Bug Summary

File:llvm/lib/Transforms/Scalar/SROA.cpp
Warning:line 2155, column 3
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 -fno-split-dwarf-inlining -debugger-tuning=gdb -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-12/lib/clang/12.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/build-llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/build-llvm/include -I /build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-12/lib/clang/12.0.0/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-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/build-llvm/lib/Transforms/Scalar -fdebug-prefix-map=/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb=. -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 -o /tmp/scan-build-2020-09-26-161721-17566-1 -x c++ /build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/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!")((BeginOffset < EndOffset && "Partitions must span some bytes!"
) ? static_cast<void> (0) : __assert_fail ("BeginOffset < EndOffset && \"Partitions must span some bytes!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 389, __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()) &&(((P.SI != SE || !P.SplitTails.empty()) && "Cannot advance past the end of the slices!"
) ? static_cast<void> (0) : __assert_fail ("(P.SI != SE || !P.SplitTails.empty()) && \"Cannot advance past the end of the slices!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 458, __PRETTY_FUNCTION__))
458 "Cannot advance past the end of the slices!")(((P.SI != SE || !P.SplitTails.empty()) && "Cannot advance past the end of the slices!"
) ? static_cast<void> (0) : __assert_fail ("(P.SI != SE || !P.SplitTails.empty()) && \"Cannot advance past the end of the slices!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 458, __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 P.SplitTails.erase(llvm::remove_if(P.SplitTails,
471 [&](Slice *S) {
472 return S->endOffset() <=
473 P.EndOffset;
474 }),
475 P.SplitTails.end());
476 assert(llvm::any_of(P.SplitTails,((llvm::any_of(P.SplitTails, [&](Slice *S) { return S->
endOffset() == MaxSplitSliceEndOffset; }) && "Could not find the current max split slice offset!"
) ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 480, __PRETTY_FUNCTION__))
477 [&](Slice *S) {((llvm::any_of(P.SplitTails, [&](Slice *S) { return S->
endOffset() == MaxSplitSliceEndOffset; }) && "Could not find the current max split slice offset!"
) ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 480, __PRETTY_FUNCTION__))
478 return S->endOffset() == MaxSplitSliceEndOffset;((llvm::any_of(P.SplitTails, [&](Slice *S) { return S->
endOffset() == MaxSplitSliceEndOffset; }) && "Could not find the current max split slice offset!"
) ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 480, __PRETTY_FUNCTION__))
479 }) &&((llvm::any_of(P.SplitTails, [&](Slice *S) { return S->
endOffset() == MaxSplitSliceEndOffset; }) && "Could not find the current max split slice offset!"
) ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 480, __PRETTY_FUNCTION__))
480 "Could not find the current max split slice offset!")((llvm::any_of(P.SplitTails, [&](Slice *S) { return S->
endOffset() == MaxSplitSliceEndOffset; }) && "Could not find the current max split slice offset!"
) ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 480, __PRETTY_FUNCTION__))
;
481 assert(llvm::all_of(P.SplitTails,((llvm::all_of(P.SplitTails, [&](Slice *S) { return S->
endOffset() <= MaxSplitSliceEndOffset; }) && "Max split slice end offset is not actually the max!"
) ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 485, __PRETTY_FUNCTION__))
482 [&](Slice *S) {((llvm::all_of(P.SplitTails, [&](Slice *S) { return S->
endOffset() <= MaxSplitSliceEndOffset; }) && "Max split slice end offset is not actually the max!"
) ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 485, __PRETTY_FUNCTION__))
483 return S->endOffset() <= MaxSplitSliceEndOffset;((llvm::all_of(P.SplitTails, [&](Slice *S) { return S->
endOffset() <= MaxSplitSliceEndOffset; }) && "Max split slice end offset is not actually the max!"
) ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 485, __PRETTY_FUNCTION__))
484 }) &&((llvm::all_of(P.SplitTails, [&](Slice *S) { return S->
endOffset() <= MaxSplitSliceEndOffset; }) && "Max split slice end offset is not actually the max!"
) ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 485, __PRETTY_FUNCTION__))
485 "Max split slice end offset is not actually the max!")((llvm::all_of(P.SplitTails, [&](Slice *S) { return S->
endOffset() <= MaxSplitSliceEndOffset; }) && "Max split slice end offset is not actually the max!"
) ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 485, __PRETTY_FUNCTION__))
;
486 }
487 }
488
489 // If P.SI is already at the end, then we've cleared the split tail and
490 // now have an end iterator.
491 if (P.SI == SE) {
492 assert(P.SplitTails.empty() && "Failed to clear the split slices!")((P.SplitTails.empty() && "Failed to clear the split slices!"
) ? static_cast<void> (0) : __assert_fail ("P.SplitTails.empty() && \"Failed to clear the split slices!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 492, __PRETTY_FUNCTION__))
;
493 return;
494 }
495
496 // If we had a non-empty partition previously, set up the state for
497 // subsequent partitions.
498 if (P.SI != P.SJ) {
499 // Accumulate all the splittable slices which started in the old
500 // partition into the split list.
501 for (Slice &S : P)
502 if (S.isSplittable() && S.endOffset() > P.EndOffset) {
503 P.SplitTails.push_back(&S);
504 MaxSplitSliceEndOffset =
505 std::max(S.endOffset(), MaxSplitSliceEndOffset);
506 }
507
508 // Start from the end of the previous partition.
509 P.SI = P.SJ;
510
511 // If P.SI is now at the end, we at most have a tail of split slices.
512 if (P.SI == SE) {
513 P.BeginOffset = P.EndOffset;
514 P.EndOffset = MaxSplitSliceEndOffset;
515 return;
516 }
517
518 // If the we have split slices and the next slice is after a gap and is
519 // not splittable immediately form an empty partition for the split
520 // slices up until the next slice begins.
521 if (!P.SplitTails.empty() && P.SI->beginOffset() != P.EndOffset &&
522 !P.SI->isSplittable()) {
523 P.BeginOffset = P.EndOffset;
524 P.EndOffset = P.SI->beginOffset();
525 return;
526 }
527 }
528
529 // OK, we need to consume new slices. Set the end offset based on the
530 // current slice, and step SJ past it. The beginning offset of the
531 // partition is the beginning offset of the next slice unless we have
532 // pre-existing split slices that are continuing, in which case we begin
533 // at the prior end offset.
534 P.BeginOffset = P.SplitTails.empty() ? P.SI->beginOffset() : P.EndOffset;
535 P.EndOffset = P.SI->endOffset();
536 ++P.SJ;
537
538 // There are two strategies to form a partition based on whether the
539 // partition starts with an unsplittable slice or a splittable slice.
540 if (!P.SI->isSplittable()) {
541 // When we're forming an unsplittable region, it must always start at
542 // the first slice and will extend through its end.
543 assert(P.BeginOffset == P.SI->beginOffset())((P.BeginOffset == P.SI->beginOffset()) ? static_cast<void
> (0) : __assert_fail ("P.BeginOffset == P.SI->beginOffset()"
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 543, __PRETTY_FUNCTION__))
;
544
545 // Form a partition including all of the overlapping slices with this
546 // unsplittable slice.
547 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
548 if (!P.SJ->isSplittable())
549 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
550 ++P.SJ;
551 }
552
553 // We have a partition across a set of overlapping unsplittable
554 // partitions.
555 return;
556 }
557
558 // If we're starting with a splittable slice, then we need to form
559 // a synthetic partition spanning it and any other overlapping splittable
560 // splices.
561 assert(P.SI->isSplittable() && "Forming a splittable partition!")((P.SI->isSplittable() && "Forming a splittable partition!"
) ? static_cast<void> (0) : __assert_fail ("P.SI->isSplittable() && \"Forming a splittable partition!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 561, __PRETTY_FUNCTION__))
;
562
563 // Collect all of the overlapping splittable slices.
564 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset &&
565 P.SJ->isSplittable()) {
566 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
567 ++P.SJ;
568 }
569
570 // Back upiP.EndOffset if we ended the span early when encountering an
571 // unsplittable slice. This synthesizes the early end offset of
572 // a partition spanning only splittable slices.
573 if (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
574 assert(!P.SJ->isSplittable())((!P.SJ->isSplittable()) ? static_cast<void> (0) : __assert_fail
("!P.SJ->isSplittable()", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 574, __PRETTY_FUNCTION__))
;
575 P.EndOffset = P.SJ->beginOffset();
576 }
577 }
578
579public:
580 bool operator==(const partition_iterator &RHS) const {
581 assert(SE == RHS.SE &&((SE == RHS.SE && "End iterators don't match between compared partition iterators!"
) ? static_cast<void> (0) : __assert_fail ("SE == RHS.SE && \"End iterators don't match between compared partition iterators!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 582, __PRETTY_FUNCTION__))
582 "End iterators don't match between compared partition iterators!")((SE == RHS.SE && "End iterators don't match between compared partition iterators!"
) ? static_cast<void> (0) : __assert_fail ("SE == RHS.SE && \"End iterators don't match between compared partition iterators!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 582, __PRETTY_FUNCTION__))
;
583
584 // The observed positions of partitions is marked by the P.SI iterator and
585 // the emptiness of the split slices. The latter is only relevant when
586 // P.SI == SE, as the end iterator will additionally have an empty split
587 // slices list, but the prior may have the same P.SI and a tail of split
588 // slices.
589 if (P.SI == RHS.P.SI && P.SplitTails.empty() == RHS.P.SplitTails.empty()) {
590 assert(P.SJ == RHS.P.SJ &&((P.SJ == RHS.P.SJ && "Same set of slices formed two different sized partitions!"
) ? static_cast<void> (0) : __assert_fail ("P.SJ == RHS.P.SJ && \"Same set of slices formed two different sized partitions!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 591, __PRETTY_FUNCTION__))
591 "Same set of slices formed two different sized partitions!")((P.SJ == RHS.P.SJ && "Same set of slices formed two different sized partitions!"
) ? static_cast<void> (0) : __assert_fail ("P.SJ == RHS.P.SJ && \"Same set of slices formed two different sized partitions!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 591, __PRETTY_FUNCTION__))
;
592 assert(P.SplitTails.size() == RHS.P.SplitTails.size() &&((P.SplitTails.size() == RHS.P.SplitTails.size() && "Same slice position with differently sized non-empty split "
"slice tails!") ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 594, __PRETTY_FUNCTION__))
593 "Same slice position with differently sized non-empty split "((P.SplitTails.size() == RHS.P.SplitTails.size() && "Same slice position with differently sized non-empty split "
"slice tails!") ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 594, __PRETTY_FUNCTION__))
594 "slice tails!")((P.SplitTails.size() == RHS.P.SplitTails.size() && "Same slice position with differently sized non-empty split "
"slice tails!") ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 594, __PRETTY_FUNCTION__))
;
595 return true;
596 }
597 return false;
598 }
599
600 partition_iterator &operator++() {
601 advance();
602 return *this;
603 }
604
605 Partition &operator*() { return P; }
606};
607
608/// A forward range over the partitions of the alloca's slices.
609///
610/// This accesses an iterator range over the partitions of the alloca's
611/// slices. It computes these partitions on the fly based on the overlapping
612/// offsets of the slices and the ability to split them. It will visit "empty"
613/// partitions to cover regions of the alloca only accessed via split
614/// slices.
615iterator_range<AllocaSlices::partition_iterator> AllocaSlices::partitions() {
616 return make_range(partition_iterator(begin(), end()),
617 partition_iterator(end(), end()));
618}
619
620static Value *foldSelectInst(SelectInst &SI) {
621 // If the condition being selected on is a constant or the same value is
622 // being selected between, fold the select. Yes this does (rarely) happen
623 // early on.
624 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
625 return SI.getOperand(1 + CI->isZero());
626 if (SI.getOperand(1) == SI.getOperand(2))
627 return SI.getOperand(1);
628
629 return nullptr;
630}
631
632/// A helper that folds a PHI node or a select.
633static Value *foldPHINodeOrSelectInst(Instruction &I) {
634 if (PHINode *PN = dyn_cast<PHINode>(&I)) {
635 // If PN merges together the same value, return that value.
636 return PN->hasConstantValue();
637 }
638 return foldSelectInst(cast<SelectInst>(I));
639}
640
641/// Builder for the alloca slices.
642///
643/// This class builds a set of alloca slices by recursively visiting the uses
644/// of an alloca and making a slice for each load and store at each offset.
645class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
646 friend class PtrUseVisitor<SliceBuilder>;
647 friend class InstVisitor<SliceBuilder>;
648
649 using Base = PtrUseVisitor<SliceBuilder>;
650
651 const uint64_t AllocSize;
652 AllocaSlices &AS;
653
654 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
655 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
656
657 /// Set to de-duplicate dead instructions found in the use walk.
658 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
659
660public:
661 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &AS)
662 : PtrUseVisitor<SliceBuilder>(DL),
663 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType()).getFixedSize()),
664 AS(AS) {}
665
666private:
667 void markAsDead(Instruction &I) {
668 if (VisitedDeadInsts.insert(&I).second)
669 AS.DeadUsers.push_back(&I);
670 }
671
672 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
673 bool IsSplittable = false) {
674 // Completely skip uses which have a zero size or start either before or
675 // past the end of the allocation.
676 if (Size == 0 || Offset.uge(AllocSize)) {
677 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)
678 << 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)
679 << " 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)
680 << 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)
681 << " 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)
682 << " 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)
;
683 return markAsDead(I);
684 }
685
686 uint64_t BeginOffset = Offset.getZExtValue();
687 uint64_t EndOffset = BeginOffset + Size;
688
689 // Clamp the end offset to the end of the allocation. Note that this is
690 // formulated to handle even the case where "BeginOffset + Size" overflows.
691 // This may appear superficially to be something we could ignore entirely,
692 // but that is not so! There may be widened loads or PHI-node uses where
693 // some instructions are dead but not others. We can't completely ignore
694 // them, and so have to record at least the information here.
695 assert(AllocSize >= BeginOffset)((AllocSize >= BeginOffset) ? static_cast<void> (0) :
__assert_fail ("AllocSize >= BeginOffset", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 695, __PRETTY_FUNCTION__))
; // Established above.
696 if (Size > AllocSize - BeginOffset) {
697 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)
698 << 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)
699 << " 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)
700 << " 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)
701 << " 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)
;
702 EndOffset = AllocSize;
703 }
704
705 AS.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
706 }
707
708 void visitBitCastInst(BitCastInst &BC) {
709 if (BC.use_empty())
710 return markAsDead(BC);
711
712 return Base::visitBitCastInst(BC);
713 }
714
715 void visitAddrSpaceCastInst(AddrSpaceCastInst &ASC) {
716 if (ASC.use_empty())
717 return markAsDead(ASC);
718
719 return Base::visitAddrSpaceCastInst(ASC);
720 }
721
722 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
723 if (GEPI.use_empty())
724 return markAsDead(GEPI);
725
726 if (SROAStrictInbounds && GEPI.isInBounds()) {
727 // FIXME: This is a manually un-factored variant of the basic code inside
728 // of GEPs with checking of the inbounds invariant specified in the
729 // langref in a very strict sense. If we ever want to enable
730 // SROAStrictInbounds, this code should be factored cleanly into
731 // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
732 // by writing out the code here where we have the underlying allocation
733 // size readily available.
734 APInt GEPOffset = Offset;
735 const DataLayout &DL = GEPI.getModule()->getDataLayout();
736 for (gep_type_iterator GTI = gep_type_begin(GEPI),
737 GTE = gep_type_end(GEPI);
738 GTI != GTE; ++GTI) {
739 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
740 if (!OpC)
741 break;
742
743 // Handle a struct index, which adds its field offset to the pointer.
744 if (StructType *STy = GTI.getStructTypeOrNull()) {
745 unsigned ElementIdx = OpC->getZExtValue();
746 const StructLayout *SL = DL.getStructLayout(STy);
747 GEPOffset +=
748 APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
749 } else {
750 // For array or vector indices, scale the index by the size of the
751 // type.
752 APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
753 GEPOffset +=
754 Index *
755 APInt(Offset.getBitWidth(),
756 DL.getTypeAllocSize(GTI.getIndexedType()).getFixedSize());
757 }
758
759 // If this index has computed an intermediate pointer which is not
760 // inbounds, then the result of the GEP is a poison value and we can
761 // delete it and all uses.
762 if (GEPOffset.ugt(AllocSize))
763 return markAsDead(GEPI);
764 }
765 }
766
767 return Base::visitGetElementPtrInst(GEPI);
768 }
769
770 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
771 uint64_t Size, bool IsVolatile) {
772 // We allow splitting of non-volatile loads and stores where the type is an
773 // integer type. These may be used to implement 'memcpy' or other "transfer
774 // of bits" patterns.
775 bool IsSplittable = Ty->isIntegerTy() && !IsVolatile;
776
777 insertUse(I, Offset, Size, IsSplittable);
778 }
779
780 void visitLoadInst(LoadInst &LI) {
781 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&(((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
"All simple FCA loads should have been pre-split") ? static_cast
<void> (0) : __assert_fail ("(!LI.isSimple() || LI.getType()->isSingleValueType()) && \"All simple FCA loads should have been pre-split\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 782, __PRETTY_FUNCTION__))
782 "All simple FCA loads should have been pre-split")(((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
"All simple FCA loads should have been pre-split") ? static_cast
<void> (0) : __assert_fail ("(!LI.isSimple() || LI.getType()->isSingleValueType()) && \"All simple FCA loads should have been pre-split\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 782, __PRETTY_FUNCTION__))
;
783
784 if (!IsOffsetKnown)
785 return PI.setAborted(&LI);
786
787 if (LI.isVolatile() &&
788 LI.getPointerAddressSpace() != DL.getAllocaAddrSpace())
789 return PI.setAborted(&LI);
790
791 if (isa<ScalableVectorType>(LI.getType()))
792 return PI.setAborted(&LI);
793
794 uint64_t Size = DL.getTypeStoreSize(LI.getType()).getFixedSize();
795 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
796 }
797
798 void visitStoreInst(StoreInst &SI) {
799 Value *ValOp = SI.getValueOperand();
800 if (ValOp == *U)
801 return PI.setEscapedAndAborted(&SI);
802 if (!IsOffsetKnown)
803 return PI.setAborted(&SI);
804
805 if (SI.isVolatile() &&
806 SI.getPointerAddressSpace() != DL.getAllocaAddrSpace())
807 return PI.setAborted(&SI);
808
809 if (isa<ScalableVectorType>(ValOp->getType()))
810 return PI.setAborted(&SI);
811
812 uint64_t Size = DL.getTypeStoreSize(ValOp->getType()).getFixedSize();
813
814 // If this memory access can be shown to *statically* extend outside the
815 // bounds of the allocation, it's behavior is undefined, so simply
816 // ignore it. Note that this is more strict than the generic clamping
817 // behavior of insertUse. We also try to handle cases which might run the
818 // risk of overflow.
819 // FIXME: We should instead consider the pointer to have escaped if this
820 // function is being instrumented for addressing bugs or race conditions.
821 if (Size > AllocSize || Offset.ugt(AllocSize - Size)) {
822 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)
823 << 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)
824 << 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)
825 << " 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)
826 << " 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)
;
827 return markAsDead(SI);
828 }
829
830 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&(((!SI.isSimple() || ValOp->getType()->isSingleValueType
()) && "All simple FCA stores should have been pre-split"
) ? static_cast<void> (0) : __assert_fail ("(!SI.isSimple() || ValOp->getType()->isSingleValueType()) && \"All simple FCA stores should have been pre-split\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 831, __PRETTY_FUNCTION__))
831 "All simple FCA stores should have been pre-split")(((!SI.isSimple() || ValOp->getType()->isSingleValueType
()) && "All simple FCA stores should have been pre-split"
) ? static_cast<void> (0) : __assert_fail ("(!SI.isSimple() || ValOp->getType()->isSingleValueType()) && \"All simple FCA stores should have been pre-split\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 831, __PRETTY_FUNCTION__))
;
832 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
833 }
834
835 void visitMemSetInst(MemSetInst &II) {
836 assert(II.getRawDest() == *U && "Pointer use is not the destination?")((II.getRawDest() == *U && "Pointer use is not the destination?"
) ? static_cast<void> (0) : __assert_fail ("II.getRawDest() == *U && \"Pointer use is not the destination?\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 836, __PRETTY_FUNCTION__))
;
837 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
838 if ((Length && Length->getValue() == 0) ||
839 (IsOffsetKnown && Offset.uge(AllocSize)))
840 // Zero-length mem transfer intrinsics can be ignored entirely.
841 return markAsDead(II);
842
843 if (!IsOffsetKnown)
844 return PI.setAborted(&II);
845
846 // Don't replace this with a store with a different address space. TODO:
847 // Use a store with the casted new alloca?
848 if (II.isVolatile() && II.getDestAddressSpace() != DL.getAllocaAddrSpace())
849 return PI.setAborted(&II);
850
851 insertUse(II, Offset, Length ? Length->getLimitedValue()
852 : AllocSize - Offset.getLimitedValue(),
853 (bool)Length);
854 }
855
856 void visitMemTransferInst(MemTransferInst &II) {
857 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
858 if (Length && Length->getValue() == 0)
859 // Zero-length mem transfer intrinsics can be ignored entirely.
860 return markAsDead(II);
861
862 // Because we can visit these intrinsics twice, also check to see if the
863 // first time marked this instruction as dead. If so, skip it.
864 if (VisitedDeadInsts.count(&II))
865 return;
866
867 if (!IsOffsetKnown)
868 return PI.setAborted(&II);
869
870 // Don't replace this with a load/store with a different address space.
871 // TODO: Use a store with the casted new alloca?
872 if (II.isVolatile() &&
873 (II.getDestAddressSpace() != DL.getAllocaAddrSpace() ||
874 II.getSourceAddressSpace() != DL.getAllocaAddrSpace()))
875 return PI.setAborted(&II);
876
877 // This side of the transfer is completely out-of-bounds, and so we can
878 // nuke the entire transfer. However, we also need to nuke the other side
879 // if already added to our partitions.
880 // FIXME: Yet another place we really should bypass this when
881 // instrumenting for ASan.
882 if (Offset.uge(AllocSize)) {
883 SmallDenseMap<Instruction *, unsigned>::iterator MTPI =
884 MemTransferSliceMap.find(&II);
885 if (MTPI != MemTransferSliceMap.end())
886 AS.Slices[MTPI->second].kill();
887 return markAsDead(II);
888 }
889
890 uint64_t RawOffset = Offset.getLimitedValue();
891 uint64_t Size = Length ? Length->getLimitedValue() : AllocSize - RawOffset;
892
893 // Check for the special case where the same exact value is used for both
894 // source and dest.
895 if (*U == II.getRawDest() && *U == II.getRawSource()) {
896 // For non-volatile transfers this is a no-op.
897 if (!II.isVolatile())
898 return markAsDead(II);
899
900 return insertUse(II, Offset, Size, /*IsSplittable=*/false);
901 }
902
903 // If we have seen both source and destination for a mem transfer, then
904 // they both point to the same alloca.
905 bool Inserted;
906 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
907 std::tie(MTPI, Inserted) =
908 MemTransferSliceMap.insert(std::make_pair(&II, AS.Slices.size()));
909 unsigned PrevIdx = MTPI->second;
910 if (!Inserted) {
911 Slice &PrevP = AS.Slices[PrevIdx];
912
913 // Check if the begin offsets match and this is a non-volatile transfer.
914 // In that case, we can completely elide the transfer.
915 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
916 PrevP.kill();
917 return markAsDead(II);
918 }
919
920 // Otherwise we have an offset transfer within the same alloca. We can't
921 // split those.
922 PrevP.makeUnsplittable();
923 }
924
925 // Insert the use now that we've fixed up the splittable nature.
926 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
927
928 // Check that we ended up with a valid index in the map.
929 assert(AS.Slices[PrevIdx].getUse()->getUser() == &II &&((AS.Slices[PrevIdx].getUse()->getUser() == &II &&
"Map index doesn't point back to a slice with this user.") ?
static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 930, __PRETTY_FUNCTION__))
930 "Map index doesn't point back to a slice with this user.")((AS.Slices[PrevIdx].getUse()->getUser() == &II &&
"Map index doesn't point back to a slice with this user.") ?
static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 930, __PRETTY_FUNCTION__))
;
931 }
932
933 // Disable SRoA for any intrinsics except for lifetime invariants.
934 // FIXME: What about debug intrinsics? This matches old behavior, but
935 // doesn't make sense.
936 void visitIntrinsicInst(IntrinsicInst &II) {
937 if (II.isDroppable()) {
938 AS.DeadUseIfPromotable.push_back(U);
939 return;
940 }
941
942 if (!IsOffsetKnown)
943 return PI.setAborted(&II);
944
945 if (II.isLifetimeStartOrEnd()) {
946 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
947 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
948 Length->getLimitedValue());
949 insertUse(II, Offset, Size, true);
950 return;
951 }
952
953 Base::visitIntrinsicInst(II);
954 }
955
956 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
957 // We consider any PHI or select that results in a direct load or store of
958 // the same offset to be a viable use for slicing purposes. These uses
959 // are considered unsplittable and the size is the maximum loaded or stored
960 // size.
961 SmallPtrSet<Instruction *, 4> Visited;
962 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
963 Visited.insert(Root);
964 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
965 const DataLayout &DL = Root->getModule()->getDataLayout();
966 // If there are no loads or stores, the access is dead. We mark that as
967 // a size zero access.
968 Size = 0;
969 do {
970 Instruction *I, *UsedI;
971 std::tie(UsedI, I) = Uses.pop_back_val();
972
973 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
974 Size = std::max(Size,
975 DL.getTypeStoreSize(LI->getType()).getFixedSize());
976 continue;
977 }
978 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
979 Value *Op = SI->getOperand(0);
980 if (Op == UsedI)
981 return SI;
982 Size = std::max(Size,
983 DL.getTypeStoreSize(Op->getType()).getFixedSize());
984 continue;
985 }
986
987 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
988 if (!GEP->hasAllZeroIndices())
989 return GEP;
990 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
991 !isa<SelectInst>(I) && !isa<AddrSpaceCastInst>(I)) {
992 return I;
993 }
994
995 for (User *U : I->users())
996 if (Visited.insert(cast<Instruction>(U)).second)
997 Uses.push_back(std::make_pair(I, cast<Instruction>(U)));
998 } while (!Uses.empty());
999
1000 return nullptr;
1001 }
1002
1003 void visitPHINodeOrSelectInst(Instruction &I) {
1004 assert(isa<PHINode>(I) || isa<SelectInst>(I))((isa<PHINode>(I) || isa<SelectInst>(I)) ? static_cast
<void> (0) : __assert_fail ("isa<PHINode>(I) || isa<SelectInst>(I)"
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1004, __PRETTY_FUNCTION__))
;
1005 if (I.use_empty())
1006 return markAsDead(I);
1007
1008 // TODO: We could use SimplifyInstruction here to fold PHINodes and
1009 // SelectInsts. However, doing so requires to change the current
1010 // dead-operand-tracking mechanism. For instance, suppose neither loading
1011 // from %U nor %other traps. Then "load (select undef, %U, %other)" does not
1012 // trap either. However, if we simply replace %U with undef using the
1013 // current dead-operand-tracking mechanism, "load (select undef, undef,
1014 // %other)" may trap because the select may return the first operand
1015 // "undef".
1016 if (Value *Result = foldPHINodeOrSelectInst(I)) {
1017 if (Result == *U)
1018 // If the result of the constant fold will be the pointer, recurse
1019 // through the PHI/select as if we had RAUW'ed it.
1020 enqueueUsers(I);
1021 else
1022 // Otherwise the operand to the PHI/select is dead, and we can replace
1023 // it with undef.
1024 AS.DeadOperands.push_back(U);
1025
1026 return;
1027 }
1028
1029 if (!IsOffsetKnown)
1030 return PI.setAborted(&I);
1031
1032 // See if we already have computed info on this node.
1033 uint64_t &Size = PHIOrSelectSizes[&I];
1034 if (!Size) {
1035 // This is a new PHI/Select, check for an unsafe use of it.
1036 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&I, Size))
1037 return PI.setAborted(UnsafeI);
1038 }
1039
1040 // For PHI and select operands outside the alloca, we can't nuke the entire
1041 // phi or select -- the other side might still be relevant, so we special
1042 // case them here and use a separate structure to track the operands
1043 // themselves which should be replaced with undef.
1044 // FIXME: This should instead be escaped in the event we're instrumenting
1045 // for address sanitization.
1046 if (Offset.uge(AllocSize)) {
1047 AS.DeadOperands.push_back(U);
1048 return;
1049 }
1050
1051 insertUse(I, Offset, Size);
1052 }
1053
1054 void visitPHINode(PHINode &PN) { visitPHINodeOrSelectInst(PN); }
1055
1056 void visitSelectInst(SelectInst &SI) { visitPHINodeOrSelectInst(SI); }
1057
1058 /// Disable SROA entirely if there are unhandled users of the alloca.
1059 void visitInstruction(Instruction &I) { PI.setAborted(&I); }
1060};
1061
1062AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
1063 :
1064#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1065 AI(AI),
1066#endif
1067 PointerEscapingInstr(nullptr) {
1068 SliceBuilder PB(DL, AI, *this);
1069 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
1070 if (PtrI.isEscaped() || PtrI.isAborted()) {
1071 // FIXME: We should sink the escape vs. abort info into the caller nicely,
1072 // possibly by just storing the PtrInfo in the AllocaSlices.
1073 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
1074 : PtrI.getAbortingInst();
1075 assert(PointerEscapingInstr && "Did not track a bad instruction")((PointerEscapingInstr && "Did not track a bad instruction"
) ? static_cast<void> (0) : __assert_fail ("PointerEscapingInstr && \"Did not track a bad instruction\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1075, __PRETTY_FUNCTION__))
;
1076 return;
1077 }
1078
1079 Slices.erase(
1080 llvm::remove_if(Slices, [](const Slice &S) { return S.isDead(); }),
1081 Slices.end());
1082
1083 // Sort the uses. This arranges for the offsets to be in ascending order,
1084 // and the sizes to be in descending order.
1085 std::stable_sort(Slices.begin(), Slices.end());
1086}
1087
1088#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1089
1090void AllocaSlices::print(raw_ostream &OS, const_iterator I,
1091 StringRef Indent) const {
1092 printSlice(OS, I, Indent);
1093 OS << "\n";
1094 printUse(OS, I, Indent);
1095}
1096
1097void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
1098 StringRef Indent) const {
1099 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
1100 << " slice #" << (I - begin())
1101 << (I->isSplittable() ? " (splittable)" : "");
1102}
1103
1104void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
1105 StringRef Indent) const {
1106 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
1107}
1108
1109void AllocaSlices::print(raw_ostream &OS) const {
1110 if (PointerEscapingInstr) {
1111 OS << "Can't analyze slices for alloca: " << AI << "\n"
1112 << " A pointer to this alloca escaped by:\n"
1113 << " " << *PointerEscapingInstr << "\n";
1114 return;
1115 }
1116
1117 OS << "Slices of alloca: " << AI << "\n";
1118 for (const_iterator I = begin(), E = end(); I != E; ++I)
1119 print(OS, I);
1120}
1121
1122LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void AllocaSlices::dump(const_iterator I) const {
1123 print(dbgs(), I);
1124}
1125LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void AllocaSlices::dump() const { print(dbgs()); }
1126
1127#endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1128
1129/// Walk the range of a partitioning looking for a common type to cover this
1130/// sequence of slices.
1131static Type *findCommonType(AllocaSlices::const_iterator B,
1132 AllocaSlices::const_iterator E,
1133 uint64_t EndOffset) {
1134 Type *Ty = nullptr;
1135 bool TyIsCommon = true;
1136 IntegerType *ITy = nullptr;
1137
1138 // Note that we need to look at *every* alloca slice's Use to ensure we
1139 // always get consistent results regardless of the order of slices.
1140 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1141 Use *U = I->getUse();
1142 if (isa<IntrinsicInst>(*U->getUser()))
1143 continue;
1144 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1145 continue;
1146
1147 Type *UserTy = nullptr;
1148 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1149 UserTy = LI->getType();
1150 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1151 UserTy = SI->getValueOperand()->getType();
1152 }
1153
1154 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1155 // If the type is larger than the partition, skip it. We only encounter
1156 // this for split integer operations where we want to use the type of the
1157 // entity causing the split. Also skip if the type is not a byte width
1158 // multiple.
1159 if (UserITy->getBitWidth() % 8 != 0 ||
1160 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1161 continue;
1162
1163 // Track the largest bitwidth integer type used in this way in case there
1164 // is no common type.
1165 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1166 ITy = UserITy;
1167 }
1168
1169 // To avoid depending on the order of slices, Ty and TyIsCommon must not
1170 // depend on types skipped above.
1171 if (!UserTy || (Ty && Ty != UserTy))
1172 TyIsCommon = false; // Give up on anything but an iN type.
1173 else
1174 Ty = UserTy;
1175 }
1176
1177 return TyIsCommon ? Ty : ITy;
1178}
1179
1180/// PHI instructions that use an alloca and are subsequently loaded can be
1181/// rewritten to load both input pointers in the pred blocks and then PHI the
1182/// results, allowing the load of the alloca to be promoted.
1183/// From this:
1184/// %P2 = phi [i32* %Alloca, i32* %Other]
1185/// %V = load i32* %P2
1186/// to:
1187/// %V1 = load i32* %Alloca -> will be mem2reg'd
1188/// ...
1189/// %V2 = load i32* %Other
1190/// ...
1191/// %V = phi [i32 %V1, i32 %V2]
1192///
1193/// We can do this to a select if its only uses are loads and if the operands
1194/// to the select can be loaded unconditionally.
1195///
1196/// FIXME: This should be hoisted into a generic utility, likely in
1197/// Transforms/Util/Local.h
1198static bool isSafePHIToSpeculate(PHINode &PN) {
1199 const DataLayout &DL = PN.getModule()->getDataLayout();
1200
1201 // For now, we can only do this promotion if the load is in the same block
1202 // as the PHI, and if there are no stores between the phi and load.
1203 // TODO: Allow recursive phi users.
1204 // TODO: Allow stores.
1205 BasicBlock *BB = PN.getParent();
1206 Align MaxAlign;
1207 uint64_t APWidth = DL.getIndexTypeSizeInBits(PN.getType());
1208 APInt MaxSize(APWidth, 0);
1209 bool HaveLoad = false;
1210 for (User *U : PN.users()) {
1211 LoadInst *LI = dyn_cast<LoadInst>(U);
1212 if (!LI || !LI->isSimple())
1213 return false;
1214
1215 // For now we only allow loads in the same block as the PHI. This is
1216 // a common case that happens when instcombine merges two loads through
1217 // a PHI.
1218 if (LI->getParent() != BB)
1219 return false;
1220
1221 // Ensure that there are no instructions between the PHI and the load that
1222 // could store.
1223 for (BasicBlock::iterator BBI(PN); &*BBI != LI; ++BBI)
1224 if (BBI->mayWriteToMemory())
1225 return false;
1226
1227 uint64_t Size = DL.getTypeStoreSize(LI->getType()).getFixedSize();
1228 MaxAlign = std::max(MaxAlign, LI->getAlign());
1229 MaxSize = MaxSize.ult(Size) ? APInt(APWidth, Size) : MaxSize;
1230 HaveLoad = true;
1231 }
1232
1233 if (!HaveLoad)
1234 return false;
1235
1236 // We can only transform this if it is safe to push the loads into the
1237 // predecessor blocks. The only thing to watch out for is that we can't put
1238 // a possibly trapping load in the predecessor if it is a critical edge.
1239 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1240 Instruction *TI = PN.getIncomingBlock(Idx)->getTerminator();
1241 Value *InVal = PN.getIncomingValue(Idx);
1242
1243 // If the value is produced by the terminator of the predecessor (an
1244 // invoke) or it has side-effects, there is no valid place to put a load
1245 // in the predecessor.
1246 if (TI == InVal || TI->mayHaveSideEffects())
1247 return false;
1248
1249 // If the predecessor has a single successor, then the edge isn't
1250 // critical.
1251 if (TI->getNumSuccessors() == 1)
1252 continue;
1253
1254 // If this pointer is always safe to load, or if we can prove that there
1255 // is already a load in the block, then we can move the load to the pred
1256 // block.
1257 if (isSafeToLoadUnconditionally(InVal, MaxAlign, MaxSize, DL, TI))
1258 continue;
1259
1260 return false;
1261 }
1262
1263 return true;
1264}
1265
1266static void speculatePHINodeLoads(PHINode &PN) {
1267 LLVM_DEBUG(dbgs() << " original: " << PN << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << PN <<
"\n"; } } while (false)
;
1268
1269 LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1270 Type *LoadTy = SomeLoad->getType();
1271 IRBuilderTy PHIBuilder(&PN);
1272 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1273 PN.getName() + ".sroa.speculated");
1274
1275 // Get the AA tags and alignment to use from one of the loads. It does not
1276 // matter which one we get and if any differ.
1277 AAMDNodes AATags;
1278 SomeLoad->getAAMetadata(AATags);
1279 Align Alignment = SomeLoad->getAlign();
1280
1281 // Rewrite all loads of the PN to use the new PHI.
1282 while (!PN.use_empty()) {
1283 LoadInst *LI = cast<LoadInst>(PN.user_back());
1284 LI->replaceAllUsesWith(NewPN);
1285 LI->eraseFromParent();
1286 }
1287
1288 // Inject loads into all of the pred blocks.
1289 DenseMap<BasicBlock*, Value*> InjectedLoads;
1290 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1291 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1292 Value *InVal = PN.getIncomingValue(Idx);
1293
1294 // A PHI node is allowed to have multiple (duplicated) entries for the same
1295 // basic block, as long as the value is the same. So if we already injected
1296 // a load in the predecessor, then we should reuse the same load for all
1297 // duplicated entries.
1298 if (Value* V = InjectedLoads.lookup(Pred)) {
1299 NewPN->addIncoming(V, Pred);
1300 continue;
1301 }
1302
1303 Instruction *TI = Pred->getTerminator();
1304 IRBuilderTy PredBuilder(TI);
1305
1306 LoadInst *Load = PredBuilder.CreateAlignedLoad(
1307 LoadTy, InVal, Alignment,
1308 (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1309 ++NumLoadsSpeculated;
1310 if (AATags)
1311 Load->setAAMetadata(AATags);
1312 NewPN->addIncoming(Load, Pred);
1313 InjectedLoads[Pred] = Load;
1314 }
1315
1316 LLVM_DEBUG(dbgs() << " speculated to: " << *NewPN << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " speculated to: " <<
*NewPN << "\n"; } } while (false)
;
1317 PN.eraseFromParent();
1318}
1319
1320/// Select instructions that use an alloca and are subsequently loaded can be
1321/// rewritten to load both input pointers and then select between the result,
1322/// allowing the load of the alloca to be promoted.
1323/// From this:
1324/// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1325/// %V = load i32* %P2
1326/// to:
1327/// %V1 = load i32* %Alloca -> will be mem2reg'd
1328/// %V2 = load i32* %Other
1329/// %V = select i1 %cond, i32 %V1, i32 %V2
1330///
1331/// We can do this to a select if its only uses are loads and if the operand
1332/// to the select can be loaded unconditionally.
1333static bool isSafeSelectToSpeculate(SelectInst &SI) {
1334 Value *TValue = SI.getTrueValue();
1335 Value *FValue = SI.getFalseValue();
1336 const DataLayout &DL = SI.getModule()->getDataLayout();
1337
1338 for (User *U : SI.users()) {
1339 LoadInst *LI = dyn_cast<LoadInst>(U);
1340 if (!LI || !LI->isSimple())
1341 return false;
1342
1343 // Both operands to the select need to be dereferenceable, either
1344 // absolutely (e.g. allocas) or at this point because we can see other
1345 // accesses to it.
1346 if (!isSafeToLoadUnconditionally(TValue, LI->getType(),
1347 LI->getAlign(), DL, LI))
1348 return false;
1349 if (!isSafeToLoadUnconditionally(FValue, LI->getType(),
1350 LI->getAlign(), DL, LI))
1351 return false;
1352 }
1353
1354 return true;
1355}
1356
1357static void speculateSelectInstLoads(SelectInst &SI) {
1358 LLVM_DEBUG(dbgs() << " original: " << SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << SI <<
"\n"; } } while (false)
;
1359
1360 IRBuilderTy IRB(&SI);
1361 Value *TV = SI.getTrueValue();
1362 Value *FV = SI.getFalseValue();
1363 // Replace the loads of the select with a select of two loads.
1364 while (!SI.use_empty()) {
1365 LoadInst *LI = cast<LoadInst>(SI.user_back());
1366 assert(LI->isSimple() && "We only speculate simple loads")((LI->isSimple() && "We only speculate simple loads"
) ? static_cast<void> (0) : __assert_fail ("LI->isSimple() && \"We only speculate simple loads\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1366, __PRETTY_FUNCTION__))
;
1367
1368 IRB.SetInsertPoint(LI);
1369 LoadInst *TL = IRB.CreateLoad(LI->getType(), TV,
1370 LI->getName() + ".sroa.speculate.load.true");
1371 LoadInst *FL = IRB.CreateLoad(LI->getType(), FV,
1372 LI->getName() + ".sroa.speculate.load.false");
1373 NumLoadsSpeculated += 2;
1374
1375 // Transfer alignment and AA info if present.
1376 TL->setAlignment(LI->getAlign());
1377 FL->setAlignment(LI->getAlign());
1378
1379 AAMDNodes Tags;
1380 LI->getAAMetadata(Tags);
1381 if (Tags) {
1382 TL->setAAMetadata(Tags);
1383 FL->setAAMetadata(Tags);
1384 }
1385
1386 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1387 LI->getName() + ".sroa.speculated");
1388
1389 LLVM_DEBUG(dbgs() << " speculated to: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " speculated to: " <<
*V << "\n"; } } while (false)
;
1390 LI->replaceAllUsesWith(V);
1391 LI->eraseFromParent();
1392 }
1393 SI.eraseFromParent();
1394}
1395
1396/// Build a GEP out of a base pointer and indices.
1397///
1398/// This will return the BasePtr if that is valid, or build a new GEP
1399/// instruction using the IRBuilder if GEP-ing is needed.
1400static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1401 SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1402 if (Indices.empty())
1403 return BasePtr;
1404
1405 // A single zero index is a no-op, so check for this and avoid building a GEP
1406 // in that case.
1407 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1408 return BasePtr;
1409
1410 return IRB.CreateInBoundsGEP(BasePtr->getType()->getPointerElementType(),
1411 BasePtr, Indices, NamePrefix + "sroa_idx");
1412}
1413
1414/// Get a natural GEP off of the BasePtr walking through Ty toward
1415/// TargetTy without changing the offset of the pointer.
1416///
1417/// This routine assumes we've already established a properly offset GEP with
1418/// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1419/// zero-indices down through type layers until we find one the same as
1420/// TargetTy. If we can't find one with the same type, we at least try to use
1421/// one with the same size. If none of that works, we just produce the GEP as
1422/// indicated by Indices to have the correct offset.
1423static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1424 Value *BasePtr, Type *Ty, Type *TargetTy,
1425 SmallVectorImpl<Value *> &Indices,
1426 Twine NamePrefix) {
1427 if (Ty == TargetTy)
1428 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1429
1430 // Offset size to use for the indices.
1431 unsigned OffsetSize = DL.getIndexTypeSizeInBits(BasePtr->getType());
1432
1433 // See if we can descend into a struct and locate a field with the correct
1434 // type.
1435 unsigned NumLayers = 0;
1436 Type *ElementTy = Ty;
1437 do {
1438 if (ElementTy->isPointerTy())
1439 break;
1440
1441 if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1442 ElementTy = ArrayTy->getElementType();
1443 Indices.push_back(IRB.getIntN(OffsetSize, 0));
1444 } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1445 ElementTy = VectorTy->getElementType();
1446 Indices.push_back(IRB.getInt32(0));
1447 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1448 if (STy->element_begin() == STy->element_end())
1449 break; // Nothing left to descend into.
1450 ElementTy = *STy->element_begin();
1451 Indices.push_back(IRB.getInt32(0));
1452 } else {
1453 break;
1454 }
1455 ++NumLayers;
1456 } while (ElementTy != TargetTy);
1457 if (ElementTy != TargetTy)
1458 Indices.erase(Indices.end() - NumLayers, Indices.end());
1459
1460 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1461}
1462
1463/// Recursively compute indices for a natural GEP.
1464///
1465/// This is the recursive step for getNaturalGEPWithOffset that walks down the
1466/// element types adding appropriate indices for the GEP.
1467static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1468 Value *Ptr, Type *Ty, APInt &Offset,
1469 Type *TargetTy,
1470 SmallVectorImpl<Value *> &Indices,
1471 Twine NamePrefix) {
1472 if (Offset == 0)
1473 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices,
1474 NamePrefix);
1475
1476 // We can't recurse through pointer types.
1477 if (Ty->isPointerTy())
1478 return nullptr;
1479
1480 // We try to analyze GEPs over vectors here, but note that these GEPs are
1481 // extremely poorly defined currently. The long-term goal is to remove GEPing
1482 // over a vector from the IR completely.
1483 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1484 unsigned ElementSizeInBits =
1485 DL.getTypeSizeInBits(VecTy->getScalarType()).getFixedSize();
1486 if (ElementSizeInBits % 8 != 0) {
1487 // GEPs over non-multiple of 8 size vector elements are invalid.
1488 return nullptr;
1489 }
1490 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1491 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1492 if (NumSkippedElements.ugt(cast<FixedVectorType>(VecTy)->getNumElements()))
1493 return nullptr;
1494 Offset -= NumSkippedElements * ElementSize;
1495 Indices.push_back(IRB.getInt(NumSkippedElements));
1496 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1497 Offset, TargetTy, Indices, NamePrefix);
1498 }
1499
1500 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1501 Type *ElementTy = ArrTy->getElementType();
1502 APInt ElementSize(Offset.getBitWidth(),
1503 DL.getTypeAllocSize(ElementTy).getFixedSize());
1504 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1505 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1506 return nullptr;
1507
1508 Offset -= NumSkippedElements * ElementSize;
1509 Indices.push_back(IRB.getInt(NumSkippedElements));
1510 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1511 Indices, NamePrefix);
1512 }
1513
1514 StructType *STy = dyn_cast<StructType>(Ty);
1515 if (!STy)
1516 return nullptr;
1517
1518 const StructLayout *SL = DL.getStructLayout(STy);
1519 uint64_t StructOffset = Offset.getZExtValue();
1520 if (StructOffset >= SL->getSizeInBytes())
1521 return nullptr;
1522 unsigned Index = SL->getElementContainingOffset(StructOffset);
1523 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1524 Type *ElementTy = STy->getElementType(Index);
1525 if (Offset.uge(DL.getTypeAllocSize(ElementTy).getFixedSize()))
1526 return nullptr; // The offset points into alignment padding.
1527
1528 Indices.push_back(IRB.getInt32(Index));
1529 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1530 Indices, NamePrefix);
1531}
1532
1533/// Get a natural GEP from a base pointer to a particular offset and
1534/// resulting in a particular type.
1535///
1536/// The goal is to produce a "natural" looking GEP that works with the existing
1537/// composite types to arrive at the appropriate offset and element type for
1538/// a pointer. TargetTy is the element type the returned GEP should point-to if
1539/// possible. We recurse by decreasing Offset, adding the appropriate index to
1540/// Indices, and setting Ty to the result subtype.
1541///
1542/// If no natural GEP can be constructed, this function returns null.
1543static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1544 Value *Ptr, APInt Offset, Type *TargetTy,
1545 SmallVectorImpl<Value *> &Indices,
1546 Twine NamePrefix) {
1547 PointerType *Ty = cast<PointerType>(Ptr->getType());
29
The object is a 'PointerType'
1548
1549 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1550 // an i8.
1551 if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
30
Assuming the condition is false
1552 return nullptr;
1553
1554 Type *ElementTy = Ty->getElementType();
1555 if (!ElementTy->isSized())
31
Assuming the condition is false
32
Taking false branch
1556 return nullptr; // We can't GEP through an unsized element.
1557 if (isa<ScalableVectorType>(ElementTy))
33
Assuming 'ElementTy' is not a 'ScalableVectorType'
34
Taking false branch
1558 return nullptr;
1559 APInt ElementSize(Offset.getBitWidth(),
1560 DL.getTypeAllocSize(ElementTy).getFixedSize());
1561 if (ElementSize == 0)
35
Taking false branch
1562 return nullptr; // Zero-length arrays can't help us build a natural GEP.
1563 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1564
1565 Offset -= NumSkippedElements * ElementSize;
1566 Indices.push_back(IRB.getInt(NumSkippedElements));
1567 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
36
Value assigned to field 'VecTy', which participates in a condition later
37
Value assigned to field 'IntTy', which participates in a condition later
1568 Indices, NamePrefix);
1569}
1570
1571/// Compute an adjusted pointer from Ptr by Offset bytes where the
1572/// resulting pointer has PointerTy.
1573///
1574/// This tries very hard to compute a "natural" GEP which arrives at the offset
1575/// and produces the pointer type desired. Where it cannot, it will try to use
1576/// the natural GEP to arrive at the offset and bitcast to the type. Where that
1577/// fails, it will try to use an existing i8* and GEP to the byte offset and
1578/// bitcast to the type.
1579///
1580/// The strategy for finding the more natural GEPs is to peel off layers of the
1581/// pointer, walking back through bit casts and GEPs, searching for a base
1582/// pointer from which we can compute a natural GEP with the desired
1583/// properties. The algorithm tries to fold as many constant indices into
1584/// a single GEP as possible, thus making each GEP more independent of the
1585/// surrounding code.
1586static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1587 APInt Offset, Type *PointerTy, Twine NamePrefix) {
1588 // Even though we don't look through PHI nodes, we could be called on an
1589 // instruction in an unreachable block, which may be on a cycle.
1590 SmallPtrSet<Value *, 4> Visited;
1591 Visited.insert(Ptr);
1592 SmallVector<Value *, 4> Indices;
1593
1594 // We may end up computing an offset pointer that has the wrong type. If we
1595 // never are able to compute one directly that has the correct type, we'll
1596 // fall back to it, so keep it and the base it was computed from around here.
1597 Value *OffsetPtr = nullptr;
1598 Value *OffsetBasePtr;
1599
1600 // Remember any i8 pointer we come across to re-use if we need to do a raw
1601 // byte offset.
1602 Value *Int8Ptr = nullptr;
1603 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1604
1605 PointerType *TargetPtrTy = cast<PointerType>(PointerTy);
24
'PointerTy' is a 'PointerType'
1606 Type *TargetTy = TargetPtrTy->getElementType();
1607
1608 // As `addrspacecast` is , `Ptr` (the storage pointer) may have different
1609 // address space from the expected `PointerTy` (the pointer to be used).
1610 // Adjust the pointer type based the original storage pointer.
1611 auto AS = cast<PointerType>(Ptr->getType())->getAddressSpace();
25
The object is a 'PointerType'
1612 PointerTy = TargetTy->getPointerTo(AS);
1613
1614 do {
1615 // First fold any existing GEPs into the offset.
1616 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
26
Assuming 'Ptr' is not a 'GEPOperator'
27
Loop condition is false. Execution continues on line 1627
1617 APInt GEPOffset(Offset.getBitWidth(), 0);
1618 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1619 break;
1620 Offset += GEPOffset;
1621 Ptr = GEP->getPointerOperand();
1622 if (!Visited.insert(Ptr).second)
1623 break;
1624 }
1625
1626 // See if we can perform a natural GEP here.
1627 Indices.clear();
1628 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
28
Calling 'getNaturalGEPWithOffset'
38
Returning from 'getNaturalGEPWithOffset'
39
Assuming 'P' is null
40
Taking false branch
1629 Indices, NamePrefix)) {
1630 // If we have a new natural pointer at the offset, clear out any old
1631 // offset pointer we computed. Unless it is the base pointer or
1632 // a non-instruction, we built a GEP we don't need. Zap it.
1633 if (OffsetPtr && OffsetPtr != OffsetBasePtr)
1634 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr)) {
1635 assert(I->use_empty() && "Built a GEP with uses some how!")((I->use_empty() && "Built a GEP with uses some how!"
) ? static_cast<void> (0) : __assert_fail ("I->use_empty() && \"Built a GEP with uses some how!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1635, __PRETTY_FUNCTION__))
;
1636 I->eraseFromParent();
1637 }
1638 OffsetPtr = P;
1639 OffsetBasePtr = Ptr;
1640 // If we also found a pointer of the right type, we're done.
1641 if (P->getType() == PointerTy)
1642 break;
1643 }
1644
1645 // Stash this pointer if we've found an i8*.
1646 if (Ptr->getType()->isIntegerTy(8)) {
41
Assuming the condition is false
42
Taking false branch
1647 Int8Ptr = Ptr;
1648 Int8PtrOffset = Offset;
1649 }
1650
1651 // Peel off a layer of the pointer and update the offset appropriately.
1652 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
43
Taking false branch
1653 Ptr = cast<Operator>(Ptr)->getOperand(0);
1654 } else if (GlobalAlias *GA
44.1
'GA' is null
= dyn_cast<GlobalAlias>(Ptr)) {
44
Assuming 'Ptr' is not a 'GlobalAlias'
45
Taking false branch
1655 if (GA->isInterposable())
1656 break;
1657 Ptr = GA->getAliasee();
1658 } else {
1659 break;
46
Execution continues on line 1664
1660 }
1661 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!")((Ptr->getType()->isPointerTy() && "Unexpected operand type!"
) ? static_cast<void> (0) : __assert_fail ("Ptr->getType()->isPointerTy() && \"Unexpected operand type!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1661, __PRETTY_FUNCTION__))
;
1662 } while (Visited.insert(Ptr).second);
1663
1664 if (!OffsetPtr
46.1
'OffsetPtr' is null
) {
47
Taking true branch
1665 if (!Int8Ptr
47.1
'Int8Ptr' is null
) {
48
Taking true branch
1666 Int8Ptr = IRB.CreateBitCast(
1667 Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1668 NamePrefix + "sroa_raw_cast");
1669 Int8PtrOffset = Offset;
1670 }
1671
1672 OffsetPtr = Int8PtrOffset == 0
49
'?' condition is false
1673 ? Int8Ptr
1674 : IRB.CreateInBoundsGEP(IRB.getInt8Ty(), Int8Ptr,
1675 IRB.getInt(Int8PtrOffset),
1676 NamePrefix + "sroa_raw_idx");
1677 }
1678 Ptr = OffsetPtr;
1679
1680 // On the off chance we were targeting i8*, guard the bitcast here.
1681 if (cast<PointerType>(Ptr->getType()) != TargetPtrTy) {
50
The object is a 'PointerType'
51
Assuming the condition is false
52
Taking false branch
1682 Ptr = IRB.CreatePointerBitCastOrAddrSpaceCast(Ptr,
1683 TargetPtrTy,
1684 NamePrefix + "sroa_cast");
1685 }
1686
1687 return Ptr;
1688}
1689
1690/// Compute the adjusted alignment for a load or store from an offset.
1691static Align getAdjustedAlignment(Instruction *I, uint64_t Offset) {
1692 return commonAlignment(getLoadStoreAlignment(I), Offset);
1693}
1694
1695/// Test whether we can convert a value from the old to the new type.
1696///
1697/// This predicate should be used to guard calls to convertValue in order to
1698/// ensure that we only try to convert viable values. The strategy is that we
1699/// will peel off single element struct and array wrappings to get to an
1700/// underlying value, and convert that value.
1701static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1702 if (OldTy == NewTy)
1703 return true;
1704
1705 // For integer types, we can't handle any bit-width differences. This would
1706 // break both vector conversions with extension and introduce endianness
1707 // issues when in conjunction with loads and stores.
1708 if (isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) {
1709 assert(cast<IntegerType>(OldTy)->getBitWidth() !=((cast<IntegerType>(OldTy)->getBitWidth() != cast<
IntegerType>(NewTy)->getBitWidth() && "We can't have the same bitwidth for different int types"
) ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1711, __PRETTY_FUNCTION__))
1710 cast<IntegerType>(NewTy)->getBitWidth() &&((cast<IntegerType>(OldTy)->getBitWidth() != cast<
IntegerType>(NewTy)->getBitWidth() && "We can't have the same bitwidth for different int types"
) ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1711, __PRETTY_FUNCTION__))
1711 "We can't have the same bitwidth for different int types")((cast<IntegerType>(OldTy)->getBitWidth() != cast<
IntegerType>(NewTy)->getBitWidth() && "We can't have the same bitwidth for different int types"
) ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1711, __PRETTY_FUNCTION__))
;
1712 return false;
1713 }
1714
1715 if (DL.getTypeSizeInBits(NewTy).getFixedSize() !=
1716 DL.getTypeSizeInBits(OldTy).getFixedSize())
1717 return false;
1718 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1719 return false;
1720
1721 // We can convert pointers to integers and vice-versa. Same for vectors
1722 // of pointers and integers.
1723 OldTy = OldTy->getScalarType();
1724 NewTy = NewTy->getScalarType();
1725 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1726 if (NewTy->isPointerTy() && OldTy->isPointerTy()) {
1727 unsigned OldAS = OldTy->getPointerAddressSpace();
1728 unsigned NewAS = NewTy->getPointerAddressSpace();
1729 // Convert pointers if they are pointers from the same address space or
1730 // different integral (not non-integral) address spaces with the same
1731 // pointer size.
1732 return OldAS == NewAS ||
1733 (!DL.isNonIntegralAddressSpace(OldAS) &&
1734 !DL.isNonIntegralAddressSpace(NewAS) &&
1735 DL.getPointerSize(OldAS) == DL.getPointerSize(NewAS));
1736 }
1737
1738 // We can convert integers to integral pointers, but not to non-integral
1739 // pointers.
1740 if (OldTy->isIntegerTy())
1741 return !DL.isNonIntegralPointerType(NewTy);
1742
1743 // We can convert integral pointers to integers, but non-integral pointers
1744 // need to remain pointers.
1745 if (!DL.isNonIntegralPointerType(OldTy))
1746 return NewTy->isIntegerTy();
1747
1748 return false;
1749 }
1750
1751 return true;
1752}
1753
1754/// Generic routine to convert an SSA value to a value of a different
1755/// type.
1756///
1757/// This will try various different casting techniques, such as bitcasts,
1758/// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1759/// two types for viability with this routine.
1760static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1761 Type *NewTy) {
1762 Type *OldTy = V->getType();
1763 assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type")((canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type"
) ? static_cast<void> (0) : __assert_fail ("canConvertValue(DL, OldTy, NewTy) && \"Value not convertable to type\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1763, __PRETTY_FUNCTION__))
;
1764
1765 if (OldTy == NewTy)
1766 return V;
1767
1768 assert(!(isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) &&((!(isa<IntegerType>(OldTy) && isa<IntegerType
>(NewTy)) && "Integer types must be the exact same to convert."
) ? static_cast<void> (0) : __assert_fail ("!(isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) && \"Integer types must be the exact same to convert.\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1769, __PRETTY_FUNCTION__))
1769 "Integer types must be the exact same to convert.")((!(isa<IntegerType>(OldTy) && isa<IntegerType
>(NewTy)) && "Integer types must be the exact same to convert."
) ? static_cast<void> (0) : __assert_fail ("!(isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) && \"Integer types must be the exact same to convert.\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1769, __PRETTY_FUNCTION__))
;
1770
1771 // See if we need inttoptr for this type pair. May require additional bitcast.
1772 if (OldTy->isIntOrIntVectorTy() && NewTy->isPtrOrPtrVectorTy()) {
1773 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1774 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1775 // Expand <4 x i32> to <2 x i8*> --> <4 x i32> to <2 x i64> to <2 x i8*>
1776 // Directly handle i64 to i8*
1777 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1778 NewTy);
1779 }
1780
1781 // See if we need ptrtoint for this type pair. May require additional bitcast.
1782 if (OldTy->isPtrOrPtrVectorTy() && NewTy->isIntOrIntVectorTy()) {
1783 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1784 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1785 // Expand <2 x i8*> to <4 x i32> --> <2 x i8*> to <2 x i64> to <4 x i32>
1786 // Expand i8* to i64 --> i8* to i64 to i64
1787 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1788 NewTy);
1789 }
1790
1791 if (OldTy->isPtrOrPtrVectorTy() && NewTy->isPtrOrPtrVectorTy()) {
1792 unsigned OldAS = OldTy->getPointerAddressSpace();
1793 unsigned NewAS = NewTy->getPointerAddressSpace();
1794 // To convert pointers with different address spaces (they are already
1795 // checked convertible, i.e. they have the same pointer size), so far we
1796 // cannot use `bitcast` (which has restrict on the same address space) or
1797 // `addrspacecast` (which is not always no-op casting). Instead, use a pair
1798 // of no-op `ptrtoint`/`inttoptr` casts through an integer with the same bit
1799 // size.
1800 if (OldAS != NewAS) {
1801 assert(DL.getPointerSize(OldAS) == DL.getPointerSize(NewAS))((DL.getPointerSize(OldAS) == DL.getPointerSize(NewAS)) ? static_cast
<void> (0) : __assert_fail ("DL.getPointerSize(OldAS) == DL.getPointerSize(NewAS)"
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1801, __PRETTY_FUNCTION__))
;
1802 return IRB.CreateIntToPtr(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1803 NewTy);
1804 }
1805 }
1806
1807 return IRB.CreateBitCast(V, NewTy);
1808}
1809
1810/// Test whether the given slice use can be promoted to a vector.
1811///
1812/// This function is called to test each entry in a partition which is slated
1813/// for a single slice.
1814static bool isVectorPromotionViableForSlice(Partition &P, const Slice &S,
1815 VectorType *Ty,
1816 uint64_t ElementSize,
1817 const DataLayout &DL) {
1818 // First validate the slice offsets.
1819 uint64_t BeginOffset =
1820 std::max(S.beginOffset(), P.beginOffset()) - P.beginOffset();
1821 uint64_t BeginIndex = BeginOffset / ElementSize;
1822 if (BeginIndex * ElementSize != BeginOffset ||
1823 BeginIndex >= cast<FixedVectorType>(Ty)->getNumElements())
1824 return false;
1825 uint64_t EndOffset =
1826 std::min(S.endOffset(), P.endOffset()) - P.beginOffset();
1827 uint64_t EndIndex = EndOffset / ElementSize;
1828 if (EndIndex * ElementSize != EndOffset ||
1829 EndIndex > cast<FixedVectorType>(Ty)->getNumElements())
1830 return false;
1831
1832 assert(EndIndex > BeginIndex && "Empty vector!")((EndIndex > BeginIndex && "Empty vector!") ? static_cast
<void> (0) : __assert_fail ("EndIndex > BeginIndex && \"Empty vector!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1832, __PRETTY_FUNCTION__))
;
1833 uint64_t NumElements = EndIndex - BeginIndex;
1834 Type *SliceTy = (NumElements == 1)
1835 ? Ty->getElementType()
1836 : FixedVectorType::get(Ty->getElementType(), NumElements);
1837
1838 Type *SplitIntTy =
1839 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1840
1841 Use *U = S.getUse();
1842
1843 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1844 if (MI->isVolatile())
1845 return false;
1846 if (!S.isSplittable())
1847 return false; // Skip any unsplittable intrinsics.
1848 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1849 if (!II->isLifetimeStartOrEnd() && !II->isDroppable())
1850 return false;
1851 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1852 // Disable vector promotion when there are loads or stores of an FCA.
1853 return false;
1854 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1855 if (LI->isVolatile())
1856 return false;
1857 Type *LTy = LI->getType();
1858 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1859 assert(LTy->isIntegerTy())((LTy->isIntegerTy()) ? static_cast<void> (0) : __assert_fail
("LTy->isIntegerTy()", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1859, __PRETTY_FUNCTION__))
;
1860 LTy = SplitIntTy;
1861 }
1862 if (!canConvertValue(DL, SliceTy, LTy))
1863 return false;
1864 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1865 if (SI->isVolatile())
1866 return false;
1867 Type *STy = SI->getValueOperand()->getType();
1868 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1869 assert(STy->isIntegerTy())((STy->isIntegerTy()) ? static_cast<void> (0) : __assert_fail
("STy->isIntegerTy()", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1869, __PRETTY_FUNCTION__))
;
1870 STy = SplitIntTy;
1871 }
1872 if (!canConvertValue(DL, STy, SliceTy))
1873 return false;
1874 } else {
1875 return false;
1876 }
1877
1878 return true;
1879}
1880
1881/// Test whether the given alloca partitioning and range of slices can be
1882/// promoted to a vector.
1883///
1884/// This is a quick test to check whether we can rewrite a particular alloca
1885/// partition (and its newly formed alloca) into a vector alloca with only
1886/// whole-vector loads and stores such that it could be promoted to a vector
1887/// SSA value. We only can ensure this for a limited set of operations, and we
1888/// don't want to do the rewrites unless we are confident that the result will
1889/// be promotable, so we have an early test here.
1890static VectorType *isVectorPromotionViable(Partition &P, const DataLayout &DL) {
1891 // Collect the candidate types for vector-based promotion. Also track whether
1892 // we have different element types.
1893 SmallVector<VectorType *, 4> CandidateTys;
1894 Type *CommonEltTy = nullptr;
1895 bool HaveCommonEltTy = true;
1896 auto CheckCandidateType = [&](Type *Ty) {
1897 if (auto *VTy = dyn_cast<VectorType>(Ty)) {
1898 // Return if bitcast to vectors is different for total size in bits.
1899 if (!CandidateTys.empty()) {
1900 VectorType *V = CandidateTys[0];
1901 if (DL.getTypeSizeInBits(VTy).getFixedSize() !=
1902 DL.getTypeSizeInBits(V).getFixedSize()) {
1903 CandidateTys.clear();
1904 return;
1905 }
1906 }
1907 CandidateTys.push_back(VTy);
1908 if (!CommonEltTy)
1909 CommonEltTy = VTy->getElementType();
1910 else if (CommonEltTy != VTy->getElementType())
1911 HaveCommonEltTy = false;
1912 }
1913 };
1914 // Consider any loads or stores that are the exact size of the slice.
1915 for (const Slice &S : P)
1916 if (S.beginOffset() == P.beginOffset() &&
1917 S.endOffset() == P.endOffset()) {
1918 if (auto *LI = dyn_cast<LoadInst>(S.getUse()->getUser()))
1919 CheckCandidateType(LI->getType());
1920 else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser()))
1921 CheckCandidateType(SI->getValueOperand()->getType());
1922 }
1923
1924 // If we didn't find a vector type, nothing to do here.
1925 if (CandidateTys.empty())
1926 return nullptr;
1927
1928 // Remove non-integer vector types if we had multiple common element types.
1929 // FIXME: It'd be nice to replace them with integer vector types, but we can't
1930 // do that until all the backends are known to produce good code for all
1931 // integer vector types.
1932 if (!HaveCommonEltTy) {
1933 CandidateTys.erase(
1934 llvm::remove_if(CandidateTys,
1935 [](VectorType *VTy) {
1936 return !VTy->getElementType()->isIntegerTy();
1937 }),
1938 CandidateTys.end());
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() ==((DL.getTypeSizeInBits(RHSTy).getFixedSize() == DL.getTypeSizeInBits
(LHSTy).getFixedSize() && "Cannot have vector types of different sizes!"
) ? static_cast<void> (0) : __assert_fail ("DL.getTypeSizeInBits(RHSTy).getFixedSize() == DL.getTypeSizeInBits(LHSTy).getFixedSize() && \"Cannot have vector types of different sizes!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1950, __PRETTY_FUNCTION__))
1949 DL.getTypeSizeInBits(LHSTy).getFixedSize() &&((DL.getTypeSizeInBits(RHSTy).getFixedSize() == DL.getTypeSizeInBits
(LHSTy).getFixedSize() && "Cannot have vector types of different sizes!"
) ? static_cast<void> (0) : __assert_fail ("DL.getTypeSizeInBits(RHSTy).getFixedSize() == DL.getTypeSizeInBits(LHSTy).getFixedSize() && \"Cannot have vector types of different sizes!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1950, __PRETTY_FUNCTION__))
1950 "Cannot have vector types of different sizes!")((DL.getTypeSizeInBits(RHSTy).getFixedSize() == DL.getTypeSizeInBits
(LHSTy).getFixedSize() && "Cannot have vector types of different sizes!"
) ? static_cast<void> (0) : __assert_fail ("DL.getTypeSizeInBits(RHSTy).getFixedSize() == DL.getTypeSizeInBits(LHSTy).getFixedSize() && \"Cannot have vector types of different sizes!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1950, __PRETTY_FUNCTION__))
;
1951 assert(RHSTy->getElementType()->isIntegerTy() &&((RHSTy->getElementType()->isIntegerTy() && "All non-integer types eliminated!"
) ? static_cast<void> (0) : __assert_fail ("RHSTy->getElementType()->isIntegerTy() && \"All non-integer types eliminated!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1952, __PRETTY_FUNCTION__))
1952 "All non-integer types eliminated!")((RHSTy->getElementType()->isIntegerTy() && "All non-integer types eliminated!"
) ? static_cast<void> (0) : __assert_fail ("RHSTy->getElementType()->isIntegerTy() && \"All non-integer types eliminated!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1952, __PRETTY_FUNCTION__))
;
1953 assert(LHSTy->getElementType()->isIntegerTy() &&((LHSTy->getElementType()->isIntegerTy() && "All non-integer types eliminated!"
) ? static_cast<void> (0) : __assert_fail ("LHSTy->getElementType()->isIntegerTy() && \"All non-integer types eliminated!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1954, __PRETTY_FUNCTION__))
1954 "All non-integer types eliminated!")((LHSTy->getElementType()->isIntegerTy() && "All non-integer types eliminated!"
) ? static_cast<void> (0) : __assert_fail ("LHSTy->getElementType()->isIntegerTy() && \"All non-integer types eliminated!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1954, __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 &&((VTy->getElementType() == CommonEltTy && "Unaccounted for element type!"
) ? static_cast<void> (0) : __assert_fail ("VTy->getElementType() == CommonEltTy && \"Unaccounted for element type!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1968, __PRETTY_FUNCTION__))
1968 "Unaccounted for element type!")((VTy->getElementType() == CommonEltTy && "Unaccounted for element type!"
) ? static_cast<void> (0) : __assert_fail ("VTy->getElementType() == CommonEltTy && \"Unaccounted for element type!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1968, __PRETTY_FUNCTION__))
;
1969 assert(VTy == CandidateTys[0] &&((VTy == CandidateTys[0] && "Different vector types with the same element type!"
) ? static_cast<void> (0) : __assert_fail ("VTy == CandidateTys[0] && \"Different vector types with the same element type!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1970, __PRETTY_FUNCTION__))
1970 "Different vector types with the same element type!")((VTy == CandidateTys[0] && "Different vector types with the same element type!"
) ? static_cast<void> (0) : __assert_fail ("VTy == CandidateTys[0] && \"Different vector types with the same element type!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1970, __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 &&(((DL.getTypeSizeInBits(VTy).getFixedSize() % 8) == 0 &&
"vector size not a multiple of element size?") ? static_cast
<void> (0) : __assert_fail ("(DL.getTypeSizeInBits(VTy).getFixedSize() % 8) == 0 && \"vector size not a multiple of element size?\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1986, __PRETTY_FUNCTION__))
1986 "vector size not a multiple of element size?")(((DL.getTypeSizeInBits(VTy).getFixedSize() % 8) == 0 &&
"vector size not a multiple of element size?") ? static_cast
<void> (0) : __assert_fail ("(DL.getTypeSizeInBits(VTy).getFixedSize() % 8) == 0 && \"vector size not a multiple of element size?\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 1986, __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 =
2124 P.begin() != P.end() ? false : DL.isLegalInteger(SizeInBits);
2125
2126 for (const Slice &S : P)
2127 if (!isIntegerWideningViableForSlice(S, P.beginOffset(), AllocaTy, DL,
2128 WholeAllocaOp))
2129 return false;
2130
2131 for (const Slice *S : P.splitSliceTails())
2132 if (!isIntegerWideningViableForSlice(*S, P.beginOffset(), AllocaTy, DL,
2133 WholeAllocaOp))
2134 return false;
2135
2136 return WholeAllocaOp;
2137}
2138
2139static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
2140 IntegerType *Ty, uint64_t Offset,
2141 const Twine &Name) {
2142 LLVM_DEBUG(dbgs() << " start: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " start: " << *V <<
"\n"; } } while (false)
;
60
Assuming 'DebugFlag' is false
61
Loop condition is false. Exiting loop
2143 IntegerType *IntTy = cast<IntegerType>(V->getType());
62
The object is a 'IntegerType'
2144 assert(DL.getTypeStoreSize(Ty).getFixedSize() + Offset <=((DL.getTypeStoreSize(Ty).getFixedSize() + Offset <= DL.getTypeStoreSize
(IntTy).getFixedSize() && "Element extends past full value"
) ? static_cast<void> (0) : __assert_fail ("DL.getTypeStoreSize(Ty).getFixedSize() + Offset <= DL.getTypeStoreSize(IntTy).getFixedSize() && \"Element extends past full value\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2146, __PRETTY_FUNCTION__))
63
Assuming the condition is true
64
'?' condition is true
2145 DL.getTypeStoreSize(IntTy).getFixedSize() &&((DL.getTypeStoreSize(Ty).getFixedSize() + Offset <= DL.getTypeStoreSize
(IntTy).getFixedSize() && "Element extends past full value"
) ? static_cast<void> (0) : __assert_fail ("DL.getTypeStoreSize(Ty).getFixedSize() + Offset <= DL.getTypeStoreSize(IntTy).getFixedSize() && \"Element extends past full value\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2146, __PRETTY_FUNCTION__))
2146 "Element extends past full value")((DL.getTypeStoreSize(Ty).getFixedSize() + Offset <= DL.getTypeStoreSize
(IntTy).getFixedSize() && "Element extends past full value"
) ? static_cast<void> (0) : __assert_fail ("DL.getTypeStoreSize(Ty).getFixedSize() + Offset <= DL.getTypeStoreSize(IntTy).getFixedSize() && \"Element extends past full value\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2146, __PRETTY_FUNCTION__))
;
2147 uint64_t ShAmt = 8 * Offset;
2148 if (DL.isBigEndian())
65
Assuming the condition is false
66
Taking false branch
2149 ShAmt = 8 * (DL.getTypeStoreSize(IntTy).getFixedSize() -
2150 DL.getTypeStoreSize(Ty).getFixedSize() - Offset);
2151 if (ShAmt) {
67
Assuming 'ShAmt' is 0
68
Taking false branch
2152 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2153 LLVM_DEBUG(dbgs() << " shifted: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " shifted: " << *V <<
"\n"; } } while (false)
;
2154 }
2155 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&((Ty->getBitWidth() <= IntTy->getBitWidth() &&
"Cannot extract to a larger integer!") ? static_cast<void
> (0) : __assert_fail ("Ty->getBitWidth() <= IntTy->getBitWidth() && \"Cannot extract to a larger integer!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2156, __PRETTY_FUNCTION__))
69
Called C++ object pointer is null
2156 "Cannot extract to a larger integer!")((Ty->getBitWidth() <= IntTy->getBitWidth() &&
"Cannot extract to a larger integer!") ? static_cast<void
> (0) : __assert_fail ("Ty->getBitWidth() <= IntTy->getBitWidth() && \"Cannot extract to a larger integer!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2156, __PRETTY_FUNCTION__))
;
2157 if (Ty != IntTy) {
2158 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2159 LLVM_DEBUG(dbgs() << " trunced: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " trunced: " << *V <<
"\n"; } } while (false)
;
2160 }
2161 return V;
2162}
2163
2164static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
2165 Value *V, uint64_t Offset, const Twine &Name) {
2166 IntegerType *IntTy = cast<IntegerType>(Old->getType());
2167 IntegerType *Ty = cast<IntegerType>(V->getType());
2168 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&((Ty->getBitWidth() <= IntTy->getBitWidth() &&
"Cannot insert a larger integer!") ? static_cast<void>
(0) : __assert_fail ("Ty->getBitWidth() <= IntTy->getBitWidth() && \"Cannot insert a larger integer!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2169, __PRETTY_FUNCTION__))
2169 "Cannot insert a larger integer!")((Ty->getBitWidth() <= IntTy->getBitWidth() &&
"Cannot insert a larger integer!") ? static_cast<void>
(0) : __assert_fail ("Ty->getBitWidth() <= IntTy->getBitWidth() && \"Cannot insert a larger integer!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2169, __PRETTY_FUNCTION__))
;
2170 LLVM_DEBUG(dbgs() << " start: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " start: " << *V <<
"\n"; } } while (false)
;
2171 if (Ty != IntTy) {
2172 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2173 LLVM_DEBUG(dbgs() << " extended: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " extended: " << *V <<
"\n"; } } while (false)
;
2174 }
2175 assert(DL.getTypeStoreSize(Ty).getFixedSize() + Offset <=((DL.getTypeStoreSize(Ty).getFixedSize() + Offset <= DL.getTypeStoreSize
(IntTy).getFixedSize() && "Element store outside of alloca store"
) ? static_cast<void> (0) : __assert_fail ("DL.getTypeStoreSize(Ty).getFixedSize() + Offset <= DL.getTypeStoreSize(IntTy).getFixedSize() && \"Element store outside of alloca store\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2177, __PRETTY_FUNCTION__))
2176 DL.getTypeStoreSize(IntTy).getFixedSize() &&((DL.getTypeStoreSize(Ty).getFixedSize() + Offset <= DL.getTypeStoreSize
(IntTy).getFixedSize() && "Element store outside of alloca store"
) ? static_cast<void> (0) : __assert_fail ("DL.getTypeStoreSize(Ty).getFixedSize() + Offset <= DL.getTypeStoreSize(IntTy).getFixedSize() && \"Element store outside of alloca store\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2177, __PRETTY_FUNCTION__))
2177 "Element store outside of alloca store")((DL.getTypeStoreSize(Ty).getFixedSize() + Offset <= DL.getTypeStoreSize
(IntTy).getFixedSize() && "Element store outside of alloca store"
) ? static_cast<void> (0) : __assert_fail ("DL.getTypeStoreSize(Ty).getFixedSize() + Offset <= DL.getTypeStoreSize(IntTy).getFixedSize() && \"Element store outside of alloca store\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2177, __PRETTY_FUNCTION__))
;
2178 uint64_t ShAmt = 8 * Offset;
2179 if (DL.isBigEndian())
2180 ShAmt = 8 * (DL.getTypeStoreSize(IntTy).getFixedSize() -
2181 DL.getTypeStoreSize(Ty).getFixedSize() - Offset);
2182 if (ShAmt) {
2183 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2184 LLVM_DEBUG(dbgs() << " shifted: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " shifted: " << *V <<
"\n"; } } while (false)
;
2185 }
2186
2187 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2188 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2189 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2190 LLVM_DEBUG(dbgs() << " masked: " << *Old << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " masked: " << *Old <<
"\n"; } } while (false)
;
2191 V = IRB.CreateOr(Old, V, Name + ".insert");
2192 LLVM_DEBUG(dbgs() << " inserted: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " inserted: " << *V <<
"\n"; } } while (false)
;
2193 }
2194 return V;
2195}
2196
2197static Value *extractVector(IRBuilderTy &IRB, Value *V, unsigned BeginIndex,
2198 unsigned EndIndex, const Twine &Name) {
2199 auto *VecTy = cast<FixedVectorType>(V->getType());
2200 unsigned NumElements = EndIndex - BeginIndex;
2201 assert(NumElements <= VecTy->getNumElements() && "Too many elements!")((NumElements <= VecTy->getNumElements() && "Too many elements!"
) ? static_cast<void> (0) : __assert_fail ("NumElements <= VecTy->getNumElements() && \"Too many elements!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2201, __PRETTY_FUNCTION__))
;
2202
2203 if (NumElements == VecTy->getNumElements())
2204 return V;
2205
2206 if (NumElements == 1) {
2207 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2208 Name + ".extract");
2209 LLVM_DEBUG(dbgs() << " extract: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " extract: " << *V <<
"\n"; } } while (false)
;
2210 return V;
2211 }
2212
2213 SmallVector<int, 8> Mask;
2214 Mask.reserve(NumElements);
2215 for (unsigned i = BeginIndex; i != EndIndex; ++i)
2216 Mask.push_back(i);
2217 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()), Mask,
2218 Name + ".extract");
2219 LLVM_DEBUG(dbgs() << " shuffle: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " shuffle: " << *V <<
"\n"; } } while (false)
;
2220 return V;
2221}
2222
2223static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
2224 unsigned BeginIndex, const Twine &Name) {
2225 VectorType *VecTy = cast<VectorType>(Old->getType());
2226 assert(VecTy && "Can only insert a vector into a vector")((VecTy && "Can only insert a vector into a vector") ?
static_cast<void> (0) : __assert_fail ("VecTy && \"Can only insert a vector into a vector\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2226, __PRETTY_FUNCTION__))
;
2227
2228 VectorType *Ty = dyn_cast<VectorType>(V->getType());
2229 if (!Ty) {
2230 // Single element to insert.
2231 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2232 Name + ".insert");
2233 LLVM_DEBUG(dbgs() << " insert: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " insert: " << *V <<
"\n"; } } while (false)
;
2234 return V;
2235 }
2236
2237 assert(cast<FixedVectorType>(Ty)->getNumElements() <=((cast<FixedVectorType>(Ty)->getNumElements() <= cast
<FixedVectorType>(VecTy)->getNumElements() &&
"Too many elements!") ? static_cast<void> (0) : __assert_fail
("cast<FixedVectorType>(Ty)->getNumElements() <= cast<FixedVectorType>(VecTy)->getNumElements() && \"Too many elements!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2239, __PRETTY_FUNCTION__))
2238 cast<FixedVectorType>(VecTy)->getNumElements() &&((cast<FixedVectorType>(Ty)->getNumElements() <= cast
<FixedVectorType>(VecTy)->getNumElements() &&
"Too many elements!") ? static_cast<void> (0) : __assert_fail
("cast<FixedVectorType>(Ty)->getNumElements() <= cast<FixedVectorType>(VecTy)->getNumElements() && \"Too many elements!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2239, __PRETTY_FUNCTION__))
2239 "Too many elements!")((cast<FixedVectorType>(Ty)->getNumElements() <= cast
<FixedVectorType>(VecTy)->getNumElements() &&
"Too many elements!") ? static_cast<void> (0) : __assert_fail
("cast<FixedVectorType>(Ty)->getNumElements() <= cast<FixedVectorType>(VecTy)->getNumElements() && \"Too many elements!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2239, __PRETTY_FUNCTION__))
;
2240 if (cast<FixedVectorType>(Ty)->getNumElements() ==
2241 cast<FixedVectorType>(VecTy)->getNumElements()) {
2242 assert(V->getType() == VecTy && "Vector type mismatch")((V->getType() == VecTy && "Vector type mismatch")
? static_cast<void> (0) : __assert_fail ("V->getType() == VecTy && \"Vector type mismatch\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2242, __PRETTY_FUNCTION__))
;
2243 return V;
2244 }
2245 unsigned EndIndex = BeginIndex + cast<FixedVectorType>(Ty)->getNumElements();
2246
2247 // When inserting a smaller vector into the larger to store, we first
2248 // use a shuffle vector to widen it with undef elements, and then
2249 // a second shuffle vector to select between the loaded vector and the
2250 // incoming vector.
2251 SmallVector<Constant *, 8> Mask;
2252 Mask.reserve(cast<FixedVectorType>(VecTy)->getNumElements());
2253 for (unsigned i = 0; i != cast<FixedVectorType>(VecTy)->getNumElements(); ++i)
2254 if (i >= BeginIndex && i < EndIndex)
2255 Mask.push_back(IRB.getInt32(i - BeginIndex));
2256 else
2257 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2258 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2259 ConstantVector::get(Mask), Name + ".expand");
2260 LLVM_DEBUG(dbgs() << " shuffle: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " shuffle: " << *V <<
"\n"; } } while (false)
;
2261
2262 Mask.clear();
2263 for (unsigned i = 0; i != cast<FixedVectorType>(VecTy)->getNumElements(); ++i)
2264 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
2265
2266 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
2267
2268 LLVM_DEBUG(dbgs() << " blend: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " blend: " << *V <<
"\n"; } } while (false)
;
2269 return V;
2270}
2271
2272/// Visitor to rewrite instructions using p particular slice of an alloca
2273/// to use a new alloca.
2274///
2275/// Also implements the rewriting to vector-based accesses when the partition
2276/// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2277/// lives here.
2278class llvm::sroa::AllocaSliceRewriter
2279 : public InstVisitor<AllocaSliceRewriter, bool> {
2280 // Befriend the base class so it can delegate to private visit methods.
2281 friend class InstVisitor<AllocaSliceRewriter, bool>;
2282
2283 using Base = InstVisitor<AllocaSliceRewriter, bool>;
2284
2285 const DataLayout &DL;
2286 AllocaSlices &AS;
2287 SROA &Pass;
2288 AllocaInst &OldAI, &NewAI;
2289 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2290 Type *NewAllocaTy;
2291
2292 // This is a convenience and flag variable that will be null unless the new
2293 // alloca's integer operations should be widened to this integer type due to
2294 // passing isIntegerWideningViable above. If it is non-null, the desired
2295 // integer type will be stored here for easy access during rewriting.
2296 IntegerType *IntTy;
2297
2298 // If we are rewriting an alloca partition which can be written as pure
2299 // vector operations, we stash extra information here. When VecTy is
2300 // non-null, we have some strict guarantees about the rewritten alloca:
2301 // - The new alloca is exactly the size of the vector type here.
2302 // - The accesses all either map to the entire vector or to a single
2303 // element.
2304 // - The set of accessing instructions is only one of those handled above
2305 // in isVectorPromotionViable. Generally these are the same access kinds
2306 // which are promotable via mem2reg.
2307 VectorType *VecTy;
2308 Type *ElementTy;
2309 uint64_t ElementSize;
2310
2311 // The original offset of the slice currently being rewritten relative to
2312 // the original alloca.
2313 uint64_t BeginOffset = 0;
2314 uint64_t EndOffset = 0;
2315
2316 // The new offsets of the slice currently being rewritten relative to the
2317 // original alloca.
2318 uint64_t NewBeginOffset = 0, NewEndOffset = 0;
2319
2320 uint64_t SliceSize = 0;
2321 bool IsSplittable = false;
2322 bool IsSplit = false;
2323 Use *OldUse = nullptr;
2324 Instruction *OldPtr = nullptr;
2325
2326 // Track post-rewrite users which are PHI nodes and Selects.
2327 SmallSetVector<PHINode *, 8> &PHIUsers;
2328 SmallSetVector<SelectInst *, 8> &SelectUsers;
2329
2330 // Utility IR builder, whose name prefix is setup for each visited use, and
2331 // the insertion point is set to point to the user.
2332 IRBuilderTy IRB;
2333
2334public:
2335 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &AS, SROA &Pass,
2336 AllocaInst &OldAI, AllocaInst &NewAI,
2337 uint64_t NewAllocaBeginOffset,
2338 uint64_t NewAllocaEndOffset, bool IsIntegerPromotable,
2339 VectorType *PromotableVecTy,
2340 SmallSetVector<PHINode *, 8> &PHIUsers,
2341 SmallSetVector<SelectInst *, 8> &SelectUsers)
2342 : DL(DL), AS(AS), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2343 NewAllocaBeginOffset(NewAllocaBeginOffset),
2344 NewAllocaEndOffset(NewAllocaEndOffset),
2345 NewAllocaTy(NewAI.getAllocatedType()),
2346 IntTy(
2347 IsIntegerPromotable
2348 ? Type::getIntNTy(NewAI.getContext(),
2349 DL.getTypeSizeInBits(NewAI.getAllocatedType())
2350 .getFixedSize())
2351 : nullptr),
2352 VecTy(PromotableVecTy),
2353 ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2354 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy).getFixedSize() / 8
2355 : 0),
2356 PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2357 IRB(NewAI.getContext(), ConstantFolder()) {
2358 if (VecTy) {
2359 assert((DL.getTypeSizeInBits(ElementTy).getFixedSize() % 8) == 0 &&(((DL.getTypeSizeInBits(ElementTy).getFixedSize() % 8) == 0 &&
"Only multiple-of-8 sized vector elements are viable") ? static_cast
<void> (0) : __assert_fail ("(DL.getTypeSizeInBits(ElementTy).getFixedSize() % 8) == 0 && \"Only multiple-of-8 sized vector elements are viable\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2360, __PRETTY_FUNCTION__))
2360 "Only multiple-of-8 sized vector elements are viable")(((DL.getTypeSizeInBits(ElementTy).getFixedSize() % 8) == 0 &&
"Only multiple-of-8 sized vector elements are viable") ? static_cast
<void> (0) : __assert_fail ("(DL.getTypeSizeInBits(ElementTy).getFixedSize() % 8) == 0 && \"Only multiple-of-8 sized vector elements are viable\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2360, __PRETTY_FUNCTION__))
;
2361 ++NumVectorized;
2362 }
2363 assert((!IntTy && !VecTy) || (IntTy && !VecTy) || (!IntTy && VecTy))(((!IntTy && !VecTy) || (IntTy && !VecTy) || (
!IntTy && VecTy)) ? static_cast<void> (0) : __assert_fail
("(!IntTy && !VecTy) || (IntTy && !VecTy) || (!IntTy && VecTy)"
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2363, __PRETTY_FUNCTION__))
;
2364 }
2365
2366 bool visit(AllocaSlices::const_iterator I) {
2367 bool CanSROA = true;
2368 BeginOffset = I->beginOffset();
2369 EndOffset = I->endOffset();
2370 IsSplittable = I->isSplittable();
2371 IsSplit =
2372 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2373 LLVM_DEBUG(dbgs() << " rewriting " << (IsSplit ? "split " : ""))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " rewriting " << (IsSplit ?
"split " : ""); } } while (false)
;
2374 LLVM_DEBUG(AS.printSlice(dbgs(), I, ""))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { AS.printSlice(dbgs(), I, ""); } } while (false)
;
2375 LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "\n"; } } while (false)
;
2376
2377 // Compute the intersecting offset range.
2378 assert(BeginOffset < NewAllocaEndOffset)((BeginOffset < NewAllocaEndOffset) ? static_cast<void>
(0) : __assert_fail ("BeginOffset < NewAllocaEndOffset", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2378, __PRETTY_FUNCTION__))
;
2379 assert(EndOffset > NewAllocaBeginOffset)((EndOffset > NewAllocaBeginOffset) ? static_cast<void>
(0) : __assert_fail ("EndOffset > NewAllocaBeginOffset", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2379, __PRETTY_FUNCTION__))
;
2380 NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2381 NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2382
2383 SliceSize = NewEndOffset - NewBeginOffset;
2384
2385 OldUse = I->getUse();
2386 OldPtr = cast<Instruction>(OldUse->get());
2387
2388 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2389 IRB.SetInsertPoint(OldUserI);
2390 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2391 IRB.getInserter().SetNamePrefix(
2392 Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2393
2394 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2395 if (VecTy || IntTy)
2396 assert(CanSROA)((CanSROA) ? static_cast<void> (0) : __assert_fail ("CanSROA"
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2396, __PRETTY_FUNCTION__))
;
2397 return CanSROA;
2398 }
2399
2400private:
2401 // Make sure the other visit overloads are visible.
2402 using Base::visit;
2403
2404 // Every instruction which can end up as a user must have a rewrite rule.
2405 bool visitInstruction(Instruction &I) {
2406 LLVM_DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " !!!! Cannot rewrite: " <<
I << "\n"; } } while (false)
;
2407 llvm_unreachable("No rewrite rule for this instruction!")::llvm::llvm_unreachable_internal("No rewrite rule for this instruction!"
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2407)
;
2408 }
2409
2410 Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2411 // Note that the offset computation can use BeginOffset or NewBeginOffset
2412 // interchangeably for unsplit slices.
2413 assert(IsSplit || BeginOffset == NewBeginOffset)((IsSplit || BeginOffset == NewBeginOffset) ? static_cast<
void> (0) : __assert_fail ("IsSplit || BeginOffset == NewBeginOffset"
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2413, __PRETTY_FUNCTION__))
;
2414 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2415
2416#ifndef NDEBUG
2417 StringRef OldName = OldPtr->getName();
2418 // Skip through the last '.sroa.' component of the name.
2419 size_t LastSROAPrefix = OldName.rfind(".sroa.");
2420 if (LastSROAPrefix != StringRef::npos) {
2421 OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2422 // Look for an SROA slice index.
2423 size_t IndexEnd = OldName.find_first_not_of("0123456789");
2424 if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2425 // Strip the index and look for the offset.
2426 OldName = OldName.substr(IndexEnd + 1);
2427 size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2428 if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2429 // Strip the offset.
2430 OldName = OldName.substr(OffsetEnd + 1);
2431 }
2432 }
2433 // Strip any SROA suffixes as well.
2434 OldName = OldName.substr(0, OldName.find(".sroa_"));
2435#endif
2436
2437 return getAdjustedPtr(IRB, DL, &NewAI,
2438 APInt(DL.getIndexTypeSizeInBits(PointerTy), Offset),
2439 PointerTy,
2440#ifndef NDEBUG
2441 Twine(OldName) + "."
2442#else
2443 Twine()
2444#endif
2445 );
2446 }
2447
2448 /// Compute suitable alignment to access this slice of the *new*
2449 /// alloca.
2450 ///
2451 /// You can optionally pass a type to this routine and if that type's ABI
2452 /// alignment is itself suitable, this will return zero.
2453 Align getSliceAlign() {
2454 return commonAlignment(NewAI.getAlign(),
2455 NewBeginOffset - NewAllocaBeginOffset);
2456 }
2457
2458 unsigned getIndex(uint64_t Offset) {
2459 assert(VecTy && "Can only call getIndex when rewriting a vector")((VecTy && "Can only call getIndex when rewriting a vector"
) ? static_cast<void> (0) : __assert_fail ("VecTy && \"Can only call getIndex when rewriting a vector\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2459, __PRETTY_FUNCTION__))
;
2460 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2461 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds")((RelOffset / ElementSize < (4294967295U) && "Index out of bounds"
) ? static_cast<void> (0) : __assert_fail ("RelOffset / ElementSize < UINT32_MAX && \"Index out of bounds\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2461, __PRETTY_FUNCTION__))
;
2462 uint32_t Index = RelOffset / ElementSize;
2463 assert(Index * ElementSize == RelOffset)((Index * ElementSize == RelOffset) ? static_cast<void>
(0) : __assert_fail ("Index * ElementSize == RelOffset", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2463, __PRETTY_FUNCTION__))
;
2464 return Index;
2465 }
2466
2467 void deleteIfTriviallyDead(Value *V) {
2468 Instruction *I = cast<Instruction>(V);
2469 if (isInstructionTriviallyDead(I))
2470 Pass.DeadInsts.insert(I);
2471 }
2472
2473 Value *rewriteVectorizedLoadInst() {
2474 unsigned BeginIndex = getIndex(NewBeginOffset);
2475 unsigned EndIndex = getIndex(NewEndOffset);
2476 assert(EndIndex > BeginIndex && "Empty vector!")((EndIndex > BeginIndex && "Empty vector!") ? static_cast
<void> (0) : __assert_fail ("EndIndex > BeginIndex && \"Empty vector!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2476, __PRETTY_FUNCTION__))
;
2477
2478 Value *V = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2479 NewAI.getAlign(), "load");
2480 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2481 }
2482
2483 Value *rewriteIntegerLoad(LoadInst &LI) {
2484 assert(IntTy && "We cannot insert an integer to the alloca")((IntTy && "We cannot insert an integer to the alloca"
) ? static_cast<void> (0) : __assert_fail ("IntTy && \"We cannot insert an integer to the alloca\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2484, __PRETTY_FUNCTION__))
;
2485 assert(!LI.isVolatile())((!LI.isVolatile()) ? static_cast<void> (0) : __assert_fail
("!LI.isVolatile()", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2485, __PRETTY_FUNCTION__))
;
2486 Value *V = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2487 NewAI.getAlign(), "load");
2488 V = convertValue(DL, IRB, V, IntTy);
2489 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset")((NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset"
) ? static_cast<void> (0) : __assert_fail ("NewBeginOffset >= NewAllocaBeginOffset && \"Out of bounds offset\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2489, __PRETTY_FUNCTION__))
;
2490 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2491 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset) {
2492 IntegerType *ExtractTy = Type::getIntNTy(LI.getContext(), SliceSize * 8);
2493 V = extractInteger(DL, IRB, V, ExtractTy, Offset, "extract");
2494 }
2495 // It is possible that the extracted type is not the load type. This
2496 // happens if there is a load past the end of the alloca, and as
2497 // a consequence the slice is narrower but still a candidate for integer
2498 // lowering. To handle this case, we just zero extend the extracted
2499 // integer.
2500 assert(cast<IntegerType>(LI.getType())->getBitWidth() >= SliceSize * 8 &&((cast<IntegerType>(LI.getType())->getBitWidth() >=
SliceSize * 8 && "Can only handle an extract for an overly wide load"
) ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2501, __PRETTY_FUNCTION__))
2501 "Can only handle an extract for an overly wide load")((cast<IntegerType>(LI.getType())->getBitWidth() >=
SliceSize * 8 && "Can only handle an extract for an overly wide load"
) ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2501, __PRETTY_FUNCTION__))
;
2502 if (cast<IntegerType>(LI.getType())->getBitWidth() > SliceSize * 8)
2503 V = IRB.CreateZExt(V, LI.getType());
2504 return V;
2505 }
2506
2507 bool visitLoadInst(LoadInst &LI) {
2508 LLVM_DEBUG(dbgs() << " original: " << LI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << LI <<
"\n"; } } while (false)
;
2509 Value *OldOp = LI.getOperand(0);
2510 assert(OldOp == OldPtr)((OldOp == OldPtr) ? static_cast<void> (0) : __assert_fail
("OldOp == OldPtr", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2510, __PRETTY_FUNCTION__))
;
2511
2512 AAMDNodes AATags;
2513 LI.getAAMetadata(AATags);
2514
2515 unsigned AS = LI.getPointerAddressSpace();
2516
2517 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2518 : LI.getType();
2519 const bool IsLoadPastEnd =
2520 DL.getTypeStoreSize(TargetTy).getFixedSize() > SliceSize;
2521 bool IsPtrAdjusted = false;
2522 Value *V;
2523 if (VecTy) {
2524 V = rewriteVectorizedLoadInst();
2525 } else if (IntTy && LI.getType()->isIntegerTy()) {
2526 V = rewriteIntegerLoad(LI);
2527 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2528 NewEndOffset == NewAllocaEndOffset &&
2529 (canConvertValue(DL, NewAllocaTy, TargetTy) ||
2530 (IsLoadPastEnd && NewAllocaTy->isIntegerTy() &&
2531 TargetTy->isIntegerTy()))) {
2532 LoadInst *NewLI = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2533 NewAI.getAlign(), LI.isVolatile(),
2534 LI.getName());
2535 if (AATags)
2536 NewLI->setAAMetadata(AATags);
2537 if (LI.isVolatile())
2538 NewLI->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
2539 if (NewLI->isAtomic())
2540 NewLI->setAlignment(LI.getAlign());
2541
2542 // Any !nonnull metadata or !range metadata on the old load is also valid
2543 // on the new load. This is even true in some cases even when the loads
2544 // are different types, for example by mapping !nonnull metadata to
2545 // !range metadata by modeling the null pointer constant converted to the
2546 // integer type.
2547 // FIXME: Add support for range metadata here. Currently the utilities
2548 // for this don't propagate range metadata in trivial cases from one
2549 // integer load to another, don't handle non-addrspace-0 null pointers
2550 // correctly, and don't have any support for mapping ranges as the
2551 // integer type becomes winder or narrower.
2552 if (MDNode *N = LI.getMetadata(LLVMContext::MD_nonnull))
2553 copyNonnullMetadata(LI, N, *NewLI);
2554
2555 // Try to preserve nonnull metadata
2556 V = NewLI;
2557
2558 // If this is an integer load past the end of the slice (which means the
2559 // bytes outside the slice are undef or this load is dead) just forcibly
2560 // fix the integer size with correct handling of endianness.
2561 if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2562 if (auto *TITy = dyn_cast<IntegerType>(TargetTy))
2563 if (AITy->getBitWidth() < TITy->getBitWidth()) {
2564 V = IRB.CreateZExt(V, TITy, "load.ext");
2565 if (DL.isBigEndian())
2566 V = IRB.CreateShl(V, TITy->getBitWidth() - AITy->getBitWidth(),
2567 "endian_shift");
2568 }
2569 } else {
2570 Type *LTy = TargetTy->getPointerTo(AS);
2571 LoadInst *NewLI =
2572 IRB.CreateAlignedLoad(TargetTy, getNewAllocaSlicePtr(IRB, LTy),
2573 getSliceAlign(), LI.isVolatile(), LI.getName());
2574 if (AATags)
2575 NewLI->setAAMetadata(AATags);
2576 if (LI.isVolatile())
2577 NewLI->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
2578
2579 V = NewLI;
2580 IsPtrAdjusted = true;
2581 }
2582 V = convertValue(DL, IRB, V, TargetTy);
2583
2584 if (IsSplit) {
2585 assert(!LI.isVolatile())((!LI.isVolatile()) ? static_cast<void> (0) : __assert_fail
("!LI.isVolatile()", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2585, __PRETTY_FUNCTION__))
;
2586 assert(LI.getType()->isIntegerTy() &&((LI.getType()->isIntegerTy() && "Only integer type loads and stores are split"
) ? static_cast<void> (0) : __assert_fail ("LI.getType()->isIntegerTy() && \"Only integer type loads and stores are split\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2587, __PRETTY_FUNCTION__))
2587 "Only integer type loads and stores are split")((LI.getType()->isIntegerTy() && "Only integer type loads and stores are split"
) ? static_cast<void> (0) : __assert_fail ("LI.getType()->isIntegerTy() && \"Only integer type loads and stores are split\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2587, __PRETTY_FUNCTION__))
;
2588 assert(SliceSize < DL.getTypeStoreSize(LI.getType()).getFixedSize() &&((SliceSize < DL.getTypeStoreSize(LI.getType()).getFixedSize
() && "Split load isn't smaller than original load") ?
static_cast<void> (0) : __assert_fail ("SliceSize < DL.getTypeStoreSize(LI.getType()).getFixedSize() && \"Split load isn't smaller than original load\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2589, __PRETTY_FUNCTION__))
2589 "Split load isn't smaller than original load")((SliceSize < DL.getTypeStoreSize(LI.getType()).getFixedSize
() && "Split load isn't smaller than original load") ?
static_cast<void> (0) : __assert_fail ("SliceSize < DL.getTypeStoreSize(LI.getType()).getFixedSize() && \"Split load isn't smaller than original load\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2589, __PRETTY_FUNCTION__))
;
2590 assert(DL.typeSizeEqualsStoreSize(LI.getType()) &&((DL.typeSizeEqualsStoreSize(LI.getType()) && "Non-byte-multiple bit width"
) ? static_cast<void> (0) : __assert_fail ("DL.typeSizeEqualsStoreSize(LI.getType()) && \"Non-byte-multiple bit width\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2591, __PRETTY_FUNCTION__))
2591 "Non-byte-multiple bit width")((DL.typeSizeEqualsStoreSize(LI.getType()) && "Non-byte-multiple bit width"
) ? static_cast<void> (0) : __assert_fail ("DL.typeSizeEqualsStoreSize(LI.getType()) && \"Non-byte-multiple bit width\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2591, __PRETTY_FUNCTION__))
;
2592 // Move the insertion point just past the load so that we can refer to it.
2593 IRB.SetInsertPoint(&*std::next(BasicBlock::iterator(&LI)));
2594 // Create a placeholder value with the same type as LI to use as the
2595 // basis for the new value. This allows us to replace the uses of LI with
2596 // the computed value, and then replace the placeholder with LI, leaving
2597 // LI only used for this computation.
2598 Value *Placeholder = new LoadInst(
2599 LI.getType(), UndefValue::get(LI.getType()->getPointerTo(AS)), "",
2600 false, Align(1));
2601 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset - BeginOffset,
2602 "insert");
2603 LI.replaceAllUsesWith(V);
2604 Placeholder->replaceAllUsesWith(&LI);
2605 Placeholder->deleteValue();
2606 } else {
2607 LI.replaceAllUsesWith(V);
2608 }
2609
2610 Pass.DeadInsts.insert(&LI);
2611 deleteIfTriviallyDead(OldOp);
2612 LLVM_DEBUG(dbgs() << " to: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *V <<
"\n"; } } while (false)
;
2613 return !LI.isVolatile() && !IsPtrAdjusted;
2614 }
2615
2616 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp,
2617 AAMDNodes AATags) {
2618 if (V->getType() != VecTy) {
2619 unsigned BeginIndex = getIndex(NewBeginOffset);
2620 unsigned EndIndex = getIndex(NewEndOffset);
2621 assert(EndIndex > BeginIndex && "Empty vector!")((EndIndex > BeginIndex && "Empty vector!") ? static_cast
<void> (0) : __assert_fail ("EndIndex > BeginIndex && \"Empty vector!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2621, __PRETTY_FUNCTION__))
;
2622 unsigned NumElements = EndIndex - BeginIndex;
2623 assert(NumElements <= cast<FixedVectorType>(VecTy)->getNumElements() &&((NumElements <= cast<FixedVectorType>(VecTy)->getNumElements
() && "Too many elements!") ? static_cast<void>
(0) : __assert_fail ("NumElements <= cast<FixedVectorType>(VecTy)->getNumElements() && \"Too many elements!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2624, __PRETTY_FUNCTION__))
2624 "Too many elements!")((NumElements <= cast<FixedVectorType>(VecTy)->getNumElements
() && "Too many elements!") ? static_cast<void>
(0) : __assert_fail ("NumElements <= cast<FixedVectorType>(VecTy)->getNumElements() && \"Too many elements!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2624, __PRETTY_FUNCTION__))
;
2625 Type *SliceTy = (NumElements == 1)
2626 ? ElementTy
2627 : FixedVectorType::get(ElementTy, NumElements);
2628 if (V->getType() != SliceTy)
2629 V = convertValue(DL, IRB, V, SliceTy);
2630
2631 // Mix in the existing elements.
2632 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2633 NewAI.getAlign(), "load");
2634 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2635 }
2636 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlign());
2637 if (AATags)
2638 Store->setAAMetadata(AATags);
2639 Pass.DeadInsts.insert(&SI);
2640
2641 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *Store <<
"\n"; } } while (false)
;
2642 return true;
2643 }
2644
2645 bool rewriteIntegerStore(Value *V, StoreInst &SI, AAMDNodes AATags) {
2646 assert(IntTy && "We cannot extract an integer from the alloca")((IntTy && "We cannot extract an integer from the alloca"
) ? static_cast<void> (0) : __assert_fail ("IntTy && \"We cannot extract an integer from the alloca\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2646, __PRETTY_FUNCTION__))
;
2647 assert(!SI.isVolatile())((!SI.isVolatile()) ? static_cast<void> (0) : __assert_fail
("!SI.isVolatile()", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2647, __PRETTY_FUNCTION__))
;
2648 if (DL.getTypeSizeInBits(V->getType()).getFixedSize() !=
2649 IntTy->getBitWidth()) {
2650 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2651 NewAI.getAlign(), "oldload");
2652 Old = convertValue(DL, IRB, Old, IntTy);
2653 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset")((BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset"
) ? static_cast<void> (0) : __assert_fail ("BeginOffset >= NewAllocaBeginOffset && \"Out of bounds offset\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2653, __PRETTY_FUNCTION__))
;
2654 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2655 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset, "insert");
2656 }
2657 V = convertValue(DL, IRB, V, NewAllocaTy);
2658 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlign());
2659 Store->copyMetadata(SI, {LLVMContext::MD_mem_parallel_loop_access,
2660 LLVMContext::MD_access_group});
2661 if (AATags)
2662 Store->setAAMetadata(AATags);
2663 Pass.DeadInsts.insert(&SI);
2664 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *Store <<
"\n"; } } while (false)
;
2665 return true;
2666 }
2667
2668 bool visitStoreInst(StoreInst &SI) {
2669 LLVM_DEBUG(dbgs() << " original: " << SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << SI <<
"\n"; } } while (false)
;
2670 Value *OldOp = SI.getOperand(1);
2671 assert(OldOp == OldPtr)((OldOp == OldPtr) ? static_cast<void> (0) : __assert_fail
("OldOp == OldPtr", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2671, __PRETTY_FUNCTION__))
;
2672
2673 AAMDNodes AATags;
2674 SI.getAAMetadata(AATags);
2675
2676 Value *V = SI.getValueOperand();
2677
2678 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2679 // alloca that should be re-examined after promoting this alloca.
2680 if (V->getType()->isPointerTy())
2681 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2682 Pass.PostPromotionWorklist.insert(AI);
2683
2684 if (SliceSize < DL.getTypeStoreSize(V->getType()).getFixedSize()) {
2685 assert(!SI.isVolatile())((!SI.isVolatile()) ? static_cast<void> (0) : __assert_fail
("!SI.isVolatile()", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2685, __PRETTY_FUNCTION__))
;
2686 assert(V->getType()->isIntegerTy() &&((V->getType()->isIntegerTy() && "Only integer type loads and stores are split"
) ? static_cast<void> (0) : __assert_fail ("V->getType()->isIntegerTy() && \"Only integer type loads and stores are split\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2687, __PRETTY_FUNCTION__))
2687 "Only integer type loads and stores are split")((V->getType()->isIntegerTy() && "Only integer type loads and stores are split"
) ? static_cast<void> (0) : __assert_fail ("V->getType()->isIntegerTy() && \"Only integer type loads and stores are split\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2687, __PRETTY_FUNCTION__))
;
2688 assert(DL.typeSizeEqualsStoreSize(V->getType()) &&((DL.typeSizeEqualsStoreSize(V->getType()) && "Non-byte-multiple bit width"
) ? static_cast<void> (0) : __assert_fail ("DL.typeSizeEqualsStoreSize(V->getType()) && \"Non-byte-multiple bit width\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2689, __PRETTY_FUNCTION__))
2689 "Non-byte-multiple bit width")((DL.typeSizeEqualsStoreSize(V->getType()) && "Non-byte-multiple bit width"
) ? static_cast<void> (0) : __assert_fail ("DL.typeSizeEqualsStoreSize(V->getType()) && \"Non-byte-multiple bit width\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2689, __PRETTY_FUNCTION__))
;
2690 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2691 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset - BeginOffset,
2692 "extract");
2693 }
2694
2695 if (VecTy)
2696 return rewriteVectorizedStoreInst(V, SI, OldOp, AATags);
2697 if (IntTy && V->getType()->isIntegerTy())
2698 return rewriteIntegerStore(V, SI, AATags);
2699
2700 const bool IsStorePastEnd =
2701 DL.getTypeStoreSize(V->getType()).getFixedSize() > SliceSize;
2702 StoreInst *NewSI;
2703 if (NewBeginOffset == NewAllocaBeginOffset &&
2704 NewEndOffset == NewAllocaEndOffset &&
2705 (canConvertValue(DL, V->getType(), NewAllocaTy) ||
2706 (IsStorePastEnd && NewAllocaTy->isIntegerTy() &&
2707 V->getType()->isIntegerTy()))) {
2708 // If this is an integer store past the end of slice (and thus the bytes
2709 // past that point are irrelevant or this is unreachable), truncate the
2710 // value prior to storing.
2711 if (auto *VITy = dyn_cast<IntegerType>(V->getType()))
2712 if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2713 if (VITy->getBitWidth() > AITy->getBitWidth()) {
2714 if (DL.isBigEndian())
2715 V = IRB.CreateLShr(V, VITy->getBitWidth() - AITy->getBitWidth(),
2716 "endian_shift");
2717 V = IRB.CreateTrunc(V, AITy, "load.trunc");
2718 }
2719
2720 V = convertValue(DL, IRB, V, NewAllocaTy);
2721 NewSI =
2722 IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlign(), SI.isVolatile());
2723 } else {
2724 unsigned AS = SI.getPointerAddressSpace();
2725 Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo(AS));
2726 NewSI =
2727 IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(), SI.isVolatile());
2728 }
2729 NewSI->copyMetadata(SI, {LLVMContext::MD_mem_parallel_loop_access,
2730 LLVMContext::MD_access_group});
2731 if (AATags)
2732 NewSI->setAAMetadata(AATags);
2733 if (SI.isVolatile())
2734 NewSI->setAtomic(SI.getOrdering(), SI.getSyncScopeID());
2735 if (NewSI->isAtomic())
2736 NewSI->setAlignment(SI.getAlign());
2737 Pass.DeadInsts.insert(&SI);
2738 deleteIfTriviallyDead(OldOp);
2739
2740 LLVM_DEBUG(dbgs() << " to: " << *NewSI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *NewSI <<
"\n"; } } while (false)
;
2741 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2742 }
2743
2744 /// Compute an integer value from splatting an i8 across the given
2745 /// number of bytes.
2746 ///
2747 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2748 /// call this routine.
2749 /// FIXME: Heed the advice above.
2750 ///
2751 /// \param V The i8 value to splat.
2752 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2753 Value *getIntegerSplat(Value *V, unsigned Size) {
2754 assert(Size > 0 && "Expected a positive number of bytes.")((Size > 0 && "Expected a positive number of bytes."
) ? static_cast<void> (0) : __assert_fail ("Size > 0 && \"Expected a positive number of bytes.\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2754, __PRETTY_FUNCTION__))
;
2755 IntegerType *VTy = cast<IntegerType>(V->getType());
2756 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte")((VTy->getBitWidth() == 8 && "Expected an i8 value for the byte"
) ? static_cast<void> (0) : __assert_fail ("VTy->getBitWidth() == 8 && \"Expected an i8 value for the byte\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2756, __PRETTY_FUNCTION__))
;
2757 if (Size == 1)
2758 return V;
2759
2760 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size * 8);
2761 V = IRB.CreateMul(
2762 IRB.CreateZExt(V, SplatIntTy, "zext"),
2763 ConstantExpr::getUDiv(
2764 Constant::getAllOnesValue(SplatIntTy),
2765 ConstantExpr::getZExt(Constant::getAllOnesValue(V->getType()),
2766 SplatIntTy)),
2767 "isplat");
2768 return V;
2769 }
2770
2771 /// Compute a vector splat for a given element value.
2772 Value *getVectorSplat(Value *V, unsigned NumElements) {
2773 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2774 LLVM_DEBUG(dbgs() << " splat: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " splat: " << *V <<
"\n"; } } while (false)
;
2775 return V;
2776 }
2777
2778 bool visitMemSetInst(MemSetInst &II) {
2779 LLVM_DEBUG(dbgs() << " original: " << II << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << II <<
"\n"; } } while (false)
;
2780 assert(II.getRawDest() == OldPtr)((II.getRawDest() == OldPtr) ? static_cast<void> (0) : __assert_fail
("II.getRawDest() == OldPtr", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2780, __PRETTY_FUNCTION__))
;
2781
2782 AAMDNodes AATags;
2783 II.getAAMetadata(AATags);
2784
2785 // If the memset has a variable size, it cannot be split, just adjust the
2786 // pointer to the new alloca.
2787 if (!isa<Constant>(II.getLength())) {
2788 assert(!IsSplit)((!IsSplit) ? static_cast<void> (0) : __assert_fail ("!IsSplit"
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2788, __PRETTY_FUNCTION__))
;
2789 assert(NewBeginOffset == BeginOffset)((NewBeginOffset == BeginOffset) ? static_cast<void> (0
) : __assert_fail ("NewBeginOffset == BeginOffset", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2789, __PRETTY_FUNCTION__))
;
2790 II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2791 II.setDestAlignment(getSliceAlign());
2792
2793 deleteIfTriviallyDead(OldPtr);
2794 return false;
2795 }
2796
2797 // Record this instruction for deletion.
2798 Pass.DeadInsts.insert(&II);
2799
2800 Type *AllocaTy = NewAI.getAllocatedType();
2801 Type *ScalarTy = AllocaTy->getScalarType();
2802
2803 const bool CanContinue = [&]() {
2804 if (VecTy || IntTy)
2805 return true;
2806 if (BeginOffset > NewAllocaBeginOffset ||
2807 EndOffset < NewAllocaEndOffset)
2808 return false;
2809 auto *C = cast<ConstantInt>(II.getLength());
2810 if (C->getBitWidth() > 64)
2811 return false;
2812 const auto Len = C->getZExtValue();
2813 auto *Int8Ty = IntegerType::getInt8Ty(NewAI.getContext());
2814 auto *SrcTy = FixedVectorType::get(Int8Ty, Len);
2815 return canConvertValue(DL, SrcTy, AllocaTy) &&
2816 DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy).getFixedSize());
2817 }();
2818
2819 // If this doesn't map cleanly onto the alloca type, and that type isn't
2820 // a single value type, just emit a memset.
2821 if (!CanContinue) {
2822 Type *SizeTy = II.getLength()->getType();
2823 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2824 CallInst *New = IRB.CreateMemSet(
2825 getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2826 MaybeAlign(getSliceAlign()), II.isVolatile());
2827 if (AATags)
2828 New->setAAMetadata(AATags);
2829 LLVM_DEBUG(dbgs() << " to: " << *New << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *New <<
"\n"; } } while (false)
;
2830 return false;
2831 }
2832
2833 // If we can represent this as a simple value, we have to build the actual
2834 // value to store, which requires expanding the byte present in memset to
2835 // a sensible representation for the alloca type. This is essentially
2836 // splatting the byte to a sufficiently wide integer, splatting it across
2837 // any desired vector width, and bitcasting to the final type.
2838 Value *V;
2839
2840 if (VecTy) {
2841 // If this is a memset of a vectorized alloca, insert it.
2842 assert(ElementTy == ScalarTy)((ElementTy == ScalarTy) ? static_cast<void> (0) : __assert_fail
("ElementTy == ScalarTy", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2842, __PRETTY_FUNCTION__))
;
2843
2844 unsigned BeginIndex = getIndex(NewBeginOffset);
2845 unsigned EndIndex = getIndex(NewEndOffset);
2846 assert(EndIndex > BeginIndex && "Empty vector!")((EndIndex > BeginIndex && "Empty vector!") ? static_cast
<void> (0) : __assert_fail ("EndIndex > BeginIndex && \"Empty vector!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2846, __PRETTY_FUNCTION__))
;
2847 unsigned NumElements = EndIndex - BeginIndex;
2848 assert(NumElements <= cast<FixedVectorType>(VecTy)->getNumElements() &&((NumElements <= cast<FixedVectorType>(VecTy)->getNumElements
() && "Too many elements!") ? static_cast<void>
(0) : __assert_fail ("NumElements <= cast<FixedVectorType>(VecTy)->getNumElements() && \"Too many elements!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2849, __PRETTY_FUNCTION__))
2849 "Too many elements!")((NumElements <= cast<FixedVectorType>(VecTy)->getNumElements
() && "Too many elements!") ? static_cast<void>
(0) : __assert_fail ("NumElements <= cast<FixedVectorType>(VecTy)->getNumElements() && \"Too many elements!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2849, __PRETTY_FUNCTION__))
;
2850
2851 Value *Splat = getIntegerSplat(
2852 II.getValue(), DL.getTypeSizeInBits(ElementTy).getFixedSize() / 8);
2853 Splat = convertValue(DL, IRB, Splat, ElementTy);
2854 if (NumElements > 1)
2855 Splat = getVectorSplat(Splat, NumElements);
2856
2857 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2858 NewAI.getAlign(), "oldload");
2859 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2860 } else if (IntTy) {
2861 // If this is a memset on an alloca where we can widen stores, insert the
2862 // set integer.
2863 assert(!II.isVolatile())((!II.isVolatile()) ? static_cast<void> (0) : __assert_fail
("!II.isVolatile()", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2863, __PRETTY_FUNCTION__))
;
2864
2865 uint64_t Size = NewEndOffset - NewBeginOffset;
2866 V = getIntegerSplat(II.getValue(), Size);
2867
2868 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2869 EndOffset != NewAllocaBeginOffset)) {
2870 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2871 NewAI.getAlign(), "oldload");
2872 Old = convertValue(DL, IRB, Old, IntTy);
2873 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2874 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2875 } else {
2876 assert(V->getType() == IntTy &&((V->getType() == IntTy && "Wrong type for an alloca wide integer!"
) ? static_cast<void> (0) : __assert_fail ("V->getType() == IntTy && \"Wrong type for an alloca wide integer!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2877, __PRETTY_FUNCTION__))
2877 "Wrong type for an alloca wide integer!")((V->getType() == IntTy && "Wrong type for an alloca wide integer!"
) ? static_cast<void> (0) : __assert_fail ("V->getType() == IntTy && \"Wrong type for an alloca wide integer!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2877, __PRETTY_FUNCTION__))
;
2878 }
2879 V = convertValue(DL, IRB, V, AllocaTy);
2880 } else {
2881 // Established these invariants above.
2882 assert(NewBeginOffset == NewAllocaBeginOffset)((NewBeginOffset == NewAllocaBeginOffset) ? static_cast<void
> (0) : __assert_fail ("NewBeginOffset == NewAllocaBeginOffset"
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2882, __PRETTY_FUNCTION__))
;
2883 assert(NewEndOffset == NewAllocaEndOffset)((NewEndOffset == NewAllocaEndOffset) ? static_cast<void>
(0) : __assert_fail ("NewEndOffset == NewAllocaEndOffset", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2883, __PRETTY_FUNCTION__))
;
2884
2885 V = getIntegerSplat(II.getValue(),
2886 DL.getTypeSizeInBits(ScalarTy).getFixedSize() / 8);
2887 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2888 V = getVectorSplat(
2889 V, cast<FixedVectorType>(AllocaVecTy)->getNumElements());
2890
2891 V = convertValue(DL, IRB, V, AllocaTy);
2892 }
2893
2894 StoreInst *New =
2895 IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlign(), II.isVolatile());
2896 if (AATags)
2897 New->setAAMetadata(AATags);
2898 LLVM_DEBUG(dbgs() << " to: " << *New << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *New <<
"\n"; } } while (false)
;
2899 return !II.isVolatile();
2900 }
2901
2902 bool visitMemTransferInst(MemTransferInst &II) {
2903 // Rewriting of memory transfer instructions can be a bit tricky. We break
2904 // them into two categories: split intrinsics and unsplit intrinsics.
2905
2906 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
2907
2908 AAMDNodes AATags;
2909 II.getAAMetadata(AATags);
2910
2911 bool IsDest = &II.getRawDestUse() == OldUse;
2912 assert((IsDest && II.getRawDest() == OldPtr) ||(((IsDest && II.getRawDest() == OldPtr) || (!IsDest &&
II.getRawSource() == OldPtr)) ? static_cast<void> (0) :
__assert_fail ("(IsDest && II.getRawDest() == OldPtr) || (!IsDest && II.getRawSource() == OldPtr)"
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2913, __PRETTY_FUNCTION__))
3
Assuming 'IsDest' is false
4
Assuming the condition is true
5
'?' condition is true
2913 (!IsDest && II.getRawSource() == OldPtr))(((IsDest && II.getRawDest() == OldPtr) || (!IsDest &&
II.getRawSource() == OldPtr)) ? static_cast<void> (0) :
__assert_fail ("(IsDest && II.getRawDest() == OldPtr) || (!IsDest && II.getRawSource() == OldPtr)"
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2913, __PRETTY_FUNCTION__))
;
2914
2915 MaybeAlign SliceAlign = getSliceAlign();
2916
2917 // For unsplit intrinsics, we simply modify the source and destination
2918 // pointers in place. This isn't just an optimization, it is a matter of
2919 // correctness. With unsplit intrinsics we may be dealing with transfers
2920 // within a single alloca before SROA ran, or with transfers that have
2921 // a variable length. We may also be dealing with memmove instead of
2922 // memcpy, and so simply updating the pointers is the necessary for us to
2923 // update both source and dest of a single call.
2924 if (!IsSplittable) {
6
Assuming field 'IsSplittable' is true
7
Taking false branch
2925 Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2926 if (IsDest) {
2927 II.setDest(AdjustedPtr);
2928 II.setDestAlignment(SliceAlign);
2929 }
2930 else {
2931 II.setSource(AdjustedPtr);
2932 II.setSourceAlignment(SliceAlign);
2933 }
2934
2935 LLVM_DEBUG(dbgs() << " to: " << II << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << II <<
"\n"; } } while (false)
;
2936 deleteIfTriviallyDead(OldPtr);
2937 return false;
2938 }
2939 // For split transfer intrinsics we have an incredibly useful assurance:
2940 // the source and destination do not reside within the same alloca, and at
2941 // least one of them does not escape. This means that we can replace
2942 // memmove with memcpy, and we don't need to worry about all manner of
2943 // downsides to splitting and transforming the operations.
2944
2945 // If this doesn't map cleanly onto the alloca type, and that type isn't
2946 // a single value type, just emit a memcpy.
2947 bool EmitMemCpy =
2948 !VecTy && !IntTy &&
8
Assuming field 'VecTy' is non-null
2949 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2950 SliceSize !=
2951 DL.getTypeStoreSize(NewAI.getAllocatedType()).getFixedSize() ||
2952 !NewAI.getAllocatedType()->isSingleValueType());
2953
2954 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2955 // size hasn't been shrunk based on analysis of the viable range, this is
2956 // a no-op.
2957 if (EmitMemCpy
8.1
'EmitMemCpy' is false
&& &OldAI == &NewAI) {
2958 // Ensure the start lines up.
2959 assert(NewBeginOffset == BeginOffset)((NewBeginOffset == BeginOffset) ? static_cast<void> (0
) : __assert_fail ("NewBeginOffset == BeginOffset", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2959, __PRETTY_FUNCTION__))
;
2960
2961 // Rewrite the size as needed.
2962 if (NewEndOffset != EndOffset)
2963 II.setLength(ConstantInt::get(II.getLength()->getType(),
2964 NewEndOffset - NewBeginOffset));
2965 return false;
2966 }
2967 // Record this instruction for deletion.
2968 Pass.DeadInsts.insert(&II);
2969
2970 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2971 // alloca that should be re-examined after rewriting this instruction.
2972 Value *OtherPtr = IsDest
8.2
'IsDest' is false
? II.getRawSource() : II.getRawDest();
9
'?' condition is false
2973 if (AllocaInst *AI
10.1
'AI' is null
=
11
Taking false branch
2974 dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
10
Assuming the object is not a 'AllocaInst'
2975 assert(AI != &OldAI && AI != &NewAI &&((AI != &OldAI && AI != &NewAI && "Splittable transfers cannot reach the same alloca on both ends."
) ? static_cast<void> (0) : __assert_fail ("AI != &OldAI && AI != &NewAI && \"Splittable transfers cannot reach the same alloca on both ends.\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2976, __PRETTY_FUNCTION__))
2976 "Splittable transfers cannot reach the same alloca on both ends.")((AI != &OldAI && AI != &NewAI && "Splittable transfers cannot reach the same alloca on both ends."
) ? static_cast<void> (0) : __assert_fail ("AI != &OldAI && AI != &NewAI && \"Splittable transfers cannot reach the same alloca on both ends.\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 2976, __PRETTY_FUNCTION__))
;
2977 Pass.Worklist.insert(AI);
2978 }
2979
2980 Type *OtherPtrTy = OtherPtr->getType();
2981 unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2982
2983 // Compute the relative offset for the other pointer within the transfer.
2984 unsigned OffsetWidth = DL.getIndexSizeInBits(OtherAS);
2985 APInt OtherOffset(OffsetWidth, NewBeginOffset - BeginOffset);
2986 Align OtherAlign =
2987 (IsDest
11.1
'IsDest' is false
? II.getSourceAlign() : II.getDestAlign()).valueOrOne();
12
'?' condition is false
2988 OtherAlign =
2989 commonAlignment(OtherAlign, OtherOffset.zextOrTrunc(64).getZExtValue());
2990
2991 if (EmitMemCpy
12.1
'EmitMemCpy' is false
) {
13
Taking false branch
2992 // Compute the other pointer, folding as much as possible to produce
2993 // a single, simple GEP in most cases.
2994 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2995 OtherPtr->getName() + ".");
2996
2997 Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2998 Type *SizeTy = II.getLength()->getType();
2999 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
3000
3001 Value *DestPtr, *SrcPtr;
3002 MaybeAlign DestAlign, SrcAlign;
3003 // Note: IsDest is true iff we're copying into the new alloca slice
3004 if (IsDest) {
3005 DestPtr = OurPtr;
3006 DestAlign = SliceAlign;
3007 SrcPtr = OtherPtr;
3008 SrcAlign = OtherAlign;
3009 } else {
3010 DestPtr = OtherPtr;
3011 DestAlign = OtherAlign;
3012 SrcPtr = OurPtr;
3013 SrcAlign = SliceAlign;
3014 }
3015 CallInst *New = IRB.CreateMemCpy(DestPtr, DestAlign, SrcPtr, SrcAlign,
3016 Size, II.isVolatile());
3017 if (AATags)
3018 New->setAAMetadata(AATags);
3019 LLVM_DEBUG(dbgs() << " to: " << *New << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *New <<
"\n"; } } while (false)
;
3020 return false;
3021 }
3022
3023 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
14
Assuming field 'NewBeginOffset' is not equal to field 'NewAllocaBeginOffset'
3024 NewEndOffset == NewAllocaEndOffset;
3025 uint64_t Size = NewEndOffset - NewBeginOffset;
3026 unsigned BeginIndex = VecTy
14.1
Field 'VecTy' is non-null
? getIndex(NewBeginOffset) : 0;
15
'?' condition is true
3027 unsigned EndIndex = VecTy
15.1
Field 'VecTy' is non-null
? getIndex(NewEndOffset) : 0;
16
'?' condition is true
3028 unsigned NumElements = EndIndex - BeginIndex;
3029 IntegerType *SubIntTy =
19
'SubIntTy' initialized to a null pointer value
3030 IntTy ? Type::getIntNTy(IntTy->getContext(), Size * 8) : nullptr;
17
Assuming field 'IntTy' is null
18
'?' condition is false
3031
3032 // Reset the other pointer type to match the register type we're going to
3033 // use, but using the address space of the original other pointer.
3034 Type *OtherTy;
3035 if (VecTy
19.1
Field 'VecTy' is non-null
&& !IsWholeAlloca
19.2
'IsWholeAlloca' is false
) {
20
Taking true branch
3036 if (NumElements == 1)
21
Assuming 'NumElements' is not equal to 1
22
Taking false branch
3037 OtherTy = VecTy->getElementType();
3038 else
3039 OtherTy = FixedVectorType::get(VecTy->getElementType(), NumElements);
3040 } else if (IntTy && !IsWholeAlloca) {
3041 OtherTy = SubIntTy;
3042 } else {
3043 OtherTy = NewAllocaTy;
3044 }
3045 OtherPtrTy = OtherTy->getPointerTo(OtherAS);
3046
3047 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
23
Calling 'getAdjustedPtr'
53
Returning from 'getAdjustedPtr'
3048 OtherPtr->getName() + ".");
3049 MaybeAlign SrcAlign = OtherAlign;
3050 Value *DstPtr = &NewAI;
3051 MaybeAlign DstAlign = SliceAlign;
3052 if (!IsDest
53.1
'IsDest' is false
) {
54
Taking true branch
3053 std::swap(SrcPtr, DstPtr);
3054 std::swap(SrcAlign, DstAlign);
3055 }
3056
3057 Value *Src;
3058 if (VecTy && !IsWholeAlloca && !IsDest) {
55
Assuming field 'VecTy' is null
3059 Src = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
3060 NewAI.getAlign(), "load");
3061 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
3062 } else if (IntTy && !IsWholeAlloca
56.1
'IsWholeAlloca' is false
&& !IsDest
56.2
'IsDest' is false
) {
56
Assuming field 'IntTy' is non-null
57
Taking true branch
3063 Src = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
3064 NewAI.getAlign(), "load");
3065 Src = convertValue(DL, IRB, Src, IntTy);
3066 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
3067 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
58
Passing null pointer value via 4th parameter 'Ty'
59
Calling 'extractInteger'
3068 } else {
3069 LoadInst *Load = IRB.CreateAlignedLoad(OtherTy, SrcPtr, SrcAlign,
3070 II.isVolatile(), "copyload");
3071 if (AATags)
3072 Load->setAAMetadata(AATags);
3073 Src = Load;
3074 }
3075
3076 if (VecTy && !IsWholeAlloca && IsDest) {
3077 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
3078 NewAI.getAlign(), "oldload");
3079 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
3080 } else if (IntTy && !IsWholeAlloca && IsDest) {
3081 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
3082 NewAI.getAlign(), "oldload");
3083 Old = convertValue(DL, IRB, Old, IntTy);
3084 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
3085 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
3086 Src = convertValue(DL, IRB, Src, NewAllocaTy);
3087 }
3088
3089 StoreInst *Store = cast<StoreInst>(
3090 IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
3091 if (AATags)
3092 Store->setAAMetadata(AATags);
3093 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *Store <<
"\n"; } } while (false)
;
3094 return !II.isVolatile();
3095 }
3096
3097 bool visitIntrinsicInst(IntrinsicInst &II) {
3098 assert((II.isLifetimeStartOrEnd() || II.isDroppable()) &&(((II.isLifetimeStartOrEnd() || II.isDroppable()) && "Unexpected intrinsic!"
) ? static_cast<void> (0) : __assert_fail ("(II.isLifetimeStartOrEnd() || II.isDroppable()) && \"Unexpected intrinsic!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3099, __PRETTY_FUNCTION__))
3099 "Unexpected intrinsic!")(((II.isLifetimeStartOrEnd() || II.isDroppable()) && "Unexpected intrinsic!"
) ? static_cast<void> (0) : __assert_fail ("(II.isLifetimeStartOrEnd() || II.isDroppable()) && \"Unexpected intrinsic!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3099, __PRETTY_FUNCTION__))
;
3100 LLVM_DEBUG(dbgs() << " original: " << II << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << II <<
"\n"; } } while (false)
;
3101
3102 // Record this instruction for deletion.
3103 Pass.DeadInsts.insert(&II);
3104
3105 if (II.isDroppable()) {
3106 assert(II.getIntrinsicID() == Intrinsic::assume && "Expected assume")((II.getIntrinsicID() == Intrinsic::assume && "Expected assume"
) ? static_cast<void> (0) : __assert_fail ("II.getIntrinsicID() == Intrinsic::assume && \"Expected assume\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3106, __PRETTY_FUNCTION__))
;
3107 // TODO For now we forget assumed information, this can be improved.
3108 OldPtr->dropDroppableUsesIn(II);
3109 return true;
3110 }
3111
3112 assert(II.getArgOperand(1) == OldPtr)((II.getArgOperand(1) == OldPtr) ? static_cast<void> (0
) : __assert_fail ("II.getArgOperand(1) == OldPtr", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3112, __PRETTY_FUNCTION__))
;
3113 // Lifetime intrinsics are only promotable if they cover the whole alloca.
3114 // Therefore, we drop lifetime intrinsics which don't cover the whole
3115 // alloca.
3116 // (In theory, intrinsics which partially cover an alloca could be
3117 // promoted, but PromoteMemToReg doesn't handle that case.)
3118 // FIXME: Check whether the alloca is promotable before dropping the
3119 // lifetime intrinsics?
3120 if (NewBeginOffset != NewAllocaBeginOffset ||
3121 NewEndOffset != NewAllocaEndOffset)
3122 return true;
3123
3124 ConstantInt *Size =
3125 ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
3126 NewEndOffset - NewBeginOffset);
3127 // Lifetime intrinsics always expect an i8* so directly get such a pointer
3128 // for the new alloca slice.
3129 Type *PointerTy = IRB.getInt8PtrTy(OldPtr->getType()->getPointerAddressSpace());
3130 Value *Ptr = getNewAllocaSlicePtr(IRB, PointerTy);
3131 Value *New;
3132 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
3133 New = IRB.CreateLifetimeStart(Ptr, Size);
3134 else
3135 New = IRB.CreateLifetimeEnd(Ptr, Size);
3136
3137 (void)New;
3138 LLVM_DEBUG(dbgs() << " to: " << *New << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *New <<
"\n"; } } while (false)
;
3139
3140 return true;
3141 }
3142
3143 void fixLoadStoreAlign(Instruction &Root) {
3144 // This algorithm implements the same visitor loop as
3145 // hasUnsafePHIOrSelectUse, and fixes the alignment of each load
3146 // or store found.
3147 SmallPtrSet<Instruction *, 4> Visited;
3148 SmallVector<Instruction *, 4> Uses;
3149 Visited.insert(&Root);
3150 Uses.push_back(&Root);
3151 do {
3152 Instruction *I = Uses.pop_back_val();
3153
3154 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3155 LI->setAlignment(std::min(LI->getAlign(), getSliceAlign()));
3156 continue;
3157 }
3158 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3159 SI->setAlignment(std::min(SI->getAlign(), getSliceAlign()));
3160 continue;
3161 }
3162
3163 assert(isa<BitCastInst>(I) || isa<AddrSpaceCastInst>(I) ||((isa<BitCastInst>(I) || isa<AddrSpaceCastInst>(I
) || isa<PHINode>(I) || isa<SelectInst>(I) || isa
<GetElementPtrInst>(I)) ? static_cast<void> (0) :
__assert_fail ("isa<BitCastInst>(I) || isa<AddrSpaceCastInst>(I) || isa<PHINode>(I) || isa<SelectInst>(I) || isa<GetElementPtrInst>(I)"
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3165, __PRETTY_FUNCTION__))
3164 isa<PHINode>(I) || isa<SelectInst>(I) ||((isa<BitCastInst>(I) || isa<AddrSpaceCastInst>(I
) || isa<PHINode>(I) || isa<SelectInst>(I) || isa
<GetElementPtrInst>(I)) ? static_cast<void> (0) :
__assert_fail ("isa<BitCastInst>(I) || isa<AddrSpaceCastInst>(I) || isa<PHINode>(I) || isa<SelectInst>(I) || isa<GetElementPtrInst>(I)"
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3165, __PRETTY_FUNCTION__))
3165 isa<GetElementPtrInst>(I))((isa<BitCastInst>(I) || isa<AddrSpaceCastInst>(I
) || isa<PHINode>(I) || isa<SelectInst>(I) || isa
<GetElementPtrInst>(I)) ? static_cast<void> (0) :
__assert_fail ("isa<BitCastInst>(I) || isa<AddrSpaceCastInst>(I) || isa<PHINode>(I) || isa<SelectInst>(I) || isa<GetElementPtrInst>(I)"
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3165, __PRETTY_FUNCTION__))
;
3166 for (User *U : I->users())
3167 if (Visited.insert(cast<Instruction>(U)).second)
3168 Uses.push_back(cast<Instruction>(U));
3169 } while (!Uses.empty());
3170 }
3171
3172 bool visitPHINode(PHINode &PN) {
3173 LLVM_DEBUG(dbgs() << " original: " << PN << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << PN <<
"\n"; } } while (false)
;
3174 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable")((BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable"
) ? static_cast<void> (0) : __assert_fail ("BeginOffset >= NewAllocaBeginOffset && \"PHIs are unsplittable\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3174, __PRETTY_FUNCTION__))
;
3175 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable")((EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable"
) ? static_cast<void> (0) : __assert_fail ("EndOffset <= NewAllocaEndOffset && \"PHIs are unsplittable\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3175, __PRETTY_FUNCTION__))
;
3176
3177 // We would like to compute a new pointer in only one place, but have it be
3178 // as local as possible to the PHI. To do that, we re-use the location of
3179 // the old pointer, which necessarily must be in the right position to
3180 // dominate the PHI.
3181 IRBuilderBase::InsertPointGuard Guard(IRB);
3182 if (isa<PHINode>(OldPtr))
3183 IRB.SetInsertPoint(&*OldPtr->getParent()->getFirstInsertionPt());
3184 else
3185 IRB.SetInsertPoint(OldPtr);
3186 IRB.SetCurrentDebugLocation(OldPtr->getDebugLoc());
3187
3188 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3189 // Replace the operands which were using the old pointer.
3190 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
3191
3192 LLVM_DEBUG(dbgs() << " to: " << PN << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << PN <<
"\n"; } } while (false)
;
3193 deleteIfTriviallyDead(OldPtr);
3194
3195 // Fix the alignment of any loads or stores using this PHI node.
3196 fixLoadStoreAlign(PN);
3197
3198 // PHIs can't be promoted on their own, but often can be speculated. We
3199 // check the speculation outside of the rewriter so that we see the
3200 // fully-rewritten alloca.
3201 PHIUsers.insert(&PN);
3202 return true;
3203 }
3204
3205 bool visitSelectInst(SelectInst &SI) {
3206 LLVM_DEBUG(dbgs() << " original: " << SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << SI <<
"\n"; } } while (false)
;
3207 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&(((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr
) && "Pointer isn't an operand!") ? static_cast<void
> (0) : __assert_fail ("(SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) && \"Pointer isn't an operand!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3208, __PRETTY_FUNCTION__))
3208 "Pointer isn't an operand!")(((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr
) && "Pointer isn't an operand!") ? static_cast<void
> (0) : __assert_fail ("(SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) && \"Pointer isn't an operand!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3208, __PRETTY_FUNCTION__))
;
3209 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable")((BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable"
) ? static_cast<void> (0) : __assert_fail ("BeginOffset >= NewAllocaBeginOffset && \"Selects are unsplittable\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3209, __PRETTY_FUNCTION__))
;
3210 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable")((EndOffset <= NewAllocaEndOffset && "Selects are unsplittable"
) ? static_cast<void> (0) : __assert_fail ("EndOffset <= NewAllocaEndOffset && \"Selects are unsplittable\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3210, __PRETTY_FUNCTION__))
;
3211
3212 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3213 // Replace the operands which were using the old pointer.
3214 if (SI.getOperand(1) == OldPtr)
3215 SI.setOperand(1, NewPtr);
3216 if (SI.getOperand(2) == OldPtr)
3217 SI.setOperand(2, NewPtr);
3218
3219 LLVM_DEBUG(dbgs() << " to: " << SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << SI <<
"\n"; } } while (false)
;
3220 deleteIfTriviallyDead(OldPtr);
3221
3222 // Fix the alignment of any loads or stores using this select.
3223 fixLoadStoreAlign(SI);
3224
3225 // Selects can't be promoted on their own, but often can be speculated. We
3226 // check the speculation outside of the rewriter so that we see the
3227 // fully-rewritten alloca.
3228 SelectUsers.insert(&SI);
3229 return true;
3230 }
3231};
3232
3233namespace {
3234
3235/// Visitor to rewrite aggregate loads and stores as scalar.
3236///
3237/// This pass aggressively rewrites all aggregate loads and stores on
3238/// a particular pointer (or any pointer derived from it which we can identify)
3239/// with scalar loads and stores.
3240class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
3241 // Befriend the base class so it can delegate to private visit methods.
3242 friend class InstVisitor<AggLoadStoreRewriter, bool>;
3243
3244 /// Queue of pointer uses to analyze and potentially rewrite.
3245 SmallVector<Use *, 8> Queue;
3246
3247 /// Set to prevent us from cycling with phi nodes and loops.
3248 SmallPtrSet<User *, 8> Visited;
3249
3250 /// The current pointer use being rewritten. This is used to dig up the used
3251 /// value (as opposed to the user).
3252 Use *U = nullptr;
3253
3254 /// Used to calculate offsets, and hence alignment, of subobjects.
3255 const DataLayout &DL;
3256
3257public:
3258 AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
3259
3260 /// Rewrite loads and stores through a pointer and all pointers derived from
3261 /// it.
3262 bool rewrite(Instruction &I) {
3263 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)
;
3264 enqueueUsers(I);
3265 bool Changed = false;
3266 while (!Queue.empty()) {
3267 U = Queue.pop_back_val();
3268 Changed |= visit(cast<Instruction>(U->getUser()));
3269 }
3270 return Changed;
3271 }
3272
3273private:
3274 /// Enqueue all the users of the given instruction for further processing.
3275 /// This uses a set to de-duplicate users.
3276 void enqueueUsers(Instruction &I) {
3277 for (Use &U : I.uses())
3278 if (Visited.insert(U.getUser()).second)
3279 Queue.push_back(&U);
3280 }
3281
3282 // Conservative default is to not rewrite anything.
3283 bool visitInstruction(Instruction &I) { return false; }
3284
3285 /// Generic recursive split emission class.
3286 template <typename Derived> class OpSplitter {
3287 protected:
3288 /// The builder used to form new instructions.
3289 IRBuilderTy IRB;
3290
3291 /// The indices which to be used with insert- or extractvalue to select the
3292 /// appropriate value within the aggregate.
3293 SmallVector<unsigned, 4> Indices;
3294
3295 /// The indices to a GEP instruction which will move Ptr to the correct slot
3296 /// within the aggregate.
3297 SmallVector<Value *, 4> GEPIndices;
3298
3299 /// The base pointer of the original op, used as a base for GEPing the
3300 /// split operations.
3301 Value *Ptr;
3302
3303 /// The base pointee type being GEPed into.
3304 Type *BaseTy;
3305
3306 /// Known alignment of the base pointer.
3307 Align BaseAlign;
3308
3309 /// To calculate offset of each component so we can correctly deduce
3310 /// alignments.
3311 const DataLayout &DL;
3312
3313 /// Initialize the splitter with an insertion point, Ptr and start with a
3314 /// single zero GEP index.
3315 OpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
3316 Align BaseAlign, const DataLayout &DL)
3317 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr),
3318 BaseTy(BaseTy), BaseAlign(BaseAlign), DL(DL) {}
3319
3320 public:
3321 /// Generic recursive split emission routine.
3322 ///
3323 /// This method recursively splits an aggregate op (load or store) into
3324 /// scalar or vector ops. It splits recursively until it hits a single value
3325 /// and emits that single value operation via the template argument.
3326 ///
3327 /// The logic of this routine relies on GEPs and insertvalue and
3328 /// extractvalue all operating with the same fundamental index list, merely
3329 /// formatted differently (GEPs need actual values).
3330 ///
3331 /// \param Ty The type being split recursively into smaller ops.
3332 /// \param Agg The aggregate value being built up or stored, depending on
3333 /// whether this is splitting a load or a store respectively.
3334 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3335 if (Ty->isSingleValueType()) {
3336 unsigned Offset = DL.getIndexedOffsetInType(BaseTy, GEPIndices);
3337 return static_cast<Derived *>(this)->emitFunc(
3338 Ty, Agg, commonAlignment(BaseAlign, Offset), Name);
3339 }
3340
3341 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3342 unsigned OldSize = Indices.size();
3343 (void)OldSize;
3344 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3345 ++Idx) {
3346 assert(Indices.size() == OldSize && "Did not return to the old size")((Indices.size() == OldSize && "Did not return to the old size"
) ? static_cast<void> (0) : __assert_fail ("Indices.size() == OldSize && \"Did not return to the old size\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3346, __PRETTY_FUNCTION__))
;
3347 Indices.push_back(Idx);
3348 GEPIndices.push_back(IRB.getInt32(Idx));
3349 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3350 GEPIndices.pop_back();
3351 Indices.pop_back();
3352 }
3353 return;
3354 }
3355
3356 if (StructType *STy = dyn_cast<StructType>(Ty)) {
3357 unsigned OldSize = Indices.size();
3358 (void)OldSize;
3359 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3360 ++Idx) {
3361 assert(Indices.size() == OldSize && "Did not return to the old size")((Indices.size() == OldSize && "Did not return to the old size"
) ? static_cast<void> (0) : __assert_fail ("Indices.size() == OldSize && \"Did not return to the old size\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3361, __PRETTY_FUNCTION__))
;
3362 Indices.push_back(Idx);
3363 GEPIndices.push_back(IRB.getInt32(Idx));
3364 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3365 GEPIndices.pop_back();
3366 Indices.pop_back();
3367 }
3368 return;
3369 }
3370
3371 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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3371)
;
3372 }
3373 };
3374
3375 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3376 AAMDNodes AATags;
3377
3378 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
3379 AAMDNodes AATags, Align BaseAlign, const DataLayout &DL)
3380 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr, BaseTy, BaseAlign,
3381 DL),
3382 AATags(AATags) {}
3383
3384 /// Emit a leaf load of a single value. This is called at the leaves of the
3385 /// recursive emission to actually load values.
3386 void emitFunc(Type *Ty, Value *&Agg, Align Alignment, const Twine &Name) {
3387 assert(Ty->isSingleValueType())((Ty->isSingleValueType()) ? static_cast<void> (0) :
__assert_fail ("Ty->isSingleValueType()", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3387, __PRETTY_FUNCTION__))
;
3388 // Load the single value and insert it using the indices.
3389 Value *GEP =
3390 IRB.CreateInBoundsGEP(BaseTy, Ptr, GEPIndices, Name + ".gep");
3391 LoadInst *Load =
3392 IRB.CreateAlignedLoad(Ty, GEP, Alignment, Name + ".load");
3393 if (AATags)
3394 Load->setAAMetadata(AATags);
3395 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3396 LLVM_DEBUG(dbgs() << " to: " << *Load << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *Load <<
"\n"; } } while (false)
;
3397 }
3398 };
3399
3400 bool visitLoadInst(LoadInst &LI) {
3401 assert(LI.getPointerOperand() == *U)((LI.getPointerOperand() == *U) ? static_cast<void> (0)
: __assert_fail ("LI.getPointerOperand() == *U", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3401, __PRETTY_FUNCTION__))
;
3402 if (!LI.isSimple() || LI.getType()->isSingleValueType())
3403 return false;
3404
3405 // We have an aggregate being loaded, split it apart.
3406 LLVM_DEBUG(dbgs() << " original: " << LI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << LI <<
"\n"; } } while (false)
;
3407 AAMDNodes AATags;
3408 LI.getAAMetadata(AATags);
3409 LoadOpSplitter Splitter(&LI, *U, LI.getType(), AATags,
3410 getAdjustedAlignment(&LI, 0), DL);
3411 Value *V = UndefValue::get(LI.getType());
3412 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3413 Visited.erase(&LI);
3414 LI.replaceAllUsesWith(V);
3415 LI.eraseFromParent();
3416 return true;
3417 }
3418
3419 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3420 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
3421 AAMDNodes AATags, Align BaseAlign, const DataLayout &DL)
3422 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr, BaseTy, BaseAlign,
3423 DL),
3424 AATags(AATags) {}
3425 AAMDNodes AATags;
3426 /// Emit a leaf store of a single value. This is called at the leaves of the
3427 /// recursive emission to actually produce stores.
3428 void emitFunc(Type *Ty, Value *&Agg, Align Alignment, const Twine &Name) {
3429 assert(Ty->isSingleValueType())((Ty->isSingleValueType()) ? static_cast<void> (0) :
__assert_fail ("Ty->isSingleValueType()", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3429, __PRETTY_FUNCTION__))
;
3430 // Extract the single value and store it using the indices.
3431 //
3432 // The gep and extractvalue values are factored out of the CreateStore
3433 // call to make the output independent of the argument evaluation order.
3434 Value *ExtractValue =
3435 IRB.CreateExtractValue(Agg, Indices, Name + ".extract");
3436 Value *InBoundsGEP =
3437 IRB.CreateInBoundsGEP(BaseTy, Ptr, GEPIndices, Name + ".gep");
3438 StoreInst *Store =
3439 IRB.CreateAlignedStore(ExtractValue, InBoundsGEP, Alignment);
3440 if (AATags)
3441 Store->setAAMetadata(AATags);
3442 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *Store <<
"\n"; } } while (false)
;
3443 }
3444 };
3445
3446 bool visitStoreInst(StoreInst &SI) {
3447 if (!SI.isSimple() || SI.getPointerOperand() != *U)
3448 return false;
3449 Value *V = SI.getValueOperand();
3450 if (V->getType()->isSingleValueType())
3451 return false;
3452
3453 // We have an aggregate being stored, split it apart.
3454 LLVM_DEBUG(dbgs() << " original: " << SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << SI <<
"\n"; } } while (false)
;
3455 AAMDNodes AATags;
3456 SI.getAAMetadata(AATags);
3457 StoreOpSplitter Splitter(&SI, *U, V->getType(), AATags,
3458 getAdjustedAlignment(&SI, 0), DL);
3459 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3460 Visited.erase(&SI);
3461 SI.eraseFromParent();
3462 return true;
3463 }
3464
3465 bool visitBitCastInst(BitCastInst &BC) {
3466 enqueueUsers(BC);
3467 return false;
3468 }
3469
3470 bool visitAddrSpaceCastInst(AddrSpaceCastInst &ASC) {
3471 enqueueUsers(ASC);
3472 return false;
3473 }
3474
3475 // Fold gep (select cond, ptr1, ptr2) => select cond, gep(ptr1), gep(ptr2)
3476 bool foldGEPSelect(GetElementPtrInst &GEPI) {
3477 if (!GEPI.hasAllConstantIndices())
3478 return false;
3479
3480 SelectInst *Sel = cast<SelectInst>(GEPI.getPointerOperand());
3481
3482 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)
3483 << "\n original: " << *Seldo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Rewriting gep(select) -> select(gep):"
<< "\n original: " << *Sel << "\n "
<< GEPI; } } while (false)
3484 << "\n " << GEPI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Rewriting gep(select) -> select(gep):"
<< "\n original: " << *Sel << "\n "
<< GEPI; } } while (false)
;
3485
3486 IRBuilderTy Builder(&GEPI);
3487 SmallVector<Value *, 4> Index(GEPI.idx_begin(), GEPI.idx_end());
3488 bool IsInBounds = GEPI.isInBounds();
3489
3490 Value *True = Sel->getTrueValue();
3491 Value *NTrue =
3492 IsInBounds
3493 ? Builder.CreateInBoundsGEP(True, Index,
3494 True->getName() + ".sroa.gep")
3495 : Builder.CreateGEP(True, Index, True->getName() + ".sroa.gep");
3496
3497 Value *False = Sel->getFalseValue();
3498
3499 Value *NFalse =
3500 IsInBounds
3501 ? Builder.CreateInBoundsGEP(False, Index,
3502 False->getName() + ".sroa.gep")
3503 : Builder.CreateGEP(False, Index, False->getName() + ".sroa.gep");
3504
3505 Value *NSel = Builder.CreateSelect(Sel->getCondition(), NTrue, NFalse,
3506 Sel->getName() + ".sroa.sel");
3507 Visited.erase(&GEPI);
3508 GEPI.replaceAllUsesWith(NSel);
3509 GEPI.eraseFromParent();
3510 Instruction *NSelI = cast<Instruction>(NSel);
3511 Visited.insert(NSelI);
3512 enqueueUsers(*NSelI);
3513
3514 LLVM_DEBUG(dbgs() << "\n to: " << *NTruedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "\n to: " << *NTrue
<< "\n " << *NFalse << "\n "
<< *NSel << '\n'; } } while (false)
3515 << "\n " << *NFalsedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "\n to: " << *NTrue
<< "\n " << *NFalse << "\n "
<< *NSel << '\n'; } } while (false)
3516 << "\n " << *NSel << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "\n to: " << *NTrue
<< "\n " << *NFalse << "\n "
<< *NSel << '\n'; } } while (false)
;
3517
3518 return true;
3519 }
3520
3521 // Fold gep (phi ptr1, ptr2) => phi gep(ptr1), gep(ptr2)
3522 bool foldGEPPhi(GetElementPtrInst &GEPI) {
3523 if (!GEPI.hasAllConstantIndices())
3524 return false;
3525
3526 PHINode *PHI = cast<PHINode>(GEPI.getPointerOperand());
3527 if (GEPI.getParent() != PHI->getParent() ||
3528 llvm::any_of(PHI->incoming_values(), [](Value *In)
3529 { Instruction *I = dyn_cast<Instruction>(In);
3530 return !I || isa<GetElementPtrInst>(I) || isa<PHINode>(I) ||
3531 succ_empty(I->getParent()) ||
3532 !I->getParent()->isLegalToHoistInto();
3533 }))
3534 return false;
3535
3536 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)
3537 << "\n original: " << *PHIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Rewriting gep(phi) -> phi(gep):"
<< "\n original: " << *PHI << "\n "
<< GEPI << "\n to: "; } } while (false)
3538 << "\n " << GEPIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Rewriting gep(phi) -> phi(gep):"
<< "\n original: " << *PHI << "\n "
<< GEPI << "\n to: "; } } while (false)
3539 << "\n to: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Rewriting gep(phi) -> phi(gep):"
<< "\n original: " << *PHI << "\n "
<< GEPI << "\n to: "; } } while (false)
;
3540
3541 SmallVector<Value *, 4> Index(GEPI.idx_begin(), GEPI.idx_end());
3542 bool IsInBounds = GEPI.isInBounds();
3543 IRBuilderTy PHIBuilder(GEPI.getParent()->getFirstNonPHI());
3544 PHINode *NewPN = PHIBuilder.CreatePHI(GEPI.getType(),
3545 PHI->getNumIncomingValues(),
3546 PHI->getName() + ".sroa.phi");
3547 for (unsigned I = 0, E = PHI->getNumIncomingValues(); I != E; ++I) {
3548 Instruction *In = cast<Instruction>(PHI->getIncomingValue(I));
3549
3550 IRBuilderTy B(In->getParent(), std::next(In->getIterator()));
3551 Value *NewVal = IsInBounds
3552 ? B.CreateInBoundsGEP(In, Index, In->getName() + ".sroa.gep")
3553 : B.CreateGEP(In, Index, In->getName() + ".sroa.gep");
3554 NewPN->addIncoming(NewVal, PHI->getIncomingBlock(I));
3555 }
3556
3557 Visited.erase(&GEPI);
3558 GEPI.replaceAllUsesWith(NewPN);
3559 GEPI.eraseFromParent();
3560 Visited.insert(NewPN);
3561 enqueueUsers(*NewPN);
3562
3563 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)
3564 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)
3565 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)
;
3566
3567 return true;
3568 }
3569
3570 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3571 if (isa<SelectInst>(GEPI.getPointerOperand()) &&
3572 foldGEPSelect(GEPI))
3573 return true;
3574
3575 if (isa<PHINode>(GEPI.getPointerOperand()) &&
3576 foldGEPPhi(GEPI))
3577 return true;
3578
3579 enqueueUsers(GEPI);
3580 return false;
3581 }
3582
3583 bool visitPHINode(PHINode &PN) {
3584 enqueueUsers(PN);
3585 return false;
3586 }
3587
3588 bool visitSelectInst(SelectInst &SI) {
3589 enqueueUsers(SI);
3590 return false;
3591 }
3592};
3593
3594} // end anonymous namespace
3595
3596/// Strip aggregate type wrapping.
3597///
3598/// This removes no-op aggregate types wrapping an underlying type. It will
3599/// strip as many layers of types as it can without changing either the type
3600/// size or the allocated size.
3601static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
3602 if (Ty->isSingleValueType())
3603 return Ty;
3604
3605 uint64_t AllocSize = DL.getTypeAllocSize(Ty).getFixedSize();
3606 uint64_t TypeSize = DL.getTypeSizeInBits(Ty).getFixedSize();
3607
3608 Type *InnerTy;
3609 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3610 InnerTy = ArrTy->getElementType();
3611 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3612 const StructLayout *SL = DL.getStructLayout(STy);
3613 unsigned Index = SL->getElementContainingOffset(0);
3614 InnerTy = STy->getElementType(Index);
3615 } else {
3616 return Ty;
3617 }
3618
3619 if (AllocSize > DL.getTypeAllocSize(InnerTy).getFixedSize() ||
3620 TypeSize > DL.getTypeSizeInBits(InnerTy).getFixedSize())
3621 return Ty;
3622
3623 return stripAggregateTypeWrapping(DL, InnerTy);
3624}
3625
3626/// Try to find a partition of the aggregate type passed in for a given
3627/// offset and size.
3628///
3629/// This recurses through the aggregate type and tries to compute a subtype
3630/// based on the offset and size. When the offset and size span a sub-section
3631/// of an array, it will even compute a new array type for that sub-section,
3632/// and the same for structs.
3633///
3634/// Note that this routine is very strict and tries to find a partition of the
3635/// type which produces the *exact* right offset and size. It is not forgiving
3636/// when the size or offset cause either end of type-based partition to be off.
3637/// Also, this is a best-effort routine. It is reasonable to give up and not
3638/// return a type if necessary.
3639static Type *getTypePartition(const DataLayout &DL, Type *Ty, uint64_t Offset,
3640 uint64_t Size) {
3641 if (Offset == 0 && DL.getTypeAllocSize(Ty).getFixedSize() == Size)
3642 return stripAggregateTypeWrapping(DL, Ty);
3643 if (Offset > DL.getTypeAllocSize(Ty).getFixedSize() ||
3644 (DL.getTypeAllocSize(Ty).getFixedSize() - Offset) < Size)
3645 return nullptr;
3646
3647 if (isa<ArrayType>(Ty) || isa<VectorType>(Ty)) {
3648 Type *ElementTy;
3649 uint64_t TyNumElements;
3650 if (auto *AT = dyn_cast<ArrayType>(Ty)) {
3651 ElementTy = AT->getElementType();
3652 TyNumElements = AT->getNumElements();
3653 } else {
3654 // FIXME: This isn't right for vectors with non-byte-sized or
3655 // non-power-of-two sized elements.
3656 auto *VT = cast<FixedVectorType>(Ty);
3657 ElementTy = VT->getElementType();
3658 TyNumElements = VT->getNumElements();
3659 }
3660 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy).getFixedSize();
3661 uint64_t NumSkippedElements = Offset / ElementSize;
3662 if (NumSkippedElements >= TyNumElements)
3663 return nullptr;
3664 Offset -= NumSkippedElements * ElementSize;
3665
3666 // First check if we need to recurse.
3667 if (Offset > 0 || Size < ElementSize) {
3668 // Bail if the partition ends in a different array element.
3669 if ((Offset + Size) > ElementSize)
3670 return nullptr;
3671 // Recurse through the element type trying to peel off offset bytes.
3672 return getTypePartition(DL, ElementTy, Offset, Size);
3673 }
3674 assert(Offset == 0)((Offset == 0) ? static_cast<void> (0) : __assert_fail (
"Offset == 0", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3674, __PRETTY_FUNCTION__))
;
3675
3676 if (Size == ElementSize)
3677 return stripAggregateTypeWrapping(DL, ElementTy);
3678 assert(Size > ElementSize)((Size > ElementSize) ? static_cast<void> (0) : __assert_fail
("Size > ElementSize", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3678, __PRETTY_FUNCTION__))
;
3679 uint64_t NumElements = Size / ElementSize;
3680 if (NumElements * ElementSize != Size)
3681 return nullptr;
3682 return ArrayType::get(ElementTy, NumElements);
3683 }
3684
3685 StructType *STy = dyn_cast<StructType>(Ty);
3686 if (!STy)
3687 return nullptr;
3688
3689 const StructLayout *SL = DL.getStructLayout(STy);
3690 if (Offset >= SL->getSizeInBytes())
3691 return nullptr;
3692 uint64_t EndOffset = Offset + Size;
3693 if (EndOffset > SL->getSizeInBytes())
3694 return nullptr;
3695
3696 unsigned Index = SL->getElementContainingOffset(Offset);
3697 Offset -= SL->getElementOffset(Index);
3698
3699 Type *ElementTy = STy->getElementType(Index);
3700 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy).getFixedSize();
3701 if (Offset >= ElementSize)
3702 return nullptr; // The offset points into alignment padding.
3703
3704 // See if any partition must be contained by the element.
3705 if (Offset > 0 || Size < ElementSize) {
3706 if ((Offset + Size) > ElementSize)
3707 return nullptr;
3708 return getTypePartition(DL, ElementTy, Offset, Size);
3709 }
3710 assert(Offset == 0)((Offset == 0) ? static_cast<void> (0) : __assert_fail (
"Offset == 0", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3710, __PRETTY_FUNCTION__))
;
3711
3712 if (Size == ElementSize)
3713 return stripAggregateTypeWrapping(DL, ElementTy);
3714
3715 StructType::element_iterator EI = STy->element_begin() + Index,
3716 EE = STy->element_end();
3717 if (EndOffset < SL->getSizeInBytes()) {
3718 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3719 if (Index == EndIndex)
3720 return nullptr; // Within a single element and its padding.
3721
3722 // Don't try to form "natural" types if the elements don't line up with the
3723 // expected size.
3724 // FIXME: We could potentially recurse down through the last element in the
3725 // sub-struct to find a natural end point.
3726 if (SL->getElementOffset(EndIndex) != EndOffset)
3727 return nullptr;
3728
3729 assert(Index < EndIndex)((Index < EndIndex) ? static_cast<void> (0) : __assert_fail
("Index < EndIndex", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3729, __PRETTY_FUNCTION__))
;
3730 EE = STy->element_begin() + EndIndex;
3731 }
3732
3733 // Try to build up a sub-structure.
3734 StructType *SubTy =
3735 StructType::get(STy->getContext(), makeArrayRef(EI, EE), STy->isPacked());
3736 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3737 if (Size != SubSL->getSizeInBytes())
3738 return nullptr; // The sub-struct doesn't have quite the size needed.
3739
3740 return SubTy;
3741}
3742
3743/// Pre-split loads and stores to simplify rewriting.
3744///
3745/// We want to break up the splittable load+store pairs as much as
3746/// possible. This is important to do as a preprocessing step, as once we
3747/// start rewriting the accesses to partitions of the alloca we lose the
3748/// necessary information to correctly split apart paired loads and stores
3749/// which both point into this alloca. The case to consider is something like
3750/// the following:
3751///
3752/// %a = alloca [12 x i8]
3753/// %gep1 = getelementptr [12 x i8]* %a, i32 0, i32 0
3754/// %gep2 = getelementptr [12 x i8]* %a, i32 0, i32 4
3755/// %gep3 = getelementptr [12 x i8]* %a, i32 0, i32 8
3756/// %iptr1 = bitcast i8* %gep1 to i64*
3757/// %iptr2 = bitcast i8* %gep2 to i64*
3758/// %fptr1 = bitcast i8* %gep1 to float*
3759/// %fptr2 = bitcast i8* %gep2 to float*
3760/// %fptr3 = bitcast i8* %gep3 to float*
3761/// store float 0.0, float* %fptr1
3762/// store float 1.0, float* %fptr2
3763/// %v = load i64* %iptr1
3764/// store i64 %v, i64* %iptr2
3765/// %f1 = load float* %fptr2
3766/// %f2 = load float* %fptr3
3767///
3768/// Here we want to form 3 partitions of the alloca, each 4 bytes large, and
3769/// promote everything so we recover the 2 SSA values that should have been
3770/// there all along.
3771///
3772/// \returns true if any changes are made.
3773bool SROA::presplitLoadsAndStores(AllocaInst &AI, AllocaSlices &AS) {
3774 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)
;
3775
3776 // Track the loads and stores which are candidates for pre-splitting here, in
3777 // the order they first appear during the partition scan. These give stable
3778 // iteration order and a basis for tracking which loads and stores we
3779 // actually split.
3780 SmallVector<LoadInst *, 4> Loads;
3781 SmallVector<StoreInst *, 4> Stores;
3782
3783 // We need to accumulate the splits required of each load or store where we
3784 // can find them via a direct lookup. This is important to cross-check loads
3785 // and stores against each other. We also track the slice so that we can kill
3786 // all the slices that end up split.
3787 struct SplitOffsets {
3788 Slice *S;
3789 std::vector<uint64_t> Splits;
3790 };
3791 SmallDenseMap<Instruction *, SplitOffsets, 8> SplitOffsetsMap;
3792
3793 // Track loads out of this alloca which cannot, for any reason, be pre-split.
3794 // This is important as we also cannot pre-split stores of those loads!
3795 // FIXME: This is all pretty gross. It means that we can be more aggressive
3796 // in pre-splitting when the load feeding the store happens to come from
3797 // a separate alloca. Put another way, the effectiveness of SROA would be
3798 // decreased by a frontend which just concatenated all of its local allocas
3799 // into one big flat alloca. But defeating such patterns is exactly the job
3800 // SROA is tasked with! Sadly, to not have this discrepancy we would have
3801 // change store pre-splitting to actually force pre-splitting of the load
3802 // that feeds it *and all stores*. That makes pre-splitting much harder, but
3803 // maybe it would make it more principled?
3804 SmallPtrSet<LoadInst *, 8> UnsplittableLoads;
3805
3806 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)
;
3807 for (auto &P : AS.partitions()) {
3808 for (Slice &S : P) {
3809 Instruction *I = cast<Instruction>(S.getUse()->getUser());
3810 if (!S.isSplittable() || S.endOffset() <= P.endOffset()) {
3811 // If this is a load we have to track that it can't participate in any
3812 // pre-splitting. If this is a store of a load we have to track that
3813 // that load also can't participate in any pre-splitting.
3814 if (auto *LI = dyn_cast<LoadInst>(I))
3815 UnsplittableLoads.insert(LI);
3816 else if (auto *SI = dyn_cast<StoreInst>(I))
3817 if (auto *LI = dyn_cast<LoadInst>(SI->getValueOperand()))
3818 UnsplittableLoads.insert(LI);
3819 continue;
3820 }
3821 assert(P.endOffset() > S.beginOffset() &&((P.endOffset() > S.beginOffset() && "Empty or backwards partition!"
) ? static_cast<void> (0) : __assert_fail ("P.endOffset() > S.beginOffset() && \"Empty or backwards partition!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3822, __PRETTY_FUNCTION__))
3822 "Empty or backwards partition!")((P.endOffset() > S.beginOffset() && "Empty or backwards partition!"
) ? static_cast<void> (0) : __assert_fail ("P.endOffset() > S.beginOffset() && \"Empty or backwards partition!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3822, __PRETTY_FUNCTION__))
;
3823
3824 // Determine if this is a pre-splittable slice.
3825 if (auto *LI = dyn_cast<LoadInst>(I)) {
3826 assert(!LI->isVolatile() && "Cannot split volatile loads!")((!LI->isVolatile() && "Cannot split volatile loads!"
) ? static_cast<void> (0) : __assert_fail ("!LI->isVolatile() && \"Cannot split volatile loads!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3826, __PRETTY_FUNCTION__))
;
3827
3828 // The load must be used exclusively to store into other pointers for
3829 // us to be able to arbitrarily pre-split it. The stores must also be
3830 // simple to avoid changing semantics.
3831 auto IsLoadSimplyStored = [](LoadInst *LI) {
3832 for (User *LU : LI->users()) {
3833 auto *SI = dyn_cast<StoreInst>(LU);
3834 if (!SI || !SI->isSimple())
3835 return false;
3836 }
3837 return true;
3838 };
3839 if (!IsLoadSimplyStored(LI)) {
3840 UnsplittableLoads.insert(LI);
3841 continue;
3842 }
3843
3844 Loads.push_back(LI);
3845 } else if (auto *SI = dyn_cast<StoreInst>(I)) {
3846 if (S.getUse() != &SI->getOperandUse(SI->getPointerOperandIndex()))
3847 // Skip stores *of* pointers. FIXME: This shouldn't even be possible!
3848 continue;
3849 auto *StoredLoad = dyn_cast<LoadInst>(SI->getValueOperand());
3850 if (!StoredLoad || !StoredLoad->isSimple())
3851 continue;
3852 assert(!SI->isVolatile() && "Cannot split volatile stores!")((!SI->isVolatile() && "Cannot split volatile stores!"
) ? static_cast<void> (0) : __assert_fail ("!SI->isVolatile() && \"Cannot split volatile stores!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3852, __PRETTY_FUNCTION__))
;
3853
3854 Stores.push_back(SI);
3855 } else {
3856 // Other uses cannot be pre-split.
3857 continue;
3858 }
3859
3860 // Record the initial split.
3861 LLVM_DEBUG(dbgs() << " Candidate: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Candidate: " << *I <<
"\n"; } } while (false)
;
3862 auto &Offsets = SplitOffsetsMap[I];
3863 assert(Offsets.Splits.empty() &&((Offsets.Splits.empty() && "Should not have splits the first time we see an instruction!"
) ? static_cast<void> (0) : __assert_fail ("Offsets.Splits.empty() && \"Should not have splits the first time we see an instruction!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3864, __PRETTY_FUNCTION__))
3864 "Should not have splits the first time we see an instruction!")((Offsets.Splits.empty() && "Should not have splits the first time we see an instruction!"
) ? static_cast<void> (0) : __assert_fail ("Offsets.Splits.empty() && \"Should not have splits the first time we see an instruction!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3864, __PRETTY_FUNCTION__))
;
3865 Offsets.S = &S;
3866 Offsets.Splits.push_back(P.endOffset() - S.beginOffset());
3867 }
3868
3869 // Now scan the already split slices, and add a split for any of them which
3870 // we're going to pre-split.
3871 for (Slice *S : P.splitSliceTails()) {
3872 auto SplitOffsetsMapI =
3873 SplitOffsetsMap.find(cast<Instruction>(S->getUse()->getUser()));
3874 if (SplitOffsetsMapI == SplitOffsetsMap.end())
3875 continue;
3876 auto &Offsets = SplitOffsetsMapI->second;
3877
3878 assert(Offsets.S == S && "Found a mismatched slice!")((Offsets.S == S && "Found a mismatched slice!") ? static_cast
<void> (0) : __assert_fail ("Offsets.S == S && \"Found a mismatched slice!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3878, __PRETTY_FUNCTION__))
;
3879 assert(!Offsets.Splits.empty() &&((!Offsets.Splits.empty() && "Cannot have an empty set of splits on the second partition!"
) ? static_cast<void> (0) : __assert_fail ("!Offsets.Splits.empty() && \"Cannot have an empty set of splits on the second partition!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3880, __PRETTY_FUNCTION__))
3880 "Cannot have an empty set of splits on the second partition!")((!Offsets.Splits.empty() && "Cannot have an empty set of splits on the second partition!"
) ? static_cast<void> (0) : __assert_fail ("!Offsets.Splits.empty() && \"Cannot have an empty set of splits on the second partition!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3880, __PRETTY_FUNCTION__))
;
3881 assert(Offsets.Splits.back() ==((Offsets.Splits.back() == P.beginOffset() - Offsets.S->beginOffset
() && "Previous split does not end where this one begins!"
) ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3883, __PRETTY_FUNCTION__))
3882 P.beginOffset() - Offsets.S->beginOffset() &&((Offsets.Splits.back() == P.beginOffset() - Offsets.S->beginOffset
() && "Previous split does not end where this one begins!"
) ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3883, __PRETTY_FUNCTION__))
3883 "Previous split does not end where this one begins!")((Offsets.Splits.back() == P.beginOffset() - Offsets.S->beginOffset
() && "Previous split does not end where this one begins!"
) ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3883, __PRETTY_FUNCTION__))
;
3884
3885 // Record each split. The last partition's end isn't needed as the size
3886 // of the slice dictates that.
3887 if (S->endOffset() > P.endOffset())
3888 Offsets.Splits.push_back(P.endOffset() - Offsets.S->beginOffset());
3889 }
3890 }
3891
3892 // We may have split loads where some of their stores are split stores. For
3893 // such loads and stores, we can only pre-split them if their splits exactly
3894 // match relative to their starting offset. We have to verify this prior to
3895 // any rewriting.
3896 Stores.erase(
3897 llvm::remove_if(Stores,
3898 [&UnsplittableLoads, &SplitOffsetsMap](StoreInst *SI) {
3899 // Lookup the load we are storing in our map of split
3900 // offsets.
3901 auto *LI = cast<LoadInst>(SI->getValueOperand());
3902 // If it was completely unsplittable, then we're done,
3903 // and this store can't be pre-split.
3904 if (UnsplittableLoads.count(LI))
3905 return true;
3906
3907 auto LoadOffsetsI = SplitOffsetsMap.find(LI);
3908 if (LoadOffsetsI == SplitOffsetsMap.end())
3909 return false; // Unrelated loads are definitely safe.
3910 auto &LoadOffsets = LoadOffsetsI->second;
3911
3912 // Now lookup the store's offsets.
3913 auto &StoreOffsets = SplitOffsetsMap[SI];
3914
3915 // If the relative offsets of each split in the load and
3916 // store match exactly, then we can split them and we
3917 // don't need to remove them here.
3918 if (LoadOffsets.Splits == StoreOffsets.Splits)
3919 return false;
3920
3921 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Mismatched splits for load and store:\n"
<< " " << *LI << "\n" << " "
<< *SI << "\n"; } } while (false)
3922 dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Mismatched splits for load and store:\n"
<< " " << *LI << "\n" << " "
<< *SI << "\n"; } } while (false)
3923 << " 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)
3924 << " " << *LI << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Mismatched splits for load and store:\n"
<< " " << *LI << "\n" << " "
<< *SI << "\n"; } } while (false)
3925 << " " << *SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Mismatched splits for load and store:\n"
<< " " << *LI << "\n" << " "
<< *SI << "\n"; } } while (false)
;
3926
3927 // We've found a store and load that we need to split
3928 // with mismatched relative splits. Just give up on them
3929 // and remove both instructions from our list of
3930 // candidates.
3931 UnsplittableLoads.insert(LI);
3932 return true;
3933 }),
3934 Stores.end());
3935 // Now we have to go *back* through all the stores, because a later store may
3936 // have caused an earlier store's load to become unsplittable and if it is
3937 // unsplittable for the later store, then we can't rely on it being split in
3938 // the earlier store either.
3939 Stores.erase(llvm::remove_if(Stores,
3940 [&UnsplittableLoads](StoreInst *SI) {
3941 auto *LI =
3942 cast<LoadInst>(SI->getValueOperand());
3943 return UnsplittableLoads.count(LI);
3944 }),
3945 Stores.end());
3946 // Once we've established all the loads that can't be split for some reason,
3947 // filter any that made it into our list out.
3948 Loads.erase(llvm::remove_if(Loads,
3949 [&UnsplittableLoads](LoadInst *LI) {
3950 return UnsplittableLoads.count(LI);
3951 }),
3952 Loads.end());
3953
3954 // If no loads or stores are left, there is no pre-splitting to be done for
3955 // this alloca.
3956 if (Loads.empty() && Stores.empty())
3957 return false;
3958
3959 // From here on, we can't fail and will be building new accesses, so rig up
3960 // an IR builder.
3961 IRBuilderTy IRB(&AI);
3962
3963 // Collect the new slices which we will merge into the alloca slices.
3964 SmallVector<Slice, 4> NewSlices;
3965
3966 // Track any allocas we end up splitting loads and stores for so we iterate
3967 // on them.
3968 SmallPtrSet<AllocaInst *, 4> ResplitPromotableAllocas;
3969
3970 // At this point, we have collected all of the loads and stores we can
3971 // pre-split, and the specific splits needed for them. We actually do the
3972 // splitting in a specific order in order to handle when one of the loads in
3973 // the value operand to one of the stores.
3974 //
3975 // First, we rewrite all of the split loads, and just accumulate each split
3976 // load in a parallel structure. We also build the slices for them and append
3977 // them to the alloca slices.
3978 SmallDenseMap<LoadInst *, std::vector<LoadInst *>, 1> SplitLoadsMap;
3979 std::vector<LoadInst *> SplitLoads;
3980 const DataLayout &DL = AI.getModule()->getDataLayout();
3981 for (LoadInst *LI : Loads) {
3982 SplitLoads.clear();
3983
3984 IntegerType *Ty = cast<IntegerType>(LI->getType());
3985 uint64_t LoadSize = Ty->getBitWidth() / 8;
3986 assert(LoadSize > 0 && "Cannot have a zero-sized integer load!")((LoadSize > 0 && "Cannot have a zero-sized integer load!"
) ? static_cast<void> (0) : __assert_fail ("LoadSize > 0 && \"Cannot have a zero-sized integer load!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3986, __PRETTY_FUNCTION__))
;
3987
3988 auto &Offsets = SplitOffsetsMap[LI];
3989 assert(LoadSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&((LoadSize == Offsets.S->endOffset() - Offsets.S->beginOffset
() && "Slice size should always match load size exactly!"
) ? static_cast<void> (0) : __assert_fail ("LoadSize == Offsets.S->endOffset() - Offsets.S->beginOffset() && \"Slice size should always match load size exactly!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3990, __PRETTY_FUNCTION__))
3990 "Slice size should always match load size exactly!")((LoadSize == Offsets.S->endOffset() - Offsets.S->beginOffset
() && "Slice size should always match load size exactly!"
) ? static_cast<void> (0) : __assert_fail ("LoadSize == Offsets.S->endOffset() - Offsets.S->beginOffset() && \"Slice size should always match load size exactly!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3990, __PRETTY_FUNCTION__))
;
3991 uint64_t BaseOffset = Offsets.S->beginOffset();
3992 assert(BaseOffset + LoadSize > BaseOffset &&((BaseOffset + LoadSize > BaseOffset && "Cannot represent alloca access size using 64-bit integers!"
) ? static_cast<void> (0) : __assert_fail ("BaseOffset + LoadSize > BaseOffset && \"Cannot represent alloca access size using 64-bit integers!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3993, __PRETTY_FUNCTION__))
3993 "Cannot represent alloca access size using 64-bit integers!")((BaseOffset + LoadSize > BaseOffset && "Cannot represent alloca access size using 64-bit integers!"
) ? static_cast<void> (0) : __assert_fail ("BaseOffset + LoadSize > BaseOffset && \"Cannot represent alloca access size using 64-bit integers!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 3993, __PRETTY_FUNCTION__))
;
3994
3995 Instruction *BasePtr = cast<Instruction>(LI->getPointerOperand());
3996 IRB.SetInsertPoint(LI);
3997
3998 LLVM_DEBUG(dbgs() << " Splitting load: " << *LI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Splitting load: " << *LI
<< "\n"; } } while (false)
;
3999
4000 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
4001 int Idx = 0, Size = Offsets.Splits.size();
4002 for (;;) {
4003 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
4004 auto AS = LI->getPointerAddressSpace();
4005 auto *PartPtrTy = PartTy->getPointerTo(AS);
4006 LoadInst *PLoad = IRB.CreateAlignedLoad(
4007 PartTy,
4008 getAdjustedPtr(IRB, DL, BasePtr,
4009 APInt(DL.getIndexSizeInBits(AS), PartOffset),
4010 PartPtrTy, BasePtr->getName() + "."),
4011 getAdjustedAlignment(LI, PartOffset),
4012 /*IsVolatile*/ false, LI->getName());
4013 PLoad->copyMetadata(*LI, {LLVMContext::MD_mem_parallel_loop_access,
4014 LLVMContext::MD_access_group});
4015
4016 // Append this load onto the list of split loads so we can find it later
4017 // to rewrite the stores.
4018 SplitLoads.push_back(PLoad);
4019
4020 // Now build a new slice for the alloca.
4021 NewSlices.push_back(
4022 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
4023 &PLoad->getOperandUse(PLoad->getPointerOperandIndex()),
4024 /*IsSplittable*/ false));
4025 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)
4026 << ", " << NewSlices.back().endOffset()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " new slice [" << NewSlices
.back().beginOffset() << ", " << NewSlices.back()
.endOffset() << "): " << *PLoad << "\n"; } }
while (false)
4027 << "): " << *PLoad << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " new slice [" << NewSlices
.back().beginOffset() << ", " << NewSlices.back()
.endOffset() << "): " << *PLoad << "\n"; } }
while (false)
;
4028
4029 // See if we've handled all the splits.
4030 if (Idx >= Size)
4031 break;
4032
4033 // Setup the next partition.
4034 PartOffset = Offsets.Splits[Idx];
4035 ++Idx;
4036 PartSize = (Idx < Size ? Offsets.Splits[Idx] : LoadSize) - PartOffset;
4037 }
4038
4039 // Now that we have the split loads, do the slow walk over all uses of the
4040 // load and rewrite them as split stores, or save the split loads to use
4041 // below if the store is going to be split there anyways.
4042 bool DeferredStores = false;
4043 for (User *LU : LI->users()) {
4044 StoreInst *SI = cast<StoreInst>(LU);
4045 if (!Stores.empty() && SplitOffsetsMap.count(SI)) {
4046 DeferredStores = true;
4047 LLVM_DEBUG(dbgs() << " Deferred splitting of store: " << *SIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Deferred splitting of store: "
<< *SI << "\n"; } } while (false)
4048 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Deferred splitting of store: "
<< *SI << "\n"; } } while (false)
;
4049 continue;
4050 }
4051
4052 Value *StoreBasePtr = SI->getPointerOperand();
4053 IRB.SetInsertPoint(SI);
4054
4055 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)
;
4056
4057 for (int Idx = 0, Size = SplitLoads.size(); Idx < Size; ++Idx) {
4058 LoadInst *PLoad = SplitLoads[Idx];
4059 uint64_t PartOffset = Idx == 0 ? 0 : Offsets.Splits[Idx - 1];
4060 auto *PartPtrTy =
4061 PLoad->getType()->getPointerTo(SI->getPointerAddressSpace());
4062
4063 auto AS = SI->getPointerAddressSpace();
4064 StoreInst *PStore = IRB.CreateAlignedStore(
4065 PLoad,
4066 getAdjustedPtr(IRB, DL, StoreBasePtr,
4067 APInt(DL.getIndexSizeInBits(AS), PartOffset),
4068 PartPtrTy, StoreBasePtr->getName() + "."),
4069 getAdjustedAlignment(SI, PartOffset),
4070 /*IsVolatile*/ false);
4071 PStore->copyMetadata(*LI, {LLVMContext::MD_mem_parallel_loop_access,
4072 LLVMContext::MD_access_group});
4073 LLVM_DEBUG(dbgs() << " +" << PartOffset << ":" << *PStore << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " +" << PartOffset <<
":" << *PStore << "\n"; } } while (false)
;
4074 }
4075
4076 // We want to immediately iterate on any allocas impacted by splitting
4077 // this store, and we have to track any promotable alloca (indicated by
4078 // a direct store) as needing to be resplit because it is no longer
4079 // promotable.
4080 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(StoreBasePtr)) {
4081 ResplitPromotableAllocas.insert(OtherAI);
4082 Worklist.insert(OtherAI);
4083 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
4084 StoreBasePtr->stripInBoundsOffsets())) {
4085 Worklist.insert(OtherAI);
4086 }
4087
4088 // Mark the original store as dead.
4089 DeadInsts.insert(SI);
4090 }
4091
4092 // Save the split loads if there are deferred stores among the users.
4093 if (DeferredStores)
4094 SplitLoadsMap.insert(std::make_pair(LI, std::move(SplitLoads)));
4095
4096 // Mark the original load as dead and kill the original slice.
4097 DeadInsts.insert(LI);
4098 Offsets.S->kill();
4099 }
4100
4101 // Second, we rewrite all of the split stores. At this point, we know that
4102 // all loads from this alloca have been split already. For stores of such
4103 // loads, we can simply look up the pre-existing split loads. For stores of
4104 // other loads, we split those loads first and then write split stores of
4105 // them.
4106 for (StoreInst *SI : Stores) {
4107 auto *LI = cast<LoadInst>(SI->getValueOperand());
4108 IntegerType *Ty = cast<IntegerType>(LI->getType());
4109 uint64_t StoreSize = Ty->getBitWidth() / 8;
4110 assert(StoreSize > 0 && "Cannot have a zero-sized integer store!")((StoreSize > 0 && "Cannot have a zero-sized integer store!"
) ? static_cast<void> (0) : __assert_fail ("StoreSize > 0 && \"Cannot have a zero-sized integer store!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4110, __PRETTY_FUNCTION__))
;
4111
4112 auto &Offsets = SplitOffsetsMap[SI];
4113 assert(StoreSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&((StoreSize == Offsets.S->endOffset() - Offsets.S->beginOffset
() && "Slice size should always match load size exactly!"
) ? static_cast<void> (0) : __assert_fail ("StoreSize == Offsets.S->endOffset() - Offsets.S->beginOffset() && \"Slice size should always match load size exactly!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4114, __PRETTY_FUNCTION__))
4114 "Slice size should always match load size exactly!")((StoreSize == Offsets.S->endOffset() - Offsets.S->beginOffset
() && "Slice size should always match load size exactly!"
) ? static_cast<void> (0) : __assert_fail ("StoreSize == Offsets.S->endOffset() - Offsets.S->beginOffset() && \"Slice size should always match load size exactly!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4114, __PRETTY_FUNCTION__))
;
4115 uint64_t BaseOffset = Offsets.S->beginOffset();
4116 assert(BaseOffset + StoreSize > BaseOffset &&((BaseOffset + StoreSize > BaseOffset && "Cannot represent alloca access size using 64-bit integers!"
) ? static_cast<void> (0) : __assert_fail ("BaseOffset + StoreSize > BaseOffset && \"Cannot represent alloca access size using 64-bit integers!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4117, __PRETTY_FUNCTION__))
4117 "Cannot represent alloca access size using 64-bit integers!")((BaseOffset + StoreSize > BaseOffset && "Cannot represent alloca access size using 64-bit integers!"
) ? static_cast<void> (0) : __assert_fail ("BaseOffset + StoreSize > BaseOffset && \"Cannot represent alloca access size using 64-bit integers!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4117, __PRETTY_FUNCTION__))
;
4118
4119 Value *LoadBasePtr = LI->getPointerOperand();
4120 Instruction *StoreBasePtr = cast<Instruction>(SI->getPointerOperand());
4121
4122 LLVM_DEBUG(dbgs() << " Splitting store: " << *SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Splitting store: " << *SI
<< "\n"; } } while (false)
;
4123
4124 // Check whether we have an already split load.
4125 auto SplitLoadsMapI = SplitLoadsMap.find(LI);
4126 std::vector<LoadInst *> *SplitLoads = nullptr;
4127 if (SplitLoadsMapI != SplitLoadsMap.end()) {
4128 SplitLoads = &SplitLoadsMapI->second;
4129 assert(SplitLoads->size() == Offsets.Splits.size() + 1 &&((SplitLoads->size() == Offsets.Splits.size() + 1 &&
"Too few split loads for the number of splits in the store!"
) ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4130, __PRETTY_FUNCTION__))
4130 "Too few split loads for the number of splits in the store!")((SplitLoads->size() == Offsets.Splits.size() + 1 &&
"Too few split loads for the number of splits in the store!"
) ? static_cast<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-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4130, __PRETTY_FUNCTION__))
;
4131 } else {
4132 LLVM_DEBUG(dbgs() << " of load: " << *LI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " of load: " << *LI
<< "\n"; } } while (false)
;
4133 }
4134
4135 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
4136 int Idx = 0, Size = Offsets.Splits.size();
4137 for (;;) {
4138 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
4139 auto *LoadPartPtrTy = PartTy->getPointerTo(LI->getPointerAddressSpace());
4140 auto *StorePartPtrTy = PartTy->getPointerTo(SI->getPointerAddressSpace());
4141
4142 // Either lookup a split load or create one.
4143 LoadInst *PLoad;
4144 if (SplitLoads) {
4145 PLoad = (*SplitLoads)[Idx];
4146 } else {
4147 IRB.SetInsertPoint(LI);
4148 auto AS = LI->getPointerAddressSpace();
4149 PLoad = IRB.CreateAlignedLoad(
4150 PartTy,
4151 getAdjustedPtr(IRB, DL, LoadBasePtr,
4152 APInt(DL.getIndexSizeInBits(AS), PartOffset),
4153 LoadPartPtrTy, LoadBasePtr->getName() + "."),
4154 getAdjustedAlignment(LI, PartOffset),
4155 /*IsVolatile*/ false, LI->getName());
4156 }
4157
4158 // And store this partition.
4159 IRB.SetInsertPoint(SI);
4160 auto AS = SI->getPointerAddressSpace();
4161 StoreInst *PStore = IRB.CreateAlignedStore(
4162 PLoad,
4163 getAdjustedPtr(IRB, DL, StoreBasePtr,
4164 APInt(DL.getIndexSizeInBits(AS), PartOffset),
4165 StorePartPtrTy, StoreBasePtr->getName() + "."),
4166 getAdjustedAlignment(SI, PartOffset),
4167 /*IsVolatile*/ false);
4168
4169 // Now build a new slice for the alloca.
4170 NewSlices.push_back(
4171 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
4172 &PStore->getOperandUse(PStore->getPointerOperandIndex()),
4173 /*IsSplittable*/ false));
4174 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)
4175 << ", " << NewSlices.back().endOffset()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " new slice [" << NewSlices
.back().beginOffset() << ", " << NewSlices.back()
.endOffset() << "): " << *PStore << "\n"; }
} while (false)
4176 << "): " << *PStore << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " new slice [" << NewSlices
.back().beginOffset() << ", " << NewSlices.back()
.endOffset() << "): " << *PStore << "\n"; }
} while (false)
;
4177 if (!SplitLoads) {
4178 LLVM_DEBUG(dbgs() << " of split load: " << *PLoad << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " of split load: " << *
PLoad << "\n"; } } while (false)
;
4179 }
4180
4181 // See if we've finished all the splits.
4182 if (Idx >= Size)
4183 break;
4184
4185 // Setup the next partition.
4186 PartOffset = Offsets.Splits[Idx];
4187 ++Idx;
4188 PartSize = (Idx < Size ? Offsets.Splits[Idx] : StoreSize) - PartOffset;
4189 }
4190
4191 // We want to immediately iterate on any allocas impacted by splitting
4192 // this load, which is only relevant if it isn't a load of this alloca and
4193 // thus we didn't already split the loads above. We also have to keep track
4194 // of any promotable allocas we split loads on as they can no longer be
4195 // promoted.
4196 if (!SplitLoads) {
4197 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(LoadBasePtr)) {
4198 assert(OtherAI != &AI && "We can't re-split our own alloca!")((OtherAI != &AI && "We can't re-split our own alloca!"
) ? static_cast<void> (0) : __assert_fail ("OtherAI != &AI && \"We can't re-split our own alloca!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4198, __PRETTY_FUNCTION__))
;
4199 ResplitPromotableAllocas.insert(OtherAI);
4200 Worklist.insert(OtherAI);
4201 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
4202 LoadBasePtr->stripInBoundsOffsets())) {
4203 assert(OtherAI != &AI && "We can't re-split our own alloca!")((OtherAI != &AI && "We can't re-split our own alloca!"
) ? static_cast<void> (0) : __assert_fail ("OtherAI != &AI && \"We can't re-split our own alloca!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4203, __PRETTY_FUNCTION__))
;
4204 Worklist.insert(OtherAI);
4205 }
4206 }
4207
4208 // Mark the original store as dead now that we've split it up and kill its
4209 // slice. Note that we leave the original load in place unless this store
4210 // was its only use. It may in turn be split up if it is an alloca load
4211 // for some other alloca, but it may be a normal load. This may introduce
4212 // redundant loads, but where those can be merged the rest of the optimizer
4213 // should handle the merging, and this uncovers SSA splits which is more
4214 // important. In practice, the original loads will almost always be fully
4215 // split and removed eventually, and the splits will be merged by any
4216 // trivial CSE, including instcombine.
4217 if (LI->hasOneUse()) {
4218 assert(*LI->user_begin() == SI && "Single use isn't this store!")((*LI->user_begin() == SI && "Single use isn't this store!"
) ? static_cast<void> (0) : __assert_fail ("*LI->user_begin() == SI && \"Single use isn't this store!\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4218, __PRETTY_FUNCTION__))
;
4219 DeadInsts.insert(LI);
4220 }
4221 DeadInsts.insert(SI);
4222 Offsets.S->kill();
4223 }
4224
4225 // Remove the killed slices that have ben pre-split.
4226 AS.erase(llvm::remove_if(AS, [](const Slice &S) { return S.isDead(); }),
4227 AS.end());
4228
4229 // Insert our new slices. This will sort and merge them into the sorted
4230 // sequence.
4231 AS.insert(NewSlices);
4232
4233 LLVM_DEBUG(dbgs() << " Pre-split slices:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Pre-split slices:\n"; } } while
(false)
;
4234#ifndef NDEBUG
4235 for (auto I = AS.begin(), E = AS.end(); I != E; ++I)
4236 LLVM_DEBUG(AS.print(dbgs(), I, " "))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { AS.print(dbgs(), I, " "); } } while (false)
;
4237#endif
4238
4239 // Finally, don't try to promote any allocas that new require re-splitting.
4240 // They have already been added to the worklist above.
4241 PromotableAllocas.erase(
4242 llvm::remove_if(
4243 PromotableAllocas,
4244 [&](AllocaInst *AI) { return ResplitPromotableAllocas.count(AI); }),
4245 PromotableAllocas.end());
4246
4247 return true;
4248}
4249
4250/// Rewrite an alloca partition's users.
4251///
4252/// This routine drives both of the rewriting goals of the SROA pass. It tries
4253/// to rewrite uses of an alloca partition to be conducive for SSA value
4254/// promotion. If the partition needs a new, more refined alloca, this will
4255/// build that new alloca, preserving as much type information as possible, and
4256/// rewrite the uses of the old alloca to point at the new one and have the
4257/// appropriate new offsets. It also evaluates how successful the rewrite was
4258/// at enabling promotion and if it was successful queues the alloca to be
4259/// promoted.
4260AllocaInst *SROA::rewritePartition(AllocaInst &AI, AllocaSlices &AS,
4261 Partition &P) {
4262 // Try to compute a friendly type for this partition of the alloca. This
4263 // won't always succeed, in which case we fall back to a legal integer type
4264 // or an i8 array of an appropriate size.
4265 Type *SliceTy = nullptr;
4266 const DataLayout &DL = AI.getModule()->getDataLayout();
4267 if (Type *CommonUseTy = findCommonType(P.begin(), P.end(), P.endOffset()))
4268 if (DL.getTypeAllocSize(CommonUseTy).getFixedSize() >= P.size())
4269 SliceTy = CommonUseTy;
4270 if (!SliceTy)
4271 if (Type *TypePartitionTy = getTypePartition(DL, AI.getAllocatedType(),
4272 P.beginOffset(), P.size()))
4273 SliceTy = TypePartitionTy;
4274 if ((!SliceTy || (SliceTy->isArrayTy() &&
4275 SliceTy->getArrayElementType()->isIntegerTy())) &&
4276 DL.isLegalInteger(P.size() * 8))
4277 SliceTy = Type::getIntNTy(*C, P.size() * 8);
4278 if (!SliceTy)
4279 SliceTy = ArrayType::get(Type::getInt8Ty(*C), P.size());
4280 assert(DL.getTypeAllocSize(SliceTy).getFixedSize() >= P.size())((DL.getTypeAllocSize(SliceTy).getFixedSize() >= P.size())
? static_cast<void> (0) : __assert_fail ("DL.getTypeAllocSize(SliceTy).getFixedSize() >= P.size()"
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4280, __PRETTY_FUNCTION__))
;
4281
4282 bool IsIntegerPromotable = isIntegerWideningViable(P, SliceTy, DL);
4283
4284 VectorType *VecTy =
4285 IsIntegerPromotable ? nullptr : isVectorPromotionViable(P, DL);
4286 if (VecTy)
4287 SliceTy = VecTy;
4288
4289 // Check for the case where we're going to rewrite to a new alloca of the
4290 // exact same type as the original, and with the same access offsets. In that
4291 // case, re-use the existing alloca, but still run through the rewriter to
4292 // perform phi and select speculation.
4293 // P.beginOffset() can be non-zero even with the same type in a case with
4294 // out-of-bounds access (e.g. @PR35657 function in SROA/basictest.ll).
4295 AllocaInst *NewAI;
4296 if (SliceTy == AI.getAllocatedType() && P.beginOffset() == 0) {
4297 NewAI = &AI;
4298 // FIXME: We should be able to bail at this point with "nothing changed".
4299 // FIXME: We might want to defer PHI speculation until after here.
4300 // FIXME: return nullptr;
4301 } else {
4302 // Make sure the alignment is compatible with P.beginOffset().
4303 const Align Alignment = commonAlignment(AI.getAlign(), P.beginOffset());
4304 // If we will get at least this much alignment from the type alone, leave
4305 // the alloca's alignment unconstrained.
4306 const bool IsUnconstrained = Alignment <= DL.getABITypeAlign(SliceTy);
4307 NewAI = new AllocaInst(
4308 SliceTy, AI.getType()->getAddressSpace(), nullptr,
4309 IsUnconstrained ? DL.getPrefTypeAlign(SliceTy) : Alignment,
4310 AI.getName() + ".sroa." + Twine(P.begin() - AS.begin()), &AI);
4311 // Copy the old AI debug location over to the new one.
4312 NewAI->setDebugLoc(AI.getDebugLoc());
4313 ++NumNewAllocas;
4314 }
4315
4316 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)
4317 << "[" << P.beginOffset() << "," << P.endOffset()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "Rewriting alloca partition " <<
"[" << P.beginOffset() << "," << P.endOffset
() << ") to: " << *NewAI << "\n"; } } while
(false)
4318 << ") to: " << *NewAI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "Rewriting alloca partition " <<
"[" << P.beginOffset() << "," << P.endOffset
() << ") to: " << *NewAI << "\n"; } } while
(false)
;
4319
4320 // Track the high watermark on the worklist as it is only relevant for
4321 // promoted allocas. We will reset it to this point if the alloca is not in
4322 // fact scheduled for promotion.
4323 unsigned PPWOldSize = PostPromotionWorklist.size();
4324 unsigned NumUses = 0;
4325 SmallSetVector<PHINode *, 8> PHIUsers;
4326 SmallSetVector<SelectInst *, 8> SelectUsers;
4327
4328 AllocaSliceRewriter Rewriter(DL, AS, *this, AI, *NewAI, P.beginOffset(),
4329 P.endOffset(), IsIntegerPromotable, VecTy,
4330 PHIUsers, SelectUsers);
4331 bool Promotable = true;
4332 for (Slice *S : P.splitSliceTails()) {
4333 Promotable &= Rewriter.visit(S);
4334 ++NumUses;
4335 }
4336 for (Slice &S : P) {
4337 Promotable &= Rewriter.visit(&S);
4338 ++NumUses;
4339 }
4340
4341 NumAllocaPartitionUses += NumUses;
4342 MaxUsesPerAllocaPartition.updateMax(NumUses);
4343
4344 // Now that we've processed all the slices in the new partition, check if any
4345 // PHIs or Selects would block promotion.
4346 for (PHINode *PHI : PHIUsers)
4347 if (!isSafePHIToSpeculate(*PHI)) {
4348 Promotable = false;
4349 PHIUsers.clear();
4350 SelectUsers.clear();
4351 break;
4352 }
4353
4354 for (SelectInst *Sel : SelectUsers)
4355 if (!isSafeSelectToSpeculate(*Sel)) {
4356 Promotable = false;
4357 PHIUsers.clear();
4358 SelectUsers.clear();
4359 break;
4360 }
4361
4362 if (Promotable) {
4363 for (Use *U : AS.getDeadUsesIfPromotable()) {
4364 auto *OldInst = dyn_cast<Instruction>(U->get());
4365 Value::dropDroppableUse(*U);
4366 if (OldInst)
4367 if (isInstructionTriviallyDead(OldInst))
4368 DeadInsts.insert(OldInst);
4369 }
4370 if (PHIUsers.empty() && SelectUsers.empty()) {
4371 // Promote the alloca.
4372 PromotableAllocas.push_back(NewAI);
4373 } else {
4374 // If we have either PHIs or Selects to speculate, add them to those
4375 // worklists and re-queue the new alloca so that we promote in on the
4376 // next iteration.
4377 for (PHINode *PHIUser : PHIUsers)
4378 SpeculatablePHIs.insert(PHIUser);
4379 for (SelectInst *SelectUser : SelectUsers)
4380 SpeculatableSelects.insert(SelectUser);
4381 Worklist.insert(NewAI);
4382 }
4383 } else {
4384 // Drop any post-promotion work items if promotion didn't happen.
4385 while (PostPromotionWorklist.size() > PPWOldSize)
4386 PostPromotionWorklist.pop_back();
4387
4388 // We couldn't promote and we didn't create a new partition, nothing
4389 // happened.
4390 if (NewAI == &AI)
4391 return nullptr;
4392
4393 // If we can't promote the alloca, iterate on it to check for new
4394 // refinements exposed by splitting the current alloca. Don't iterate on an
4395 // alloca which didn't actually change and didn't get promoted.
4396 Worklist.insert(NewAI);
4397 }
4398
4399 return NewAI;
4400}
4401
4402/// Walks the slices of an alloca and form partitions based on them,
4403/// rewriting each of their uses.
4404bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &AS) {
4405 if (AS.begin() == AS.end())
4406 return false;
4407
4408 unsigned NumPartitions = 0;
4409 bool Changed = false;
4410 const DataLayout &DL = AI.getModule()->getDataLayout();
4411
4412 // First try to pre-split loads and stores.
4413 Changed |= presplitLoadsAndStores(AI, AS);
4414
4415 // Now that we have identified any pre-splitting opportunities,
4416 // mark loads and stores unsplittable except for the following case.
4417 // We leave a slice splittable if all other slices are disjoint or fully
4418 // included in the slice, such as whole-alloca loads and stores.
4419 // If we fail to split these during pre-splitting, we want to force them
4420 // to be rewritten into a partition.
4421 bool IsSorted = true;
4422
4423 uint64_t AllocaSize =
4424 DL.getTypeAllocSize(AI.getAllocatedType()).getFixedSize();
4425 const uint64_t MaxBitVectorSize = 1024;
4426 if (AllocaSize <= MaxBitVectorSize) {
4427 // If a byte boundary is included in any load or store, a slice starting or
4428 // ending at the boundary is not splittable.
4429 SmallBitVector SplittableOffset(AllocaSize + 1, true);
4430 for (Slice &S : AS)
4431 for (unsigned O = S.beginOffset() + 1;
4432 O < S.endOffset() && O < AllocaSize; O++)
4433 SplittableOffset.reset(O);
4434
4435 for (Slice &S : AS) {
4436 if (!S.isSplittable())
4437 continue;
4438
4439 if ((S.beginOffset() > AllocaSize || SplittableOffset[S.beginOffset()]) &&
4440 (S.endOffset() > AllocaSize || SplittableOffset[S.endOffset()]))
4441 continue;
4442
4443 if (isa<LoadInst>(S.getUse()->getUser()) ||
4444 isa<StoreInst>(S.getUse()->getUser())) {
4445 S.makeUnsplittable();
4446 IsSorted = false;
4447 }
4448 }
4449 }
4450 else {
4451 // We only allow whole-alloca splittable loads and stores
4452 // for a large alloca to avoid creating too large BitVector.
4453 for (Slice &S : AS) {
4454 if (!S.isSplittable())
4455 continue;
4456
4457 if (S.beginOffset() == 0 && S.endOffset() >= AllocaSize)
4458 continue;
4459
4460 if (isa<LoadInst>(S.getUse()->getUser()) ||
4461 isa<StoreInst>(S.getUse()->getUser())) {
4462 S.makeUnsplittable();
4463 IsSorted = false;
4464 }
4465 }
4466 }
4467
4468 if (!IsSorted)
4469 llvm::sort(AS);
4470
4471 /// Describes the allocas introduced by rewritePartition in order to migrate
4472 /// the debug info.
4473 struct Fragment {
4474 AllocaInst *Alloca;
4475 uint64_t Offset;
4476 uint64_t Size;
4477 Fragment(AllocaInst *AI, uint64_t O, uint64_t S)
4478 : Alloca(AI), Offset(O), Size(S) {}
4479 };
4480 SmallVector<Fragment, 4> Fragments;
4481
4482 // Rewrite each partition.
4483 for (auto &P : AS.partitions()) {
4484 if (AllocaInst *NewAI = rewritePartition(AI, AS, P)) {
4485 Changed = true;
4486 if (NewAI != &AI) {
4487 uint64_t SizeOfByte = 8;
4488 uint64_t AllocaSize =
4489 DL.getTypeSizeInBits(NewAI->getAllocatedType()).getFixedSize();
4490 // Don't include any padding.
4491 uint64_t Size = std::min(AllocaSize, P.size() * SizeOfByte);
4492 Fragments.push_back(Fragment(NewAI, P.beginOffset() * SizeOfByte, Size));
4493 }
4494 }
4495 ++NumPartitions;
4496 }
4497
4498 NumAllocaPartitions += NumPartitions;
4499 MaxPartitionsPerAlloca.updateMax(NumPartitions);
4500
4501 // Migrate debug information from the old alloca to the new alloca(s)
4502 // and the individual partitions.
4503 TinyPtrVector<DbgVariableIntrinsic *> DbgDeclares = FindDbgAddrUses(&AI);
4504 for (DbgVariableIntrinsic *DbgDeclare : DbgDeclares) {
4505 auto *Expr = DbgDeclare->getExpression();
4506 DIBuilder DIB(*AI.getModule(), /*AllowUnresolved*/ false);
4507 uint64_t AllocaSize =
4508 DL.getTypeSizeInBits(AI.getAllocatedType()).getFixedSize();
4509 for (auto Fragment : Fragments) {
4510 // Create a fragment expression describing the new partition or reuse AI's
4511 // expression if there is only one partition.
4512 auto *FragmentExpr = Expr;
4513 if (Fragment.Size < AllocaSize || Expr->isFragment()) {
4514 // If this alloca is already a scalar replacement of a larger aggregate,
4515 // Fragment.Offset describes the offset inside the scalar.
4516 auto ExprFragment = Expr->getFragmentInfo();
4517 uint64_t Offset = ExprFragment ? ExprFragment->OffsetInBits : 0;
4518 uint64_t Start = Offset + Fragment.Offset;
4519 uint64_t Size = Fragment.Size;
4520 if (ExprFragment) {
4521 uint64_t AbsEnd =
4522 ExprFragment->OffsetInBits + ExprFragment->SizeInBits;
4523 if (Start >= AbsEnd)
4524 // No need to describe a SROAed padding.
4525 continue;
4526 Size = std::min(Size, AbsEnd - Start);
4527 }
4528 // The new, smaller fragment is stenciled out from the old fragment.
4529 if (auto OrigFragment = FragmentExpr->getFragmentInfo()) {
4530 assert(Start >= OrigFragment->OffsetInBits &&((Start >= OrigFragment->OffsetInBits && "new fragment is outside of original fragment"
) ? static_cast<void> (0) : __assert_fail ("Start >= OrigFragment->OffsetInBits && \"new fragment is outside of original fragment\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4531, __PRETTY_FUNCTION__))
4531 "new fragment is outside of original fragment")((Start >= OrigFragment->OffsetInBits && "new fragment is outside of original fragment"
) ? static_cast<void> (0) : __assert_fail ("Start >= OrigFragment->OffsetInBits && \"new fragment is outside of original fragment\""
, "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/SROA.cpp"
, 4531, __PRETTY_FUNCTION__))
;
4532 Start -= OrigFragment->OffsetInBits;
4533 }
4534
4535 // The alloca may be larger than the variable.
4536 auto VarSize = DbgDeclare->getVariable()->getSizeInBits();
4537 if (VarSize) {
4538 if (Size > *VarSize)
4539 Size = *VarSize;
4540 if (Size == 0 || Start + Size > *VarSize)
4541 continue;
4542 }
4543
4544 // Avoid creating a fragment expression that covers the entire variable.
4545 if (!VarSize || *VarSize != Size) {
4546 if (auto E =
4547 DIExpression::createFragmentExpression(Expr, Start, Size))
4548 FragmentExpr = *E;
4549 else
4550 continue;
4551 }
4552 }
4553
4554 // Remove any existing intrinsics on the new alloca describing
4555 // the variable fragment.
4556 for (DbgVariableIntrinsic *OldDII : FindDbgAddrUses(Fragment.Alloca)) {
4557 auto SameVariableFragment = [](const DbgVariableIntrinsic *LHS,
4558 const DbgVariableIntrinsic *RHS) {
4559 return LHS->getVariable() == RHS->getVariable() &&
4560 LHS->getDebugLoc()->getInlinedAt() ==
4561 RHS->getDebugLoc()->getInlinedAt();
4562 };
4563 if (SameVariableFragment(OldDII, DbgDeclare))
4564 OldDII->eraseFromParent();
4565 }
4566
4567 DIB.insertDeclare(Fragment.Alloca, DbgDeclare->getVariable(), FragmentExpr,
4568 DbgDeclare->getDebugLoc(), &AI);
4569 }
4570 }
4571 return Changed;
4572}
4573
4574/// Clobber a use with undef, deleting the used value if it becomes dead.
4575void SROA::clobberUse(Use &U) {
4576 Value *OldV = U;
4577 // Replace the use with an undef value.
4578 U = UndefValue::get(OldV->getType());
4579
4580 // Check for this making an instruction dead. We have to garbage collect
4581 // all the dead instructions to ensure the uses of any alloca end up being
4582 // minimal.
4583 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
4584 if (isInstructionTriviallyDead(OldI)) {
4585 DeadInsts.insert(OldI);
4586 }
4587}
4588
4589/// Analyze an alloca for SROA.
4590///
4591/// This analyzes the alloca to ensure we can reason about it, builds
4592/// the slices of the alloca, and then hands it off to be split and
4593/// rewritten as needed.
4594bool SROA::runOnAlloca(AllocaInst &AI) {
4595 LLVM_DEBUG(dbgs() << "SROA alloca: " << AI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "SROA alloca: " << AI <<
"\n"; } } while (false)
;
4596 ++NumAllocasAnalyzed;
4597
4598 // Special case dead allocas, as they're trivial.
4599 if (AI.use_empty()) {
4600 AI.eraseFromParent();
4601 return true;
4602 }
4603 const DataLayout &DL = AI.getModule()->getDataLayout();
4604
4605 // Skip alloca forms that this analysis can't handle.
4606 auto *AT = AI.getAllocatedType();
4607 if (AI.isArrayAllocation() || !AT->isSized() || isa<ScalableVectorType>(AT) ||
4608 DL.getTypeAllocSize(AT).getFixedSize() == 0)
4609 return false;
4610
4611 bool Changed = false;
4612
4613 // First, split any FCA loads and stores touching this alloca to promote
4614 // better splitting and promotion opportunities.
4615 AggLoadStoreRewriter AggRewriter(DL);
4616 Changed |= AggRewriter.rewrite(AI);
4617
4618 // Build the slices using a recursive instruction-visiting builder.
4619 AllocaSlices AS(DL, AI);
4620 LLVM_DEBUG(AS.print(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { AS.print(dbgs()); } } while (false)
;
4621 if (AS.isEscaped())
4622 return Changed;
4623
4624 // Delete all the dead users of this alloca before splitting and rewriting it.
4625 for (Instruction *DeadUser : AS.getDeadUsers()) {
4626 // Free up everything used by this instruction.
4627 for (Use &DeadOp : DeadUser->operands())
4628 clobberUse(DeadOp);
4629
4630 // Now replace the uses of this instruction.
4631 DeadUser->replaceAllUsesWith(UndefValue::get(DeadUser->getType()));
4632
4633 // And mark it for deletion.
4634 DeadInsts.insert(DeadUser);
4635 Changed = true;
4636 }
4637 for (Use *DeadOp : AS.getDeadOperands()) {
4638 clobberUse(*DeadOp);
4639 Changed = true;
4640 }
4641
4642 // No slices to split. Leave the dead alloca for a later pass to clean up.
4643 if (AS.begin() == AS.end())
4644 return Changed;
4645
4646 Changed |= splitAlloca(AI, AS);
4647
4648 LLVM_DEBUG(dbgs() << " Speculating PHIs\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Speculating PHIs\n"; } } while
(false)
;
4649 while (!SpeculatablePHIs.empty())
4650 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
4651
4652 LLVM_DEBUG(dbgs() << " Speculating Selects\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Speculating Selects\n"; } } while
(false)
;
4653 while (!SpeculatableSelects.empty())
4654 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
4655
4656 return Changed;
4657}
4658
4659/// Delete the dead instructions accumulated in this run.
4660///
4661/// Recursively deletes the dead instructions we've accumulated. This is done
4662/// at the very end to maximize locality of the recursive delete and to
4663/// minimize the problems of invalidated instruction pointers as such pointers
4664/// are used heavily in the intermediate stages of the algorithm.
4665///
4666/// We also record the alloca instructions deleted here so that they aren't
4667/// subsequently handed to mem2reg to promote.
4668bool SROA::deleteDeadInstructions(
4669 SmallPtrSetImpl<AllocaInst *> &DeletedAllocas) {
4670 bool Changed = false;
4671 while (!DeadInsts.empty()) {
4672 Instruction *I = DeadInsts.pop_back_val();
4673 LLVM_DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "Deleting dead instruction: " <<
*I << "\n"; } } while (false)
;
4674
4675 // If the instruction is an alloca, find the possible dbg.declare connected
4676 // to it, and remove it too. We must do this before calling RAUW or we will
4677 // not be able to find it.
4678 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
4679 DeletedAllocas.insert(AI);
4680 for (DbgVariableIntrinsic *OldDII : FindDbgAddrUses(AI))
4681 OldDII->eraseFromParent();
4682 }
4683
4684 I->replaceAllUsesWith(UndefValue::get(I->getType()));
4685
4686 for (Use &Operand : I->operands())
4687 if (Instruction *U = dyn_cast<Instruction>(Operand)) {
4688 // Zero out the operand and see if it becomes trivially dead.
4689 Operand = nullptr;
4690 if (isInstructionTriviallyDead(U))
4691 DeadInsts.insert(U);
4692 }
4693
4694 ++NumDeleted;
4695 I->eraseFromParent();
4696 Changed = true;
4697 }
4698 return Changed;
4699}
4700
4701/// Promote the allocas, using the best available technique.
4702///
4703/// This attempts to promote whatever allocas have been identified as viable in
4704/// the PromotableAllocas list. If that list is empty, there is nothing to do.
4705/// This function returns whether any promotion occurred.
4706bool SROA::promoteAllocas(Function &F) {
4707 if (PromotableAllocas.empty())
4708 return false;
4709
4710 NumPromoted += PromotableAllocas.size();
4711
4712 LLVM_DEBUG(dbgs() << "Promoting allocas with mem2reg...\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "Promoting allocas with mem2reg...\n"
; } } while (false)
;
4713 PromoteMemToReg(PromotableAllocas, *DT, AC);
4714 PromotableAllocas.clear();
4715 return true;
4716}
4717
4718PreservedAnalyses SROA::runImpl(Function &F, DominatorTree &RunDT,
4719 AssumptionCache &RunAC) {
4720 LLVM_DEBUG(dbgs() << "SROA function: " << F.getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "SROA function: " << F.getName
() << "\n"; } } while (false)
;
4721 C = &F.getContext();
4722 DT = &RunDT;
4723 AC = &RunAC;
4724
4725 BasicBlock &EntryBB = F.getEntryBlock();
4726 for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
4727 I != E; ++I) {
4728 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
4729 if (isa<ScalableVectorType>(AI->getAllocatedType())) {
4730 if (isAllocaPromotable(AI))
4731 PromotableAllocas.push_back(AI);
4732 } else {
4733 Worklist.insert(AI);
4734 }
4735 }
4736 }
4737
4738 bool Changed = false;
4739 // A set of deleted alloca instruction pointers which should be removed from
4740 // the list of promotable allocas.
4741 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
4742
4743 do {
4744 while (!Worklist.empty()) {
4745 Changed |= runOnAlloca(*Worklist.pop_back_val());
4746 Changed |= deleteDeadInstructions(DeletedAllocas);
4747
4748 // Remove the deleted allocas from various lists so that we don't try to
4749 // continue processing them.
4750 if (!DeletedAllocas.empty()) {
4751 auto IsInSet = [&](AllocaInst *AI) { return DeletedAllocas.count(AI); };
4752 Worklist.remove_if(IsInSet);
4753 PostPromotionWorklist.remove_if(IsInSet);
4754 PromotableAllocas.erase(llvm::remove_if(PromotableAllocas, IsInSet),
4755 PromotableAllocas.end());
4756 DeletedAllocas.clear();
4757 }
4758 }
4759
4760 Changed |= promoteAllocas(F);
4761
4762 Worklist = PostPromotionWorklist;
4763 PostPromotionWorklist.clear();
4764 } while (!Worklist.empty());
4765
4766 if (!Changed)
4767 return PreservedAnalyses::all();
4768
4769 PreservedAnalyses PA;
4770 PA.preserveSet<CFGAnalyses>();
4771 PA.preserve<GlobalsAA>();
4772 return PA;
4773}
4774
4775PreservedAnalyses SROA::run(Function &F, FunctionAnalysisManager &AM) {
4776 return runImpl(F, AM.getResult<DominatorTreeAnalysis>(F),
4777 AM.getResult<AssumptionAnalysis>(F));
4778}
4779
4780/// A legacy pass for the legacy pass manager that wraps the \c SROA pass.
4781///
4782/// This is in the llvm namespace purely to allow it to be a friend of the \c
4783/// SROA pass.
4784class llvm::sroa::SROALegacyPass : public FunctionPass {
4785 /// The SROA implementation.
4786 SROA Impl;
4787
4788public:
4789 static char ID;
4790
4791 SROALegacyPass() : FunctionPass(ID) {
4792 initializeSROALegacyPassPass(*PassRegistry::getPassRegistry());
4793 }
4794
4795 bool runOnFunction(Function &F) override {
4796 if (skipFunction(F))
4797 return false;
4798
4799 auto PA = Impl.runImpl(
4800 F, getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
4801 getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F));
4802 return !PA.areAllPreserved();
4803 }
4804
4805 void getAnalysisUsage(AnalysisUsage &AU) const override {
4806 AU.addRequired<AssumptionCacheTracker>();
4807 AU.addRequired<DominatorTreeWrapperPass>();
4808 AU.addPreserved<GlobalsAAWrapperPass>();
4809 AU.setPreservesCFG();
4810 }
4811
4812 StringRef getPassName() const override { return "SROA"; }
4813};
4814
4815char SROALegacyPass::ID = 0;
4816
4817FunctionPass *llvm::createSROAPass() { return new SROALegacyPass(); }
4818
4819INITIALIZE_PASS_BEGIN(SROALegacyPass, "sroa",static void *initializeSROALegacyPassPassOnce(PassRegistry &
Registry) {
4820 "Scalar Replacement Of Aggregates", false, false)static void *initializeSROALegacyPassPassOnce(PassRegistry &
Registry) {
4821INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry);
4822INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);
4823INITIALIZE_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)); }
4824 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)); }