Bug Summary

File:lib/Transforms/Scalar/SROA.cpp
Warning:line 2093, column 3
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -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 -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mthread-model posix -fmath-errno -masm-verbose -mconstructor-aliases -munwind-tables -fuse-init-array -target-cpu x86-64 -dwarf-column-info -debugger-tuning=gdb -momit-leaf-frame-pointer -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-9/lib/clang/9.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-9~svn362543/build-llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-9~svn362543/build-llvm/include -I /build/llvm-toolchain-snapshot-9~svn362543/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/include/clang/9.0.0/include/ -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-9/lib/clang/9.0.0/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++11 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-9~svn362543/build-llvm/lib/Transforms/Scalar -fdebug-prefix-map=/build/llvm-toolchain-snapshot-9~svn362543=. -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -stack-protector 2 -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -o /tmp/scan-build-2019-06-05-060531-1271-1 -x c++ /build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp -faddrsig
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/Transforms/Utils/Local.h"
45#include "llvm/Config/llvm-config.h"
46#include "llvm/IR/BasicBlock.h"
47#include "llvm/IR/Constant.h"
48#include "llvm/IR/ConstantFolder.h"
49#include "llvm/IR/Constants.h"
50#include "llvm/IR/DIBuilder.h"
51#include "llvm/IR/DataLayout.h"
52#include "llvm/IR/DebugInfoMetadata.h"
53#include "llvm/IR/DerivedTypes.h"
54#include "llvm/IR/Dominators.h"
55#include "llvm/IR/Function.h"
56#include "llvm/IR/GetElementPtrTypeIterator.h"
57#include "llvm/IR/GlobalAlias.h"
58#include "llvm/IR/IRBuilder.h"
59#include "llvm/IR/InstVisitor.h"
60#include "llvm/IR/InstrTypes.h"
61#include "llvm/IR/Instruction.h"
62#include "llvm/IR/Instructions.h"
63#include "llvm/IR/IntrinsicInst.h"
64#include "llvm/IR/Intrinsics.h"
65#include "llvm/IR/LLVMContext.h"
66#include "llvm/IR/Metadata.h"
67#include "llvm/IR/Module.h"
68#include "llvm/IR/Operator.h"
69#include "llvm/IR/PassManager.h"
70#include "llvm/IR/Type.h"
71#include "llvm/IR/Use.h"
72#include "llvm/IR/User.h"
73#include "llvm/IR/Value.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/PromoteMemToReg.h"
84#include <algorithm>
85#include <cassert>
86#include <chrono>
87#include <cstddef>
88#include <cstdint>
89#include <cstring>
90#include <iterator>
91#include <string>
92#include <tuple>
93#include <utility>
94#include <vector>
95
96#ifndef NDEBUG
97// We only use this for a debug check.
98#include <random>
99#endif
100
101using namespace llvm;
102using namespace llvm::sroa;
103
104#define DEBUG_TYPE"sroa" "sroa"
105
106STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement")static llvm::Statistic NumAllocasAnalyzed = {"sroa", "NumAllocasAnalyzed"
, "Number of allocas analyzed for replacement", {0}, {false}}
;
107STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed")static llvm::Statistic NumAllocaPartitions = {"sroa", "NumAllocaPartitions"
, "Number of alloca partitions formed", {0}, {false}}
;
108STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions per alloca")static llvm::Statistic MaxPartitionsPerAlloca = {"sroa", "MaxPartitionsPerAlloca"
, "Maximum number of partitions per alloca", {0}, {false}}
;
109STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten")static llvm::Statistic NumAllocaPartitionUses = {"sroa", "NumAllocaPartitionUses"
, "Number of alloca partition uses rewritten", {0}, {false}}
;
110STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition")static llvm::Statistic MaxUsesPerAllocaPartition = {"sroa", "MaxUsesPerAllocaPartition"
, "Maximum number of uses of a partition", {0}, {false}}
;
111STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced")static llvm::Statistic NumNewAllocas = {"sroa", "NumNewAllocas"
, "Number of new, smaller allocas introduced", {0}, {false}}
;
112STATISTIC(NumPromoted, "Number of allocas promoted to SSA values")static llvm::Statistic NumPromoted = {"sroa", "NumPromoted", "Number of allocas promoted to SSA values"
, {0}, {false}}
;
113STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion")static llvm::Statistic NumLoadsSpeculated = {"sroa", "NumLoadsSpeculated"
, "Number of loads speculated to allow promotion", {0}, {false
}}
;
114STATISTIC(NumDeleted, "Number of instructions deleted")static llvm::Statistic NumDeleted = {"sroa", "NumDeleted", "Number of instructions deleted"
, {0}, {false}}
;
115STATISTIC(NumVectorized, "Number of vectorized aggregates")static llvm::Statistic NumVectorized = {"sroa", "NumVectorized"
, "Number of vectorized aggregates", {0}, {false}}
;
116
117/// Hidden option to enable randomly shuffling the slices to help uncover
118/// instability in their order.
119static cl::opt<bool> SROARandomShuffleSlices("sroa-random-shuffle-slices",
120 cl::init(false), cl::Hidden);
121
122/// Hidden option to experiment with completely strict handling of inbounds
123/// GEPs.
124static cl::opt<bool> SROAStrictInbounds("sroa-strict-inbounds", cl::init(false),
125 cl::Hidden);
126
127namespace {
128
129/// A custom IRBuilder inserter which prefixes all names, but only in
130/// Assert builds.
131class IRBuilderPrefixedInserter : public IRBuilderDefaultInserter {
132 std::string Prefix;
133
134 const Twine getNameWithPrefix(const Twine &Name) const {
135 return Name.isTriviallyEmpty() ? Name : Prefix + Name;
136 }
137
138public:
139 void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
140
141protected:
142 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
143 BasicBlock::iterator InsertPt) const {
144 IRBuilderDefaultInserter::InsertHelper(I, getNameWithPrefix(Name), BB,
145 InsertPt);
146 }
147};
148
149/// Provide a type for IRBuilder that drops names in release builds.
150using IRBuilderTy = IRBuilder<ConstantFolder, IRBuilderPrefixedInserter>;
151
152/// A used slice of an alloca.
153///
154/// This structure represents a slice of an alloca used by some instruction. It
155/// stores both the begin and end offsets of this use, a pointer to the use
156/// itself, and a flag indicating whether we can classify the use as splittable
157/// or not when forming partitions of the alloca.
158class Slice {
159 /// The beginning offset of the range.
160 uint64_t BeginOffset = 0;
161
162 /// The ending offset, not included in the range.
163 uint64_t EndOffset = 0;
164
165 /// Storage for both the use of this slice and whether it can be
166 /// split.
167 PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
168
169public:
170 Slice() = default;
171
172 Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
173 : BeginOffset(BeginOffset), EndOffset(EndOffset),
174 UseAndIsSplittable(U, IsSplittable) {}
175
176 uint64_t beginOffset() const { return BeginOffset; }
177 uint64_t endOffset() const { return EndOffset; }
178
179 bool isSplittable() const { return UseAndIsSplittable.getInt(); }
180 void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
181
182 Use *getUse() const { return UseAndIsSplittable.getPointer(); }
183
184 bool isDead() const { return getUse() == nullptr; }
185 void kill() { UseAndIsSplittable.setPointer(nullptr); }
186
187 /// Support for ordering ranges.
188 ///
189 /// This provides an ordering over ranges such that start offsets are
190 /// always increasing, and within equal start offsets, the end offsets are
191 /// decreasing. Thus the spanning range comes first in a cluster with the
192 /// same start position.
193 bool operator<(const Slice &RHS) const {
194 if (beginOffset() < RHS.beginOffset())
195 return true;
196 if (beginOffset() > RHS.beginOffset())
197 return false;
198 if (isSplittable() != RHS.isSplittable())
199 return !isSplittable();
200 if (endOffset() > RHS.endOffset())
201 return true;
202 return false;
203 }
204
205 /// Support comparison with a single offset to allow binary searches.
206 friend LLVM_ATTRIBUTE_UNUSED__attribute__((__unused__)) bool operator<(const Slice &LHS,
207 uint64_t RHSOffset) {
208 return LHS.beginOffset() < RHSOffset;
209 }
210 friend LLVM_ATTRIBUTE_UNUSED__attribute__((__unused__)) bool operator<(uint64_t LHSOffset,
211 const Slice &RHS) {
212 return LHSOffset < RHS.beginOffset();
213 }
214
215 bool operator==(const Slice &RHS) const {
216 return isSplittable() == RHS.isSplittable() &&
217 beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
218 }
219 bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
220};
221
222} // end anonymous namespace
223
224/// Representation of the alloca slices.
225///
226/// This class represents the slices of an alloca which are formed by its
227/// various uses. If a pointer escapes, we can't fully build a representation
228/// for the slices used and we reflect that in this structure. The uses are
229/// stored, sorted by increasing beginning offset and with unsplittable slices
230/// starting at a particular offset before splittable slices.
231class llvm::sroa::AllocaSlices {
232public:
233 /// Construct the slices of a particular alloca.
234 AllocaSlices(const DataLayout &DL, AllocaInst &AI);
235
236 /// Test whether a pointer to the allocation escapes our analysis.
237 ///
238 /// If this is true, the slices are never fully built and should be
239 /// ignored.
240 bool isEscaped() const { return PointerEscapingInstr; }
241
242 /// Support for iterating over the slices.
243 /// @{
244 using iterator = SmallVectorImpl<Slice>::iterator;
245 using range = iterator_range<iterator>;
246
247 iterator begin() { return Slices.begin(); }
248 iterator end() { return Slices.end(); }
249
250 using const_iterator = SmallVectorImpl<Slice>::const_iterator;
251 using const_range = iterator_range<const_iterator>;
252
253 const_iterator begin() const { return Slices.begin(); }
254 const_iterator end() const { return Slices.end(); }
255 /// @}
256
257 /// Erase a range of slices.
258 void erase(iterator Start, iterator Stop) { Slices.erase(Start, Stop); }
259
260 /// Insert new slices for this alloca.
261 ///
262 /// This moves the slices into the alloca's slices collection, and re-sorts
263 /// everything so that the usual ordering properties of the alloca's slices
264 /// hold.
265 void insert(ArrayRef<Slice> NewSlices) {
266 int OldSize = Slices.size();
267 Slices.append(NewSlices.begin(), NewSlices.end());
268 auto SliceI = Slices.begin() + OldSize;
269 llvm::sort(SliceI, Slices.end());
270 std::inplace_merge(Slices.begin(), SliceI, Slices.end());
271 }
272
273 // Forward declare the iterator and range accessor for walking the
274 // partitions.
275 class partition_iterator;
276 iterator_range<partition_iterator> partitions();
277
278 /// Access the dead users for this alloca.
279 ArrayRef<Instruction *> getDeadUsers() const { return DeadUsers; }
280
281 /// Access the dead operands referring to this alloca.
282 ///
283 /// These are operands which have cannot actually be used to refer to the
284 /// alloca as they are outside its range and the user doesn't correct for
285 /// that. These mostly consist of PHI node inputs and the like which we just
286 /// need to replace with undef.
287 ArrayRef<Use *> getDeadOperands() const { return DeadOperands; }
288
289#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
290 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
291 void printSlice(raw_ostream &OS, const_iterator I,
292 StringRef Indent = " ") const;
293 void printUse(raw_ostream &OS, const_iterator I,
294 StringRef Indent = " ") const;
295 void print(raw_ostream &OS) const;
296 void dump(const_iterator I) const;
297 void dump() const;
298#endif
299
300private:
301 template <typename DerivedT, typename RetT = void> class BuilderBase;
302 class SliceBuilder;
303
304 friend class AllocaSlices::SliceBuilder;
305
306#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
307 /// Handle to alloca instruction to simplify method interfaces.
308 AllocaInst &AI;
309#endif
310
311 /// The instruction responsible for this alloca not having a known set
312 /// of slices.
313 ///
314 /// When an instruction (potentially) escapes the pointer to the alloca, we
315 /// store a pointer to that here and abort trying to form slices of the
316 /// alloca. This will be null if the alloca slices are analyzed successfully.
317 Instruction *PointerEscapingInstr;
318
319 /// The slices of the alloca.
320 ///
321 /// We store a vector of the slices formed by uses of the alloca here. This
322 /// vector is sorted by increasing begin offset, and then the unsplittable
323 /// slices before the splittable ones. See the Slice inner class for more
324 /// details.
325 SmallVector<Slice, 8> Slices;
326
327 /// Instructions which will become dead if we rewrite the alloca.
328 ///
329 /// Note that these are not separated by slice. This is because we expect an
330 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
331 /// all these instructions can simply be removed and replaced with undef as
332 /// they come from outside of the allocated space.
333 SmallVector<Instruction *, 8> DeadUsers;
334
335 /// Operands which will become dead if we rewrite the alloca.
336 ///
337 /// These are operands that in their particular use can be replaced with
338 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
339 /// to PHI nodes and the like. They aren't entirely dead (there might be
340 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
341 /// want to swap this particular input for undef to simplify the use lists of
342 /// the alloca.
343 SmallVector<Use *, 8> DeadOperands;
344};
345
346/// A partition of the slices.
347///
348/// An ephemeral representation for a range of slices which can be viewed as
349/// a partition of the alloca. This range represents a span of the alloca's
350/// memory which cannot be split, and provides access to all of the slices
351/// overlapping some part of the partition.
352///
353/// Objects of this type are produced by traversing the alloca's slices, but
354/// are only ephemeral and not persistent.
355class llvm::sroa::Partition {
356private:
357 friend class AllocaSlices;
358 friend class AllocaSlices::partition_iterator;
359
360 using iterator = AllocaSlices::iterator;
361
362 /// The beginning and ending offsets of the alloca for this
363 /// partition.
364 uint64_t BeginOffset, EndOffset;
365
366 /// The start and end iterators of this partition.
367 iterator SI, SJ;
368
369 /// A collection of split slice tails overlapping the partition.
370 SmallVector<Slice *, 4> SplitTails;
371
372 /// Raw constructor builds an empty partition starting and ending at
373 /// the given iterator.
374 Partition(iterator SI) : SI(SI), SJ(SI) {}
375
376public:
377 /// The start offset of this partition.
378 ///
379 /// All of the contained slices start at or after this offset.
380 uint64_t beginOffset() const { return BeginOffset; }
381
382 /// The end offset of this partition.
383 ///
384 /// All of the contained slices end at or before this offset.
385 uint64_t endOffset() const { return EndOffset; }
386
387 /// The size of the partition.
388 ///
389 /// Note that this can never be zero.
390 uint64_t size() const {
391 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 391, __PRETTY_FUNCTION__))
;
392 return EndOffset - BeginOffset;
393 }
394
395 /// Test whether this partition contains no slices, and merely spans
396 /// a region occupied by split slices.
397 bool empty() const { return SI == SJ; }
398
399 /// \name Iterate slices that start within the partition.
400 /// These may be splittable or unsplittable. They have a begin offset >= the
401 /// partition begin offset.
402 /// @{
403 // FIXME: We should probably define a "concat_iterator" helper and use that
404 // to stitch together pointee_iterators over the split tails and the
405 // contiguous iterators of the partition. That would give a much nicer
406 // interface here. We could then additionally expose filtered iterators for
407 // split, unsplit, and unsplittable splices based on the usage patterns.
408 iterator begin() const { return SI; }
409 iterator end() const { return SJ; }
410 /// @}
411
412 /// Get the sequence of split slice tails.
413 ///
414 /// These tails are of slices which start before this partition but are
415 /// split and overlap into the partition. We accumulate these while forming
416 /// partitions.
417 ArrayRef<Slice *> splitSliceTails() const { return SplitTails; }
418};
419
420/// An iterator over partitions of the alloca's slices.
421///
422/// This iterator implements the core algorithm for partitioning the alloca's
423/// slices. It is a forward iterator as we don't support backtracking for
424/// efficiency reasons, and re-use a single storage area to maintain the
425/// current set of split slices.
426///
427/// It is templated on the slice iterator type to use so that it can operate
428/// with either const or non-const slice iterators.
429class AllocaSlices::partition_iterator
430 : public iterator_facade_base<partition_iterator, std::forward_iterator_tag,
431 Partition> {
432 friend class AllocaSlices;
433
434 /// Most of the state for walking the partitions is held in a class
435 /// with a nice interface for examining them.
436 Partition P;
437
438 /// We need to keep the end of the slices to know when to stop.
439 AllocaSlices::iterator SE;
440
441 /// We also need to keep track of the maximum split end offset seen.
442 /// FIXME: Do we really?
443 uint64_t MaxSplitSliceEndOffset = 0;
444
445 /// Sets the partition to be empty at given iterator, and sets the
446 /// end iterator.
447 partition_iterator(AllocaSlices::iterator SI, AllocaSlices::iterator SE)
448 : P(SI), SE(SE) {
449 // If not already at the end, advance our state to form the initial
450 // partition.
451 if (SI != SE)
452 advance();
453 }
454
455 /// Advance the iterator to the next partition.
456 ///
457 /// Requires that the iterator not be at the end of the slices.
458 void advance() {
459 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 460, __PRETTY_FUNCTION__))
460 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 460, __PRETTY_FUNCTION__))
;
461
462 // Clear out any split uses which have ended.
463 if (!P.SplitTails.empty()) {
464 if (P.EndOffset >= MaxSplitSliceEndOffset) {
465 // If we've finished all splits, this is easy.
466 P.SplitTails.clear();
467 MaxSplitSliceEndOffset = 0;
468 } else {
469 // Remove the uses which have ended in the prior partition. This
470 // cannot change the max split slice end because we just checked that
471 // the prior partition ended prior to that max.
472 P.SplitTails.erase(llvm::remove_if(P.SplitTails,
473 [&](Slice *S) {
474 return S->endOffset() <=
475 P.EndOffset;
476 }),
477 P.SplitTails.end());
478 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 482, __PRETTY_FUNCTION__))
479 [&](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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 482, __PRETTY_FUNCTION__))
480 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 482, __PRETTY_FUNCTION__))
481 }) &&((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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 482, __PRETTY_FUNCTION__))
482 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 482, __PRETTY_FUNCTION__))
;
483 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 487, __PRETTY_FUNCTION__))
484 [&](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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 487, __PRETTY_FUNCTION__))
485 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 487, __PRETTY_FUNCTION__))
486 }) &&((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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 487, __PRETTY_FUNCTION__))
487 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 487, __PRETTY_FUNCTION__))
;
488 }
489 }
490
491 // If P.SI is already at the end, then we've cleared the split tail and
492 // now have an end iterator.
493 if (P.SI == SE) {
494 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 494, __PRETTY_FUNCTION__))
;
495 return;
496 }
497
498 // If we had a non-empty partition previously, set up the state for
499 // subsequent partitions.
500 if (P.SI != P.SJ) {
501 // Accumulate all the splittable slices which started in the old
502 // partition into the split list.
503 for (Slice &S : P)
504 if (S.isSplittable() && S.endOffset() > P.EndOffset) {
505 P.SplitTails.push_back(&S);
506 MaxSplitSliceEndOffset =
507 std::max(S.endOffset(), MaxSplitSliceEndOffset);
508 }
509
510 // Start from the end of the previous partition.
511 P.SI = P.SJ;
512
513 // If P.SI is now at the end, we at most have a tail of split slices.
514 if (P.SI == SE) {
515 P.BeginOffset = P.EndOffset;
516 P.EndOffset = MaxSplitSliceEndOffset;
517 return;
518 }
519
520 // If the we have split slices and the next slice is after a gap and is
521 // not splittable immediately form an empty partition for the split
522 // slices up until the next slice begins.
523 if (!P.SplitTails.empty() && P.SI->beginOffset() != P.EndOffset &&
524 !P.SI->isSplittable()) {
525 P.BeginOffset = P.EndOffset;
526 P.EndOffset = P.SI->beginOffset();
527 return;
528 }
529 }
530
531 // OK, we need to consume new slices. Set the end offset based on the
532 // current slice, and step SJ past it. The beginning offset of the
533 // partition is the beginning offset of the next slice unless we have
534 // pre-existing split slices that are continuing, in which case we begin
535 // at the prior end offset.
536 P.BeginOffset = P.SplitTails.empty() ? P.SI->beginOffset() : P.EndOffset;
537 P.EndOffset = P.SI->endOffset();
538 ++P.SJ;
539
540 // There are two strategies to form a partition based on whether the
541 // partition starts with an unsplittable slice or a splittable slice.
542 if (!P.SI->isSplittable()) {
543 // When we're forming an unsplittable region, it must always start at
544 // the first slice and will extend through its end.
545 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 545, __PRETTY_FUNCTION__))
;
546
547 // Form a partition including all of the overlapping slices with this
548 // unsplittable slice.
549 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
550 if (!P.SJ->isSplittable())
551 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
552 ++P.SJ;
553 }
554
555 // We have a partition across a set of overlapping unsplittable
556 // partitions.
557 return;
558 }
559
560 // If we're starting with a splittable slice, then we need to form
561 // a synthetic partition spanning it and any other overlapping splittable
562 // splices.
563 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 563, __PRETTY_FUNCTION__))
;
564
565 // Collect all of the overlapping splittable slices.
566 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset &&
567 P.SJ->isSplittable()) {
568 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
569 ++P.SJ;
570 }
571
572 // Back upiP.EndOffset if we ended the span early when encountering an
573 // unsplittable slice. This synthesizes the early end offset of
574 // a partition spanning only splittable slices.
575 if (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
576 assert(!P.SJ->isSplittable())((!P.SJ->isSplittable()) ? static_cast<void> (0) : __assert_fail
("!P.SJ->isSplittable()", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 576, __PRETTY_FUNCTION__))
;
577 P.EndOffset = P.SJ->beginOffset();
578 }
579 }
580
581public:
582 bool operator==(const partition_iterator &RHS) const {
583 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 584, __PRETTY_FUNCTION__))
584 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 584, __PRETTY_FUNCTION__))
;
585
586 // The observed positions of partitions is marked by the P.SI iterator and
587 // the emptiness of the split slices. The latter is only relevant when
588 // P.SI == SE, as the end iterator will additionally have an empty split
589 // slices list, but the prior may have the same P.SI and a tail of split
590 // slices.
591 if (P.SI == RHS.P.SI && P.SplitTails.empty() == RHS.P.SplitTails.empty()) {
592 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 593, __PRETTY_FUNCTION__))
593 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 593, __PRETTY_FUNCTION__))
;
594 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 596, __PRETTY_FUNCTION__))
595 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 596, __PRETTY_FUNCTION__))
596 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 596, __PRETTY_FUNCTION__))
;
597 return true;
598 }
599 return false;
600 }
601
602 partition_iterator &operator++() {
603 advance();
604 return *this;
605 }
606
607 Partition &operator*() { return P; }
608};
609
610/// A forward range over the partitions of the alloca's slices.
611///
612/// This accesses an iterator range over the partitions of the alloca's
613/// slices. It computes these partitions on the fly based on the overlapping
614/// offsets of the slices and the ability to split them. It will visit "empty"
615/// partitions to cover regions of the alloca only accessed via split
616/// slices.
617iterator_range<AllocaSlices::partition_iterator> AllocaSlices::partitions() {
618 return make_range(partition_iterator(begin(), end()),
619 partition_iterator(end(), end()));
620}
621
622static Value *foldSelectInst(SelectInst &SI) {
623 // If the condition being selected on is a constant or the same value is
624 // being selected between, fold the select. Yes this does (rarely) happen
625 // early on.
626 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
627 return SI.getOperand(1 + CI->isZero());
628 if (SI.getOperand(1) == SI.getOperand(2))
629 return SI.getOperand(1);
630
631 return nullptr;
632}
633
634/// A helper that folds a PHI node or a select.
635static Value *foldPHINodeOrSelectInst(Instruction &I) {
636 if (PHINode *PN = dyn_cast<PHINode>(&I)) {
637 // If PN merges together the same value, return that value.
638 return PN->hasConstantValue();
639 }
640 return foldSelectInst(cast<SelectInst>(I));
641}
642
643/// Builder for the alloca slices.
644///
645/// This class builds a set of alloca slices by recursively visiting the uses
646/// of an alloca and making a slice for each load and store at each offset.
647class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
648 friend class PtrUseVisitor<SliceBuilder>;
649 friend class InstVisitor<SliceBuilder>;
650
651 using Base = PtrUseVisitor<SliceBuilder>;
652
653 const uint64_t AllocSize;
654 AllocaSlices &AS;
655
656 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
657 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
658
659 /// Set to de-duplicate dead instructions found in the use walk.
660 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
661
662public:
663 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &AS)
664 : PtrUseVisitor<SliceBuilder>(DL),
665 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), AS(AS) {}
666
667private:
668 void markAsDead(Instruction &I) {
669 if (VisitedDeadInsts.insert(&I).second)
670 AS.DeadUsers.push_back(&I);
671 }
672
673 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
674 bool IsSplittable = false) {
675 // Completely skip uses which have a zero size or start either before or
676 // past the end of the allocation.
677 if (Size == 0 || Offset.uge(AllocSize)) {
678 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)
679 << 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)
680 << " 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)
681 << 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)
682 << " 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)
683 << " 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)
;
684 return markAsDead(I);
685 }
686
687 uint64_t BeginOffset = Offset.getZExtValue();
688 uint64_t EndOffset = BeginOffset + Size;
689
690 // Clamp the end offset to the end of the allocation. Note that this is
691 // formulated to handle even the case where "BeginOffset + Size" overflows.
692 // This may appear superficially to be something we could ignore entirely,
693 // but that is not so! There may be widened loads or PHI-node uses where
694 // some instructions are dead but not others. We can't completely ignore
695 // them, and so have to record at least the information here.
696 assert(AllocSize >= BeginOffset)((AllocSize >= BeginOffset) ? static_cast<void> (0) :
__assert_fail ("AllocSize >= BeginOffset", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 696, __PRETTY_FUNCTION__))
; // Established above.
697 if (Size > AllocSize - BeginOffset) {
698 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)
699 << 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)
700 << " 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)
701 << " 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)
702 << " 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)
;
703 EndOffset = AllocSize;
704 }
705
706 AS.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
707 }
708
709 void visitBitCastInst(BitCastInst &BC) {
710 if (BC.use_empty())
711 return markAsDead(BC);
712
713 return Base::visitBitCastInst(BC);
714 }
715
716 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
717 if (GEPI.use_empty())
718 return markAsDead(GEPI);
719
720 if (SROAStrictInbounds && GEPI.isInBounds()) {
721 // FIXME: This is a manually un-factored variant of the basic code inside
722 // of GEPs with checking of the inbounds invariant specified in the
723 // langref in a very strict sense. If we ever want to enable
724 // SROAStrictInbounds, this code should be factored cleanly into
725 // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
726 // by writing out the code here where we have the underlying allocation
727 // size readily available.
728 APInt GEPOffset = Offset;
729 const DataLayout &DL = GEPI.getModule()->getDataLayout();
730 for (gep_type_iterator GTI = gep_type_begin(GEPI),
731 GTE = gep_type_end(GEPI);
732 GTI != GTE; ++GTI) {
733 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
734 if (!OpC)
735 break;
736
737 // Handle a struct index, which adds its field offset to the pointer.
738 if (StructType *STy = GTI.getStructTypeOrNull()) {
739 unsigned ElementIdx = OpC->getZExtValue();
740 const StructLayout *SL = DL.getStructLayout(STy);
741 GEPOffset +=
742 APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
743 } else {
744 // For array or vector indices, scale the index by the size of the
745 // type.
746 APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
747 GEPOffset += Index * APInt(Offset.getBitWidth(),
748 DL.getTypeAllocSize(GTI.getIndexedType()));
749 }
750
751 // If this index has computed an intermediate pointer which is not
752 // inbounds, then the result of the GEP is a poison value and we can
753 // delete it and all uses.
754 if (GEPOffset.ugt(AllocSize))
755 return markAsDead(GEPI);
756 }
757 }
758
759 return Base::visitGetElementPtrInst(GEPI);
760 }
761
762 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
763 uint64_t Size, bool IsVolatile) {
764 // We allow splitting of non-volatile loads and stores where the type is an
765 // integer type. These may be used to implement 'memcpy' or other "transfer
766 // of bits" patterns.
767 bool IsSplittable = Ty->isIntegerTy() && !IsVolatile;
768
769 insertUse(I, Offset, Size, IsSplittable);
770 }
771
772 void visitLoadInst(LoadInst &LI) {
773 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 774, __PRETTY_FUNCTION__))
774 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 774, __PRETTY_FUNCTION__))
;
775
776 if (!IsOffsetKnown)
777 return PI.setAborted(&LI);
778
779 const DataLayout &DL = LI.getModule()->getDataLayout();
780 uint64_t Size = DL.getTypeStoreSize(LI.getType());
781 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
782 }
783
784 void visitStoreInst(StoreInst &SI) {
785 Value *ValOp = SI.getValueOperand();
786 if (ValOp == *U)
787 return PI.setEscapedAndAborted(&SI);
788 if (!IsOffsetKnown)
789 return PI.setAborted(&SI);
790
791 const DataLayout &DL = SI.getModule()->getDataLayout();
792 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
793
794 // If this memory access can be shown to *statically* extend outside the
795 // bounds of the allocation, it's behavior is undefined, so simply
796 // ignore it. Note that this is more strict than the generic clamping
797 // behavior of insertUse. We also try to handle cases which might run the
798 // risk of overflow.
799 // FIXME: We should instead consider the pointer to have escaped if this
800 // function is being instrumented for addressing bugs or race conditions.
801 if (Size > AllocSize || Offset.ugt(AllocSize - Size)) {
802 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)
803 << 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)
804 << 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)
805 << " 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)
806 << " 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)
;
807 return markAsDead(SI);
808 }
809
810 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 811, __PRETTY_FUNCTION__))
811 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 811, __PRETTY_FUNCTION__))
;
812 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
813 }
814
815 void visitMemSetInst(MemSetInst &II) {
816 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 816, __PRETTY_FUNCTION__))
;
817 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
818 if ((Length && Length->getValue() == 0) ||
819 (IsOffsetKnown && Offset.uge(AllocSize)))
820 // Zero-length mem transfer intrinsics can be ignored entirely.
821 return markAsDead(II);
822
823 if (!IsOffsetKnown)
824 return PI.setAborted(&II);
825
826 insertUse(II, Offset, Length ? Length->getLimitedValue()
827 : AllocSize - Offset.getLimitedValue(),
828 (bool)Length);
829 }
830
831 void visitMemTransferInst(MemTransferInst &II) {
832 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
833 if (Length && Length->getValue() == 0)
834 // Zero-length mem transfer intrinsics can be ignored entirely.
835 return markAsDead(II);
836
837 // Because we can visit these intrinsics twice, also check to see if the
838 // first time marked this instruction as dead. If so, skip it.
839 if (VisitedDeadInsts.count(&II))
840 return;
841
842 if (!IsOffsetKnown)
843 return PI.setAborted(&II);
844
845 // This side of the transfer is completely out-of-bounds, and so we can
846 // nuke the entire transfer. However, we also need to nuke the other side
847 // if already added to our partitions.
848 // FIXME: Yet another place we really should bypass this when
849 // instrumenting for ASan.
850 if (Offset.uge(AllocSize)) {
851 SmallDenseMap<Instruction *, unsigned>::iterator MTPI =
852 MemTransferSliceMap.find(&II);
853 if (MTPI != MemTransferSliceMap.end())
854 AS.Slices[MTPI->second].kill();
855 return markAsDead(II);
856 }
857
858 uint64_t RawOffset = Offset.getLimitedValue();
859 uint64_t Size = Length ? Length->getLimitedValue() : AllocSize - RawOffset;
860
861 // Check for the special case where the same exact value is used for both
862 // source and dest.
863 if (*U == II.getRawDest() && *U == II.getRawSource()) {
864 // For non-volatile transfers this is a no-op.
865 if (!II.isVolatile())
866 return markAsDead(II);
867
868 return insertUse(II, Offset, Size, /*IsSplittable=*/false);
869 }
870
871 // If we have seen both source and destination for a mem transfer, then
872 // they both point to the same alloca.
873 bool Inserted;
874 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
875 std::tie(MTPI, Inserted) =
876 MemTransferSliceMap.insert(std::make_pair(&II, AS.Slices.size()));
877 unsigned PrevIdx = MTPI->second;
878 if (!Inserted) {
879 Slice &PrevP = AS.Slices[PrevIdx];
880
881 // Check if the begin offsets match and this is a non-volatile transfer.
882 // In that case, we can completely elide the transfer.
883 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
884 PrevP.kill();
885 return markAsDead(II);
886 }
887
888 // Otherwise we have an offset transfer within the same alloca. We can't
889 // split those.
890 PrevP.makeUnsplittable();
891 }
892
893 // Insert the use now that we've fixed up the splittable nature.
894 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
895
896 // Check that we ended up with a valid index in the map.
897 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 898, __PRETTY_FUNCTION__))
898 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 898, __PRETTY_FUNCTION__))
;
899 }
900
901 // Disable SRoA for any intrinsics except for lifetime invariants.
902 // FIXME: What about debug intrinsics? This matches old behavior, but
903 // doesn't make sense.
904 void visitIntrinsicInst(IntrinsicInst &II) {
905 if (!IsOffsetKnown)
906 return PI.setAborted(&II);
907
908 if (II.isLifetimeStartOrEnd()) {
909 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
910 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
911 Length->getLimitedValue());
912 insertUse(II, Offset, Size, true);
913 return;
914 }
915
916 Base::visitIntrinsicInst(II);
917 }
918
919 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
920 // We consider any PHI or select that results in a direct load or store of
921 // the same offset to be a viable use for slicing purposes. These uses
922 // are considered unsplittable and the size is the maximum loaded or stored
923 // size.
924 SmallPtrSet<Instruction *, 4> Visited;
925 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
926 Visited.insert(Root);
927 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
928 const DataLayout &DL = Root->getModule()->getDataLayout();
929 // If there are no loads or stores, the access is dead. We mark that as
930 // a size zero access.
931 Size = 0;
932 do {
933 Instruction *I, *UsedI;
934 std::tie(UsedI, I) = Uses.pop_back_val();
935
936 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
937 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
938 continue;
939 }
940 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
941 Value *Op = SI->getOperand(0);
942 if (Op == UsedI)
943 return SI;
944 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
945 continue;
946 }
947
948 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
949 if (!GEP->hasAllZeroIndices())
950 return GEP;
951 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
952 !isa<SelectInst>(I)) {
953 return I;
954 }
955
956 for (User *U : I->users())
957 if (Visited.insert(cast<Instruction>(U)).second)
958 Uses.push_back(std::make_pair(I, cast<Instruction>(U)));
959 } while (!Uses.empty());
960
961 return nullptr;
962 }
963
964 void visitPHINodeOrSelectInst(Instruction &I) {
965 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 965, __PRETTY_FUNCTION__))
;
966 if (I.use_empty())
967 return markAsDead(I);
968
969 // TODO: We could use SimplifyInstruction here to fold PHINodes and
970 // SelectInsts. However, doing so requires to change the current
971 // dead-operand-tracking mechanism. For instance, suppose neither loading
972 // from %U nor %other traps. Then "load (select undef, %U, %other)" does not
973 // trap either. However, if we simply replace %U with undef using the
974 // current dead-operand-tracking mechanism, "load (select undef, undef,
975 // %other)" may trap because the select may return the first operand
976 // "undef".
977 if (Value *Result = foldPHINodeOrSelectInst(I)) {
978 if (Result == *U)
979 // If the result of the constant fold will be the pointer, recurse
980 // through the PHI/select as if we had RAUW'ed it.
981 enqueueUsers(I);
982 else
983 // Otherwise the operand to the PHI/select is dead, and we can replace
984 // it with undef.
985 AS.DeadOperands.push_back(U);
986
987 return;
988 }
989
990 if (!IsOffsetKnown)
991 return PI.setAborted(&I);
992
993 // See if we already have computed info on this node.
994 uint64_t &Size = PHIOrSelectSizes[&I];
995 if (!Size) {
996 // This is a new PHI/Select, check for an unsafe use of it.
997 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&I, Size))
998 return PI.setAborted(UnsafeI);
999 }
1000
1001 // For PHI and select operands outside the alloca, we can't nuke the entire
1002 // phi or select -- the other side might still be relevant, so we special
1003 // case them here and use a separate structure to track the operands
1004 // themselves which should be replaced with undef.
1005 // FIXME: This should instead be escaped in the event we're instrumenting
1006 // for address sanitization.
1007 if (Offset.uge(AllocSize)) {
1008 AS.DeadOperands.push_back(U);
1009 return;
1010 }
1011
1012 insertUse(I, Offset, Size);
1013 }
1014
1015 void visitPHINode(PHINode &PN) { visitPHINodeOrSelectInst(PN); }
1016
1017 void visitSelectInst(SelectInst &SI) { visitPHINodeOrSelectInst(SI); }
1018
1019 /// Disable SROA entirely if there are unhandled users of the alloca.
1020 void visitInstruction(Instruction &I) { PI.setAborted(&I); }
1021};
1022
1023AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
1024 :
1025#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1026 AI(AI),
1027#endif
1028 PointerEscapingInstr(nullptr) {
1029 SliceBuilder PB(DL, AI, *this);
1030 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
1031 if (PtrI.isEscaped() || PtrI.isAborted()) {
1032 // FIXME: We should sink the escape vs. abort info into the caller nicely,
1033 // possibly by just storing the PtrInfo in the AllocaSlices.
1034 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
1035 : PtrI.getAbortingInst();
1036 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1036, __PRETTY_FUNCTION__))
;
1037 return;
1038 }
1039
1040 Slices.erase(
1041 llvm::remove_if(Slices, [](const Slice &S) { return S.isDead(); }),
1042 Slices.end());
1043
1044#ifndef NDEBUG
1045 if (SROARandomShuffleSlices) {
1046 std::mt19937 MT(static_cast<unsigned>(
1047 std::chrono::system_clock::now().time_since_epoch().count()));
1048 std::shuffle(Slices.begin(), Slices.end(), MT);
1049 }
1050#endif
1051
1052 // Sort the uses. This arranges for the offsets to be in ascending order,
1053 // and the sizes to be in descending order.
1054 llvm::sort(Slices);
1055}
1056
1057#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1058
1059void AllocaSlices::print(raw_ostream &OS, const_iterator I,
1060 StringRef Indent) const {
1061 printSlice(OS, I, Indent);
1062 OS << "\n";
1063 printUse(OS, I, Indent);
1064}
1065
1066void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
1067 StringRef Indent) const {
1068 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
1069 << " slice #" << (I - begin())
1070 << (I->isSplittable() ? " (splittable)" : "");
1071}
1072
1073void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
1074 StringRef Indent) const {
1075 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
1076}
1077
1078void AllocaSlices::print(raw_ostream &OS) const {
1079 if (PointerEscapingInstr) {
1080 OS << "Can't analyze slices for alloca: " << AI << "\n"
1081 << " A pointer to this alloca escaped by:\n"
1082 << " " << *PointerEscapingInstr << "\n";
1083 return;
1084 }
1085
1086 OS << "Slices of alloca: " << AI << "\n";
1087 for (const_iterator I = begin(), E = end(); I != E; ++I)
1088 print(OS, I);
1089}
1090
1091LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void AllocaSlices::dump(const_iterator I) const {
1092 print(dbgs(), I);
1093}
1094LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void AllocaSlices::dump() const { print(dbgs()); }
1095
1096#endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1097
1098/// Walk the range of a partitioning looking for a common type to cover this
1099/// sequence of slices.
1100static Type *findCommonType(AllocaSlices::const_iterator B,
1101 AllocaSlices::const_iterator E,
1102 uint64_t EndOffset) {
1103 Type *Ty = nullptr;
1104 bool TyIsCommon = true;
1105 IntegerType *ITy = nullptr;
1106
1107 // Note that we need to look at *every* alloca slice's Use to ensure we
1108 // always get consistent results regardless of the order of slices.
1109 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1110 Use *U = I->getUse();
1111 if (isa<IntrinsicInst>(*U->getUser()))
1112 continue;
1113 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1114 continue;
1115
1116 Type *UserTy = nullptr;
1117 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1118 UserTy = LI->getType();
1119 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1120 UserTy = SI->getValueOperand()->getType();
1121 }
1122
1123 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1124 // If the type is larger than the partition, skip it. We only encounter
1125 // this for split integer operations where we want to use the type of the
1126 // entity causing the split. Also skip if the type is not a byte width
1127 // multiple.
1128 if (UserITy->getBitWidth() % 8 != 0 ||
1129 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1130 continue;
1131
1132 // Track the largest bitwidth integer type used in this way in case there
1133 // is no common type.
1134 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1135 ITy = UserITy;
1136 }
1137
1138 // To avoid depending on the order of slices, Ty and TyIsCommon must not
1139 // depend on types skipped above.
1140 if (!UserTy || (Ty && Ty != UserTy))
1141 TyIsCommon = false; // Give up on anything but an iN type.
1142 else
1143 Ty = UserTy;
1144 }
1145
1146 return TyIsCommon ? Ty : ITy;
1147}
1148
1149/// PHI instructions that use an alloca and are subsequently loaded can be
1150/// rewritten to load both input pointers in the pred blocks and then PHI the
1151/// results, allowing the load of the alloca to be promoted.
1152/// From this:
1153/// %P2 = phi [i32* %Alloca, i32* %Other]
1154/// %V = load i32* %P2
1155/// to:
1156/// %V1 = load i32* %Alloca -> will be mem2reg'd
1157/// ...
1158/// %V2 = load i32* %Other
1159/// ...
1160/// %V = phi [i32 %V1, i32 %V2]
1161///
1162/// We can do this to a select if its only uses are loads and if the operands
1163/// to the select can be loaded unconditionally.
1164///
1165/// FIXME: This should be hoisted into a generic utility, likely in
1166/// Transforms/Util/Local.h
1167static bool isSafePHIToSpeculate(PHINode &PN) {
1168 // For now, we can only do this promotion if the load is in the same block
1169 // as the PHI, and if there are no stores between the phi and load.
1170 // TODO: Allow recursive phi users.
1171 // TODO: Allow stores.
1172 BasicBlock *BB = PN.getParent();
1173 unsigned MaxAlign = 0;
1174 bool HaveLoad = false;
1175 for (User *U : PN.users()) {
1176 LoadInst *LI = dyn_cast<LoadInst>(U);
1177 if (!LI || !LI->isSimple())
1178 return false;
1179
1180 // For now we only allow loads in the same block as the PHI. This is
1181 // a common case that happens when instcombine merges two loads through
1182 // a PHI.
1183 if (LI->getParent() != BB)
1184 return false;
1185
1186 // Ensure that there are no instructions between the PHI and the load that
1187 // could store.
1188 for (BasicBlock::iterator BBI(PN); &*BBI != LI; ++BBI)
1189 if (BBI->mayWriteToMemory())
1190 return false;
1191
1192 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1193 HaveLoad = true;
1194 }
1195
1196 if (!HaveLoad)
1197 return false;
1198
1199 const DataLayout &DL = PN.getModule()->getDataLayout();
1200
1201 // We can only transform this if it is safe to push the loads into the
1202 // predecessor blocks. The only thing to watch out for is that we can't put
1203 // a possibly trapping load in the predecessor if it is a critical edge.
1204 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1205 Instruction *TI = PN.getIncomingBlock(Idx)->getTerminator();
1206 Value *InVal = PN.getIncomingValue(Idx);
1207
1208 // If the value is produced by the terminator of the predecessor (an
1209 // invoke) or it has side-effects, there is no valid place to put a load
1210 // in the predecessor.
1211 if (TI == InVal || TI->mayHaveSideEffects())
1212 return false;
1213
1214 // If the predecessor has a single successor, then the edge isn't
1215 // critical.
1216 if (TI->getNumSuccessors() == 1)
1217 continue;
1218
1219 // If this pointer is always safe to load, or if we can prove that there
1220 // is already a load in the block, then we can move the load to the pred
1221 // block.
1222 if (isSafeToLoadUnconditionally(InVal, MaxAlign, DL, TI))
1223 continue;
1224
1225 return false;
1226 }
1227
1228 return true;
1229}
1230
1231static void speculatePHINodeLoads(PHINode &PN) {
1232 LLVM_DEBUG(dbgs() << " original: " << PN << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << PN <<
"\n"; } } while (false)
;
1233
1234 LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1235 Type *LoadTy = SomeLoad->getType();
1236 IRBuilderTy PHIBuilder(&PN);
1237 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1238 PN.getName() + ".sroa.speculated");
1239
1240 // Get the AA tags and alignment to use from one of the loads. It doesn't
1241 // matter which one we get and if any differ.
1242 AAMDNodes AATags;
1243 SomeLoad->getAAMetadata(AATags);
1244 unsigned Align = SomeLoad->getAlignment();
1245
1246 // Rewrite all loads of the PN to use the new PHI.
1247 while (!PN.use_empty()) {
1248 LoadInst *LI = cast<LoadInst>(PN.user_back());
1249 LI->replaceAllUsesWith(NewPN);
1250 LI->eraseFromParent();
1251 }
1252
1253 // Inject loads into all of the pred blocks.
1254 DenseMap<BasicBlock*, Value*> InjectedLoads;
1255 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1256 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1257 Value *InVal = PN.getIncomingValue(Idx);
1258
1259 // A PHI node is allowed to have multiple (duplicated) entries for the same
1260 // basic block, as long as the value is the same. So if we already injected
1261 // a load in the predecessor, then we should reuse the same load for all
1262 // duplicated entries.
1263 if (Value* V = InjectedLoads.lookup(Pred)) {
1264 NewPN->addIncoming(V, Pred);
1265 continue;
1266 }
1267
1268 Instruction *TI = Pred->getTerminator();
1269 IRBuilderTy PredBuilder(TI);
1270
1271 LoadInst *Load = PredBuilder.CreateLoad(
1272 LoadTy, InVal,
1273 (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1274 ++NumLoadsSpeculated;
1275 Load->setAlignment(Align);
1276 if (AATags)
1277 Load->setAAMetadata(AATags);
1278 NewPN->addIncoming(Load, Pred);
1279 InjectedLoads[Pred] = Load;
1280 }
1281
1282 LLVM_DEBUG(dbgs() << " speculated to: " << *NewPN << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " speculated to: " <<
*NewPN << "\n"; } } while (false)
;
1283 PN.eraseFromParent();
1284}
1285
1286/// Select instructions that use an alloca and are subsequently loaded can be
1287/// rewritten to load both input pointers and then select between the result,
1288/// allowing the load of the alloca to be promoted.
1289/// From this:
1290/// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1291/// %V = load i32* %P2
1292/// to:
1293/// %V1 = load i32* %Alloca -> will be mem2reg'd
1294/// %V2 = load i32* %Other
1295/// %V = select i1 %cond, i32 %V1, i32 %V2
1296///
1297/// We can do this to a select if its only uses are loads and if the operand
1298/// to the select can be loaded unconditionally.
1299static bool isSafeSelectToSpeculate(SelectInst &SI) {
1300 Value *TValue = SI.getTrueValue();
1301 Value *FValue = SI.getFalseValue();
1302 const DataLayout &DL = SI.getModule()->getDataLayout();
1303
1304 for (User *U : SI.users()) {
1305 LoadInst *LI = dyn_cast<LoadInst>(U);
1306 if (!LI || !LI->isSimple())
1307 return false;
1308
1309 // Both operands to the select need to be dereferenceable, either
1310 // absolutely (e.g. allocas) or at this point because we can see other
1311 // accesses to it.
1312 if (!isSafeToLoadUnconditionally(TValue, LI->getAlignment(), DL, LI))
1313 return false;
1314 if (!isSafeToLoadUnconditionally(FValue, LI->getAlignment(), DL, LI))
1315 return false;
1316 }
1317
1318 return true;
1319}
1320
1321static void speculateSelectInstLoads(SelectInst &SI) {
1322 LLVM_DEBUG(dbgs() << " original: " << SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << SI <<
"\n"; } } while (false)
;
1323
1324 IRBuilderTy IRB(&SI);
1325 Value *TV = SI.getTrueValue();
1326 Value *FV = SI.getFalseValue();
1327 // Replace the loads of the select with a select of two loads.
1328 while (!SI.use_empty()) {
1329 LoadInst *LI = cast<LoadInst>(SI.user_back());
1330 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1330, __PRETTY_FUNCTION__))
;
1331
1332 IRB.SetInsertPoint(LI);
1333 LoadInst *TL = IRB.CreateLoad(LI->getType(), TV,
1334 LI->getName() + ".sroa.speculate.load.true");
1335 LoadInst *FL = IRB.CreateLoad(LI->getType(), FV,
1336 LI->getName() + ".sroa.speculate.load.false");
1337 NumLoadsSpeculated += 2;
1338
1339 // Transfer alignment and AA info if present.
1340 TL->setAlignment(LI->getAlignment());
1341 FL->setAlignment(LI->getAlignment());
1342
1343 AAMDNodes Tags;
1344 LI->getAAMetadata(Tags);
1345 if (Tags) {
1346 TL->setAAMetadata(Tags);
1347 FL->setAAMetadata(Tags);
1348 }
1349
1350 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1351 LI->getName() + ".sroa.speculated");
1352
1353 LLVM_DEBUG(dbgs() << " speculated to: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " speculated to: " <<
*V << "\n"; } } while (false)
;
1354 LI->replaceAllUsesWith(V);
1355 LI->eraseFromParent();
1356 }
1357 SI.eraseFromParent();
1358}
1359
1360/// Build a GEP out of a base pointer and indices.
1361///
1362/// This will return the BasePtr if that is valid, or build a new GEP
1363/// instruction using the IRBuilder if GEP-ing is needed.
1364static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1365 SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1366 if (Indices.empty())
1367 return BasePtr;
1368
1369 // A single zero index is a no-op, so check for this and avoid building a GEP
1370 // in that case.
1371 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1372 return BasePtr;
1373
1374 return IRB.CreateInBoundsGEP(BasePtr->getType()->getPointerElementType(),
1375 BasePtr, Indices, NamePrefix + "sroa_idx");
1376}
1377
1378/// Get a natural GEP off of the BasePtr walking through Ty toward
1379/// TargetTy without changing the offset of the pointer.
1380///
1381/// This routine assumes we've already established a properly offset GEP with
1382/// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1383/// zero-indices down through type layers until we find one the same as
1384/// TargetTy. If we can't find one with the same type, we at least try to use
1385/// one with the same size. If none of that works, we just produce the GEP as
1386/// indicated by Indices to have the correct offset.
1387static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1388 Value *BasePtr, Type *Ty, Type *TargetTy,
1389 SmallVectorImpl<Value *> &Indices,
1390 Twine NamePrefix) {
1391 if (Ty == TargetTy)
1392 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1393
1394 // Offset size to use for the indices.
1395 unsigned OffsetSize = DL.getIndexTypeSizeInBits(BasePtr->getType());
1396
1397 // See if we can descend into a struct and locate a field with the correct
1398 // type.
1399 unsigned NumLayers = 0;
1400 Type *ElementTy = Ty;
1401 do {
1402 if (ElementTy->isPointerTy())
1403 break;
1404
1405 if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1406 ElementTy = ArrayTy->getElementType();
1407 Indices.push_back(IRB.getIntN(OffsetSize, 0));
1408 } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1409 ElementTy = VectorTy->getElementType();
1410 Indices.push_back(IRB.getInt32(0));
1411 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1412 if (STy->element_begin() == STy->element_end())
1413 break; // Nothing left to descend into.
1414 ElementTy = *STy->element_begin();
1415 Indices.push_back(IRB.getInt32(0));
1416 } else {
1417 break;
1418 }
1419 ++NumLayers;
1420 } while (ElementTy != TargetTy);
1421 if (ElementTy != TargetTy)
1422 Indices.erase(Indices.end() - NumLayers, Indices.end());
1423
1424 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1425}
1426
1427/// Recursively compute indices for a natural GEP.
1428///
1429/// This is the recursive step for getNaturalGEPWithOffset that walks down the
1430/// element types adding appropriate indices for the GEP.
1431static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1432 Value *Ptr, Type *Ty, APInt &Offset,
1433 Type *TargetTy,
1434 SmallVectorImpl<Value *> &Indices,
1435 Twine NamePrefix) {
1436 if (Offset == 0)
1437 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices,
1438 NamePrefix);
1439
1440 // We can't recurse through pointer types.
1441 if (Ty->isPointerTy())
1442 return nullptr;
1443
1444 // We try to analyze GEPs over vectors here, but note that these GEPs are
1445 // extremely poorly defined currently. The long-term goal is to remove GEPing
1446 // over a vector from the IR completely.
1447 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1448 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1449 if (ElementSizeInBits % 8 != 0) {
1450 // GEPs over non-multiple of 8 size vector elements are invalid.
1451 return nullptr;
1452 }
1453 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1454 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1455 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1456 return nullptr;
1457 Offset -= NumSkippedElements * ElementSize;
1458 Indices.push_back(IRB.getInt(NumSkippedElements));
1459 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1460 Offset, TargetTy, Indices, NamePrefix);
1461 }
1462
1463 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1464 Type *ElementTy = ArrTy->getElementType();
1465 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1466 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1467 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1468 return nullptr;
1469
1470 Offset -= NumSkippedElements * ElementSize;
1471 Indices.push_back(IRB.getInt(NumSkippedElements));
1472 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1473 Indices, NamePrefix);
1474 }
1475
1476 StructType *STy = dyn_cast<StructType>(Ty);
1477 if (!STy)
1478 return nullptr;
1479
1480 const StructLayout *SL = DL.getStructLayout(STy);
1481 uint64_t StructOffset = Offset.getZExtValue();
1482 if (StructOffset >= SL->getSizeInBytes())
1483 return nullptr;
1484 unsigned Index = SL->getElementContainingOffset(StructOffset);
1485 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1486 Type *ElementTy = STy->getElementType(Index);
1487 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1488 return nullptr; // The offset points into alignment padding.
1489
1490 Indices.push_back(IRB.getInt32(Index));
1491 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1492 Indices, NamePrefix);
1493}
1494
1495/// Get a natural GEP from a base pointer to a particular offset and
1496/// resulting in a particular type.
1497///
1498/// The goal is to produce a "natural" looking GEP that works with the existing
1499/// composite types to arrive at the appropriate offset and element type for
1500/// a pointer. TargetTy is the element type the returned GEP should point-to if
1501/// possible. We recurse by decreasing Offset, adding the appropriate index to
1502/// Indices, and setting Ty to the result subtype.
1503///
1504/// If no natural GEP can be constructed, this function returns null.
1505static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1506 Value *Ptr, APInt Offset, Type *TargetTy,
1507 SmallVectorImpl<Value *> &Indices,
1508 Twine NamePrefix) {
1509 PointerType *Ty = cast<PointerType>(Ptr->getType());
1510
1511 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1512 // an i8.
1513 if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1514 return nullptr;
1515
1516 Type *ElementTy = Ty->getElementType();
1517 if (!ElementTy->isSized())
1518 return nullptr; // We can't GEP through an unsized element.
1519 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1520 if (ElementSize == 0)
1521 return nullptr; // Zero-length arrays can't help us build a natural GEP.
1522 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1523
1524 Offset -= NumSkippedElements * ElementSize;
1525 Indices.push_back(IRB.getInt(NumSkippedElements));
1526 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1527 Indices, NamePrefix);
1528}
1529
1530/// Compute an adjusted pointer from Ptr by Offset bytes where the
1531/// resulting pointer has PointerTy.
1532///
1533/// This tries very hard to compute a "natural" GEP which arrives at the offset
1534/// and produces the pointer type desired. Where it cannot, it will try to use
1535/// the natural GEP to arrive at the offset and bitcast to the type. Where that
1536/// fails, it will try to use an existing i8* and GEP to the byte offset and
1537/// bitcast to the type.
1538///
1539/// The strategy for finding the more natural GEPs is to peel off layers of the
1540/// pointer, walking back through bit casts and GEPs, searching for a base
1541/// pointer from which we can compute a natural GEP with the desired
1542/// properties. The algorithm tries to fold as many constant indices into
1543/// a single GEP as possible, thus making each GEP more independent of the
1544/// surrounding code.
1545static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1546 APInt Offset, Type *PointerTy, Twine NamePrefix) {
1547 // Even though we don't look through PHI nodes, we could be called on an
1548 // instruction in an unreachable block, which may be on a cycle.
1549 SmallPtrSet<Value *, 4> Visited;
1550 Visited.insert(Ptr);
1551 SmallVector<Value *, 4> Indices;
1552
1553 // We may end up computing an offset pointer that has the wrong type. If we
1554 // never are able to compute one directly that has the correct type, we'll
1555 // fall back to it, so keep it and the base it was computed from around here.
1556 Value *OffsetPtr = nullptr;
1557 Value *OffsetBasePtr;
1558
1559 // Remember any i8 pointer we come across to re-use if we need to do a raw
1560 // byte offset.
1561 Value *Int8Ptr = nullptr;
1562 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1563
1564 Type *TargetTy = PointerTy->getPointerElementType();
1565
1566 do {
1567 // First fold any existing GEPs into the offset.
1568 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1569 APInt GEPOffset(Offset.getBitWidth(), 0);
1570 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1571 break;
1572 Offset += GEPOffset;
1573 Ptr = GEP->getPointerOperand();
1574 if (!Visited.insert(Ptr).second)
1575 break;
1576 }
1577
1578 // See if we can perform a natural GEP here.
1579 Indices.clear();
1580 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1581 Indices, NamePrefix)) {
1582 // If we have a new natural pointer at the offset, clear out any old
1583 // offset pointer we computed. Unless it is the base pointer or
1584 // a non-instruction, we built a GEP we don't need. Zap it.
1585 if (OffsetPtr && OffsetPtr != OffsetBasePtr)
1586 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr)) {
1587 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1587, __PRETTY_FUNCTION__))
;
1588 I->eraseFromParent();
1589 }
1590 OffsetPtr = P;
1591 OffsetBasePtr = Ptr;
1592 // If we also found a pointer of the right type, we're done.
1593 if (P->getType() == PointerTy)
1594 return P;
1595 }
1596
1597 // Stash this pointer if we've found an i8*.
1598 if (Ptr->getType()->isIntegerTy(8)) {
1599 Int8Ptr = Ptr;
1600 Int8PtrOffset = Offset;
1601 }
1602
1603 // Peel off a layer of the pointer and update the offset appropriately.
1604 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1605 Ptr = cast<Operator>(Ptr)->getOperand(0);
1606 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1607 if (GA->isInterposable())
1608 break;
1609 Ptr = GA->getAliasee();
1610 } else {
1611 break;
1612 }
1613 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1613, __PRETTY_FUNCTION__))
;
1614 } while (Visited.insert(Ptr).second);
1615
1616 if (!OffsetPtr) {
1617 if (!Int8Ptr) {
1618 Int8Ptr = IRB.CreateBitCast(
1619 Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1620 NamePrefix + "sroa_raw_cast");
1621 Int8PtrOffset = Offset;
1622 }
1623
1624 OffsetPtr = Int8PtrOffset == 0
1625 ? Int8Ptr
1626 : IRB.CreateInBoundsGEP(IRB.getInt8Ty(), Int8Ptr,
1627 IRB.getInt(Int8PtrOffset),
1628 NamePrefix + "sroa_raw_idx");
1629 }
1630 Ptr = OffsetPtr;
1631
1632 // On the off chance we were targeting i8*, guard the bitcast here.
1633 if (Ptr->getType() != PointerTy)
1634 Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
1635
1636 return Ptr;
1637}
1638
1639/// Compute the adjusted alignment for a load or store from an offset.
1640static unsigned getAdjustedAlignment(Instruction *I, uint64_t Offset,
1641 const DataLayout &DL) {
1642 unsigned Alignment;
1643 Type *Ty;
1644 if (auto *LI = dyn_cast<LoadInst>(I)) {
1645 Alignment = LI->getAlignment();
1646 Ty = LI->getType();
1647 } else if (auto *SI = dyn_cast<StoreInst>(I)) {
1648 Alignment = SI->getAlignment();
1649 Ty = SI->getValueOperand()->getType();
1650 } else {
1651 llvm_unreachable("Only loads and stores are allowed!")::llvm::llvm_unreachable_internal("Only loads and stores are allowed!"
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1651)
;
1652 }
1653
1654 if (!Alignment)
1655 Alignment = DL.getABITypeAlignment(Ty);
1656
1657 return MinAlign(Alignment, Offset);
1658}
1659
1660/// Test whether we can convert a value from the old to the new type.
1661///
1662/// This predicate should be used to guard calls to convertValue in order to
1663/// ensure that we only try to convert viable values. The strategy is that we
1664/// will peel off single element struct and array wrappings to get to an
1665/// underlying value, and convert that value.
1666static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1667 if (OldTy == NewTy)
1668 return true;
1669
1670 // For integer types, we can't handle any bit-width differences. This would
1671 // break both vector conversions with extension and introduce endianness
1672 // issues when in conjunction with loads and stores.
1673 if (isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) {
1674 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1676, __PRETTY_FUNCTION__))
1675 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1676, __PRETTY_FUNCTION__))
1676 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1676, __PRETTY_FUNCTION__))
;
1677 return false;
1678 }
1679
1680 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1681 return false;
1682 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1683 return false;
1684
1685 // We can convert pointers to integers and vice-versa. Same for vectors
1686 // of pointers and integers.
1687 OldTy = OldTy->getScalarType();
1688 NewTy = NewTy->getScalarType();
1689 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1690 if (NewTy->isPointerTy() && OldTy->isPointerTy()) {
1691 return cast<PointerType>(NewTy)->getPointerAddressSpace() ==
1692 cast<PointerType>(OldTy)->getPointerAddressSpace();
1693 }
1694
1695 // We can convert integers to integral pointers, but not to non-integral
1696 // pointers.
1697 if (OldTy->isIntegerTy())
1698 return !DL.isNonIntegralPointerType(NewTy);
1699
1700 // We can convert integral pointers to integers, but non-integral pointers
1701 // need to remain pointers.
1702 if (!DL.isNonIntegralPointerType(OldTy))
1703 return NewTy->isIntegerTy();
1704
1705 return false;
1706 }
1707
1708 return true;
1709}
1710
1711/// Generic routine to convert an SSA value to a value of a different
1712/// type.
1713///
1714/// This will try various different casting techniques, such as bitcasts,
1715/// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1716/// two types for viability with this routine.
1717static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1718 Type *NewTy) {
1719 Type *OldTy = V->getType();
1720 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1720, __PRETTY_FUNCTION__))
;
1721
1722 if (OldTy == NewTy)
1723 return V;
1724
1725 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1726, __PRETTY_FUNCTION__))
1726 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1726, __PRETTY_FUNCTION__))
;
1727
1728 // See if we need inttoptr for this type pair. A cast involving both scalars
1729 // and vectors requires and additional bitcast.
1730 if (OldTy->isIntOrIntVectorTy() && NewTy->isPtrOrPtrVectorTy()) {
1731 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1732 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1733 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1734 NewTy);
1735
1736 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1737 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1738 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1739 NewTy);
1740
1741 return IRB.CreateIntToPtr(V, NewTy);
1742 }
1743
1744 // See if we need ptrtoint for this type pair. A cast involving both scalars
1745 // and vectors requires and additional bitcast.
1746 if (OldTy->isPtrOrPtrVectorTy() && NewTy->isIntOrIntVectorTy()) {
1747 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1748 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1749 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1750 NewTy);
1751
1752 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1753 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1754 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1755 NewTy);
1756
1757 return IRB.CreatePtrToInt(V, NewTy);
1758 }
1759
1760 return IRB.CreateBitCast(V, NewTy);
1761}
1762
1763/// Test whether the given slice use can be promoted to a vector.
1764///
1765/// This function is called to test each entry in a partition which is slated
1766/// for a single slice.
1767static bool isVectorPromotionViableForSlice(Partition &P, const Slice &S,
1768 VectorType *Ty,
1769 uint64_t ElementSize,
1770 const DataLayout &DL) {
1771 // First validate the slice offsets.
1772 uint64_t BeginOffset =
1773 std::max(S.beginOffset(), P.beginOffset()) - P.beginOffset();
1774 uint64_t BeginIndex = BeginOffset / ElementSize;
1775 if (BeginIndex * ElementSize != BeginOffset ||
1776 BeginIndex >= Ty->getNumElements())
1777 return false;
1778 uint64_t EndOffset =
1779 std::min(S.endOffset(), P.endOffset()) - P.beginOffset();
1780 uint64_t EndIndex = EndOffset / ElementSize;
1781 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1782 return false;
1783
1784 assert(EndIndex > BeginIndex && "Empty vector!")((EndIndex > BeginIndex && "Empty vector!") ? static_cast
<void> (0) : __assert_fail ("EndIndex > BeginIndex && \"Empty vector!\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1784, __PRETTY_FUNCTION__))
;
1785 uint64_t NumElements = EndIndex - BeginIndex;
1786 Type *SliceTy = (NumElements == 1)
1787 ? Ty->getElementType()
1788 : VectorType::get(Ty->getElementType(), NumElements);
1789
1790 Type *SplitIntTy =
1791 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1792
1793 Use *U = S.getUse();
1794
1795 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1796 if (MI->isVolatile())
1797 return false;
1798 if (!S.isSplittable())
1799 return false; // Skip any unsplittable intrinsics.
1800 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1801 if (!II->isLifetimeStartOrEnd())
1802 return false;
1803 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1804 // Disable vector promotion when there are loads or stores of an FCA.
1805 return false;
1806 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1807 if (LI->isVolatile())
1808 return false;
1809 Type *LTy = LI->getType();
1810 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1811 assert(LTy->isIntegerTy())((LTy->isIntegerTy()) ? static_cast<void> (0) : __assert_fail
("LTy->isIntegerTy()", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1811, __PRETTY_FUNCTION__))
;
1812 LTy = SplitIntTy;
1813 }
1814 if (!canConvertValue(DL, SliceTy, LTy))
1815 return false;
1816 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1817 if (SI->isVolatile())
1818 return false;
1819 Type *STy = SI->getValueOperand()->getType();
1820 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1821 assert(STy->isIntegerTy())((STy->isIntegerTy()) ? static_cast<void> (0) : __assert_fail
("STy->isIntegerTy()", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1821, __PRETTY_FUNCTION__))
;
1822 STy = SplitIntTy;
1823 }
1824 if (!canConvertValue(DL, STy, SliceTy))
1825 return false;
1826 } else {
1827 return false;
1828 }
1829
1830 return true;
1831}
1832
1833/// Test whether the given alloca partitioning and range of slices can be
1834/// promoted to a vector.
1835///
1836/// This is a quick test to check whether we can rewrite a particular alloca
1837/// partition (and its newly formed alloca) into a vector alloca with only
1838/// whole-vector loads and stores such that it could be promoted to a vector
1839/// SSA value. We only can ensure this for a limited set of operations, and we
1840/// don't want to do the rewrites unless we are confident that the result will
1841/// be promotable, so we have an early test here.
1842static VectorType *isVectorPromotionViable(Partition &P, const DataLayout &DL) {
1843 // Collect the candidate types for vector-based promotion. Also track whether
1844 // we have different element types.
1845 SmallVector<VectorType *, 4> CandidateTys;
1846 Type *CommonEltTy = nullptr;
1847 bool HaveCommonEltTy = true;
1848 auto CheckCandidateType = [&](Type *Ty) {
1849 if (auto *VTy = dyn_cast<VectorType>(Ty)) {
1850 CandidateTys.push_back(VTy);
1851 if (!CommonEltTy)
1852 CommonEltTy = VTy->getElementType();
1853 else if (CommonEltTy != VTy->getElementType())
1854 HaveCommonEltTy = false;
1855 }
1856 };
1857 // Consider any loads or stores that are the exact size of the slice.
1858 for (const Slice &S : P)
1859 if (S.beginOffset() == P.beginOffset() &&
1860 S.endOffset() == P.endOffset()) {
1861 if (auto *LI = dyn_cast<LoadInst>(S.getUse()->getUser()))
1862 CheckCandidateType(LI->getType());
1863 else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser()))
1864 CheckCandidateType(SI->getValueOperand()->getType());
1865 }
1866
1867 // If we didn't find a vector type, nothing to do here.
1868 if (CandidateTys.empty())
1869 return nullptr;
1870
1871 // Remove non-integer vector types if we had multiple common element types.
1872 // FIXME: It'd be nice to replace them with integer vector types, but we can't
1873 // do that until all the backends are known to produce good code for all
1874 // integer vector types.
1875 if (!HaveCommonEltTy) {
1876 CandidateTys.erase(
1877 llvm::remove_if(CandidateTys,
1878 [](VectorType *VTy) {
1879 return !VTy->getElementType()->isIntegerTy();
1880 }),
1881 CandidateTys.end());
1882
1883 // If there were no integer vector types, give up.
1884 if (CandidateTys.empty())
1885 return nullptr;
1886
1887 // Rank the remaining candidate vector types. This is easy because we know
1888 // they're all integer vectors. We sort by ascending number of elements.
1889 auto RankVectorTypes = [&DL](VectorType *RHSTy, VectorType *LHSTy) {
1890 (void)DL;
1891 assert(DL.getTypeSizeInBits(RHSTy) == DL.getTypeSizeInBits(LHSTy) &&((DL.getTypeSizeInBits(RHSTy) == DL.getTypeSizeInBits(LHSTy) &&
"Cannot have vector types of different sizes!") ? static_cast
<void> (0) : __assert_fail ("DL.getTypeSizeInBits(RHSTy) == DL.getTypeSizeInBits(LHSTy) && \"Cannot have vector types of different sizes!\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1892, __PRETTY_FUNCTION__))
1892 "Cannot have vector types of different sizes!")((DL.getTypeSizeInBits(RHSTy) == DL.getTypeSizeInBits(LHSTy) &&
"Cannot have vector types of different sizes!") ? static_cast
<void> (0) : __assert_fail ("DL.getTypeSizeInBits(RHSTy) == DL.getTypeSizeInBits(LHSTy) && \"Cannot have vector types of different sizes!\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1892, __PRETTY_FUNCTION__))
;
1893 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1894, __PRETTY_FUNCTION__))
1894 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1894, __PRETTY_FUNCTION__))
;
1895 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1896, __PRETTY_FUNCTION__))
1896 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1896, __PRETTY_FUNCTION__))
;
1897 return RHSTy->getNumElements() < LHSTy->getNumElements();
1898 };
1899 llvm::sort(CandidateTys, RankVectorTypes);
1900 CandidateTys.erase(
1901 std::unique(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes),
1902 CandidateTys.end());
1903 } else {
1904// The only way to have the same element type in every vector type is to
1905// have the same vector type. Check that and remove all but one.
1906#ifndef NDEBUG
1907 for (VectorType *VTy : CandidateTys) {
1908 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1909, __PRETTY_FUNCTION__))
1909 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1909, __PRETTY_FUNCTION__))
;
1910 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1911, __PRETTY_FUNCTION__))
1911 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1911, __PRETTY_FUNCTION__))
;
1912 }
1913#endif
1914 CandidateTys.resize(1);
1915 }
1916
1917 // Try each vector type, and return the one which works.
1918 auto CheckVectorTypeForPromotion = [&](VectorType *VTy) {
1919 uint64_t ElementSize = DL.getTypeSizeInBits(VTy->getElementType());
1920
1921 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1922 // that aren't byte sized.
1923 if (ElementSize % 8)
1924 return false;
1925 assert((DL.getTypeSizeInBits(VTy) % 8) == 0 &&(((DL.getTypeSizeInBits(VTy) % 8) == 0 && "vector size not a multiple of element size?"
) ? static_cast<void> (0) : __assert_fail ("(DL.getTypeSizeInBits(VTy) % 8) == 0 && \"vector size not a multiple of element size?\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1926, __PRETTY_FUNCTION__))
1926 "vector size not a multiple of element size?")(((DL.getTypeSizeInBits(VTy) % 8) == 0 && "vector size not a multiple of element size?"
) ? static_cast<void> (0) : __assert_fail ("(DL.getTypeSizeInBits(VTy) % 8) == 0 && \"vector size not a multiple of element size?\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 1926, __PRETTY_FUNCTION__))
;
1927 ElementSize /= 8;
1928
1929 for (const Slice &S : P)
1930 if (!isVectorPromotionViableForSlice(P, S, VTy, ElementSize, DL))
1931 return false;
1932
1933 for (const Slice *S : P.splitSliceTails())
1934 if (!isVectorPromotionViableForSlice(P, *S, VTy, ElementSize, DL))
1935 return false;
1936
1937 return true;
1938 };
1939 for (VectorType *VTy : CandidateTys)
1940 if (CheckVectorTypeForPromotion(VTy))
1941 return VTy;
1942
1943 return nullptr;
1944}
1945
1946/// Test whether a slice of an alloca is valid for integer widening.
1947///
1948/// This implements the necessary checking for the \c isIntegerWideningViable
1949/// test below on a single slice of the alloca.
1950static bool isIntegerWideningViableForSlice(const Slice &S,
1951 uint64_t AllocBeginOffset,
1952 Type *AllocaTy,
1953 const DataLayout &DL,
1954 bool &WholeAllocaOp) {
1955 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1956
1957 uint64_t RelBegin = S.beginOffset() - AllocBeginOffset;
1958 uint64_t RelEnd = S.endOffset() - AllocBeginOffset;
1959
1960 // We can't reasonably handle cases where the load or store extends past
1961 // the end of the alloca's type and into its padding.
1962 if (RelEnd > Size)
1963 return false;
1964
1965 Use *U = S.getUse();
1966
1967 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1968 if (LI->isVolatile())
1969 return false;
1970 // We can't handle loads that extend past the allocated memory.
1971 if (DL.getTypeStoreSize(LI->getType()) > Size)
1972 return false;
1973 // So far, AllocaSliceRewriter does not support widening split slice tails
1974 // in rewriteIntegerLoad.
1975 if (S.beginOffset() < AllocBeginOffset)
1976 return false;
1977 // Note that we don't count vector loads or stores as whole-alloca
1978 // operations which enable integer widening because we would prefer to use
1979 // vector widening instead.
1980 if (!isa<VectorType>(LI->getType()) && RelBegin == 0 && RelEnd == Size)
1981 WholeAllocaOp = true;
1982 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1983 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1984 return false;
1985 } else if (RelBegin != 0 || RelEnd != Size ||
1986 !canConvertValue(DL, AllocaTy, LI->getType())) {
1987 // Non-integer loads need to be convertible from the alloca type so that
1988 // they are promotable.
1989 return false;
1990 }
1991 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1992 Type *ValueTy = SI->getValueOperand()->getType();
1993 if (SI->isVolatile())
1994 return false;
1995 // We can't handle stores that extend past the allocated memory.
1996 if (DL.getTypeStoreSize(ValueTy) > Size)
1997 return false;
1998 // So far, AllocaSliceRewriter does not support widening split slice tails
1999 // in rewriteIntegerStore.
2000 if (S.beginOffset() < AllocBeginOffset)
2001 return false;
2002 // Note that we don't count vector loads or stores as whole-alloca
2003 // operations which enable integer widening because we would prefer to use
2004 // vector widening instead.
2005 if (!isa<VectorType>(ValueTy) && RelBegin == 0 && RelEnd == Size)
2006 WholeAllocaOp = true;
2007 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
2008 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
2009 return false;
2010 } else if (RelBegin != 0 || RelEnd != Size ||
2011 !canConvertValue(DL, ValueTy, AllocaTy)) {
2012 // Non-integer stores need to be convertible to the alloca type so that
2013 // they are promotable.
2014 return false;
2015 }
2016 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
2017 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
2018 return false;
2019 if (!S.isSplittable())
2020 return false; // Skip any unsplittable intrinsics.
2021 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
2022 if (!II->isLifetimeStartOrEnd())
2023 return false;
2024 } else {
2025 return false;
2026 }
2027
2028 return true;
2029}
2030
2031/// Test whether the given alloca partition's integer operations can be
2032/// widened to promotable ones.
2033///
2034/// This is a quick test to check whether we can rewrite the integer loads and
2035/// stores to a particular alloca into wider loads and stores and be able to
2036/// promote the resulting alloca.
2037static bool isIntegerWideningViable(Partition &P, Type *AllocaTy,
2038 const DataLayout &DL) {
2039 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
2040 // Don't create integer types larger than the maximum bitwidth.
2041 if (SizeInBits > IntegerType::MAX_INT_BITS)
2042 return false;
2043
2044 // Don't try to handle allocas with bit-padding.
2045 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
2046 return false;
2047
2048 // We need to ensure that an integer type with the appropriate bitwidth can
2049 // be converted to the alloca type, whatever that is. We don't want to force
2050 // the alloca itself to have an integer type if there is a more suitable one.
2051 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
2052 if (!canConvertValue(DL, AllocaTy, IntTy) ||
2053 !canConvertValue(DL, IntTy, AllocaTy))
2054 return false;
2055
2056 // While examining uses, we ensure that the alloca has a covering load or
2057 // store. We don't want to widen the integer operations only to fail to
2058 // promote due to some other unsplittable entry (which we may make splittable
2059 // later). However, if there are only splittable uses, go ahead and assume
2060 // that we cover the alloca.
2061 // FIXME: We shouldn't consider split slices that happen to start in the
2062 // partition here...
2063 bool WholeAllocaOp =
2064 P.begin() != P.end() ? false : DL.isLegalInteger(SizeInBits);
2065
2066 for (const Slice &S : P)
2067 if (!isIntegerWideningViableForSlice(S, P.beginOffset(), AllocaTy, DL,
2068 WholeAllocaOp))
2069 return false;
2070
2071 for (const Slice *S : P.splitSliceTails())
2072 if (!isIntegerWideningViableForSlice(*S, P.beginOffset(), AllocaTy, DL,
2073 WholeAllocaOp))
2074 return false;
2075
2076 return WholeAllocaOp;
2077}
2078
2079static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
2080 IntegerType *Ty, uint64_t Offset,
2081 const Twine &Name) {
2082 LLVM_DEBUG(dbgs() << " start: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " start: " << *V <<
"\n"; } } while (false)
;
30
Assuming 'DebugFlag' is 0
31
Loop condition is false. Exiting loop
2083 IntegerType *IntTy = cast<IntegerType>(V->getType());
2084 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&((DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(
IntTy) && "Element extends past full value") ? static_cast
<void> (0) : __assert_fail ("DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) && \"Element extends past full value\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2085, __PRETTY_FUNCTION__))
32
Assuming the condition is true
33
'?' condition is true
2085 "Element extends past full value")((DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(
IntTy) && "Element extends past full value") ? static_cast
<void> (0) : __assert_fail ("DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) && \"Element extends past full value\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2085, __PRETTY_FUNCTION__))
;
2086 uint64_t ShAmt = 8 * Offset;
2087 if (DL.isBigEndian())
34
Assuming the condition is false
35
Taking false branch
2088 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2089 if (ShAmt) {
36
Assuming 'ShAmt' is 0
37
Taking false branch
2090 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2091 LLVM_DEBUG(dbgs() << " shifted: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " shifted: " << *V <<
"\n"; } } while (false)
;
2092 }
2093 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2094, __PRETTY_FUNCTION__))
38
Called C++ object pointer is null
2094 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2094, __PRETTY_FUNCTION__))
;
2095 if (Ty != IntTy) {
2096 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2097 LLVM_DEBUG(dbgs() << " trunced: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " trunced: " << *V <<
"\n"; } } while (false)
;
2098 }
2099 return V;
2100}
2101
2102static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
2103 Value *V, uint64_t Offset, const Twine &Name) {
2104 IntegerType *IntTy = cast<IntegerType>(Old->getType());
2105 IntegerType *Ty = cast<IntegerType>(V->getType());
2106 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2107, __PRETTY_FUNCTION__))
2107 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2107, __PRETTY_FUNCTION__))
;
2108 LLVM_DEBUG(dbgs() << " start: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " start: " << *V <<
"\n"; } } while (false)
;
2109 if (Ty != IntTy) {
2110 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2111 LLVM_DEBUG(dbgs() << " extended: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " extended: " << *V <<
"\n"; } } while (false)
;
2112 }
2113 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&((DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(
IntTy) && "Element store outside of alloca store") ? static_cast
<void> (0) : __assert_fail ("DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) && \"Element store outside of alloca store\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2114, __PRETTY_FUNCTION__))
2114 "Element store outside of alloca store")((DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(
IntTy) && "Element store outside of alloca store") ? static_cast
<void> (0) : __assert_fail ("DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) && \"Element store outside of alloca store\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2114, __PRETTY_FUNCTION__))
;
2115 uint64_t ShAmt = 8 * Offset;
2116 if (DL.isBigEndian())
2117 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2118 if (ShAmt) {
2119 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2120 LLVM_DEBUG(dbgs() << " shifted: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " shifted: " << *V <<
"\n"; } } while (false)
;
2121 }
2122
2123 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2124 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2125 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2126 LLVM_DEBUG(dbgs() << " masked: " << *Old << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " masked: " << *Old <<
"\n"; } } while (false)
;
2127 V = IRB.CreateOr(Old, V, Name + ".insert");
2128 LLVM_DEBUG(dbgs() << " inserted: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " inserted: " << *V <<
"\n"; } } while (false)
;
2129 }
2130 return V;
2131}
2132
2133static Value *extractVector(IRBuilderTy &IRB, Value *V, unsigned BeginIndex,
2134 unsigned EndIndex, const Twine &Name) {
2135 VectorType *VecTy = cast<VectorType>(V->getType());
2136 unsigned NumElements = EndIndex - BeginIndex;
2137 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2137, __PRETTY_FUNCTION__))
;
2138
2139 if (NumElements == VecTy->getNumElements())
2140 return V;
2141
2142 if (NumElements == 1) {
2143 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2144 Name + ".extract");
2145 LLVM_DEBUG(dbgs() << " extract: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " extract: " << *V <<
"\n"; } } while (false)
;
2146 return V;
2147 }
2148
2149 SmallVector<Constant *, 8> Mask;
2150 Mask.reserve(NumElements);
2151 for (unsigned i = BeginIndex; i != EndIndex; ++i)
2152 Mask.push_back(IRB.getInt32(i));
2153 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2154 ConstantVector::get(Mask), Name + ".extract");
2155 LLVM_DEBUG(dbgs() << " shuffle: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " shuffle: " << *V <<
"\n"; } } while (false)
;
2156 return V;
2157}
2158
2159static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
2160 unsigned BeginIndex, const Twine &Name) {
2161 VectorType *VecTy = cast<VectorType>(Old->getType());
2162 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2162, __PRETTY_FUNCTION__))
;
2163
2164 VectorType *Ty = dyn_cast<VectorType>(V->getType());
2165 if (!Ty) {
2166 // Single element to insert.
2167 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2168 Name + ".insert");
2169 LLVM_DEBUG(dbgs() << " insert: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " insert: " << *V <<
"\n"; } } while (false)
;
2170 return V;
2171 }
2172
2173 assert(Ty->getNumElements() <= VecTy->getNumElements() &&((Ty->getNumElements() <= VecTy->getNumElements() &&
"Too many elements!") ? static_cast<void> (0) : __assert_fail
("Ty->getNumElements() <= VecTy->getNumElements() && \"Too many elements!\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2174, __PRETTY_FUNCTION__))
2174 "Too many elements!")((Ty->getNumElements() <= VecTy->getNumElements() &&
"Too many elements!") ? static_cast<void> (0) : __assert_fail
("Ty->getNumElements() <= VecTy->getNumElements() && \"Too many elements!\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2174, __PRETTY_FUNCTION__))
;
2175 if (Ty->getNumElements() == VecTy->getNumElements()) {
2176 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2176, __PRETTY_FUNCTION__))
;
2177 return V;
2178 }
2179 unsigned EndIndex = BeginIndex + Ty->getNumElements();
2180
2181 // When inserting a smaller vector into the larger to store, we first
2182 // use a shuffle vector to widen it with undef elements, and then
2183 // a second shuffle vector to select between the loaded vector and the
2184 // incoming vector.
2185 SmallVector<Constant *, 8> Mask;
2186 Mask.reserve(VecTy->getNumElements());
2187 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2188 if (i >= BeginIndex && i < EndIndex)
2189 Mask.push_back(IRB.getInt32(i - BeginIndex));
2190 else
2191 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2192 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2193 ConstantVector::get(Mask), Name + ".expand");
2194 LLVM_DEBUG(dbgs() << " shuffle: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " shuffle: " << *V <<
"\n"; } } while (false)
;
2195
2196 Mask.clear();
2197 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2198 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
2199
2200 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
2201
2202 LLVM_DEBUG(dbgs() << " blend: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " blend: " << *V <<
"\n"; } } while (false)
;
2203 return V;
2204}
2205
2206/// Visitor to rewrite instructions using p particular slice of an alloca
2207/// to use a new alloca.
2208///
2209/// Also implements the rewriting to vector-based accesses when the partition
2210/// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2211/// lives here.
2212class llvm::sroa::AllocaSliceRewriter
2213 : public InstVisitor<AllocaSliceRewriter, bool> {
2214 // Befriend the base class so it can delegate to private visit methods.
2215 friend class InstVisitor<AllocaSliceRewriter, bool>;
2216
2217 using Base = InstVisitor<AllocaSliceRewriter, bool>;
2218
2219 const DataLayout &DL;
2220 AllocaSlices &AS;
2221 SROA &Pass;
2222 AllocaInst &OldAI, &NewAI;
2223 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2224 Type *NewAllocaTy;
2225
2226 // This is a convenience and flag variable that will be null unless the new
2227 // alloca's integer operations should be widened to this integer type due to
2228 // passing isIntegerWideningViable above. If it is non-null, the desired
2229 // integer type will be stored here for easy access during rewriting.
2230 IntegerType *IntTy;
2231
2232 // If we are rewriting an alloca partition which can be written as pure
2233 // vector operations, we stash extra information here. When VecTy is
2234 // non-null, we have some strict guarantees about the rewritten alloca:
2235 // - The new alloca is exactly the size of the vector type here.
2236 // - The accesses all either map to the entire vector or to a single
2237 // element.
2238 // - The set of accessing instructions is only one of those handled above
2239 // in isVectorPromotionViable. Generally these are the same access kinds
2240 // which are promotable via mem2reg.
2241 VectorType *VecTy;
2242 Type *ElementTy;
2243 uint64_t ElementSize;
2244
2245 // The original offset of the slice currently being rewritten relative to
2246 // the original alloca.
2247 uint64_t BeginOffset = 0;
2248 uint64_t EndOffset = 0;
2249
2250 // The new offsets of the slice currently being rewritten relative to the
2251 // original alloca.
2252 uint64_t NewBeginOffset, NewEndOffset;
2253
2254 uint64_t SliceSize;
2255 bool IsSplittable = false;
2256 bool IsSplit = false;
2257 Use *OldUse = nullptr;
2258 Instruction *OldPtr = nullptr;
2259
2260 // Track post-rewrite users which are PHI nodes and Selects.
2261 SmallSetVector<PHINode *, 8> &PHIUsers;
2262 SmallSetVector<SelectInst *, 8> &SelectUsers;
2263
2264 // Utility IR builder, whose name prefix is setup for each visited use, and
2265 // the insertion point is set to point to the user.
2266 IRBuilderTy IRB;
2267
2268public:
2269 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &AS, SROA &Pass,
2270 AllocaInst &OldAI, AllocaInst &NewAI,
2271 uint64_t NewAllocaBeginOffset,
2272 uint64_t NewAllocaEndOffset, bool IsIntegerPromotable,
2273 VectorType *PromotableVecTy,
2274 SmallSetVector<PHINode *, 8> &PHIUsers,
2275 SmallSetVector<SelectInst *, 8> &SelectUsers)
2276 : DL(DL), AS(AS), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2277 NewAllocaBeginOffset(NewAllocaBeginOffset),
2278 NewAllocaEndOffset(NewAllocaEndOffset),
2279 NewAllocaTy(NewAI.getAllocatedType()),
2280 IntTy(IsIntegerPromotable
2281 ? Type::getIntNTy(
2282 NewAI.getContext(),
2283 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2284 : nullptr),
2285 VecTy(PromotableVecTy),
2286 ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2287 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
2288 PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2289 IRB(NewAI.getContext(), ConstantFolder()) {
2290 if (VecTy) {
2291 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&(((DL.getTypeSizeInBits(ElementTy) % 8) == 0 && "Only multiple-of-8 sized vector elements are viable"
) ? static_cast<void> (0) : __assert_fail ("(DL.getTypeSizeInBits(ElementTy) % 8) == 0 && \"Only multiple-of-8 sized vector elements are viable\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2292, __PRETTY_FUNCTION__))
2292 "Only multiple-of-8 sized vector elements are viable")(((DL.getTypeSizeInBits(ElementTy) % 8) == 0 && "Only multiple-of-8 sized vector elements are viable"
) ? static_cast<void> (0) : __assert_fail ("(DL.getTypeSizeInBits(ElementTy) % 8) == 0 && \"Only multiple-of-8 sized vector elements are viable\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2292, __PRETTY_FUNCTION__))
;
2293 ++NumVectorized;
2294 }
2295 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2295, __PRETTY_FUNCTION__))
;
2296 }
2297
2298 bool visit(AllocaSlices::const_iterator I) {
2299 bool CanSROA = true;
2300 BeginOffset = I->beginOffset();
2301 EndOffset = I->endOffset();
2302 IsSplittable = I->isSplittable();
2303 IsSplit =
2304 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2305 LLVM_DEBUG(dbgs() << " rewriting " << (IsSplit ? "split " : ""))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " rewriting " << (IsSplit ?
"split " : ""); } } while (false)
;
2306 LLVM_DEBUG(AS.printSlice(dbgs(), I, ""))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { AS.printSlice(dbgs(), I, ""); } } while (false)
;
2307 LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "\n"; } } while (false)
;
2308
2309 // Compute the intersecting offset range.
2310 assert(BeginOffset < NewAllocaEndOffset)((BeginOffset < NewAllocaEndOffset) ? static_cast<void>
(0) : __assert_fail ("BeginOffset < NewAllocaEndOffset", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2310, __PRETTY_FUNCTION__))
;
2311 assert(EndOffset > NewAllocaBeginOffset)((EndOffset > NewAllocaBeginOffset) ? static_cast<void>
(0) : __assert_fail ("EndOffset > NewAllocaBeginOffset", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2311, __PRETTY_FUNCTION__))
;
2312 NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2313 NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2314
2315 SliceSize = NewEndOffset - NewBeginOffset;
2316
2317 OldUse = I->getUse();
2318 OldPtr = cast<Instruction>(OldUse->get());
2319
2320 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2321 IRB.SetInsertPoint(OldUserI);
2322 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2323 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2324
2325 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2326 if (VecTy || IntTy)
2327 assert(CanSROA)((CanSROA) ? static_cast<void> (0) : __assert_fail ("CanSROA"
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2327, __PRETTY_FUNCTION__))
;
2328 return CanSROA;
2329 }
2330
2331private:
2332 // Make sure the other visit overloads are visible.
2333 using Base::visit;
2334
2335 // Every instruction which can end up as a user must have a rewrite rule.
2336 bool visitInstruction(Instruction &I) {
2337 LLVM_DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " !!!! Cannot rewrite: " <<
I << "\n"; } } while (false)
;
2338 llvm_unreachable("No rewrite rule for this instruction!")::llvm::llvm_unreachable_internal("No rewrite rule for this instruction!"
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2338)
;
2339 }
2340
2341 Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2342 // Note that the offset computation can use BeginOffset or NewBeginOffset
2343 // interchangeably for unsplit slices.
2344 assert(IsSplit || BeginOffset == NewBeginOffset)((IsSplit || BeginOffset == NewBeginOffset) ? static_cast<
void> (0) : __assert_fail ("IsSplit || BeginOffset == NewBeginOffset"
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2344, __PRETTY_FUNCTION__))
;
2345 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2346
2347#ifndef NDEBUG
2348 StringRef OldName = OldPtr->getName();
2349 // Skip through the last '.sroa.' component of the name.
2350 size_t LastSROAPrefix = OldName.rfind(".sroa.");
2351 if (LastSROAPrefix != StringRef::npos) {
2352 OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2353 // Look for an SROA slice index.
2354 size_t IndexEnd = OldName.find_first_not_of("0123456789");
2355 if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2356 // Strip the index and look for the offset.
2357 OldName = OldName.substr(IndexEnd + 1);
2358 size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2359 if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2360 // Strip the offset.
2361 OldName = OldName.substr(OffsetEnd + 1);
2362 }
2363 }
2364 // Strip any SROA suffixes as well.
2365 OldName = OldName.substr(0, OldName.find(".sroa_"));
2366#endif
2367
2368 return getAdjustedPtr(IRB, DL, &NewAI,
2369 APInt(DL.getIndexTypeSizeInBits(PointerTy), Offset),
2370 PointerTy,
2371#ifndef NDEBUG
2372 Twine(OldName) + "."
2373#else
2374 Twine()
2375#endif
2376 );
2377 }
2378
2379 /// Compute suitable alignment to access this slice of the *new*
2380 /// alloca.
2381 ///
2382 /// You can optionally pass a type to this routine and if that type's ABI
2383 /// alignment is itself suitable, this will return zero.
2384 unsigned getSliceAlign(Type *Ty = nullptr) {
2385 unsigned NewAIAlign = NewAI.getAlignment();
2386 if (!NewAIAlign)
2387 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2388 unsigned Align =
2389 MinAlign(NewAIAlign, NewBeginOffset - NewAllocaBeginOffset);
2390 return (Ty && Align == DL.getABITypeAlignment(Ty)) ? 0 : Align;
2391 }
2392
2393 unsigned getIndex(uint64_t Offset) {
2394 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2394, __PRETTY_FUNCTION__))
;
2395 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2396 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2396, __PRETTY_FUNCTION__))
;
2397 uint32_t Index = RelOffset / ElementSize;
2398 assert(Index * ElementSize == RelOffset)((Index * ElementSize == RelOffset) ? static_cast<void>
(0) : __assert_fail ("Index * ElementSize == RelOffset", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2398, __PRETTY_FUNCTION__))
;
2399 return Index;
2400 }
2401
2402 void deleteIfTriviallyDead(Value *V) {
2403 Instruction *I = cast<Instruction>(V);
2404 if (isInstructionTriviallyDead(I))
2405 Pass.DeadInsts.insert(I);
2406 }
2407
2408 Value *rewriteVectorizedLoadInst() {
2409 unsigned BeginIndex = getIndex(NewBeginOffset);
2410 unsigned EndIndex = getIndex(NewEndOffset);
2411 assert(EndIndex > BeginIndex && "Empty vector!")((EndIndex > BeginIndex && "Empty vector!") ? static_cast
<void> (0) : __assert_fail ("EndIndex > BeginIndex && \"Empty vector!\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2411, __PRETTY_FUNCTION__))
;
2412
2413 Value *V = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2414 NewAI.getAlignment(), "load");
2415 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2416 }
2417
2418 Value *rewriteIntegerLoad(LoadInst &LI) {
2419 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2419, __PRETTY_FUNCTION__))
;
2420 assert(!LI.isVolatile())((!LI.isVolatile()) ? static_cast<void> (0) : __assert_fail
("!LI.isVolatile()", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2420, __PRETTY_FUNCTION__))
;
2421 Value *V = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2422 NewAI.getAlignment(), "load");
2423 V = convertValue(DL, IRB, V, IntTy);
2424 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2424, __PRETTY_FUNCTION__))
;
2425 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2426 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset) {
2427 IntegerType *ExtractTy = Type::getIntNTy(LI.getContext(), SliceSize * 8);
2428 V = extractInteger(DL, IRB, V, ExtractTy, Offset, "extract");
2429 }
2430 // It is possible that the extracted type is not the load type. This
2431 // happens if there is a load past the end of the alloca, and as
2432 // a consequence the slice is narrower but still a candidate for integer
2433 // lowering. To handle this case, we just zero extend the extracted
2434 // integer.
2435 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2436, __PRETTY_FUNCTION__))
2436 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2436, __PRETTY_FUNCTION__))
;
2437 if (cast<IntegerType>(LI.getType())->getBitWidth() > SliceSize * 8)
2438 V = IRB.CreateZExt(V, LI.getType());
2439 return V;
2440 }
2441
2442 bool visitLoadInst(LoadInst &LI) {
2443 LLVM_DEBUG(dbgs() << " original: " << LI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << LI <<
"\n"; } } while (false)
;
2444 Value *OldOp = LI.getOperand(0);
2445 assert(OldOp == OldPtr)((OldOp == OldPtr) ? static_cast<void> (0) : __assert_fail
("OldOp == OldPtr", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2445, __PRETTY_FUNCTION__))
;
2446
2447 AAMDNodes AATags;
2448 LI.getAAMetadata(AATags);
2449
2450 unsigned AS = LI.getPointerAddressSpace();
2451
2452 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2453 : LI.getType();
2454 const bool IsLoadPastEnd = DL.getTypeStoreSize(TargetTy) > SliceSize;
2455 bool IsPtrAdjusted = false;
2456 Value *V;
2457 if (VecTy) {
2458 V = rewriteVectorizedLoadInst();
2459 } else if (IntTy && LI.getType()->isIntegerTy()) {
2460 V = rewriteIntegerLoad(LI);
2461 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2462 NewEndOffset == NewAllocaEndOffset &&
2463 (canConvertValue(DL, NewAllocaTy, TargetTy) ||
2464 (IsLoadPastEnd && NewAllocaTy->isIntegerTy() &&
2465 TargetTy->isIntegerTy()))) {
2466 LoadInst *NewLI = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2467 NewAI.getAlignment(),
2468 LI.isVolatile(), LI.getName());
2469 if (AATags)
2470 NewLI->setAAMetadata(AATags);
2471 if (LI.isVolatile())
2472 NewLI->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
2473
2474 // Any !nonnull metadata or !range metadata on the old load is also valid
2475 // on the new load. This is even true in some cases even when the loads
2476 // are different types, for example by mapping !nonnull metadata to
2477 // !range metadata by modeling the null pointer constant converted to the
2478 // integer type.
2479 // FIXME: Add support for range metadata here. Currently the utilities
2480 // for this don't propagate range metadata in trivial cases from one
2481 // integer load to another, don't handle non-addrspace-0 null pointers
2482 // correctly, and don't have any support for mapping ranges as the
2483 // integer type becomes winder or narrower.
2484 if (MDNode *N = LI.getMetadata(LLVMContext::MD_nonnull))
2485 copyNonnullMetadata(LI, N, *NewLI);
2486
2487 // Try to preserve nonnull metadata
2488 V = NewLI;
2489
2490 // If this is an integer load past the end of the slice (which means the
2491 // bytes outside the slice are undef or this load is dead) just forcibly
2492 // fix the integer size with correct handling of endianness.
2493 if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2494 if (auto *TITy = dyn_cast<IntegerType>(TargetTy))
2495 if (AITy->getBitWidth() < TITy->getBitWidth()) {
2496 V = IRB.CreateZExt(V, TITy, "load.ext");
2497 if (DL.isBigEndian())
2498 V = IRB.CreateShl(V, TITy->getBitWidth() - AITy->getBitWidth(),
2499 "endian_shift");
2500 }
2501 } else {
2502 Type *LTy = TargetTy->getPointerTo(AS);
2503 LoadInst *NewLI = IRB.CreateAlignedLoad(
2504 TargetTy, getNewAllocaSlicePtr(IRB, LTy), getSliceAlign(TargetTy),
2505 LI.isVolatile(), LI.getName());
2506 if (AATags)
2507 NewLI->setAAMetadata(AATags);
2508 if (LI.isVolatile())
2509 NewLI->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
2510
2511 V = NewLI;
2512 IsPtrAdjusted = true;
2513 }
2514 V = convertValue(DL, IRB, V, TargetTy);
2515
2516 if (IsSplit) {
2517 assert(!LI.isVolatile())((!LI.isVolatile()) ? static_cast<void> (0) : __assert_fail
("!LI.isVolatile()", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2517, __PRETTY_FUNCTION__))
;
2518 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2519, __PRETTY_FUNCTION__))
2519 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2519, __PRETTY_FUNCTION__))
;
2520 assert(SliceSize < DL.getTypeStoreSize(LI.getType()) &&((SliceSize < DL.getTypeStoreSize(LI.getType()) &&
"Split load isn't smaller than original load") ? static_cast
<void> (0) : __assert_fail ("SliceSize < DL.getTypeStoreSize(LI.getType()) && \"Split load isn't smaller than original load\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2521, __PRETTY_FUNCTION__))
2521 "Split load isn't smaller than original load")((SliceSize < DL.getTypeStoreSize(LI.getType()) &&
"Split load isn't smaller than original load") ? static_cast
<void> (0) : __assert_fail ("SliceSize < DL.getTypeStoreSize(LI.getType()) && \"Split load isn't smaller than original load\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2521, __PRETTY_FUNCTION__))
;
2522 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2523, __PRETTY_FUNCTION__))
2523 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2523, __PRETTY_FUNCTION__))
;
2524 // Move the insertion point just past the load so that we can refer to it.
2525 IRB.SetInsertPoint(&*std::next(BasicBlock::iterator(&LI)));
2526 // Create a placeholder value with the same type as LI to use as the
2527 // basis for the new value. This allows us to replace the uses of LI with
2528 // the computed value, and then replace the placeholder with LI, leaving
2529 // LI only used for this computation.
2530 Value *Placeholder = new LoadInst(
2531 LI.getType(), UndefValue::get(LI.getType()->getPointerTo(AS)));
2532 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset - BeginOffset,
2533 "insert");
2534 LI.replaceAllUsesWith(V);
2535 Placeholder->replaceAllUsesWith(&LI);
2536 Placeholder->deleteValue();
2537 } else {
2538 LI.replaceAllUsesWith(V);
2539 }
2540
2541 Pass.DeadInsts.insert(&LI);
2542 deleteIfTriviallyDead(OldOp);
2543 LLVM_DEBUG(dbgs() << " to: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *V <<
"\n"; } } while (false)
;
2544 return !LI.isVolatile() && !IsPtrAdjusted;
2545 }
2546
2547 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp,
2548 AAMDNodes AATags) {
2549 if (V->getType() != VecTy) {
2550 unsigned BeginIndex = getIndex(NewBeginOffset);
2551 unsigned EndIndex = getIndex(NewEndOffset);
2552 assert(EndIndex > BeginIndex && "Empty vector!")((EndIndex > BeginIndex && "Empty vector!") ? static_cast
<void> (0) : __assert_fail ("EndIndex > BeginIndex && \"Empty vector!\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2552, __PRETTY_FUNCTION__))
;
2553 unsigned NumElements = EndIndex - BeginIndex;
2554 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2554, __PRETTY_FUNCTION__))
;
2555 Type *SliceTy = (NumElements == 1)
2556 ? ElementTy
2557 : VectorType::get(ElementTy, NumElements);
2558 if (V->getType() != SliceTy)
2559 V = convertValue(DL, IRB, V, SliceTy);
2560
2561 // Mix in the existing elements.
2562 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2563 NewAI.getAlignment(), "load");
2564 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2565 }
2566 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2567 if (AATags)
2568 Store->setAAMetadata(AATags);
2569 Pass.DeadInsts.insert(&SI);
2570
2571 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *Store <<
"\n"; } } while (false)
;
2572 return true;
2573 }
2574
2575 bool rewriteIntegerStore(Value *V, StoreInst &SI, AAMDNodes AATags) {
2576 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2576, __PRETTY_FUNCTION__))
;
2577 assert(!SI.isVolatile())((!SI.isVolatile()) ? static_cast<void> (0) : __assert_fail
("!SI.isVolatile()", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2577, __PRETTY_FUNCTION__))
;
2578 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2579 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2580 NewAI.getAlignment(), "oldload");
2581 Old = convertValue(DL, IRB, Old, IntTy);
2582 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2582, __PRETTY_FUNCTION__))
;
2583 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2584 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset, "insert");
2585 }
2586 V = convertValue(DL, IRB, V, NewAllocaTy);
2587 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2588 Store->copyMetadata(SI, {LLVMContext::MD_mem_parallel_loop_access,
2589 LLVMContext::MD_access_group});
2590 if (AATags)
2591 Store->setAAMetadata(AATags);
2592 Pass.DeadInsts.insert(&SI);
2593 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *Store <<
"\n"; } } while (false)
;
2594 return true;
2595 }
2596
2597 bool visitStoreInst(StoreInst &SI) {
2598 LLVM_DEBUG(dbgs() << " original: " << SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << SI <<
"\n"; } } while (false)
;
2599 Value *OldOp = SI.getOperand(1);
2600 assert(OldOp == OldPtr)((OldOp == OldPtr) ? static_cast<void> (0) : __assert_fail
("OldOp == OldPtr", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2600, __PRETTY_FUNCTION__))
;
2601
2602 AAMDNodes AATags;
2603 SI.getAAMetadata(AATags);
2604
2605 Value *V = SI.getValueOperand();
2606
2607 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2608 // alloca that should be re-examined after promoting this alloca.
2609 if (V->getType()->isPointerTy())
2610 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2611 Pass.PostPromotionWorklist.insert(AI);
2612
2613 if (SliceSize < DL.getTypeStoreSize(V->getType())) {
2614 assert(!SI.isVolatile())((!SI.isVolatile()) ? static_cast<void> (0) : __assert_fail
("!SI.isVolatile()", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2614, __PRETTY_FUNCTION__))
;
2615 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2616, __PRETTY_FUNCTION__))
2616 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2616, __PRETTY_FUNCTION__))
;
2617 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2618, __PRETTY_FUNCTION__))
2618 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2618, __PRETTY_FUNCTION__))
;
2619 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2620 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset - BeginOffset,
2621 "extract");
2622 }
2623
2624 if (VecTy)
2625 return rewriteVectorizedStoreInst(V, SI, OldOp, AATags);
2626 if (IntTy && V->getType()->isIntegerTy())
2627 return rewriteIntegerStore(V, SI, AATags);
2628
2629 const bool IsStorePastEnd = DL.getTypeStoreSize(V->getType()) > SliceSize;
2630 StoreInst *NewSI;
2631 if (NewBeginOffset == NewAllocaBeginOffset &&
2632 NewEndOffset == NewAllocaEndOffset &&
2633 (canConvertValue(DL, V->getType(), NewAllocaTy) ||
2634 (IsStorePastEnd && NewAllocaTy->isIntegerTy() &&
2635 V->getType()->isIntegerTy()))) {
2636 // If this is an integer store past the end of slice (and thus the bytes
2637 // past that point are irrelevant or this is unreachable), truncate the
2638 // value prior to storing.
2639 if (auto *VITy = dyn_cast<IntegerType>(V->getType()))
2640 if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2641 if (VITy->getBitWidth() > AITy->getBitWidth()) {
2642 if (DL.isBigEndian())
2643 V = IRB.CreateLShr(V, VITy->getBitWidth() - AITy->getBitWidth(),
2644 "endian_shift");
2645 V = IRB.CreateTrunc(V, AITy, "load.trunc");
2646 }
2647
2648 V = convertValue(DL, IRB, V, NewAllocaTy);
2649 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2650 SI.isVolatile());
2651 } else {
2652 unsigned AS = SI.getPointerAddressSpace();
2653 Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo(AS));
2654 NewSI = IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(V->getType()),
2655 SI.isVolatile());
2656 }
2657 NewSI->copyMetadata(SI, {LLVMContext::MD_mem_parallel_loop_access,
2658 LLVMContext::MD_access_group});
2659 if (AATags)
2660 NewSI->setAAMetadata(AATags);
2661 if (SI.isVolatile())
2662 NewSI->setAtomic(SI.getOrdering(), SI.getSyncScopeID());
2663 Pass.DeadInsts.insert(&SI);
2664 deleteIfTriviallyDead(OldOp);
2665
2666 LLVM_DEBUG(dbgs() << " to: " << *NewSI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *NewSI <<
"\n"; } } while (false)
;
2667 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2668 }
2669
2670 /// Compute an integer value from splatting an i8 across the given
2671 /// number of bytes.
2672 ///
2673 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2674 /// call this routine.
2675 /// FIXME: Heed the advice above.
2676 ///
2677 /// \param V The i8 value to splat.
2678 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2679 Value *getIntegerSplat(Value *V, unsigned Size) {
2680 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2680, __PRETTY_FUNCTION__))
;
2681 IntegerType *VTy = cast<IntegerType>(V->getType());
2682 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2682, __PRETTY_FUNCTION__))
;
2683 if (Size == 1)
2684 return V;
2685
2686 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size * 8);
2687 V = IRB.CreateMul(
2688 IRB.CreateZExt(V, SplatIntTy, "zext"),
2689 ConstantExpr::getUDiv(
2690 Constant::getAllOnesValue(SplatIntTy),
2691 ConstantExpr::getZExt(Constant::getAllOnesValue(V->getType()),
2692 SplatIntTy)),
2693 "isplat");
2694 return V;
2695 }
2696
2697 /// Compute a vector splat for a given element value.
2698 Value *getVectorSplat(Value *V, unsigned NumElements) {
2699 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2700 LLVM_DEBUG(dbgs() << " splat: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " splat: " << *V <<
"\n"; } } while (false)
;
2701 return V;
2702 }
2703
2704 bool visitMemSetInst(MemSetInst &II) {
2705 LLVM_DEBUG(dbgs() << " original: " << II << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << II <<
"\n"; } } while (false)
;
2706 assert(II.getRawDest() == OldPtr)((II.getRawDest() == OldPtr) ? static_cast<void> (0) : __assert_fail
("II.getRawDest() == OldPtr", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2706, __PRETTY_FUNCTION__))
;
2707
2708 AAMDNodes AATags;
2709 II.getAAMetadata(AATags);
2710
2711 // If the memset has a variable size, it cannot be split, just adjust the
2712 // pointer to the new alloca.
2713 if (!isa<Constant>(II.getLength())) {
2714 assert(!IsSplit)((!IsSplit) ? static_cast<void> (0) : __assert_fail ("!IsSplit"
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2714, __PRETTY_FUNCTION__))
;
2715 assert(NewBeginOffset == BeginOffset)((NewBeginOffset == BeginOffset) ? static_cast<void> (0
) : __assert_fail ("NewBeginOffset == BeginOffset", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2715, __PRETTY_FUNCTION__))
;
2716 II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2717 II.setDestAlignment(getSliceAlign());
2718
2719 deleteIfTriviallyDead(OldPtr);
2720 return false;
2721 }
2722
2723 // Record this instruction for deletion.
2724 Pass.DeadInsts.insert(&II);
2725
2726 Type *AllocaTy = NewAI.getAllocatedType();
2727 Type *ScalarTy = AllocaTy->getScalarType();
2728
2729 const bool CanContinue = [&]() {
2730 if (VecTy || IntTy)
2731 return true;
2732 if (BeginOffset > NewAllocaBeginOffset ||
2733 EndOffset < NewAllocaEndOffset)
2734 return false;
2735 auto *C = cast<ConstantInt>(II.getLength());
2736 if (C->getBitWidth() > 64)
2737 return false;
2738 const auto Len = C->getZExtValue();
2739 auto *Int8Ty = IntegerType::getInt8Ty(NewAI.getContext());
2740 auto *SrcTy = VectorType::get(Int8Ty, Len);
2741 return canConvertValue(DL, SrcTy, AllocaTy) &&
2742 DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy));
2743 }();
2744
2745 // If this doesn't map cleanly onto the alloca type, and that type isn't
2746 // a single value type, just emit a memset.
2747 if (!CanContinue) {
2748 Type *SizeTy = II.getLength()->getType();
2749 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2750 CallInst *New = IRB.CreateMemSet(
2751 getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2752 getSliceAlign(), II.isVolatile());
2753 if (AATags)
2754 New->setAAMetadata(AATags);
2755 LLVM_DEBUG(dbgs() << " to: " << *New << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *New <<
"\n"; } } while (false)
;
2756 return false;
2757 }
2758
2759 // If we can represent this as a simple value, we have to build the actual
2760 // value to store, which requires expanding the byte present in memset to
2761 // a sensible representation for the alloca type. This is essentially
2762 // splatting the byte to a sufficiently wide integer, splatting it across
2763 // any desired vector width, and bitcasting to the final type.
2764 Value *V;
2765
2766 if (VecTy) {
2767 // If this is a memset of a vectorized alloca, insert it.
2768 assert(ElementTy == ScalarTy)((ElementTy == ScalarTy) ? static_cast<void> (0) : __assert_fail
("ElementTy == ScalarTy", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2768, __PRETTY_FUNCTION__))
;
2769
2770 unsigned BeginIndex = getIndex(NewBeginOffset);
2771 unsigned EndIndex = getIndex(NewEndOffset);
2772 assert(EndIndex > BeginIndex && "Empty vector!")((EndIndex > BeginIndex && "Empty vector!") ? static_cast
<void> (0) : __assert_fail ("EndIndex > BeginIndex && \"Empty vector!\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2772, __PRETTY_FUNCTION__))
;
2773 unsigned NumElements = EndIndex - BeginIndex;
2774 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2774, __PRETTY_FUNCTION__))
;
2775
2776 Value *Splat =
2777 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2778 Splat = convertValue(DL, IRB, Splat, ElementTy);
2779 if (NumElements > 1)
2780 Splat = getVectorSplat(Splat, NumElements);
2781
2782 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2783 NewAI.getAlignment(), "oldload");
2784 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2785 } else if (IntTy) {
2786 // If this is a memset on an alloca where we can widen stores, insert the
2787 // set integer.
2788 assert(!II.isVolatile())((!II.isVolatile()) ? static_cast<void> (0) : __assert_fail
("!II.isVolatile()", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2788, __PRETTY_FUNCTION__))
;
2789
2790 uint64_t Size = NewEndOffset - NewBeginOffset;
2791 V = getIntegerSplat(II.getValue(), Size);
2792
2793 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2794 EndOffset != NewAllocaBeginOffset)) {
2795 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2796 NewAI.getAlignment(), "oldload");
2797 Old = convertValue(DL, IRB, Old, IntTy);
2798 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2799 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2800 } else {
2801 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2802, __PRETTY_FUNCTION__))
2802 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2802, __PRETTY_FUNCTION__))
;
2803 }
2804 V = convertValue(DL, IRB, V, AllocaTy);
2805 } else {
2806 // Established these invariants above.
2807 assert(NewBeginOffset == NewAllocaBeginOffset)((NewBeginOffset == NewAllocaBeginOffset) ? static_cast<void
> (0) : __assert_fail ("NewBeginOffset == NewAllocaBeginOffset"
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2807, __PRETTY_FUNCTION__))
;
2808 assert(NewEndOffset == NewAllocaEndOffset)((NewEndOffset == NewAllocaEndOffset) ? static_cast<void>
(0) : __assert_fail ("NewEndOffset == NewAllocaEndOffset", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2808, __PRETTY_FUNCTION__))
;
2809
2810 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2811 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2812 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2813
2814 V = convertValue(DL, IRB, V, AllocaTy);
2815 }
2816
2817 StoreInst *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2818 II.isVolatile());
2819 if (AATags)
2820 New->setAAMetadata(AATags);
2821 LLVM_DEBUG(dbgs() << " to: " << *New << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *New <<
"\n"; } } while (false)
;
2822 return !II.isVolatile();
2823 }
2824
2825 bool visitMemTransferInst(MemTransferInst &II) {
2826 // Rewriting of memory transfer instructions can be a bit tricky. We break
2827 // them into two categories: split intrinsics and unsplit intrinsics.
2828
2829 LLVM_DEBUG(dbgs() << " original: " << II << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << II <<
"\n"; } } while (false)
;
1
Assuming 'DebugFlag' is 0
2
Loop condition is false. Exiting loop
2830
2831 AAMDNodes AATags;
2832 II.getAAMetadata(AATags);
2833
2834 bool IsDest = &II.getRawDestUse() == OldUse;
2835 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2836, __PRETTY_FUNCTION__))
3
Assuming 'IsDest' is 0
4
Assuming the condition is true
5
'?' condition is true
2836 (!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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2836, __PRETTY_FUNCTION__))
;
2837
2838 unsigned SliceAlign = getSliceAlign();
2839
2840 // For unsplit intrinsics, we simply modify the source and destination
2841 // pointers in place. This isn't just an optimization, it is a matter of
2842 // correctness. With unsplit intrinsics we may be dealing with transfers
2843 // within a single alloca before SROA ran, or with transfers that have
2844 // a variable length. We may also be dealing with memmove instead of
2845 // memcpy, and so simply updating the pointers is the necessary for us to
2846 // update both source and dest of a single call.
2847 if (!IsSplittable) {
6
Assuming the condition is false
7
Taking false branch
2848 Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2849 if (IsDest) {
2850 II.setDest(AdjustedPtr);
2851 II.setDestAlignment(SliceAlign);
2852 }
2853 else {
2854 II.setSource(AdjustedPtr);
2855 II.setSourceAlignment(SliceAlign);
2856 }
2857
2858 LLVM_DEBUG(dbgs() << " to: " << II << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << II <<
"\n"; } } while (false)
;
2859 deleteIfTriviallyDead(OldPtr);
2860 return false;
2861 }
2862 // For split transfer intrinsics we have an incredibly useful assurance:
2863 // the source and destination do not reside within the same alloca, and at
2864 // least one of them does not escape. This means that we can replace
2865 // memmove with memcpy, and we don't need to worry about all manner of
2866 // downsides to splitting and transforming the operations.
2867
2868 // If this doesn't map cleanly onto the alloca type, and that type isn't
2869 // a single value type, just emit a memcpy.
2870 bool EmitMemCpy =
2871 !VecTy && !IntTy &&
8
Assuming the condition is false
2872 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2873 SliceSize != DL.getTypeStoreSize(NewAI.getAllocatedType()) ||
2874 !NewAI.getAllocatedType()->isSingleValueType());
2875
2876 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2877 // size hasn't been shrunk based on analysis of the viable range, this is
2878 // a no-op.
2879 if (EmitMemCpy && &OldAI == &NewAI) {
2880 // Ensure the start lines up.
2881 assert(NewBeginOffset == BeginOffset)((NewBeginOffset == BeginOffset) ? static_cast<void> (0
) : __assert_fail ("NewBeginOffset == BeginOffset", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2881, __PRETTY_FUNCTION__))
;
2882
2883 // Rewrite the size as needed.
2884 if (NewEndOffset != EndOffset)
2885 II.setLength(ConstantInt::get(II.getLength()->getType(),
2886 NewEndOffset - NewBeginOffset));
2887 return false;
2888 }
2889 // Record this instruction for deletion.
2890 Pass.DeadInsts.insert(&II);
2891
2892 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2893 // alloca that should be re-examined after rewriting this instruction.
2894 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
9
'?' condition is false
2895 if (AllocaInst *AI =
10
Taking false branch
2896 dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2897 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2898, __PRETTY_FUNCTION__))
2898 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 2898, __PRETTY_FUNCTION__))
;
2899 Pass.Worklist.insert(AI);
2900 }
2901
2902 Type *OtherPtrTy = OtherPtr->getType();
2903 unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2904
2905 // Compute the relative offset for the other pointer within the transfer.
2906 unsigned OffsetWidth = DL.getIndexSizeInBits(OtherAS);
2907 APInt OtherOffset(OffsetWidth, NewBeginOffset - BeginOffset);
2908 unsigned OtherAlign =
2909 IsDest ? II.getSourceAlignment() : II.getDestAlignment();
11
'?' condition is false
2910 OtherAlign = MinAlign(OtherAlign ? OtherAlign : 1,
12
Assuming 'OtherAlign' is 0
13
'?' condition is false
2911 OtherOffset.zextOrTrunc(64).getZExtValue());
2912
2913 if (EmitMemCpy) {
14
Taking false branch
2914 // Compute the other pointer, folding as much as possible to produce
2915 // a single, simple GEP in most cases.
2916 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2917 OtherPtr->getName() + ".");
2918
2919 Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2920 Type *SizeTy = II.getLength()->getType();
2921 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2922
2923 Value *DestPtr, *SrcPtr;
2924 unsigned DestAlign, SrcAlign;
2925 // Note: IsDest is true iff we're copying into the new alloca slice
2926 if (IsDest) {
2927 DestPtr = OurPtr;
2928 DestAlign = SliceAlign;
2929 SrcPtr = OtherPtr;
2930 SrcAlign = OtherAlign;
2931 } else {
2932 DestPtr = OtherPtr;
2933 DestAlign = OtherAlign;
2934 SrcPtr = OurPtr;
2935 SrcAlign = SliceAlign;
2936 }
2937 CallInst *New = IRB.CreateMemCpy(DestPtr, DestAlign, SrcPtr, SrcAlign,
2938 Size, II.isVolatile());
2939 if (AATags)
2940 New->setAAMetadata(AATags);
2941 LLVM_DEBUG(dbgs() << " to: " << *New << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *New <<
"\n"; } } while (false)
;
2942 return false;
2943 }
2944
2945 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
15
Assuming the condition is false
2946 NewEndOffset == NewAllocaEndOffset;
2947 uint64_t Size = NewEndOffset - NewBeginOffset;
2948 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
16
'?' condition is true
2949 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
17
'?' condition is true
2950 unsigned NumElements = EndIndex - BeginIndex;
2951 IntegerType *SubIntTy =
20
'SubIntTy' initialized to a null pointer value
2952 IntTy ? Type::getIntNTy(IntTy->getContext(), Size * 8) : nullptr;
18
Assuming the condition is false
19
'?' condition is false
2953
2954 // Reset the other pointer type to match the register type we're going to
2955 // use, but using the address space of the original other pointer.
2956 Type *OtherTy;
2957 if (VecTy && !IsWholeAlloca) {
21
Taking true branch
2958 if (NumElements == 1)
22
Assuming 'NumElements' is not equal to 1
23
Taking false branch
2959 OtherTy = VecTy->getElementType();
2960 else
2961 OtherTy = VectorType::get(VecTy->getElementType(), NumElements);
2962 } else if (IntTy && !IsWholeAlloca) {
2963 OtherTy = SubIntTy;
2964 } else {
2965 OtherTy = NewAllocaTy;
2966 }
2967 OtherPtrTy = OtherTy->getPointerTo(OtherAS);
2968
2969 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2970 OtherPtr->getName() + ".");
2971 unsigned SrcAlign = OtherAlign;
2972 Value *DstPtr = &NewAI;
2973 unsigned DstAlign = SliceAlign;
2974 if (!IsDest) {
24
Taking true branch
2975 std::swap(SrcPtr, DstPtr);
2976 std::swap(SrcAlign, DstAlign);
2977 }
2978
2979 Value *Src;
2980 if (VecTy && !IsWholeAlloca && !IsDest) {
25
Assuming the condition is false
2981 Src = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2982 NewAI.getAlignment(), "load");
2983 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2984 } else if (IntTy && !IsWholeAlloca && !IsDest) {
26
Assuming the condition is true
27
Taking true branch
2985 Src = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2986 NewAI.getAlignment(), "load");
2987 Src = convertValue(DL, IRB, Src, IntTy);
2988 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2989 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
28
Passing null pointer value via 4th parameter 'Ty'
29
Calling 'extractInteger'
2990 } else {
2991 LoadInst *Load = IRB.CreateAlignedLoad(OtherTy, SrcPtr, SrcAlign,
2992 II.isVolatile(), "copyload");
2993 if (AATags)
2994 Load->setAAMetadata(AATags);
2995 Src = Load;
2996 }
2997
2998 if (VecTy && !IsWholeAlloca && IsDest) {
2999 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
3000 NewAI.getAlignment(), "oldload");
3001 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
3002 } else if (IntTy && !IsWholeAlloca && IsDest) {
3003 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
3004 NewAI.getAlignment(), "oldload");
3005 Old = convertValue(DL, IRB, Old, IntTy);
3006 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
3007 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
3008 Src = convertValue(DL, IRB, Src, NewAllocaTy);
3009 }
3010
3011 StoreInst *Store = cast<StoreInst>(
3012 IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
3013 if (AATags)
3014 Store->setAAMetadata(AATags);
3015 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *Store <<
"\n"; } } while (false)
;
3016 return !II.isVolatile();
3017 }
3018
3019 bool visitIntrinsicInst(IntrinsicInst &II) {
3020 assert(II.isLifetimeStartOrEnd())((II.isLifetimeStartOrEnd()) ? static_cast<void> (0) : __assert_fail
("II.isLifetimeStartOrEnd()", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3020, __PRETTY_FUNCTION__))
;
3021 LLVM_DEBUG(dbgs() << " original: " << II << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << II <<
"\n"; } } while (false)
;
3022 assert(II.getArgOperand(1) == OldPtr)((II.getArgOperand(1) == OldPtr) ? static_cast<void> (0
) : __assert_fail ("II.getArgOperand(1) == OldPtr", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3022, __PRETTY_FUNCTION__))
;
3023
3024 // Record this instruction for deletion.
3025 Pass.DeadInsts.insert(&II);
3026
3027 // Lifetime intrinsics are only promotable if they cover the whole alloca.
3028 // Therefore, we drop lifetime intrinsics which don't cover the whole
3029 // alloca.
3030 // (In theory, intrinsics which partially cover an alloca could be
3031 // promoted, but PromoteMemToReg doesn't handle that case.)
3032 // FIXME: Check whether the alloca is promotable before dropping the
3033 // lifetime intrinsics?
3034 if (NewBeginOffset != NewAllocaBeginOffset ||
3035 NewEndOffset != NewAllocaEndOffset)
3036 return true;
3037
3038 ConstantInt *Size =
3039 ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
3040 NewEndOffset - NewBeginOffset);
3041 // Lifetime intrinsics always expect an i8* so directly get such a pointer
3042 // for the new alloca slice.
3043 Type *PointerTy = IRB.getInt8PtrTy(OldPtr->getType()->getPointerAddressSpace());
3044 Value *Ptr = getNewAllocaSlicePtr(IRB, PointerTy);
3045 Value *New;
3046 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
3047 New = IRB.CreateLifetimeStart(Ptr, Size);
3048 else
3049 New = IRB.CreateLifetimeEnd(Ptr, Size);
3050
3051 (void)New;
3052 LLVM_DEBUG(dbgs() << " to: " << *New << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *New <<
"\n"; } } while (false)
;
3053
3054 return true;
3055 }
3056
3057 void fixLoadStoreAlign(Instruction &Root) {
3058 // This algorithm implements the same visitor loop as
3059 // hasUnsafePHIOrSelectUse, and fixes the alignment of each load
3060 // or store found.
3061 SmallPtrSet<Instruction *, 4> Visited;
3062 SmallVector<Instruction *, 4> Uses;
3063 Visited.insert(&Root);
3064 Uses.push_back(&Root);
3065 do {
3066 Instruction *I = Uses.pop_back_val();
3067
3068 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3069 unsigned LoadAlign = LI->getAlignment();
3070 if (!LoadAlign)
3071 LoadAlign = DL.getABITypeAlignment(LI->getType());
3072 LI->setAlignment(std::min(LoadAlign, getSliceAlign()));
3073 continue;
3074 }
3075 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3076 unsigned StoreAlign = SI->getAlignment();
3077 if (!StoreAlign) {
3078 Value *Op = SI->getOperand(0);
3079 StoreAlign = DL.getABITypeAlignment(Op->getType());
3080 }
3081 SI->setAlignment(std::min(StoreAlign, getSliceAlign()));
3082 continue;
3083 }
3084
3085 assert(isa<BitCastInst>(I) || isa<PHINode>(I) ||((isa<BitCastInst>(I) || isa<PHINode>(I) || isa<
SelectInst>(I) || isa<GetElementPtrInst>(I)) ? static_cast
<void> (0) : __assert_fail ("isa<BitCastInst>(I) || isa<PHINode>(I) || isa<SelectInst>(I) || isa<GetElementPtrInst>(I)"
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3086, __PRETTY_FUNCTION__))
3086 isa<SelectInst>(I) || isa<GetElementPtrInst>(I))((isa<BitCastInst>(I) || isa<PHINode>(I) || isa<
SelectInst>(I) || isa<GetElementPtrInst>(I)) ? static_cast
<void> (0) : __assert_fail ("isa<BitCastInst>(I) || isa<PHINode>(I) || isa<SelectInst>(I) || isa<GetElementPtrInst>(I)"
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3086, __PRETTY_FUNCTION__))
;
3087 for (User *U : I->users())
3088 if (Visited.insert(cast<Instruction>(U)).second)
3089 Uses.push_back(cast<Instruction>(U));
3090 } while (!Uses.empty());
3091 }
3092
3093 bool visitPHINode(PHINode &PN) {
3094 LLVM_DEBUG(dbgs() << " original: " << PN << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << PN <<
"\n"; } } while (false)
;
3095 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3095, __PRETTY_FUNCTION__))
;
3096 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3096, __PRETTY_FUNCTION__))
;
3097
3098 // We would like to compute a new pointer in only one place, but have it be
3099 // as local as possible to the PHI. To do that, we re-use the location of
3100 // the old pointer, which necessarily must be in the right position to
3101 // dominate the PHI.
3102 IRBuilderTy PtrBuilder(IRB);
3103 if (isa<PHINode>(OldPtr))
3104 PtrBuilder.SetInsertPoint(&*OldPtr->getParent()->getFirstInsertionPt());
3105 else
3106 PtrBuilder.SetInsertPoint(OldPtr);
3107 PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
3108
3109 Value *NewPtr = getNewAllocaSlicePtr(PtrBuilder, OldPtr->getType());
3110 // Replace the operands which were using the old pointer.
3111 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
3112
3113 LLVM_DEBUG(dbgs() << " to: " << PN << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << PN <<
"\n"; } } while (false)
;
3114 deleteIfTriviallyDead(OldPtr);
3115
3116 // Fix the alignment of any loads or stores using this PHI node.
3117 fixLoadStoreAlign(PN);
3118
3119 // PHIs can't be promoted on their own, but often can be speculated. We
3120 // check the speculation outside of the rewriter so that we see the
3121 // fully-rewritten alloca.
3122 PHIUsers.insert(&PN);
3123 return true;
3124 }
3125
3126 bool visitSelectInst(SelectInst &SI) {
3127 LLVM_DEBUG(dbgs() << " original: " << SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << SI <<
"\n"; } } while (false)
;
3128 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3129, __PRETTY_FUNCTION__))
3129 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3129, __PRETTY_FUNCTION__))
;
3130 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3130, __PRETTY_FUNCTION__))
;
3131 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3131, __PRETTY_FUNCTION__))
;
3132
3133 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3134 // Replace the operands which were using the old pointer.
3135 if (SI.getOperand(1) == OldPtr)
3136 SI.setOperand(1, NewPtr);
3137 if (SI.getOperand(2) == OldPtr)
3138 SI.setOperand(2, NewPtr);
3139
3140 LLVM_DEBUG(dbgs() << " to: " << SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << SI <<
"\n"; } } while (false)
;
3141 deleteIfTriviallyDead(OldPtr);
3142
3143 // Fix the alignment of any loads or stores using this select.
3144 fixLoadStoreAlign(SI);
3145
3146 // Selects can't be promoted on their own, but often can be speculated. We
3147 // check the speculation outside of the rewriter so that we see the
3148 // fully-rewritten alloca.
3149 SelectUsers.insert(&SI);
3150 return true;
3151 }
3152};
3153
3154namespace {
3155
3156/// Visitor to rewrite aggregate loads and stores as scalar.
3157///
3158/// This pass aggressively rewrites all aggregate loads and stores on
3159/// a particular pointer (or any pointer derived from it which we can identify)
3160/// with scalar loads and stores.
3161class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
3162 // Befriend the base class so it can delegate to private visit methods.
3163 friend class InstVisitor<AggLoadStoreRewriter, bool>;
3164
3165 /// Queue of pointer uses to analyze and potentially rewrite.
3166 SmallVector<Use *, 8> Queue;
3167
3168 /// Set to prevent us from cycling with phi nodes and loops.
3169 SmallPtrSet<User *, 8> Visited;
3170
3171 /// The current pointer use being rewritten. This is used to dig up the used
3172 /// value (as opposed to the user).
3173 Use *U;
3174
3175 /// Used to calculate offsets, and hence alignment, of subobjects.
3176 const DataLayout &DL;
3177
3178public:
3179 AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
3180
3181 /// Rewrite loads and stores through a pointer and all pointers derived from
3182 /// it.
3183 bool rewrite(Instruction &I) {
3184 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)
;
3185 enqueueUsers(I);
3186 bool Changed = false;
3187 while (!Queue.empty()) {
3188 U = Queue.pop_back_val();
3189 Changed |= visit(cast<Instruction>(U->getUser()));
3190 }
3191 return Changed;
3192 }
3193
3194private:
3195 /// Enqueue all the users of the given instruction for further processing.
3196 /// This uses a set to de-duplicate users.
3197 void enqueueUsers(Instruction &I) {
3198 for (Use &U : I.uses())
3199 if (Visited.insert(U.getUser()).second)
3200 Queue.push_back(&U);
3201 }
3202
3203 // Conservative default is to not rewrite anything.
3204 bool visitInstruction(Instruction &I) { return false; }
3205
3206 /// Generic recursive split emission class.
3207 template <typename Derived> class OpSplitter {
3208 protected:
3209 /// The builder used to form new instructions.
3210 IRBuilderTy IRB;
3211
3212 /// The indices which to be used with insert- or extractvalue to select the
3213 /// appropriate value within the aggregate.
3214 SmallVector<unsigned, 4> Indices;
3215
3216 /// The indices to a GEP instruction which will move Ptr to the correct slot
3217 /// within the aggregate.
3218 SmallVector<Value *, 4> GEPIndices;
3219
3220 /// The base pointer of the original op, used as a base for GEPing the
3221 /// split operations.
3222 Value *Ptr;
3223
3224 /// The base pointee type being GEPed into.
3225 Type *BaseTy;
3226
3227 /// Known alignment of the base pointer.
3228 unsigned BaseAlign;
3229
3230 /// To calculate offset of each component so we can correctly deduce
3231 /// alignments.
3232 const DataLayout &DL;
3233
3234 /// Initialize the splitter with an insertion point, Ptr and start with a
3235 /// single zero GEP index.
3236 OpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
3237 unsigned BaseAlign, const DataLayout &DL)
3238 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr),
3239 BaseTy(BaseTy), BaseAlign(BaseAlign), DL(DL) {}
3240
3241 public:
3242 /// Generic recursive split emission routine.
3243 ///
3244 /// This method recursively splits an aggregate op (load or store) into
3245 /// scalar or vector ops. It splits recursively until it hits a single value
3246 /// and emits that single value operation via the template argument.
3247 ///
3248 /// The logic of this routine relies on GEPs and insertvalue and
3249 /// extractvalue all operating with the same fundamental index list, merely
3250 /// formatted differently (GEPs need actual values).
3251 ///
3252 /// \param Ty The type being split recursively into smaller ops.
3253 /// \param Agg The aggregate value being built up or stored, depending on
3254 /// whether this is splitting a load or a store respectively.
3255 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3256 if (Ty->isSingleValueType()) {
3257 unsigned Offset = DL.getIndexedOffsetInType(BaseTy, GEPIndices);
3258 return static_cast<Derived *>(this)->emitFunc(
3259 Ty, Agg, MinAlign(BaseAlign, Offset), Name);
3260 }
3261
3262 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3263 unsigned OldSize = Indices.size();
3264 (void)OldSize;
3265 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3266 ++Idx) {
3267 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3267, __PRETTY_FUNCTION__))
;
3268 Indices.push_back(Idx);
3269 GEPIndices.push_back(IRB.getInt32(Idx));
3270 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3271 GEPIndices.pop_back();
3272 Indices.pop_back();
3273 }
3274 return;
3275 }
3276
3277 if (StructType *STy = dyn_cast<StructType>(Ty)) {
3278 unsigned OldSize = Indices.size();
3279 (void)OldSize;
3280 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3281 ++Idx) {
3282 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3282, __PRETTY_FUNCTION__))
;
3283 Indices.push_back(Idx);
3284 GEPIndices.push_back(IRB.getInt32(Idx));
3285 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3286 GEPIndices.pop_back();
3287 Indices.pop_back();
3288 }
3289 return;
3290 }
3291
3292 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3292)
;
3293 }
3294 };
3295
3296 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3297 AAMDNodes AATags;
3298
3299 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
3300 AAMDNodes AATags, unsigned BaseAlign, const DataLayout &DL)
3301 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr, BaseTy, BaseAlign,
3302 DL), AATags(AATags) {}
3303
3304 /// Emit a leaf load of a single value. This is called at the leaves of the
3305 /// recursive emission to actually load values.
3306 void emitFunc(Type *Ty, Value *&Agg, unsigned Align, const Twine &Name) {
3307 assert(Ty->isSingleValueType())((Ty->isSingleValueType()) ? static_cast<void> (0) :
__assert_fail ("Ty->isSingleValueType()", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3307, __PRETTY_FUNCTION__))
;
3308 // Load the single value and insert it using the indices.
3309 Value *GEP =
3310 IRB.CreateInBoundsGEP(BaseTy, Ptr, GEPIndices, Name + ".gep");
3311 LoadInst *Load = IRB.CreateAlignedLoad(Ty, GEP, Align, Name + ".load");
3312 if (AATags)
3313 Load->setAAMetadata(AATags);
3314 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3315 LLVM_DEBUG(dbgs() << " to: " << *Load << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *Load <<
"\n"; } } while (false)
;
3316 }
3317 };
3318
3319 bool visitLoadInst(LoadInst &LI) {
3320 assert(LI.getPointerOperand() == *U)((LI.getPointerOperand() == *U) ? static_cast<void> (0)
: __assert_fail ("LI.getPointerOperand() == *U", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3320, __PRETTY_FUNCTION__))
;
3321 if (!LI.isSimple() || LI.getType()->isSingleValueType())
3322 return false;
3323
3324 // We have an aggregate being loaded, split it apart.
3325 LLVM_DEBUG(dbgs() << " original: " << LI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << LI <<
"\n"; } } while (false)
;
3326 AAMDNodes AATags;
3327 LI.getAAMetadata(AATags);
3328 LoadOpSplitter Splitter(&LI, *U, LI.getType(), AATags,
3329 getAdjustedAlignment(&LI, 0, DL), DL);
3330 Value *V = UndefValue::get(LI.getType());
3331 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3332 LI.replaceAllUsesWith(V);
3333 LI.eraseFromParent();
3334 return true;
3335 }
3336
3337 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3338 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
3339 AAMDNodes AATags, unsigned BaseAlign, const DataLayout &DL)
3340 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr, BaseTy, BaseAlign,
3341 DL),
3342 AATags(AATags) {}
3343 AAMDNodes AATags;
3344 /// Emit a leaf store of a single value. This is called at the leaves of the
3345 /// recursive emission to actually produce stores.
3346 void emitFunc(Type *Ty, Value *&Agg, unsigned Align, const Twine &Name) {
3347 assert(Ty->isSingleValueType())((Ty->isSingleValueType()) ? static_cast<void> (0) :
__assert_fail ("Ty->isSingleValueType()", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3347, __PRETTY_FUNCTION__))
;
3348 // Extract the single value and store it using the indices.
3349 //
3350 // The gep and extractvalue values are factored out of the CreateStore
3351 // call to make the output independent of the argument evaluation order.
3352 Value *ExtractValue =
3353 IRB.CreateExtractValue(Agg, Indices, Name + ".extract");
3354 Value *InBoundsGEP =
3355 IRB.CreateInBoundsGEP(BaseTy, Ptr, GEPIndices, Name + ".gep");
3356 StoreInst *Store =
3357 IRB.CreateAlignedStore(ExtractValue, InBoundsGEP, Align);
3358 if (AATags)
3359 Store->setAAMetadata(AATags);
3360 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " to: " << *Store <<
"\n"; } } while (false)
;
3361 }
3362 };
3363
3364 bool visitStoreInst(StoreInst &SI) {
3365 if (!SI.isSimple() || SI.getPointerOperand() != *U)
3366 return false;
3367 Value *V = SI.getValueOperand();
3368 if (V->getType()->isSingleValueType())
3369 return false;
3370
3371 // We have an aggregate being stored, split it apart.
3372 LLVM_DEBUG(dbgs() << " original: " << SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " original: " << SI <<
"\n"; } } while (false)
;
3373 AAMDNodes AATags;
3374 SI.getAAMetadata(AATags);
3375 StoreOpSplitter Splitter(&SI, *U, V->getType(), AATags,
3376 getAdjustedAlignment(&SI, 0, DL), DL);
3377 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3378 SI.eraseFromParent();
3379 return true;
3380 }
3381
3382 bool visitBitCastInst(BitCastInst &BC) {
3383 enqueueUsers(BC);
3384 return false;
3385 }
3386
3387 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3388 enqueueUsers(GEPI);
3389 return false;
3390 }
3391
3392 bool visitPHINode(PHINode &PN) {
3393 enqueueUsers(PN);
3394 return false;
3395 }
3396
3397 bool visitSelectInst(SelectInst &SI) {
3398 enqueueUsers(SI);
3399 return false;
3400 }
3401};
3402
3403} // end anonymous namespace
3404
3405/// Strip aggregate type wrapping.
3406///
3407/// This removes no-op aggregate types wrapping an underlying type. It will
3408/// strip as many layers of types as it can without changing either the type
3409/// size or the allocated size.
3410static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
3411 if (Ty->isSingleValueType())
3412 return Ty;
3413
3414 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3415 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
3416
3417 Type *InnerTy;
3418 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3419 InnerTy = ArrTy->getElementType();
3420 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3421 const StructLayout *SL = DL.getStructLayout(STy);
3422 unsigned Index = SL->getElementContainingOffset(0);
3423 InnerTy = STy->getElementType(Index);
3424 } else {
3425 return Ty;
3426 }
3427
3428 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3429 TypeSize > DL.getTypeSizeInBits(InnerTy))
3430 return Ty;
3431
3432 return stripAggregateTypeWrapping(DL, InnerTy);
3433}
3434
3435/// Try to find a partition of the aggregate type passed in for a given
3436/// offset and size.
3437///
3438/// This recurses through the aggregate type and tries to compute a subtype
3439/// based on the offset and size. When the offset and size span a sub-section
3440/// of an array, it will even compute a new array type for that sub-section,
3441/// and the same for structs.
3442///
3443/// Note that this routine is very strict and tries to find a partition of the
3444/// type which produces the *exact* right offset and size. It is not forgiving
3445/// when the size or offset cause either end of type-based partition to be off.
3446/// Also, this is a best-effort routine. It is reasonable to give up and not
3447/// return a type if necessary.
3448static Type *getTypePartition(const DataLayout &DL, Type *Ty, uint64_t Offset,
3449 uint64_t Size) {
3450 if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
3451 return stripAggregateTypeWrapping(DL, Ty);
3452 if (Offset > DL.getTypeAllocSize(Ty) ||
3453 (DL.getTypeAllocSize(Ty) - Offset) < Size)
3454 return nullptr;
3455
3456 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3457 Type *ElementTy = SeqTy->getElementType();
3458 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3459 uint64_t NumSkippedElements = Offset / ElementSize;
3460 if (NumSkippedElements >= SeqTy->getNumElements())
3461 return nullptr;
3462 Offset -= NumSkippedElements * ElementSize;
3463
3464 // First check if we need to recurse.
3465 if (Offset > 0 || Size < ElementSize) {
3466 // Bail if the partition ends in a different array element.
3467 if ((Offset + Size) > ElementSize)
3468 return nullptr;
3469 // Recurse through the element type trying to peel off offset bytes.
3470 return getTypePartition(DL, ElementTy, Offset, Size);
3471 }
3472 assert(Offset == 0)((Offset == 0) ? static_cast<void> (0) : __assert_fail (
"Offset == 0", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3472, __PRETTY_FUNCTION__))
;
3473
3474 if (Size == ElementSize)
3475 return stripAggregateTypeWrapping(DL, ElementTy);
3476 assert(Size > ElementSize)((Size > ElementSize) ? static_cast<void> (0) : __assert_fail
("Size > ElementSize", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3476, __PRETTY_FUNCTION__))
;
3477 uint64_t NumElements = Size / ElementSize;
3478 if (NumElements * ElementSize != Size)
3479 return nullptr;
3480 return ArrayType::get(ElementTy, NumElements);
3481 }
3482
3483 StructType *STy = dyn_cast<StructType>(Ty);
3484 if (!STy)
3485 return nullptr;
3486
3487 const StructLayout *SL = DL.getStructLayout(STy);
3488 if (Offset >= SL->getSizeInBytes())
3489 return nullptr;
3490 uint64_t EndOffset = Offset + Size;
3491 if (EndOffset > SL->getSizeInBytes())
3492 return nullptr;
3493
3494 unsigned Index = SL->getElementContainingOffset(Offset);
3495 Offset -= SL->getElementOffset(Index);
3496
3497 Type *ElementTy = STy->getElementType(Index);
3498 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3499 if (Offset >= ElementSize)
3500 return nullptr; // The offset points into alignment padding.
3501
3502 // See if any partition must be contained by the element.
3503 if (Offset > 0 || Size < ElementSize) {
3504 if ((Offset + Size) > ElementSize)
3505 return nullptr;
3506 return getTypePartition(DL, ElementTy, Offset, Size);
3507 }
3508 assert(Offset == 0)((Offset == 0) ? static_cast<void> (0) : __assert_fail (
"Offset == 0", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3508, __PRETTY_FUNCTION__))
;
3509
3510 if (Size == ElementSize)
3511 return stripAggregateTypeWrapping(DL, ElementTy);
3512
3513 StructType::element_iterator EI = STy->element_begin() + Index,
3514 EE = STy->element_end();
3515 if (EndOffset < SL->getSizeInBytes()) {
3516 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3517 if (Index == EndIndex)
3518 return nullptr; // Within a single element and its padding.
3519
3520 // Don't try to form "natural" types if the elements don't line up with the
3521 // expected size.
3522 // FIXME: We could potentially recurse down through the last element in the
3523 // sub-struct to find a natural end point.
3524 if (SL->getElementOffset(EndIndex) != EndOffset)
3525 return nullptr;
3526
3527 assert(Index < EndIndex)((Index < EndIndex) ? static_cast<void> (0) : __assert_fail
("Index < EndIndex", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3527, __PRETTY_FUNCTION__))
;
3528 EE = STy->element_begin() + EndIndex;
3529 }
3530
3531 // Try to build up a sub-structure.
3532 StructType *SubTy =
3533 StructType::get(STy->getContext(), makeArrayRef(EI, EE), STy->isPacked());
3534 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3535 if (Size != SubSL->getSizeInBytes())
3536 return nullptr; // The sub-struct doesn't have quite the size needed.
3537
3538 return SubTy;
3539}
3540
3541/// Pre-split loads and stores to simplify rewriting.
3542///
3543/// We want to break up the splittable load+store pairs as much as
3544/// possible. This is important to do as a preprocessing step, as once we
3545/// start rewriting the accesses to partitions of the alloca we lose the
3546/// necessary information to correctly split apart paired loads and stores
3547/// which both point into this alloca. The case to consider is something like
3548/// the following:
3549///
3550/// %a = alloca [12 x i8]
3551/// %gep1 = getelementptr [12 x i8]* %a, i32 0, i32 0
3552/// %gep2 = getelementptr [12 x i8]* %a, i32 0, i32 4
3553/// %gep3 = getelementptr [12 x i8]* %a, i32 0, i32 8
3554/// %iptr1 = bitcast i8* %gep1 to i64*
3555/// %iptr2 = bitcast i8* %gep2 to i64*
3556/// %fptr1 = bitcast i8* %gep1 to float*
3557/// %fptr2 = bitcast i8* %gep2 to float*
3558/// %fptr3 = bitcast i8* %gep3 to float*
3559/// store float 0.0, float* %fptr1
3560/// store float 1.0, float* %fptr2
3561/// %v = load i64* %iptr1
3562/// store i64 %v, i64* %iptr2
3563/// %f1 = load float* %fptr2
3564/// %f2 = load float* %fptr3
3565///
3566/// Here we want to form 3 partitions of the alloca, each 4 bytes large, and
3567/// promote everything so we recover the 2 SSA values that should have been
3568/// there all along.
3569///
3570/// \returns true if any changes are made.
3571bool SROA::presplitLoadsAndStores(AllocaInst &AI, AllocaSlices &AS) {
3572 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)
;
3573
3574 // Track the loads and stores which are candidates for pre-splitting here, in
3575 // the order they first appear during the partition scan. These give stable
3576 // iteration order and a basis for tracking which loads and stores we
3577 // actually split.
3578 SmallVector<LoadInst *, 4> Loads;
3579 SmallVector<StoreInst *, 4> Stores;
3580
3581 // We need to accumulate the splits required of each load or store where we
3582 // can find them via a direct lookup. This is important to cross-check loads
3583 // and stores against each other. We also track the slice so that we can kill
3584 // all the slices that end up split.
3585 struct SplitOffsets {
3586 Slice *S;
3587 std::vector<uint64_t> Splits;
3588 };
3589 SmallDenseMap<Instruction *, SplitOffsets, 8> SplitOffsetsMap;
3590
3591 // Track loads out of this alloca which cannot, for any reason, be pre-split.
3592 // This is important as we also cannot pre-split stores of those loads!
3593 // FIXME: This is all pretty gross. It means that we can be more aggressive
3594 // in pre-splitting when the load feeding the store happens to come from
3595 // a separate alloca. Put another way, the effectiveness of SROA would be
3596 // decreased by a frontend which just concatenated all of its local allocas
3597 // into one big flat alloca. But defeating such patterns is exactly the job
3598 // SROA is tasked with! Sadly, to not have this discrepancy we would have
3599 // change store pre-splitting to actually force pre-splitting of the load
3600 // that feeds it *and all stores*. That makes pre-splitting much harder, but
3601 // maybe it would make it more principled?
3602 SmallPtrSet<LoadInst *, 8> UnsplittableLoads;
3603
3604 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)
;
3605 for (auto &P : AS.partitions()) {
3606 for (Slice &S : P) {
3607 Instruction *I = cast<Instruction>(S.getUse()->getUser());
3608 if (!S.isSplittable() || S.endOffset() <= P.endOffset()) {
3609 // If this is a load we have to track that it can't participate in any
3610 // pre-splitting. If this is a store of a load we have to track that
3611 // that load also can't participate in any pre-splitting.
3612 if (auto *LI = dyn_cast<LoadInst>(I))
3613 UnsplittableLoads.insert(LI);
3614 else if (auto *SI = dyn_cast<StoreInst>(I))
3615 if (auto *LI = dyn_cast<LoadInst>(SI->getValueOperand()))
3616 UnsplittableLoads.insert(LI);
3617 continue;
3618 }
3619 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3620, __PRETTY_FUNCTION__))
3620 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3620, __PRETTY_FUNCTION__))
;
3621
3622 // Determine if this is a pre-splittable slice.
3623 if (auto *LI = dyn_cast<LoadInst>(I)) {
3624 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3624, __PRETTY_FUNCTION__))
;
3625
3626 // The load must be used exclusively to store into other pointers for
3627 // us to be able to arbitrarily pre-split it. The stores must also be
3628 // simple to avoid changing semantics.
3629 auto IsLoadSimplyStored = [](LoadInst *LI) {
3630 for (User *LU : LI->users()) {
3631 auto *SI = dyn_cast<StoreInst>(LU);
3632 if (!SI || !SI->isSimple())
3633 return false;
3634 }
3635 return true;
3636 };
3637 if (!IsLoadSimplyStored(LI)) {
3638 UnsplittableLoads.insert(LI);
3639 continue;
3640 }
3641
3642 Loads.push_back(LI);
3643 } else if (auto *SI = dyn_cast<StoreInst>(I)) {
3644 if (S.getUse() != &SI->getOperandUse(SI->getPointerOperandIndex()))
3645 // Skip stores *of* pointers. FIXME: This shouldn't even be possible!
3646 continue;
3647 auto *StoredLoad = dyn_cast<LoadInst>(SI->getValueOperand());
3648 if (!StoredLoad || !StoredLoad->isSimple())
3649 continue;
3650 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3650, __PRETTY_FUNCTION__))
;
3651
3652 Stores.push_back(SI);
3653 } else {
3654 // Other uses cannot be pre-split.
3655 continue;
3656 }
3657
3658 // Record the initial split.
3659 LLVM_DEBUG(dbgs() << " Candidate: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Candidate: " << *I <<
"\n"; } } while (false)
;
3660 auto &Offsets = SplitOffsetsMap[I];
3661 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3662, __PRETTY_FUNCTION__))
3662 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3662, __PRETTY_FUNCTION__))
;
3663 Offsets.S = &S;
3664 Offsets.Splits.push_back(P.endOffset() - S.beginOffset());
3665 }
3666
3667 // Now scan the already split slices, and add a split for any of them which
3668 // we're going to pre-split.
3669 for (Slice *S : P.splitSliceTails()) {
3670 auto SplitOffsetsMapI =
3671 SplitOffsetsMap.find(cast<Instruction>(S->getUse()->getUser()));
3672 if (SplitOffsetsMapI == SplitOffsetsMap.end())
3673 continue;
3674 auto &Offsets = SplitOffsetsMapI->second;
3675
3676 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3676, __PRETTY_FUNCTION__))
;
3677 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3678, __PRETTY_FUNCTION__))
3678 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3678, __PRETTY_FUNCTION__))
;
3679 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3681, __PRETTY_FUNCTION__))
3680 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3681, __PRETTY_FUNCTION__))
3681 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3681, __PRETTY_FUNCTION__))
;
3682
3683 // Record each split. The last partition's end isn't needed as the size
3684 // of the slice dictates that.
3685 if (S->endOffset() > P.endOffset())
3686 Offsets.Splits.push_back(P.endOffset() - Offsets.S->beginOffset());
3687 }
3688 }
3689
3690 // We may have split loads where some of their stores are split stores. For
3691 // such loads and stores, we can only pre-split them if their splits exactly
3692 // match relative to their starting offset. We have to verify this prior to
3693 // any rewriting.
3694 Stores.erase(
3695 llvm::remove_if(Stores,
3696 [&UnsplittableLoads, &SplitOffsetsMap](StoreInst *SI) {
3697 // Lookup the load we are storing in our map of split
3698 // offsets.
3699 auto *LI = cast<LoadInst>(SI->getValueOperand());
3700 // If it was completely unsplittable, then we're done,
3701 // and this store can't be pre-split.
3702 if (UnsplittableLoads.count(LI))
3703 return true;
3704
3705 auto LoadOffsetsI = SplitOffsetsMap.find(LI);
3706 if (LoadOffsetsI == SplitOffsetsMap.end())
3707 return false; // Unrelated loads are definitely safe.
3708 auto &LoadOffsets = LoadOffsetsI->second;
3709
3710 // Now lookup the store's offsets.
3711 auto &StoreOffsets = SplitOffsetsMap[SI];
3712
3713 // If the relative offsets of each split in the load and
3714 // store match exactly, then we can split them and we
3715 // don't need to remove them here.
3716 if (LoadOffsets.Splits == StoreOffsets.Splits)
3717 return false;
3718
3719 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Mismatched splits for load and store:\n"
<< " " << *LI << "\n" << " "
<< *SI << "\n"; } } while (false)
3720 dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Mismatched splits for load and store:\n"
<< " " << *LI << "\n" << " "
<< *SI << "\n"; } } while (false)
3721 << " 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)
3722 << " " << *LI << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Mismatched splits for load and store:\n"
<< " " << *LI << "\n" << " "
<< *SI << "\n"; } } while (false)
3723 << " " << *SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Mismatched splits for load and store:\n"
<< " " << *LI << "\n" << " "
<< *SI << "\n"; } } while (false)
;
3724
3725 // We've found a store and load that we need to split
3726 // with mismatched relative splits. Just give up on them
3727 // and remove both instructions from our list of
3728 // candidates.
3729 UnsplittableLoads.insert(LI);
3730 return true;
3731 }),
3732 Stores.end());
3733 // Now we have to go *back* through all the stores, because a later store may
3734 // have caused an earlier store's load to become unsplittable and if it is
3735 // unsplittable for the later store, then we can't rely on it being split in
3736 // the earlier store either.
3737 Stores.erase(llvm::remove_if(Stores,
3738 [&UnsplittableLoads](StoreInst *SI) {
3739 auto *LI =
3740 cast<LoadInst>(SI->getValueOperand());
3741 return UnsplittableLoads.count(LI);
3742 }),
3743 Stores.end());
3744 // Once we've established all the loads that can't be split for some reason,
3745 // filter any that made it into our list out.
3746 Loads.erase(llvm::remove_if(Loads,
3747 [&UnsplittableLoads](LoadInst *LI) {
3748 return UnsplittableLoads.count(LI);
3749 }),
3750 Loads.end());
3751
3752 // If no loads or stores are left, there is no pre-splitting to be done for
3753 // this alloca.
3754 if (Loads.empty() && Stores.empty())
3755 return false;
3756
3757 // From here on, we can't fail and will be building new accesses, so rig up
3758 // an IR builder.
3759 IRBuilderTy IRB(&AI);
3760
3761 // Collect the new slices which we will merge into the alloca slices.
3762 SmallVector<Slice, 4> NewSlices;
3763
3764 // Track any allocas we end up splitting loads and stores for so we iterate
3765 // on them.
3766 SmallPtrSet<AllocaInst *, 4> ResplitPromotableAllocas;
3767
3768 // At this point, we have collected all of the loads and stores we can
3769 // pre-split, and the specific splits needed for them. We actually do the
3770 // splitting in a specific order in order to handle when one of the loads in
3771 // the value operand to one of the stores.
3772 //
3773 // First, we rewrite all of the split loads, and just accumulate each split
3774 // load in a parallel structure. We also build the slices for them and append
3775 // them to the alloca slices.
3776 SmallDenseMap<LoadInst *, std::vector<LoadInst *>, 1> SplitLoadsMap;
3777 std::vector<LoadInst *> SplitLoads;
3778 const DataLayout &DL = AI.getModule()->getDataLayout();
3779 for (LoadInst *LI : Loads) {
3780 SplitLoads.clear();
3781
3782 IntegerType *Ty = cast<IntegerType>(LI->getType());
3783 uint64_t LoadSize = Ty->getBitWidth() / 8;
3784 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3784, __PRETTY_FUNCTION__))
;
3785
3786 auto &Offsets = SplitOffsetsMap[LI];
3787 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3788, __PRETTY_FUNCTION__))
3788 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3788, __PRETTY_FUNCTION__))
;
3789 uint64_t BaseOffset = Offsets.S->beginOffset();
3790 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3791, __PRETTY_FUNCTION__))
3791 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3791, __PRETTY_FUNCTION__))
;
3792
3793 Instruction *BasePtr = cast<Instruction>(LI->getPointerOperand());
3794 IRB.SetInsertPoint(LI);
3795
3796 LLVM_DEBUG(dbgs() << " Splitting load: " << *LI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Splitting load: " << *LI
<< "\n"; } } while (false)
;
3797
3798 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3799 int Idx = 0, Size = Offsets.Splits.size();
3800 for (;;) {
3801 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3802 auto AS = LI->getPointerAddressSpace();
3803 auto *PartPtrTy = PartTy->getPointerTo(AS);
3804 LoadInst *PLoad = IRB.CreateAlignedLoad(
3805 PartTy,
3806 getAdjustedPtr(IRB, DL, BasePtr,
3807 APInt(DL.getIndexSizeInBits(AS), PartOffset),
3808 PartPtrTy, BasePtr->getName() + "."),
3809 getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3810 LI->getName());
3811 PLoad->copyMetadata(*LI, {LLVMContext::MD_mem_parallel_loop_access,
3812 LLVMContext::MD_access_group});
3813
3814 // Append this load onto the list of split loads so we can find it later
3815 // to rewrite the stores.
3816 SplitLoads.push_back(PLoad);
3817
3818 // Now build a new slice for the alloca.
3819 NewSlices.push_back(
3820 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3821 &PLoad->getOperandUse(PLoad->getPointerOperandIndex()),
3822 /*IsSplittable*/ false));
3823 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)
3824 << ", " << NewSlices.back().endOffset()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " new slice [" << NewSlices
.back().beginOffset() << ", " << NewSlices.back()
.endOffset() << "): " << *PLoad << "\n"; } }
while (false)
3825 << "): " << *PLoad << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " new slice [" << NewSlices
.back().beginOffset() << ", " << NewSlices.back()
.endOffset() << "): " << *PLoad << "\n"; } }
while (false)
;
3826
3827 // See if we've handled all the splits.
3828 if (Idx >= Size)
3829 break;
3830
3831 // Setup the next partition.
3832 PartOffset = Offsets.Splits[Idx];
3833 ++Idx;
3834 PartSize = (Idx < Size ? Offsets.Splits[Idx] : LoadSize) - PartOffset;
3835 }
3836
3837 // Now that we have the split loads, do the slow walk over all uses of the
3838 // load and rewrite them as split stores, or save the split loads to use
3839 // below if the store is going to be split there anyways.
3840 bool DeferredStores = false;
3841 for (User *LU : LI->users()) {
3842 StoreInst *SI = cast<StoreInst>(LU);
3843 if (!Stores.empty() && SplitOffsetsMap.count(SI)) {
3844 DeferredStores = true;
3845 LLVM_DEBUG(dbgs() << " Deferred splitting of store: " << *SIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Deferred splitting of store: "
<< *SI << "\n"; } } while (false)
3846 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Deferred splitting of store: "
<< *SI << "\n"; } } while (false)
;
3847 continue;
3848 }
3849
3850 Value *StoreBasePtr = SI->getPointerOperand();
3851 IRB.SetInsertPoint(SI);
3852
3853 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)
;
3854
3855 for (int Idx = 0, Size = SplitLoads.size(); Idx < Size; ++Idx) {
3856 LoadInst *PLoad = SplitLoads[Idx];
3857 uint64_t PartOffset = Idx == 0 ? 0 : Offsets.Splits[Idx - 1];
3858 auto *PartPtrTy =
3859 PLoad->getType()->getPointerTo(SI->getPointerAddressSpace());
3860
3861 auto AS = SI->getPointerAddressSpace();
3862 StoreInst *PStore = IRB.CreateAlignedStore(
3863 PLoad,
3864 getAdjustedPtr(IRB, DL, StoreBasePtr,
3865 APInt(DL.getIndexSizeInBits(AS), PartOffset),
3866 PartPtrTy, StoreBasePtr->getName() + "."),
3867 getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3868 PStore->copyMetadata(*LI, {LLVMContext::MD_mem_parallel_loop_access,
3869 LLVMContext::MD_access_group});
3870 LLVM_DEBUG(dbgs() << " +" << PartOffset << ":" << *PStore << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " +" << PartOffset <<
":" << *PStore << "\n"; } } while (false)
;
3871 }
3872
3873 // We want to immediately iterate on any allocas impacted by splitting
3874 // this store, and we have to track any promotable alloca (indicated by
3875 // a direct store) as needing to be resplit because it is no longer
3876 // promotable.
3877 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(StoreBasePtr)) {
3878 ResplitPromotableAllocas.insert(OtherAI);
3879 Worklist.insert(OtherAI);
3880 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3881 StoreBasePtr->stripInBoundsOffsets())) {
3882 Worklist.insert(OtherAI);
3883 }
3884
3885 // Mark the original store as dead.
3886 DeadInsts.insert(SI);
3887 }
3888
3889 // Save the split loads if there are deferred stores among the users.
3890 if (DeferredStores)
3891 SplitLoadsMap.insert(std::make_pair(LI, std::move(SplitLoads)));
3892
3893 // Mark the original load as dead and kill the original slice.
3894 DeadInsts.insert(LI);
3895 Offsets.S->kill();
3896 }
3897
3898 // Second, we rewrite all of the split stores. At this point, we know that
3899 // all loads from this alloca have been split already. For stores of such
3900 // loads, we can simply look up the pre-existing split loads. For stores of
3901 // other loads, we split those loads first and then write split stores of
3902 // them.
3903 for (StoreInst *SI : Stores) {
3904 auto *LI = cast<LoadInst>(SI->getValueOperand());
3905 IntegerType *Ty = cast<IntegerType>(LI->getType());
3906 uint64_t StoreSize = Ty->getBitWidth() / 8;
3907 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3907, __PRETTY_FUNCTION__))
;
3908
3909 auto &Offsets = SplitOffsetsMap[SI];
3910 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3911, __PRETTY_FUNCTION__))
3911 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3911, __PRETTY_FUNCTION__))
;
3912 uint64_t BaseOffset = Offsets.S->beginOffset();
3913 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3914, __PRETTY_FUNCTION__))
3914 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3914, __PRETTY_FUNCTION__))
;
3915
3916 Value *LoadBasePtr = LI->getPointerOperand();
3917 Instruction *StoreBasePtr = cast<Instruction>(SI->getPointerOperand());
3918
3919 LLVM_DEBUG(dbgs() << " Splitting store: " << *SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Splitting store: " << *SI
<< "\n"; } } while (false)
;
3920
3921 // Check whether we have an already split load.
3922 auto SplitLoadsMapI = SplitLoadsMap.find(LI);
3923 std::vector<LoadInst *> *SplitLoads = nullptr;
3924 if (SplitLoadsMapI != SplitLoadsMap.end()) {
3925 SplitLoads = &SplitLoadsMapI->second;
3926 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3927, __PRETTY_FUNCTION__))
3927 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3927, __PRETTY_FUNCTION__))
;
3928 } else {
3929 LLVM_DEBUG(dbgs() << " of load: " << *LI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " of load: " << *LI
<< "\n"; } } while (false)
;
3930 }
3931
3932 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3933 int Idx = 0, Size = Offsets.Splits.size();
3934 for (;;) {
3935 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3936 auto *LoadPartPtrTy = PartTy->getPointerTo(LI->getPointerAddressSpace());
3937 auto *StorePartPtrTy = PartTy->getPointerTo(SI->getPointerAddressSpace());
3938
3939 // Either lookup a split load or create one.
3940 LoadInst *PLoad;
3941 if (SplitLoads) {
3942 PLoad = (*SplitLoads)[Idx];
3943 } else {
3944 IRB.SetInsertPoint(LI);
3945 auto AS = LI->getPointerAddressSpace();
3946 PLoad = IRB.CreateAlignedLoad(
3947 PartTy,
3948 getAdjustedPtr(IRB, DL, LoadBasePtr,
3949 APInt(DL.getIndexSizeInBits(AS), PartOffset),
3950 LoadPartPtrTy, LoadBasePtr->getName() + "."),
3951 getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3952 LI->getName());
3953 }
3954
3955 // And store this partition.
3956 IRB.SetInsertPoint(SI);
3957 auto AS = SI->getPointerAddressSpace();
3958 StoreInst *PStore = IRB.CreateAlignedStore(
3959 PLoad,
3960 getAdjustedPtr(IRB, DL, StoreBasePtr,
3961 APInt(DL.getIndexSizeInBits(AS), PartOffset),
3962 StorePartPtrTy, StoreBasePtr->getName() + "."),
3963 getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3964
3965 // Now build a new slice for the alloca.
3966 NewSlices.push_back(
3967 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3968 &PStore->getOperandUse(PStore->getPointerOperandIndex()),
3969 /*IsSplittable*/ false));
3970 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)
3971 << ", " << NewSlices.back().endOffset()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " new slice [" << NewSlices
.back().beginOffset() << ", " << NewSlices.back()
.endOffset() << "): " << *PStore << "\n"; }
} while (false)
3972 << "): " << *PStore << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " new slice [" << NewSlices
.back().beginOffset() << ", " << NewSlices.back()
.endOffset() << "): " << *PStore << "\n"; }
} while (false)
;
3973 if (!SplitLoads) {
3974 LLVM_DEBUG(dbgs() << " of split load: " << *PLoad << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " of split load: " << *
PLoad << "\n"; } } while (false)
;
3975 }
3976
3977 // See if we've finished all the splits.
3978 if (Idx >= Size)
3979 break;
3980
3981 // Setup the next partition.
3982 PartOffset = Offsets.Splits[Idx];
3983 ++Idx;
3984 PartSize = (Idx < Size ? Offsets.Splits[Idx] : StoreSize) - PartOffset;
3985 }
3986
3987 // We want to immediately iterate on any allocas impacted by splitting
3988 // this load, which is only relevant if it isn't a load of this alloca and
3989 // thus we didn't already split the loads above. We also have to keep track
3990 // of any promotable allocas we split loads on as they can no longer be
3991 // promoted.
3992 if (!SplitLoads) {
3993 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(LoadBasePtr)) {
3994 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3994, __PRETTY_FUNCTION__))
;
3995 ResplitPromotableAllocas.insert(OtherAI);
3996 Worklist.insert(OtherAI);
3997 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3998 LoadBasePtr->stripInBoundsOffsets())) {
3999 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 3999, __PRETTY_FUNCTION__))
;
4000 Worklist.insert(OtherAI);
4001 }
4002 }
4003
4004 // Mark the original store as dead now that we've split it up and kill its
4005 // slice. Note that we leave the original load in place unless this store
4006 // was its only use. It may in turn be split up if it is an alloca load
4007 // for some other alloca, but it may be a normal load. This may introduce
4008 // redundant loads, but where those can be merged the rest of the optimizer
4009 // should handle the merging, and this uncovers SSA splits which is more
4010 // important. In practice, the original loads will almost always be fully
4011 // split and removed eventually, and the splits will be merged by any
4012 // trivial CSE, including instcombine.
4013 if (LI->hasOneUse()) {
4014 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 4014, __PRETTY_FUNCTION__))
;
4015 DeadInsts.insert(LI);
4016 }
4017 DeadInsts.insert(SI);
4018 Offsets.S->kill();
4019 }
4020
4021 // Remove the killed slices that have ben pre-split.
4022 AS.erase(llvm::remove_if(AS, [](const Slice &S) { return S.isDead(); }),
4023 AS.end());
4024
4025 // Insert our new slices. This will sort and merge them into the sorted
4026 // sequence.
4027 AS.insert(NewSlices);
4028
4029 LLVM_DEBUG(dbgs() << " Pre-split slices:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Pre-split slices:\n"; } } while
(false)
;
4030#ifndef NDEBUG
4031 for (auto I = AS.begin(), E = AS.end(); I != E; ++I)
4032 LLVM_DEBUG(AS.print(dbgs(), I, " "))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { AS.print(dbgs(), I, " "); } } while (false)
;
4033#endif
4034
4035 // Finally, don't try to promote any allocas that new require re-splitting.
4036 // They have already been added to the worklist above.
4037 PromotableAllocas.erase(
4038 llvm::remove_if(
4039 PromotableAllocas,
4040 [&](AllocaInst *AI) { return ResplitPromotableAllocas.count(AI); }),
4041 PromotableAllocas.end());
4042
4043 return true;
4044}
4045
4046/// Rewrite an alloca partition's users.
4047///
4048/// This routine drives both of the rewriting goals of the SROA pass. It tries
4049/// to rewrite uses of an alloca partition to be conducive for SSA value
4050/// promotion. If the partition needs a new, more refined alloca, this will
4051/// build that new alloca, preserving as much type information as possible, and
4052/// rewrite the uses of the old alloca to point at the new one and have the
4053/// appropriate new offsets. It also evaluates how successful the rewrite was
4054/// at enabling promotion and if it was successful queues the alloca to be
4055/// promoted.
4056AllocaInst *SROA::rewritePartition(AllocaInst &AI, AllocaSlices &AS,
4057 Partition &P) {
4058 // Try to compute a friendly type for this partition of the alloca. This
4059 // won't always succeed, in which case we fall back to a legal integer type
4060 // or an i8 array of an appropriate size.
4061 Type *SliceTy = nullptr;
4062 const DataLayout &DL = AI.getModule()->getDataLayout();
4063 if (Type *CommonUseTy = findCommonType(P.begin(), P.end(), P.endOffset()))
4064 if (DL.getTypeAllocSize(CommonUseTy) >= P.size())
4065 SliceTy = CommonUseTy;
4066 if (!SliceTy)
4067 if (Type *TypePartitionTy = getTypePartition(DL, AI.getAllocatedType(),
4068 P.beginOffset(), P.size()))
4069 SliceTy = TypePartitionTy;
4070 if ((!SliceTy || (SliceTy->isArrayTy() &&
4071 SliceTy->getArrayElementType()->isIntegerTy())) &&
4072 DL.isLegalInteger(P.size() * 8))
4073 SliceTy = Type::getIntNTy(*C, P.size() * 8);
4074 if (!SliceTy)
4075 SliceTy = ArrayType::get(Type::getInt8Ty(*C), P.size());
4076 assert(DL.getTypeAllocSize(SliceTy) >= P.size())((DL.getTypeAllocSize(SliceTy) >= P.size()) ? static_cast<
void> (0) : __assert_fail ("DL.getTypeAllocSize(SliceTy) >= P.size()"
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 4076, __PRETTY_FUNCTION__))
;
4077
4078 bool IsIntegerPromotable = isIntegerWideningViable(P, SliceTy, DL);
4079
4080 VectorType *VecTy =
4081 IsIntegerPromotable ? nullptr : isVectorPromotionViable(P, DL);
4082 if (VecTy)
4083 SliceTy = VecTy;
4084
4085 // Check for the case where we're going to rewrite to a new alloca of the
4086 // exact same type as the original, and with the same access offsets. In that
4087 // case, re-use the existing alloca, but still run through the rewriter to
4088 // perform phi and select speculation.
4089 // P.beginOffset() can be non-zero even with the same type in a case with
4090 // out-of-bounds access (e.g. @PR35657 function in SROA/basictest.ll).
4091 AllocaInst *NewAI;
4092 if (SliceTy == AI.getAllocatedType() && P.beginOffset() == 0) {
4093 NewAI = &AI;
4094 // FIXME: We should be able to bail at this point with "nothing changed".
4095 // FIXME: We might want to defer PHI speculation until after here.
4096 // FIXME: return nullptr;
4097 } else {
4098 unsigned Alignment = AI.getAlignment();
4099 if (!Alignment) {
4100 // The minimum alignment which users can rely on when the explicit
4101 // alignment is omitted or zero is that required by the ABI for this
4102 // type.
4103 Alignment = DL.getABITypeAlignment(AI.getAllocatedType());
4104 }
4105 Alignment = MinAlign(Alignment, P.beginOffset());
4106 // If we will get at least this much alignment from the type alone, leave
4107 // the alloca's alignment unconstrained.
4108 if (Alignment <= DL.getABITypeAlignment(SliceTy))
4109 Alignment = 0;
4110 NewAI = new AllocaInst(
4111 SliceTy, AI.getType()->getAddressSpace(), nullptr, Alignment,
4112 AI.getName() + ".sroa." + Twine(P.begin() - AS.begin()), &AI);
4113 // Copy the old AI debug location over to the new one.
4114 NewAI->setDebugLoc(AI.getDebugLoc());
4115 ++NumNewAllocas;
4116 }
4117
4118 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)
4119 << "[" << P.beginOffset() << "," << P.endOffset()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "Rewriting alloca partition " <<
"[" << P.beginOffset() << "," << P.endOffset
() << ") to: " << *NewAI << "\n"; } } while
(false)
4120 << ") to: " << *NewAI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "Rewriting alloca partition " <<
"[" << P.beginOffset() << "," << P.endOffset
() << ") to: " << *NewAI << "\n"; } } while
(false)
;
4121
4122 // Track the high watermark on the worklist as it is only relevant for
4123 // promoted allocas. We will reset it to this point if the alloca is not in
4124 // fact scheduled for promotion.
4125 unsigned PPWOldSize = PostPromotionWorklist.size();
4126 unsigned NumUses = 0;
4127 SmallSetVector<PHINode *, 8> PHIUsers;
4128 SmallSetVector<SelectInst *, 8> SelectUsers;
4129
4130 AllocaSliceRewriter Rewriter(DL, AS, *this, AI, *NewAI, P.beginOffset(),
4131 P.endOffset(), IsIntegerPromotable, VecTy,
4132 PHIUsers, SelectUsers);
4133 bool Promotable = true;
4134 for (Slice *S : P.splitSliceTails()) {
4135 Promotable &= Rewriter.visit(S);
4136 ++NumUses;
4137 }
4138 for (Slice &S : P) {
4139 Promotable &= Rewriter.visit(&S);
4140 ++NumUses;
4141 }
4142
4143 NumAllocaPartitionUses += NumUses;
4144 MaxUsesPerAllocaPartition.updateMax(NumUses);
4145
4146 // Now that we've processed all the slices in the new partition, check if any
4147 // PHIs or Selects would block promotion.
4148 for (PHINode *PHI : PHIUsers)
4149 if (!isSafePHIToSpeculate(*PHI)) {
4150 Promotable = false;
4151 PHIUsers.clear();
4152 SelectUsers.clear();
4153 break;
4154 }
4155
4156 for (SelectInst *Sel : SelectUsers)
4157 if (!isSafeSelectToSpeculate(*Sel)) {
4158 Promotable = false;
4159 PHIUsers.clear();
4160 SelectUsers.clear();
4161 break;
4162 }
4163
4164 if (Promotable) {
4165 if (PHIUsers.empty() && SelectUsers.empty()) {
4166 // Promote the alloca.
4167 PromotableAllocas.push_back(NewAI);
4168 } else {
4169 // If we have either PHIs or Selects to speculate, add them to those
4170 // worklists and re-queue the new alloca so that we promote in on the
4171 // next iteration.
4172 for (PHINode *PHIUser : PHIUsers)
4173 SpeculatablePHIs.insert(PHIUser);
4174 for (SelectInst *SelectUser : SelectUsers)
4175 SpeculatableSelects.insert(SelectUser);
4176 Worklist.insert(NewAI);
4177 }
4178 } else {
4179 // Drop any post-promotion work items if promotion didn't happen.
4180 while (PostPromotionWorklist.size() > PPWOldSize)
4181 PostPromotionWorklist.pop_back();
4182
4183 // We couldn't promote and we didn't create a new partition, nothing
4184 // happened.
4185 if (NewAI == &AI)
4186 return nullptr;
4187
4188 // If we can't promote the alloca, iterate on it to check for new
4189 // refinements exposed by splitting the current alloca. Don't iterate on an
4190 // alloca which didn't actually change and didn't get promoted.
4191 Worklist.insert(NewAI);
4192 }
4193
4194 return NewAI;
4195}
4196
4197/// Walks the slices of an alloca and form partitions based on them,
4198/// rewriting each of their uses.
4199bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &AS) {
4200 if (AS.begin() == AS.end())
4201 return false;
4202
4203 unsigned NumPartitions = 0;
4204 bool Changed = false;
4205 const DataLayout &DL = AI.getModule()->getDataLayout();
4206
4207 // First try to pre-split loads and stores.
4208 Changed |= presplitLoadsAndStores(AI, AS);
4209
4210 // Now that we have identified any pre-splitting opportunities,
4211 // mark loads and stores unsplittable except for the following case.
4212 // We leave a slice splittable if all other slices are disjoint or fully
4213 // included in the slice, such as whole-alloca loads and stores.
4214 // If we fail to split these during pre-splitting, we want to force them
4215 // to be rewritten into a partition.
4216 bool IsSorted = true;
4217
4218 uint64_t AllocaSize = DL.getTypeAllocSize(AI.getAllocatedType());
4219 const uint64_t MaxBitVectorSize = 1024;
4220 if (AllocaSize <= MaxBitVectorSize) {
4221 // If a byte boundary is included in any load or store, a slice starting or
4222 // ending at the boundary is not splittable.
4223 SmallBitVector SplittableOffset(AllocaSize + 1, true);
4224 for (Slice &S : AS)
4225 for (unsigned O = S.beginOffset() + 1;
4226 O < S.endOffset() && O < AllocaSize; O++)
4227 SplittableOffset.reset(O);
4228
4229 for (Slice &S : AS) {
4230 if (!S.isSplittable())
4231 continue;
4232
4233 if ((S.beginOffset() > AllocaSize || SplittableOffset[S.beginOffset()]) &&
4234 (S.endOffset() > AllocaSize || SplittableOffset[S.endOffset()]))
4235 continue;
4236
4237 if (isa<LoadInst>(S.getUse()->getUser()) ||
4238 isa<StoreInst>(S.getUse()->getUser())) {
4239 S.makeUnsplittable();
4240 IsSorted = false;
4241 }
4242 }
4243 }
4244 else {
4245 // We only allow whole-alloca splittable loads and stores
4246 // for a large alloca to avoid creating too large BitVector.
4247 for (Slice &S : AS) {
4248 if (!S.isSplittable())
4249 continue;
4250
4251 if (S.beginOffset() == 0 && S.endOffset() >= AllocaSize)
4252 continue;
4253
4254 if (isa<LoadInst>(S.getUse()->getUser()) ||
4255 isa<StoreInst>(S.getUse()->getUser())) {
4256 S.makeUnsplittable();
4257 IsSorted = false;
4258 }
4259 }
4260 }
4261
4262 if (!IsSorted)
4263 llvm::sort(AS);
4264
4265 /// Describes the allocas introduced by rewritePartition in order to migrate
4266 /// the debug info.
4267 struct Fragment {
4268 AllocaInst *Alloca;
4269 uint64_t Offset;
4270 uint64_t Size;
4271 Fragment(AllocaInst *AI, uint64_t O, uint64_t S)
4272 : Alloca(AI), Offset(O), Size(S) {}
4273 };
4274 SmallVector<Fragment, 4> Fragments;
4275
4276 // Rewrite each partition.
4277 for (auto &P : AS.partitions()) {
4278 if (AllocaInst *NewAI = rewritePartition(AI, AS, P)) {
4279 Changed = true;
4280 if (NewAI != &AI) {
4281 uint64_t SizeOfByte = 8;
4282 uint64_t AllocaSize = DL.getTypeSizeInBits(NewAI->getAllocatedType());
4283 // Don't include any padding.
4284 uint64_t Size = std::min(AllocaSize, P.size() * SizeOfByte);
4285 Fragments.push_back(Fragment(NewAI, P.beginOffset() * SizeOfByte, Size));
4286 }
4287 }
4288 ++NumPartitions;
4289 }
4290
4291 NumAllocaPartitions += NumPartitions;
4292 MaxPartitionsPerAlloca.updateMax(NumPartitions);
4293
4294 // Migrate debug information from the old alloca to the new alloca(s)
4295 // and the individual partitions.
4296 TinyPtrVector<DbgVariableIntrinsic *> DbgDeclares = FindDbgAddrUses(&AI);
4297 if (!DbgDeclares.empty()) {
4298 auto *Var = DbgDeclares.front()->getVariable();
4299 auto *Expr = DbgDeclares.front()->getExpression();
4300 auto VarSize = Var->getSizeInBits();
4301 DIBuilder DIB(*AI.getModule(), /*AllowUnresolved*/ false);
4302 uint64_t AllocaSize = DL.getTypeSizeInBits(AI.getAllocatedType());
4303 for (auto Fragment : Fragments) {
4304 // Create a fragment expression describing the new partition or reuse AI's
4305 // expression if there is only one partition.
4306 auto *FragmentExpr = Expr;
4307 if (Fragment.Size < AllocaSize || Expr->isFragment()) {
4308 // If this alloca is already a scalar replacement of a larger aggregate,
4309 // Fragment.Offset describes the offset inside the scalar.
4310 auto ExprFragment = Expr->getFragmentInfo();
4311 uint64_t Offset = ExprFragment ? ExprFragment->OffsetInBits : 0;
4312 uint64_t Start = Offset + Fragment.Offset;
4313 uint64_t Size = Fragment.Size;
4314 if (ExprFragment) {
4315 uint64_t AbsEnd =
4316 ExprFragment->OffsetInBits + ExprFragment->SizeInBits;
4317 if (Start >= AbsEnd)
4318 // No need to describe a SROAed padding.
4319 continue;
4320 Size = std::min(Size, AbsEnd - Start);
4321 }
4322 // The new, smaller fragment is stenciled out from the old fragment.
4323 if (auto OrigFragment = FragmentExpr->getFragmentInfo()) {
4324 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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 4325, __PRETTY_FUNCTION__))
4325 "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-9~svn362543/lib/Transforms/Scalar/SROA.cpp"
, 4325, __PRETTY_FUNCTION__))
;
4326 Start -= OrigFragment->OffsetInBits;
4327 }
4328
4329 // The alloca may be larger than the variable.
4330 if (VarSize) {
4331 if (Size > *VarSize)
4332 Size = *VarSize;
4333 if (Size == 0 || Start + Size > *VarSize)
4334 continue;
4335 }
4336
4337 // Avoid creating a fragment expression that covers the entire variable.
4338 if (!VarSize || *VarSize != Size) {
4339 if (auto E =
4340 DIExpression::createFragmentExpression(Expr, Start, Size))
4341 FragmentExpr = *E;
4342 else
4343 continue;
4344 }
4345 }
4346
4347 // Remove any existing intrinsics describing the same alloca.
4348 for (DbgVariableIntrinsic *OldDII : FindDbgAddrUses(Fragment.Alloca))
4349 OldDII->eraseFromParent();
4350
4351 DIB.insertDeclare(Fragment.Alloca, Var, FragmentExpr,
4352 DbgDeclares.front()->getDebugLoc(), &AI);
4353 }
4354 }
4355 return Changed;
4356}
4357
4358/// Clobber a use with undef, deleting the used value if it becomes dead.
4359void SROA::clobberUse(Use &U) {
4360 Value *OldV = U;
4361 // Replace the use with an undef value.
4362 U = UndefValue::get(OldV->getType());
4363
4364 // Check for this making an instruction dead. We have to garbage collect
4365 // all the dead instructions to ensure the uses of any alloca end up being
4366 // minimal.
4367 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
4368 if (isInstructionTriviallyDead(OldI)) {
4369 DeadInsts.insert(OldI);
4370 }
4371}
4372
4373/// Analyze an alloca for SROA.
4374///
4375/// This analyzes the alloca to ensure we can reason about it, builds
4376/// the slices of the alloca, and then hands it off to be split and
4377/// rewritten as needed.
4378bool SROA::runOnAlloca(AllocaInst &AI) {
4379 LLVM_DEBUG(dbgs() << "SROA alloca: " << AI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "SROA alloca: " << AI <<
"\n"; } } while (false)
;
4380 ++NumAllocasAnalyzed;
4381
4382 // Special case dead allocas, as they're trivial.
4383 if (AI.use_empty()) {
4384 AI.eraseFromParent();
4385 return true;
4386 }
4387 const DataLayout &DL = AI.getModule()->getDataLayout();
4388
4389 // Skip alloca forms that this analysis can't handle.
4390 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
4391 DL.getTypeAllocSize(AI.getAllocatedType()) == 0)
4392 return false;
4393
4394 bool Changed = false;
4395
4396 // First, split any FCA loads and stores touching this alloca to promote
4397 // better splitting and promotion opportunities.
4398 AggLoadStoreRewriter AggRewriter(DL);
4399 Changed |= AggRewriter.rewrite(AI);
4400
4401 // Build the slices using a recursive instruction-visiting builder.
4402 AllocaSlices AS(DL, AI);
4403 LLVM_DEBUG(AS.print(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { AS.print(dbgs()); } } while (false)
;
4404 if (AS.isEscaped())
4405 return Changed;
4406
4407 // Delete all the dead users of this alloca before splitting and rewriting it.
4408 for (Instruction *DeadUser : AS.getDeadUsers()) {
4409 // Free up everything used by this instruction.
4410 for (Use &DeadOp : DeadUser->operands())
4411 clobberUse(DeadOp);
4412
4413 // Now replace the uses of this instruction.
4414 DeadUser->replaceAllUsesWith(UndefValue::get(DeadUser->getType()));
4415
4416 // And mark it for deletion.
4417 DeadInsts.insert(DeadUser);
4418 Changed = true;
4419 }
4420 for (Use *DeadOp : AS.getDeadOperands()) {
4421 clobberUse(*DeadOp);
4422 Changed = true;
4423 }
4424
4425 // No slices to split. Leave the dead alloca for a later pass to clean up.
4426 if (AS.begin() == AS.end())
4427 return Changed;
4428
4429 Changed |= splitAlloca(AI, AS);
4430
4431 LLVM_DEBUG(dbgs() << " Speculating PHIs\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Speculating PHIs\n"; } } while
(false)
;
4432 while (!SpeculatablePHIs.empty())
4433 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
4434
4435 LLVM_DEBUG(dbgs() << " Speculating Selects\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << " Speculating Selects\n"; } } while
(false)
;
4436 while (!SpeculatableSelects.empty())
4437 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
4438
4439 return Changed;
4440}
4441
4442/// Delete the dead instructions accumulated in this run.
4443///
4444/// Recursively deletes the dead instructions we've accumulated. This is done
4445/// at the very end to maximize locality of the recursive delete and to
4446/// minimize the problems of invalidated instruction pointers as such pointers
4447/// are used heavily in the intermediate stages of the algorithm.
4448///
4449/// We also record the alloca instructions deleted here so that they aren't
4450/// subsequently handed to mem2reg to promote.
4451bool SROA::deleteDeadInstructions(
4452 SmallPtrSetImpl<AllocaInst *> &DeletedAllocas) {
4453 bool Changed = false;
4454 while (!DeadInsts.empty()) {
4455 Instruction *I = DeadInsts.pop_back_val();
4456 LLVM_DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "Deleting dead instruction: " <<
*I << "\n"; } } while (false)
;
4457
4458 // If the instruction is an alloca, find the possible dbg.declare connected
4459 // to it, and remove it too. We must do this before calling RAUW or we will
4460 // not be able to find it.
4461 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
4462 DeletedAllocas.insert(AI);
4463 for (DbgVariableIntrinsic *OldDII : FindDbgAddrUses(AI))
4464 OldDII->eraseFromParent();
4465 }
4466
4467 I->replaceAllUsesWith(UndefValue::get(I->getType()));
4468
4469 for (Use &Operand : I->operands())
4470 if (Instruction *U = dyn_cast<Instruction>(Operand)) {
4471 // Zero out the operand and see if it becomes trivially dead.
4472 Operand = nullptr;
4473 if (isInstructionTriviallyDead(U))
4474 DeadInsts.insert(U);
4475 }
4476
4477 ++NumDeleted;
4478 I->eraseFromParent();
4479 Changed = true;
4480 }
4481 return Changed;
4482}
4483
4484/// Promote the allocas, using the best available technique.
4485///
4486/// This attempts to promote whatever allocas have been identified as viable in
4487/// the PromotableAllocas list. If that list is empty, there is nothing to do.
4488/// This function returns whether any promotion occurred.
4489bool SROA::promoteAllocas(Function &F) {
4490 if (PromotableAllocas.empty())
4491 return false;
4492
4493 NumPromoted += PromotableAllocas.size();
4494
4495 LLVM_DEBUG(dbgs() << "Promoting allocas with mem2reg...\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "Promoting allocas with mem2reg...\n"
; } } while (false)
;
4496 PromoteMemToReg(PromotableAllocas, *DT, AC);
4497 PromotableAllocas.clear();
4498 return true;
4499}
4500
4501PreservedAnalyses SROA::runImpl(Function &F, DominatorTree &RunDT,
4502 AssumptionCache &RunAC) {
4503 LLVM_DEBUG(dbgs() << "SROA function: " << F.getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sroa")) { dbgs() << "SROA function: " << F.getName
() << "\n"; } } while (false)
;
4504 C = &F.getContext();
4505 DT = &RunDT;
4506 AC = &RunAC;
4507
4508 BasicBlock &EntryBB = F.getEntryBlock();
4509 for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
4510 I != E; ++I) {
4511 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
4512 Worklist.insert(AI);
4513 }
4514
4515 bool Changed = false;
4516 // A set of deleted alloca instruction pointers which should be removed from
4517 // the list of promotable allocas.
4518 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
4519
4520 do {
4521 while (!Worklist.empty()) {
4522 Changed |= runOnAlloca(*Worklist.pop_back_val());
4523 Changed |= deleteDeadInstructions(DeletedAllocas);
4524
4525 // Remove the deleted allocas from various lists so that we don't try to
4526 // continue processing them.
4527 if (!DeletedAllocas.empty()) {
4528 auto IsInSet = [&](AllocaInst *AI) { return DeletedAllocas.count(AI); };
4529 Worklist.remove_if(IsInSet);
4530 PostPromotionWorklist.remove_if(IsInSet);
4531 PromotableAllocas.erase(llvm::remove_if(PromotableAllocas, IsInSet),
4532 PromotableAllocas.end());
4533 DeletedAllocas.clear();
4534 }
4535 }
4536
4537 Changed |= promoteAllocas(F);
4538
4539 Worklist = PostPromotionWorklist;
4540 PostPromotionWorklist.clear();
4541 } while (!Worklist.empty());
4542
4543 if (!Changed)
4544 return PreservedAnalyses::all();
4545
4546 PreservedAnalyses PA;
4547 PA.preserveSet<CFGAnalyses>();
4548 PA.preserve<GlobalsAA>();
4549 return PA;
4550}
4551
4552PreservedAnalyses SROA::run(Function &F, FunctionAnalysisManager &AM) {
4553 return runImpl(F, AM.getResult<DominatorTreeAnalysis>(F),
4554 AM.getResult<AssumptionAnalysis>(F));
4555}
4556
4557/// A legacy pass for the legacy pass manager that wraps the \c SROA pass.
4558///
4559/// This is in the llvm namespace purely to allow it to be a friend of the \c
4560/// SROA pass.
4561class llvm::sroa::SROALegacyPass : public FunctionPass {
4562 /// The SROA implementation.
4563 SROA Impl;
4564
4565public:
4566 static char ID;
4567
4568 SROALegacyPass() : FunctionPass(ID) {
4569 initializeSROALegacyPassPass(*PassRegistry::getPassRegistry());
4570 }
4571
4572 bool runOnFunction(Function &F) override {
4573 if (skipFunction(F))
4574 return false;
4575
4576 auto PA = Impl.runImpl(
4577 F, getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
4578 getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F));
4579 return !PA.areAllPreserved();
4580 }
4581
4582 void getAnalysisUsage(AnalysisUsage &AU) const override {
4583 AU.addRequired<AssumptionCacheTracker>();
4584 AU.addRequired<DominatorTreeWrapperPass>();
4585 AU.addPreserved<GlobalsAAWrapperPass>();
4586 AU.setPreservesCFG();
4587 }
4588
4589 StringRef getPassName() const override { return "SROA"; }
4590};
4591
4592char SROALegacyPass::ID = 0;
4593
4594FunctionPass *llvm::createSROAPass() { return new SROALegacyPass(); }
4595
4596INITIALIZE_PASS_BEGIN(SROALegacyPass, "sroa",static void *initializeSROALegacyPassPassOnce(PassRegistry &
Registry) {
4597 "Scalar Replacement Of Aggregates", false, false)static void *initializeSROALegacyPassPassOnce(PassRegistry &
Registry) {
4598INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry);
4599INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);
4600INITIALIZE_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)); }
4601 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)); }