: return nullptr;
4333 } else {
4334 // Make sure the alignment is compatible with P.beginOffset().
4335 const Align Alignment = commonAlignment(AI.getAlign(), P.beginOffset());
4336 // If we will get at least this much alignment from the type alone, leave
4337 // the alloca's alignment unconstrained.
4338 const bool IsUnconstrained = Alignment <= DL.getABITypeAlign(SliceTy);
4339 NewAI = new AllocaInst(
4340 SliceTy, AI.getType()->getAddressSpace(), nullptr,
4341 IsUnconstrained ? DL.getPrefTypeAlign(SliceTy) : Alignment,
4342 AI.getName() + ".sroa." + Twine(P.begin() - AS.begin()), &AI);
4343 // Copy the old AI debug location over to the new one.
4344 NewAI->setDebugLoc(AI.getDebugLoc());
4345 ++NumNewAllocas;
4346 }
4347
4348 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)
4349 << "[" << P.beginOffset() << "," << P.endOffset()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "Rewriting alloca partition " <<
"[" << P.beginOffset() << "," << P.endOffset
() << ") to: " << *NewAI << "\n"; } } while
(false)
4350 << ") to: " << *NewAI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "Rewriting alloca partition " <<
"[" << P.beginOffset() << "," << P.endOffset
() << ") to: " << *NewAI << "\n"; } } while
(false)
;
4351
4352 // Track the high watermark on the worklist as it is only relevant for
4353 // promoted allocas. We will reset it to this point if the alloca is not in
4354 // fact scheduled for promotion.
4355 unsigned PPWOldSize = PostPromotionWorklist.size();
4356 unsigned NumUses = 0;
4357 SmallSetVector<PHINode *, 8> PHIUsers;
4358 SmallSetVector<SelectInst *, 8> SelectUsers;
4359
4360 AllocaSliceRewriter Rewriter(DL, AS, *this, AI, *NewAI, P.beginOffset(),
4361 P.endOffset(), IsIntegerPromotable, VecTy,
4362 PHIUsers, SelectUsers);
4363 bool Promotable = true;
4364 for (Slice *S : P.splitSliceTails()) {
4365 Promotable &= Rewriter.visit(S);
4366 ++NumUses;
4367 }
4368 for (Slice &S : P) {
4369 Promotable &= Rewriter.visit(&S);
4370 ++NumUses;
4371 }
4372
4373 NumAllocaPartitionUses += NumUses;
4374 MaxUsesPerAllocaPartition.updateMax(NumUses);
4375
4376 // Now that we've processed all the slices in the new partition, check if any
4377 // PHIs or Selects would block promotion.
4378 for (PHINode *PHI : PHIUsers)
4379 if (!isSafePHIToSpeculate(*PHI)) {
4380 Promotable = false;
4381 PHIUsers.clear();
4382 SelectUsers.clear();
4383 break;
4384 }
4385
4386 for (SelectInst *Sel : SelectUsers)
4387 if (!isSafeSelectToSpeculate(*Sel)) {
4388 Promotable = false;
4389 PHIUsers.clear();
4390 SelectUsers.clear();
4391 break;
4392 }
4393
4394 if (Promotable) {
4395 for (Use *U : AS.getDeadUsesIfPromotable()) {
4396 auto *OldInst = dyn_cast<Instruction>(U->get());
4397 Value::dropDroppableUse(*U);
4398 if (OldInst)
4399 if (isInstructionTriviallyDead(OldInst))
4400 DeadInsts.push_back(OldInst);
4401 }
4402 if (PHIUsers.empty() && SelectUsers.empty()) {
4403 // Promote the alloca.
4404 PromotableAllocas.push_back(NewAI);
4405 } else {
4406 // If we have either PHIs or Selects to speculate, add them to those
4407 // worklists and re-queue the new alloca so that we promote in on the
4408 // next iteration.
4409 for (PHINode *PHIUser : PHIUsers)
4410 SpeculatablePHIs.insert(PHIUser);
4411 for (SelectInst *SelectUser : SelectUsers)
4412 SpeculatableSelects.insert(SelectUser);
4413 Worklist.insert(NewAI);
4414 }
4415 } else {
4416 // Drop any post-promotion work items if promotion didn't happen.
4417 while (PostPromotionWorklist.size() > PPWOldSize)
4418 PostPromotionWorklist.pop_back();
4419
4420 // We couldn't promote and we didn't create a new partition, nothing
4421 // happened.
4422 if (NewAI == &AI)
4423 return nullptr;
4424
4425 // If we can't promote the alloca, iterate on it to check for new
4426 // refinements exposed by splitting the current alloca. Don't iterate on an
4427 // alloca which didn't actually change and didn't get promoted.
4428 Worklist.insert(NewAI);
4429 }
4430
4431 return NewAI;
4432}
4433
4434/// Walks the slices of an alloca and form partitions based on them,
4435/// rewriting each of their uses.
4436bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &AS) {
4437 if (AS.begin() == AS.end())
4438 return false;
4439
4440 unsigned NumPartitions = 0;
4441 bool Changed = false;
4442 const DataLayout &DL = AI.getModule()->getDataLayout();
4443
4444 // First try to pre-split loads and stores.
4445 Changed |= presplitLoadsAndStores(AI, AS);
4446
4447 // Now that we have identified any pre-splitting opportunities,
4448 // mark loads and stores unsplittable except for the following case.
4449 // We leave a slice splittable if all other slices are disjoint or fully
4450 // included in the slice, such as whole-alloca loads and stores.
4451 // If we fail to split these during pre-splitting, we want to force them
4452 // to be rewritten into a partition.
4453 bool IsSorted = true;
4454
4455 uint64_t AllocaSize =
4456 DL.getTypeAllocSize(AI.getAllocatedType()).getFixedSize();
4457 const uint64_t MaxBitVectorSize = 1024;
4458 if (AllocaSize <= MaxBitVectorSize) {
4459 // If a byte boundary is included in any load or store, a slice starting or
4460 // ending at the boundary is not splittable.
4461 SmallBitVector SplittableOffset(AllocaSize + 1, true);
4462 for (Slice &S : AS)
4463 for (unsigned O = S.beginOffset() + 1;
4464 O < S.endOffset() && O < AllocaSize; O++)
4465 SplittableOffset.reset(O);
4466
4467 for (Slice &S : AS) {
4468 if (!S.isSplittable())
4469 continue;
4470
4471 if ((S.beginOffset() > AllocaSize || SplittableOffset[S.beginOffset()]) &&
4472 (S.endOffset() > AllocaSize || SplittableOffset[S.endOffset()]))
4473 continue;
4474
4475 if (isa<LoadInst>(S.getUse()->getUser()) ||
4476 isa<StoreInst>(S.getUse()->getUser())) {
4477 S.makeUnsplittable();
4478 IsSorted = false;
4479 }
4480 }
4481 }
4482 else {
4483 // We only allow whole-alloca splittable loads and stores
4484 // for a large alloca to avoid creating too large BitVector.
4485 for (Slice &S : AS) {
4486 if (!S.isSplittable())
4487 continue;
4488
4489 if (S.beginOffset() == 0 && S.endOffset() >= AllocaSize)
4490 continue;
4491
4492 if (isa<LoadInst>(S.getUse()->getUser()) ||
4493 isa<StoreInst>(S.getUse()->getUser())) {
4494 S.makeUnsplittable();
4495 IsSorted = false;
4496 }
4497 }
4498 }
4499
4500 if (!IsSorted)
4501 llvm::sort(AS);
4502
4503 /// Describes the allocas introduced by rewritePartition in order to migrate
4504 /// the debug info.
4505 struct Fragment {
4506 AllocaInst *Alloca;
4507 uint64_t Offset;
4508 uint64_t Size;
4509 Fragment(AllocaInst *AI, uint64_t O, uint64_t S)
4510 : Alloca(AI), Offset(O), Size(S) {}
4511 };
4512 SmallVector<Fragment, 4> Fragments;
4513
4514 // Rewrite each partition.
4515 for (auto &P : AS.partitions()) {
4516 if (AllocaInst *NewAI = rewritePartition(AI, AS, P)) {
4517 Changed = true;
4518 if (NewAI != &AI) {
4519 uint64_t SizeOfByte = 8;
4520 uint64_t AllocaSize =
4521 DL.getTypeSizeInBits(NewAI->getAllocatedType()).getFixedSize();
4522 // Don't include any padding.
4523 uint64_t Size = std::min(AllocaSize, P.size() * SizeOfByte);
4524 Fragments.push_back(Fragment(NewAI, P.beginOffset() * SizeOfByte, Size));
4525 }
4526 }
4527 ++NumPartitions;
4528 }
4529
4530 NumAllocaPartitions += NumPartitions;
4531 MaxPartitionsPerAlloca.updateMax(NumPartitions);
4532
4533 // Migrate debug information from the old alloca to the new alloca(s)
4534 // and the individual partitions.
4535 TinyPtrVector<DbgVariableIntrinsic *> DbgDeclares = FindDbgAddrUses(&AI);
4536 for (DbgVariableIntrinsic *DbgDeclare : DbgDeclares) {
4537 auto *Expr = DbgDeclare->getExpression();
4538 DIBuilder DIB(*AI.getModule(), /*AllowUnresolved*/ false);
4539 uint64_t AllocaSize =
4540 DL.getTypeSizeInBits(AI.getAllocatedType()).getFixedSize();
4541 for (auto Fragment : Fragments) {
4542 // Create a fragment expression describing the new partition or reuse AI's
4543 // expression if there is only one partition.
4544 auto *FragmentExpr = Expr;
4545 if (Fragment.Size < AllocaSize || Expr->isFragment()) {
4546 // If this alloca is already a scalar replacement of a larger aggregate,
4547 // Fragment.Offset describes the offset inside the scalar.
4548 auto ExprFragment = Expr->getFragmentInfo();
4549 uint64_t Offset = ExprFragment ? ExprFragment->OffsetInBits : 0;
4550 uint64_t Start = Offset + Fragment.Offset;
4551 uint64_t Size = Fragment.Size;
4552 if (ExprFragment) {
4553 uint64_t AbsEnd =
4554 ExprFragment->OffsetInBits + ExprFragment->SizeInBits;
4555 if (Start >= AbsEnd)
4556 // No need to describe a SROAed padding.
4557 continue;
4558 Size = std::min(Size, AbsEnd - Start);
4559 }
4560 // The new, smaller fragment is stenciled out from the old fragment.
4561 if (auto OrigFragment = FragmentExpr->getFragmentInfo()) {
4562 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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4563, __extension__ __PRETTY_FUNCTION__))
4563 "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\""
, "/build/llvm-toolchain-snapshot-13~++20210613111130+5be314f79ba7/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4563, __extension__ __PRETTY_FUNCTION__))
;
4564 Start -= OrigFragment->OffsetInBits;
4565 }
4566
4567 // The alloca may be larger than the variable.
4568 auto VarSize = DbgDeclare->getVariable()->getSizeInBits();
4569 if (VarSize) {
4570 if (Size > *VarSize)
4571 Size = *VarSize;
4572 if (Size == 0 || Start + Size > *VarSize)
4573 continue;
4574 }
4575
4576 // Avoid creating a fragment expression that covers the entire variable.
4577 if (!VarSize || *VarSize != Size) {
4578 if (auto E =
4579 DIExpression::createFragmentExpression(Expr, Start, Size))
4580 FragmentExpr = *E;
4581 else
4582 continue;
4583 }
4584 }
4585
4586 // Remove any existing intrinsics on the new alloca describing
4587 // the variable fragment.
4588 for (DbgVariableIntrinsic *OldDII : FindDbgAddrUses(Fragment.Alloca)) {
4589 auto SameVariableFragment = [](const DbgVariableIntrinsic *LHS,
4590 const DbgVariableIntrinsic *RHS) {
4591 return LHS->getVariable() == RHS->getVariable() &&
4592 LHS->getDebugLoc()->getInlinedAt() ==
4593 RHS->getDebugLoc()->getInlinedAt();
4594 };
4595 if (SameVariableFragment(OldDII, DbgDeclare))
4596 OldDII->eraseFromParent();
4597 }
4598
4599 DIB.insertDeclare(Fragment.Alloca, DbgDeclare->getVariable(), FragmentExpr,
4600 DbgDeclare->getDebugLoc(), &AI);
4601 }
4602 }
4603 return Changed;
4604}
4605
4606/// Clobber a use with undef, deleting the used value if it becomes dead.
4607void SROA::clobberUse(Use &U) {
4608 Value *OldV = U;
4609 // Replace the use with an undef value.
4610 U = UndefValue::get(OldV->getType());
4611
4612 // Check for this making an instruction dead. We have to garbage collect
4613 // all the dead instructions to ensure the uses of any alloca end up being
4614 // minimal.
4615 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
4616 if (isInstructionTriviallyDead(OldI)) {
4617 DeadInsts.push_back(OldI);
4618 }
4619}
4620
4621/// Analyze an alloca for SROA.
4622///
4623/// This analyzes the alloca to ensure we can reason about it, builds
4624/// the slices of the alloca, and then hands it off to be split and
4625/// rewritten as needed.
4626bool SROA::runOnAlloca(AllocaInst &AI) {
4627 LLVM_DEBUG(dbgs() << "SROA alloca: " << AI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "SROA alloca: " << AI <<
"\n"; } } while (false)
;
4628 ++NumAllocasAnalyzed;
4629
4630 // Special case dead allocas, as they're trivial.
4631 if (AI.use_empty()) {
4632 AI.eraseFromParent();
4633 return true;
4634 }
4635 const DataLayout &DL = AI.getModule()->getDataLayout();
4636
4637 // Skip alloca forms that this analysis can't handle.
4638 auto *AT = AI.getAllocatedType();
4639 if (AI.isArrayAllocation() || !AT->isSized() || isa<ScalableVectorType>(AT) ||
4640 DL.getTypeAllocSize(AT).getFixedSize() == 0)
4641 return false;
4642
4643 bool Changed = false;
4644
4645 // First, split any FCA loads and stores touching this alloca to promote
4646 // better splitting and promotion opportunities.
4647 AggLoadStoreRewriter AggRewriter(DL);
4648 Changed |= AggRewriter.rewrite(AI);
4649
4650 // Build the slices using a recursive instruction-visiting builder.
4651 AllocaSlices AS(DL, AI);
4652 LLVM_DEBUG(AS.print(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { AS.print(dbgs()); } } while (false)
;
4653 if (AS.isEscaped())
4654 return Changed;
4655
4656 // Delete all the dead users of this alloca before splitting and rewriting it.
4657 for (Instruction *DeadUser : AS.getDeadUsers()) {
4658 // Free up everything used by this instruction.
4659 for (Use &DeadOp : DeadUser->operands())
4660 clobberUse(DeadOp);
4661
4662 // Now replace the uses of this instruction.
4663 DeadUser->replaceAllUsesWith(UndefValue::get(DeadUser->getType()));
4664
4665 // And mark it for deletion.
4666 DeadInsts.push_back(DeadUser);
4667 Changed = true;
4668 }
4669 for (Use *DeadOp : AS.getDeadOperands()) {
4670 clobberUse(*DeadOp);
4671 Changed = true;
4672 }
4673
4674 // No slices to split. Leave the dead alloca for a later pass to clean up.
4675 if (AS.begin() == AS.end())
4676 return Changed;
4677
4678 Changed |= splitAlloca(AI, AS);
4679
4680 LLVM_DEBUG(dbgs() << " Speculating PHIs\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Speculating PHIs\n"; } } while
(false)
;
4681 while (!SpeculatablePHIs.empty())
4682 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
4683
4684 LLVM_DEBUG(dbgs() << " Speculating Selects\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Speculating Selects\n"; } } while
(false)
;
4685 while (!SpeculatableSelects.empty())
4686 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
4687
4688 return Changed;
4689}
4690
4691/// Delete the dead instructions accumulated in this run.
4692///
4693/// Recursively deletes the dead instructions we've accumulated. This is done
4694/// at the very end to maximize locality of the recursive delete and to
4695/// minimize the problems of invalidated instruction pointers as such pointers
4696/// are used heavily in the intermediate stages of the algorithm.
4697///
4698/// We also record the alloca instructions deleted here so that they aren't
4699/// subsequently handed to mem2reg to promote.
4700bool SROA::deleteDeadInstructions(
4701 SmallPtrSetImpl<AllocaInst *> &DeletedAllocas) {
4702 bool Changed = false;
4703 while (!DeadInsts.empty()) {
4704 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
4705 if (!I) continue;
4706 LLVM_DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "Deleting dead instruction: " <<
*I << "\n"; } } while (false)
;
4707
4708 // If the instruction is an alloca, find the possible dbg.declare connected
4709 // to it, and remove it too. We must do this before calling RAUW or we will
4710 // not be able to find it.
4711 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
4712 DeletedAllocas.insert(AI);
4713 for (DbgVariableIntrinsic *OldDII : FindDbgAddrUses(AI))
4714 OldDII->eraseFromParent();
4715 }
4716
4717 I->replaceAllUsesWith(UndefValue::get(I->getType()));
4718
4719 for (Use &Operand : I->operands())
4720 if (Instruction *U = dyn_cast<Instruction>(Operand)) {
4721 // Zero out the operand and see if it becomes trivially dead.
4722 Operand = nullptr;
4723 if (isInstructionTriviallyDead(U))
4724 DeadInsts.push_back(U);
4725 }
4726
4727 ++NumDeleted;
4728 I->eraseFromParent();
4729 Changed = true;
4730 }
4731 return Changed;
4732}
4733
4734/// Promote the allocas, using the best available technique.
4735///
4736/// This attempts to promote whatever allocas have been identified as viable in
4737/// the PromotableAllocas list. If that list is empty, there is nothing to do.
4738/// This function returns whether any promotion occurred.
4739bool SROA::promoteAllocas(Function &F) {
4740 if (PromotableAllocas.empty())
4741 return false;
4742
4743 NumPromoted += PromotableAllocas.size();
4744
4745 LLVM_DEBUG(dbgs() << "Promoting allocas with mem2reg...\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "Promoting allocas with mem2reg...\n"
; } } while (false)
;
4746 PromoteMemToReg(PromotableAllocas, *DT, AC);
4747 PromotableAllocas.clear();
4748 return true;
4749}
4750
4751PreservedAnalyses SROA::runImpl(Function &F, DominatorTree &RunDT,
4752 AssumptionCache &RunAC) {
4753 LLVM_DEBUG(dbgs() << "SROA function: " << F.getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "SROA function: " << F.getName
() << "\n"; } } while (false)
;
4754 C = &F.getContext();
4755 DT = &RunDT;
4756 AC = &RunAC;
4757
4758 BasicBlock &EntryBB = F.getEntryBlock();
4759 for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
4760 I != E; ++I) {
4761 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
4762 if (isa<ScalableVectorType>(AI->getAllocatedType())) {
4763 if (isAllocaPromotable(AI))
4764 PromotableAllocas.push_back(AI);
4765 } else {
4766 Worklist.insert(AI);
4767 }
4768 }
4769 }
4770
4771 bool Changed = false;
4772 // A set of deleted alloca instruction pointers which should be removed from
4773 // the list of promotable allocas.
4774 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
4775
4776 do {
4777 while (!Worklist.empty()) {
4778 Changed |= runOnAlloca(*Worklist.pop_back_val());
4779 Changed |= deleteDeadInstructions(DeletedAllocas);
4780
4781 // Remove the deleted allocas from various lists so that we don't try to
4782 // continue processing them.
4783 if (!DeletedAllocas.empty()) {
4784 auto IsInSet = [&](AllocaInst *AI) { return DeletedAllocas.count(AI); };
4785 Worklist.remove_if(IsInSet);
4786 PostPromotionWorklist.remove_if(IsInSet);
4787 llvm::erase_if(PromotableAllocas, IsInSet);
4788 DeletedAllocas.clear();
4789 }
4790 }
4791
4792 Changed |= promoteAllocas(F);
4793
4794 Worklist = PostPromotionWorklist;
4795 PostPromotionWorklist.clear();
4796 } while (!Worklist.empty());
4797
4798 if (!Changed)
4799 return PreservedAnalyses::all();
4800
4801 PreservedAnalyses PA;
4802 PA.preserveSet<CFGAnalyses>();
4803 return PA;
4804}
4805
4806PreservedAnalyses SROA::run(Function &F, FunctionAnalysisManager &AM) {
4807 return runImpl(F, AM.getResult<DominatorTreeAnalysis>(F),
4808 AM.getResult<AssumptionAnalysis>(F));
4809}
4810
4811/// A legacy pass for the legacy pass manager that wraps the \c SROA pass.
4812///
4813/// This is in the llvm namespace purely to allow it to be a friend of the \c
4814/// SROA pass.
4815class llvm::sroa::SROALegacyPass : public FunctionPass {
4816 /// The SROA implementation.
4817 SROA Impl;
4818
4819public:
4820 static char ID;
4821
4822 SROALegacyPass() : FunctionPass(ID) {
4823 initializeSROALegacyPassPass(*PassRegistry::getPassRegistry());
4824 }
4825
4826 bool runOnFunction(Function &F) override {
4827 if (skipFunction(F))
4828 return false;
4829
4830 auto PA = Impl.runImpl(
4831 F, getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
4832 getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F));
4833 return !PA.areAllPreserved();
4834 }
4835
4836 void getAnalysisUsage(AnalysisUsage &AU) const override {
4837 AU.addRequired<AssumptionCacheTracker>();
4838 AU.addRequired<DominatorTreeWrapperPass>();
4839 AU.addPreserved<GlobalsAAWrapperPass>();
4840 AU.setPreservesCFG();
4841 }
4842
4843 StringRef getPassName() const override { return "SROA"; }
4844};
4845
4846char SROALegacyPass::ID = 0;
4847
4848FunctionPass *llvm::createSROAPass() { return new SROALegacyPass(); }
4849
4850INITIALIZE_PASS_BEGIN(SROALegacyPass, "sroa",static void *initializeSROALegacyPassPassOnce(PassRegistry &
Registry) {
4851 "Scalar Replacement Of Aggregates", false, false)static void *initializeSROALegacyPassPassOnce(PassRegistry &
Registry) {
4852INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry);
4853INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);
4854INITIALIZE_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)); }
4855 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)); }