File: | llvm/lib/Transforms/Vectorize/LoopVectorize.cpp |
Warning: | line 8310, column 35 Potential leak of memory pointed to by 'BlockMask' |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===// |
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 | // |
9 | // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops |
10 | // and generates target-independent LLVM-IR. |
11 | // The vectorizer uses the TargetTransformInfo analysis to estimate the costs |
12 | // of instructions in order to estimate the profitability of vectorization. |
13 | // |
14 | // The loop vectorizer combines consecutive loop iterations into a single |
15 | // 'wide' iteration. After this transformation the index is incremented |
16 | // by the SIMD vector width, and not by one. |
17 | // |
18 | // This pass has three parts: |
19 | // 1. The main loop pass that drives the different parts. |
20 | // 2. LoopVectorizationLegality - A unit that checks for the legality |
21 | // of the vectorization. |
22 | // 3. InnerLoopVectorizer - A unit that performs the actual |
23 | // widening of instructions. |
24 | // 4. LoopVectorizationCostModel - A unit that checks for the profitability |
25 | // of vectorization. It decides on the optimal vector width, which |
26 | // can be one, if vectorization is not profitable. |
27 | // |
28 | // There is a development effort going on to migrate loop vectorizer to the |
29 | // VPlan infrastructure and to introduce outer loop vectorization support (see |
30 | // docs/Proposal/VectorizationPlan.rst and |
31 | // http://lists.llvm.org/pipermail/llvm-dev/2017-December/119523.html). For this |
32 | // purpose, we temporarily introduced the VPlan-native vectorization path: an |
33 | // alternative vectorization path that is natively implemented on top of the |
34 | // VPlan infrastructure. See EnableVPlanNativePath for enabling. |
35 | // |
36 | //===----------------------------------------------------------------------===// |
37 | // |
38 | // The reduction-variable vectorization is based on the paper: |
39 | // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization. |
40 | // |
41 | // Variable uniformity checks are inspired by: |
42 | // Karrenberg, R. and Hack, S. Whole Function Vectorization. |
43 | // |
44 | // The interleaved access vectorization is based on the paper: |
45 | // Dorit Nuzman, Ira Rosen and Ayal Zaks. Auto-Vectorization of Interleaved |
46 | // Data for SIMD |
47 | // |
48 | // Other ideas/concepts are from: |
49 | // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later. |
50 | // |
51 | // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of |
52 | // Vectorizing Compilers. |
53 | // |
54 | //===----------------------------------------------------------------------===// |
55 | |
56 | #include "llvm/Transforms/Vectorize/LoopVectorize.h" |
57 | #include "LoopVectorizationPlanner.h" |
58 | #include "VPRecipeBuilder.h" |
59 | #include "VPlan.h" |
60 | #include "VPlanHCFGBuilder.h" |
61 | #include "VPlanPredicator.h" |
62 | #include "VPlanTransforms.h" |
63 | #include "llvm/ADT/APInt.h" |
64 | #include "llvm/ADT/ArrayRef.h" |
65 | #include "llvm/ADT/DenseMap.h" |
66 | #include "llvm/ADT/DenseMapInfo.h" |
67 | #include "llvm/ADT/Hashing.h" |
68 | #include "llvm/ADT/MapVector.h" |
69 | #include "llvm/ADT/None.h" |
70 | #include "llvm/ADT/Optional.h" |
71 | #include "llvm/ADT/STLExtras.h" |
72 | #include "llvm/ADT/SmallPtrSet.h" |
73 | #include "llvm/ADT/SmallVector.h" |
74 | #include "llvm/ADT/Statistic.h" |
75 | #include "llvm/ADT/StringRef.h" |
76 | #include "llvm/ADT/Twine.h" |
77 | #include "llvm/ADT/iterator_range.h" |
78 | #include "llvm/Analysis/AssumptionCache.h" |
79 | #include "llvm/Analysis/BasicAliasAnalysis.h" |
80 | #include "llvm/Analysis/BlockFrequencyInfo.h" |
81 | #include "llvm/Analysis/CFG.h" |
82 | #include "llvm/Analysis/CodeMetrics.h" |
83 | #include "llvm/Analysis/DemandedBits.h" |
84 | #include "llvm/Analysis/GlobalsModRef.h" |
85 | #include "llvm/Analysis/LoopAccessAnalysis.h" |
86 | #include "llvm/Analysis/LoopAnalysisManager.h" |
87 | #include "llvm/Analysis/LoopInfo.h" |
88 | #include "llvm/Analysis/LoopIterator.h" |
89 | #include "llvm/Analysis/MemorySSA.h" |
90 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
91 | #include "llvm/Analysis/ProfileSummaryInfo.h" |
92 | #include "llvm/Analysis/ScalarEvolution.h" |
93 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" |
94 | #include "llvm/Analysis/TargetLibraryInfo.h" |
95 | #include "llvm/Analysis/TargetTransformInfo.h" |
96 | #include "llvm/Analysis/VectorUtils.h" |
97 | #include "llvm/IR/Attributes.h" |
98 | #include "llvm/IR/BasicBlock.h" |
99 | #include "llvm/IR/CFG.h" |
100 | #include "llvm/IR/Constant.h" |
101 | #include "llvm/IR/Constants.h" |
102 | #include "llvm/IR/DataLayout.h" |
103 | #include "llvm/IR/DebugInfoMetadata.h" |
104 | #include "llvm/IR/DebugLoc.h" |
105 | #include "llvm/IR/DerivedTypes.h" |
106 | #include "llvm/IR/DiagnosticInfo.h" |
107 | #include "llvm/IR/Dominators.h" |
108 | #include "llvm/IR/Function.h" |
109 | #include "llvm/IR/IRBuilder.h" |
110 | #include "llvm/IR/InstrTypes.h" |
111 | #include "llvm/IR/Instruction.h" |
112 | #include "llvm/IR/Instructions.h" |
113 | #include "llvm/IR/IntrinsicInst.h" |
114 | #include "llvm/IR/Intrinsics.h" |
115 | #include "llvm/IR/LLVMContext.h" |
116 | #include "llvm/IR/Metadata.h" |
117 | #include "llvm/IR/Module.h" |
118 | #include "llvm/IR/Operator.h" |
119 | #include "llvm/IR/Type.h" |
120 | #include "llvm/IR/Use.h" |
121 | #include "llvm/IR/User.h" |
122 | #include "llvm/IR/Value.h" |
123 | #include "llvm/IR/ValueHandle.h" |
124 | #include "llvm/IR/Verifier.h" |
125 | #include "llvm/InitializePasses.h" |
126 | #include "llvm/Pass.h" |
127 | #include "llvm/Support/Casting.h" |
128 | #include "llvm/Support/CommandLine.h" |
129 | #include "llvm/Support/Compiler.h" |
130 | #include "llvm/Support/Debug.h" |
131 | #include "llvm/Support/ErrorHandling.h" |
132 | #include "llvm/Support/InstructionCost.h" |
133 | #include "llvm/Support/MathExtras.h" |
134 | #include "llvm/Support/raw_ostream.h" |
135 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
136 | #include "llvm/Transforms/Utils/InjectTLIMappings.h" |
137 | #include "llvm/Transforms/Utils/LoopSimplify.h" |
138 | #include "llvm/Transforms/Utils/LoopUtils.h" |
139 | #include "llvm/Transforms/Utils/LoopVersioning.h" |
140 | #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" |
141 | #include "llvm/Transforms/Utils/SizeOpts.h" |
142 | #include "llvm/Transforms/Vectorize/LoopVectorizationLegality.h" |
143 | #include <algorithm> |
144 | #include <cassert> |
145 | #include <cstdint> |
146 | #include <cstdlib> |
147 | #include <functional> |
148 | #include <iterator> |
149 | #include <limits> |
150 | #include <memory> |
151 | #include <string> |
152 | #include <tuple> |
153 | #include <utility> |
154 | |
155 | using namespace llvm; |
156 | |
157 | #define LV_NAME"loop-vectorize" "loop-vectorize" |
158 | #define DEBUG_TYPE"loop-vectorize" LV_NAME"loop-vectorize" |
159 | |
160 | #ifndef NDEBUG |
161 | const char VerboseDebug[] = DEBUG_TYPE"loop-vectorize" "-verbose"; |
162 | #endif |
163 | |
164 | /// @{ |
165 | /// Metadata attribute names |
166 | const char LLVMLoopVectorizeFollowupAll[] = "llvm.loop.vectorize.followup_all"; |
167 | const char LLVMLoopVectorizeFollowupVectorized[] = |
168 | "llvm.loop.vectorize.followup_vectorized"; |
169 | const char LLVMLoopVectorizeFollowupEpilogue[] = |
170 | "llvm.loop.vectorize.followup_epilogue"; |
171 | /// @} |
172 | |
173 | STATISTIC(LoopsVectorized, "Number of loops vectorized")static llvm::Statistic LoopsVectorized = {"loop-vectorize", "LoopsVectorized" , "Number of loops vectorized"}; |
174 | STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization")static llvm::Statistic LoopsAnalyzed = {"loop-vectorize", "LoopsAnalyzed" , "Number of loops analyzed for vectorization"}; |
175 | STATISTIC(LoopsEpilogueVectorized, "Number of epilogues vectorized")static llvm::Statistic LoopsEpilogueVectorized = {"loop-vectorize" , "LoopsEpilogueVectorized", "Number of epilogues vectorized" }; |
176 | |
177 | static cl::opt<bool> EnableEpilogueVectorization( |
178 | "enable-epilogue-vectorization", cl::init(true), cl::Hidden, |
179 | cl::desc("Enable vectorization of epilogue loops.")); |
180 | |
181 | static cl::opt<unsigned> EpilogueVectorizationForceVF( |
182 | "epilogue-vectorization-force-VF", cl::init(1), cl::Hidden, |
183 | cl::desc("When epilogue vectorization is enabled, and a value greater than " |
184 | "1 is specified, forces the given VF for all applicable epilogue " |
185 | "loops.")); |
186 | |
187 | static cl::opt<unsigned> EpilogueVectorizationMinVF( |
188 | "epilogue-vectorization-minimum-VF", cl::init(16), cl::Hidden, |
189 | cl::desc("Only loops with vectorization factor equal to or larger than " |
190 | "the specified value are considered for epilogue vectorization.")); |
191 | |
192 | /// Loops with a known constant trip count below this number are vectorized only |
193 | /// if no scalar iteration overheads are incurred. |
194 | static cl::opt<unsigned> TinyTripCountVectorThreshold( |
195 | "vectorizer-min-trip-count", cl::init(16), cl::Hidden, |
196 | cl::desc("Loops with a constant trip count that is smaller than this " |
197 | "value are vectorized only if no scalar iteration overheads " |
198 | "are incurred.")); |
199 | |
200 | // Option prefer-predicate-over-epilogue indicates that an epilogue is undesired, |
201 | // that predication is preferred, and this lists all options. I.e., the |
202 | // vectorizer will try to fold the tail-loop (epilogue) into the vector body |
203 | // and predicate the instructions accordingly. If tail-folding fails, there are |
204 | // different fallback strategies depending on these values: |
205 | namespace PreferPredicateTy { |
206 | enum Option { |
207 | ScalarEpilogue = 0, |
208 | PredicateElseScalarEpilogue, |
209 | PredicateOrDontVectorize |
210 | }; |
211 | } // namespace PreferPredicateTy |
212 | |
213 | static cl::opt<PreferPredicateTy::Option> PreferPredicateOverEpilogue( |
214 | "prefer-predicate-over-epilogue", |
215 | cl::init(PreferPredicateTy::ScalarEpilogue), |
216 | cl::Hidden, |
217 | cl::desc("Tail-folding and predication preferences over creating a scalar " |
218 | "epilogue loop."), |
219 | cl::values(clEnumValN(PreferPredicateTy::ScalarEpilogue,llvm::cl::OptionEnumValue { "scalar-epilogue", int(PreferPredicateTy ::ScalarEpilogue), "Don't tail-predicate loops, create scalar epilogue" } |
220 | "scalar-epilogue",llvm::cl::OptionEnumValue { "scalar-epilogue", int(PreferPredicateTy ::ScalarEpilogue), "Don't tail-predicate loops, create scalar epilogue" } |
221 | "Don't tail-predicate loops, create scalar epilogue")llvm::cl::OptionEnumValue { "scalar-epilogue", int(PreferPredicateTy ::ScalarEpilogue), "Don't tail-predicate loops, create scalar epilogue" }, |
222 | clEnumValN(PreferPredicateTy::PredicateElseScalarEpilogue,llvm::cl::OptionEnumValue { "predicate-else-scalar-epilogue", int(PreferPredicateTy::PredicateElseScalarEpilogue), "prefer tail-folding, create scalar epilogue if tail " "folding fails." } |
223 | "predicate-else-scalar-epilogue",llvm::cl::OptionEnumValue { "predicate-else-scalar-epilogue", int(PreferPredicateTy::PredicateElseScalarEpilogue), "prefer tail-folding, create scalar epilogue if tail " "folding fails." } |
224 | "prefer tail-folding, create scalar epilogue if tail "llvm::cl::OptionEnumValue { "predicate-else-scalar-epilogue", int(PreferPredicateTy::PredicateElseScalarEpilogue), "prefer tail-folding, create scalar epilogue if tail " "folding fails." } |
225 | "folding fails.")llvm::cl::OptionEnumValue { "predicate-else-scalar-epilogue", int(PreferPredicateTy::PredicateElseScalarEpilogue), "prefer tail-folding, create scalar epilogue if tail " "folding fails." }, |
226 | clEnumValN(PreferPredicateTy::PredicateOrDontVectorize,llvm::cl::OptionEnumValue { "predicate-dont-vectorize", int(PreferPredicateTy ::PredicateOrDontVectorize), "prefers tail-folding, don't attempt vectorization if " "tail-folding fails." } |
227 | "predicate-dont-vectorize",llvm::cl::OptionEnumValue { "predicate-dont-vectorize", int(PreferPredicateTy ::PredicateOrDontVectorize), "prefers tail-folding, don't attempt vectorization if " "tail-folding fails." } |
228 | "prefers tail-folding, don't attempt vectorization if "llvm::cl::OptionEnumValue { "predicate-dont-vectorize", int(PreferPredicateTy ::PredicateOrDontVectorize), "prefers tail-folding, don't attempt vectorization if " "tail-folding fails." } |
229 | "tail-folding fails.")llvm::cl::OptionEnumValue { "predicate-dont-vectorize", int(PreferPredicateTy ::PredicateOrDontVectorize), "prefers tail-folding, don't attempt vectorization if " "tail-folding fails." })); |
230 | |
231 | static cl::opt<bool> MaximizeBandwidth( |
232 | "vectorizer-maximize-bandwidth", cl::init(false), cl::Hidden, |
233 | cl::desc("Maximize bandwidth when selecting vectorization factor which " |
234 | "will be determined by the smallest type in loop.")); |
235 | |
236 | static cl::opt<bool> EnableInterleavedMemAccesses( |
237 | "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden, |
238 | cl::desc("Enable vectorization on interleaved memory accesses in a loop")); |
239 | |
240 | /// An interleave-group may need masking if it resides in a block that needs |
241 | /// predication, or in order to mask away gaps. |
242 | static cl::opt<bool> EnableMaskedInterleavedMemAccesses( |
243 | "enable-masked-interleaved-mem-accesses", cl::init(false), cl::Hidden, |
244 | cl::desc("Enable vectorization on masked interleaved memory accesses in a loop")); |
245 | |
246 | static cl::opt<unsigned> TinyTripCountInterleaveThreshold( |
247 | "tiny-trip-count-interleave-threshold", cl::init(128), cl::Hidden, |
248 | cl::desc("We don't interleave loops with a estimated constant trip count " |
249 | "below this number")); |
250 | |
251 | static cl::opt<unsigned> ForceTargetNumScalarRegs( |
252 | "force-target-num-scalar-regs", cl::init(0), cl::Hidden, |
253 | cl::desc("A flag that overrides the target's number of scalar registers.")); |
254 | |
255 | static cl::opt<unsigned> ForceTargetNumVectorRegs( |
256 | "force-target-num-vector-regs", cl::init(0), cl::Hidden, |
257 | cl::desc("A flag that overrides the target's number of vector registers.")); |
258 | |
259 | static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor( |
260 | "force-target-max-scalar-interleave", cl::init(0), cl::Hidden, |
261 | cl::desc("A flag that overrides the target's max interleave factor for " |
262 | "scalar loops.")); |
263 | |
264 | static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor( |
265 | "force-target-max-vector-interleave", cl::init(0), cl::Hidden, |
266 | cl::desc("A flag that overrides the target's max interleave factor for " |
267 | "vectorized loops.")); |
268 | |
269 | static cl::opt<unsigned> ForceTargetInstructionCost( |
270 | "force-target-instruction-cost", cl::init(0), cl::Hidden, |
271 | cl::desc("A flag that overrides the target's expected cost for " |
272 | "an instruction to a single constant value. Mostly " |
273 | "useful for getting consistent testing.")); |
274 | |
275 | static cl::opt<bool> ForceTargetSupportsScalableVectors( |
276 | "force-target-supports-scalable-vectors", cl::init(false), cl::Hidden, |
277 | cl::desc( |
278 | "Pretend that scalable vectors are supported, even if the target does " |
279 | "not support them. This flag should only be used for testing.")); |
280 | |
281 | static cl::opt<unsigned> SmallLoopCost( |
282 | "small-loop-cost", cl::init(20), cl::Hidden, |
283 | cl::desc( |
284 | "The cost of a loop that is considered 'small' by the interleaver.")); |
285 | |
286 | static cl::opt<bool> LoopVectorizeWithBlockFrequency( |
287 | "loop-vectorize-with-block-frequency", cl::init(true), cl::Hidden, |
288 | cl::desc("Enable the use of the block frequency analysis to access PGO " |
289 | "heuristics minimizing code growth in cold regions and being more " |
290 | "aggressive in hot regions.")); |
291 | |
292 | // Runtime interleave loops for load/store throughput. |
293 | static cl::opt<bool> EnableLoadStoreRuntimeInterleave( |
294 | "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden, |
295 | cl::desc( |
296 | "Enable runtime interleaving until load/store ports are saturated")); |
297 | |
298 | /// Interleave small loops with scalar reductions. |
299 | static cl::opt<bool> InterleaveSmallLoopScalarReduction( |
300 | "interleave-small-loop-scalar-reduction", cl::init(false), cl::Hidden, |
301 | cl::desc("Enable interleaving for loops with small iteration counts that " |
302 | "contain scalar reductions to expose ILP.")); |
303 | |
304 | /// The number of stores in a loop that are allowed to need predication. |
305 | static cl::opt<unsigned> NumberOfStoresToPredicate( |
306 | "vectorize-num-stores-pred", cl::init(1), cl::Hidden, |
307 | cl::desc("Max number of stores to be predicated behind an if.")); |
308 | |
309 | static cl::opt<bool> EnableIndVarRegisterHeur( |
310 | "enable-ind-var-reg-heur", cl::init(true), cl::Hidden, |
311 | cl::desc("Count the induction variable only once when interleaving")); |
312 | |
313 | static cl::opt<bool> EnableCondStoresVectorization( |
314 | "enable-cond-stores-vec", cl::init(true), cl::Hidden, |
315 | cl::desc("Enable if predication of stores during vectorization.")); |
316 | |
317 | static cl::opt<unsigned> MaxNestedScalarReductionIC( |
318 | "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden, |
319 | cl::desc("The maximum interleave count to use when interleaving a scalar " |
320 | "reduction in a nested loop.")); |
321 | |
322 | static cl::opt<bool> |
323 | PreferInLoopReductions("prefer-inloop-reductions", cl::init(false), |
324 | cl::Hidden, |
325 | cl::desc("Prefer in-loop vector reductions, " |
326 | "overriding the targets preference.")); |
327 | |
328 | static cl::opt<bool> PreferPredicatedReductionSelect( |
329 | "prefer-predicated-reduction-select", cl::init(false), cl::Hidden, |
330 | cl::desc( |
331 | "Prefer predicating a reduction operation over an after loop select.")); |
332 | |
333 | cl::opt<bool> EnableVPlanNativePath( |
334 | "enable-vplan-native-path", cl::init(false), cl::Hidden, |
335 | cl::desc("Enable VPlan-native vectorization path with " |
336 | "support for outer loop vectorization.")); |
337 | |
338 | // FIXME: Remove this switch once we have divergence analysis. Currently we |
339 | // assume divergent non-backedge branches when this switch is true. |
340 | cl::opt<bool> EnableVPlanPredication( |
341 | "enable-vplan-predication", cl::init(false), cl::Hidden, |
342 | cl::desc("Enable VPlan-native vectorization path predicator with " |
343 | "support for outer loop vectorization.")); |
344 | |
345 | // This flag enables the stress testing of the VPlan H-CFG construction in the |
346 | // VPlan-native vectorization path. It must be used in conjuction with |
347 | // -enable-vplan-native-path. -vplan-verify-hcfg can also be used to enable the |
348 | // verification of the H-CFGs built. |
349 | static cl::opt<bool> VPlanBuildStressTest( |
350 | "vplan-build-stress-test", cl::init(false), cl::Hidden, |
351 | cl::desc( |
352 | "Build VPlan for every supported loop nest in the function and bail " |
353 | "out right after the build (stress test the VPlan H-CFG construction " |
354 | "in the VPlan-native vectorization path).")); |
355 | |
356 | cl::opt<bool> llvm::EnableLoopInterleaving( |
357 | "interleave-loops", cl::init(true), cl::Hidden, |
358 | cl::desc("Enable loop interleaving in Loop vectorization passes")); |
359 | cl::opt<bool> llvm::EnableLoopVectorization( |
360 | "vectorize-loops", cl::init(true), cl::Hidden, |
361 | cl::desc("Run the Loop vectorization passes")); |
362 | |
363 | /// A helper function that returns the type of loaded or stored value. |
364 | static Type *getMemInstValueType(Value *I) { |
365 | assert((isa<LoadInst>(I) || isa<StoreInst>(I)) &&(((isa<LoadInst>(I) || isa<StoreInst>(I)) && "Expected Load or Store instruction") ? static_cast<void> (0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Expected Load or Store instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 366, __PRETTY_FUNCTION__)) |
366 | "Expected Load or Store instruction")(((isa<LoadInst>(I) || isa<StoreInst>(I)) && "Expected Load or Store instruction") ? static_cast<void> (0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Expected Load or Store instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 366, __PRETTY_FUNCTION__)); |
367 | if (auto *LI = dyn_cast<LoadInst>(I)) |
368 | return LI->getType(); |
369 | return cast<StoreInst>(I)->getValueOperand()->getType(); |
370 | } |
371 | |
372 | /// A helper function that returns true if the given type is irregular. The |
373 | /// type is irregular if its allocated size doesn't equal the store size of an |
374 | /// element of the corresponding vector type at the given vectorization factor. |
375 | static bool hasIrregularType(Type *Ty, const DataLayout &DL, ElementCount VF) { |
376 | // Determine if an array of VF elements of type Ty is "bitcast compatible" |
377 | // with a <VF x Ty> vector. |
378 | if (VF.isVector()) { |
379 | auto *VectorTy = VectorType::get(Ty, VF); |
380 | return TypeSize::get(VF.getKnownMinValue() * |
381 | DL.getTypeAllocSize(Ty).getFixedValue(), |
382 | VF.isScalable()) != DL.getTypeStoreSize(VectorTy); |
383 | } |
384 | |
385 | // If the vectorization factor is one, we just check if an array of type Ty |
386 | // requires padding between elements. |
387 | return DL.getTypeAllocSizeInBits(Ty) != DL.getTypeSizeInBits(Ty); |
388 | } |
389 | |
390 | /// A helper function that returns the reciprocal of the block probability of |
391 | /// predicated blocks. If we return X, we are assuming the predicated block |
392 | /// will execute once for every X iterations of the loop header. |
393 | /// |
394 | /// TODO: We should use actual block probability here, if available. Currently, |
395 | /// we always assume predicated blocks have a 50% chance of executing. |
396 | static unsigned getReciprocalPredBlockProb() { return 2; } |
397 | |
398 | /// A helper function that returns an integer or floating-point constant with |
399 | /// value C. |
400 | static Constant *getSignedIntOrFpConstant(Type *Ty, int64_t C) { |
401 | return Ty->isIntegerTy() ? ConstantInt::getSigned(Ty, C) |
402 | : ConstantFP::get(Ty, C); |
403 | } |
404 | |
405 | /// Returns "best known" trip count for the specified loop \p L as defined by |
406 | /// the following procedure: |
407 | /// 1) Returns exact trip count if it is known. |
408 | /// 2) Returns expected trip count according to profile data if any. |
409 | /// 3) Returns upper bound estimate if it is known. |
410 | /// 4) Returns None if all of the above failed. |
411 | static Optional<unsigned> getSmallBestKnownTC(ScalarEvolution &SE, Loop *L) { |
412 | // Check if exact trip count is known. |
413 | if (unsigned ExpectedTC = SE.getSmallConstantTripCount(L)) |
414 | return ExpectedTC; |
415 | |
416 | // Check if there is an expected trip count available from profile data. |
417 | if (LoopVectorizeWithBlockFrequency) |
418 | if (auto EstimatedTC = getLoopEstimatedTripCount(L)) |
419 | return EstimatedTC; |
420 | |
421 | // Check if upper bound estimate is known. |
422 | if (unsigned ExpectedTC = SE.getSmallConstantMaxTripCount(L)) |
423 | return ExpectedTC; |
424 | |
425 | return None; |
426 | } |
427 | |
428 | // Forward declare GeneratedRTChecks. |
429 | class GeneratedRTChecks; |
430 | |
431 | namespace llvm { |
432 | |
433 | /// InnerLoopVectorizer vectorizes loops which contain only one basic |
434 | /// block to a specified vectorization factor (VF). |
435 | /// This class performs the widening of scalars into vectors, or multiple |
436 | /// scalars. This class also implements the following features: |
437 | /// * It inserts an epilogue loop for handling loops that don't have iteration |
438 | /// counts that are known to be a multiple of the vectorization factor. |
439 | /// * It handles the code generation for reduction variables. |
440 | /// * Scalarization (implementation using scalars) of un-vectorizable |
441 | /// instructions. |
442 | /// InnerLoopVectorizer does not perform any vectorization-legality |
443 | /// checks, and relies on the caller to check for the different legality |
444 | /// aspects. The InnerLoopVectorizer relies on the |
445 | /// LoopVectorizationLegality class to provide information about the induction |
446 | /// and reduction variables that were found to a given vectorization factor. |
447 | class InnerLoopVectorizer { |
448 | public: |
449 | InnerLoopVectorizer(Loop *OrigLoop, PredicatedScalarEvolution &PSE, |
450 | LoopInfo *LI, DominatorTree *DT, |
451 | const TargetLibraryInfo *TLI, |
452 | const TargetTransformInfo *TTI, AssumptionCache *AC, |
453 | OptimizationRemarkEmitter *ORE, ElementCount VecWidth, |
454 | unsigned UnrollFactor, LoopVectorizationLegality *LVL, |
455 | LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI, |
456 | ProfileSummaryInfo *PSI, GeneratedRTChecks &RTChecks) |
457 | : OrigLoop(OrigLoop), PSE(PSE), LI(LI), DT(DT), TLI(TLI), TTI(TTI), |
458 | AC(AC), ORE(ORE), VF(VecWidth), UF(UnrollFactor), |
459 | Builder(PSE.getSE()->getContext()), Legal(LVL), Cost(CM), BFI(BFI), |
460 | PSI(PSI), RTChecks(RTChecks) { |
461 | // Query this against the original loop and save it here because the profile |
462 | // of the original loop header may change as the transformation happens. |
463 | OptForSizeBasedOnProfile = llvm::shouldOptimizeForSize( |
464 | OrigLoop->getHeader(), PSI, BFI, PGSOQueryType::IRPass); |
465 | } |
466 | |
467 | virtual ~InnerLoopVectorizer() = default; |
468 | |
469 | /// Create a new empty loop that will contain vectorized instructions later |
470 | /// on, while the old loop will be used as the scalar remainder. Control flow |
471 | /// is generated around the vectorized (and scalar epilogue) loops consisting |
472 | /// of various checks and bypasses. Return the pre-header block of the new |
473 | /// loop. |
474 | /// In the case of epilogue vectorization, this function is overriden to |
475 | /// handle the more complex control flow around the loops. |
476 | virtual BasicBlock *createVectorizedLoopSkeleton(); |
477 | |
478 | /// Widen a single instruction within the innermost loop. |
479 | void widenInstruction(Instruction &I, VPValue *Def, VPUser &Operands, |
480 | VPTransformState &State); |
481 | |
482 | /// Widen a single call instruction within the innermost loop. |
483 | void widenCallInstruction(CallInst &I, VPValue *Def, VPUser &ArgOperands, |
484 | VPTransformState &State); |
485 | |
486 | /// Widen a single select instruction within the innermost loop. |
487 | void widenSelectInstruction(SelectInst &I, VPValue *VPDef, VPUser &Operands, |
488 | bool InvariantCond, VPTransformState &State); |
489 | |
490 | /// Fix the vectorized code, taking care of header phi's, live-outs, and more. |
491 | void fixVectorizedLoop(VPTransformState &State); |
492 | |
493 | // Return true if any runtime check is added. |
494 | bool areSafetyChecksAdded() { return AddedSafetyChecks; } |
495 | |
496 | /// A type for vectorized values in the new loop. Each value from the |
497 | /// original loop, when vectorized, is represented by UF vector values in the |
498 | /// new unrolled loop, where UF is the unroll factor. |
499 | using VectorParts = SmallVector<Value *, 2>; |
500 | |
501 | /// Vectorize a single GetElementPtrInst based on information gathered and |
502 | /// decisions taken during planning. |
503 | void widenGEP(GetElementPtrInst *GEP, VPValue *VPDef, VPUser &Indices, |
504 | unsigned UF, ElementCount VF, bool IsPtrLoopInvariant, |
505 | SmallBitVector &IsIndexLoopInvariant, VPTransformState &State); |
506 | |
507 | /// Vectorize a single PHINode in a block. This method handles the induction |
508 | /// variable canonicalization. It supports both VF = 1 for unrolled loops and |
509 | /// arbitrary length vectors. |
510 | void widenPHIInstruction(Instruction *PN, RecurrenceDescriptor *RdxDesc, |
511 | VPValue *StartV, VPValue *Def, |
512 | VPTransformState &State); |
513 | |
514 | /// A helper function to scalarize a single Instruction in the innermost loop. |
515 | /// Generates a sequence of scalar instances for each lane between \p MinLane |
516 | /// and \p MaxLane, times each part between \p MinPart and \p MaxPart, |
517 | /// inclusive. Uses the VPValue operands from \p Operands instead of \p |
518 | /// Instr's operands. |
519 | void scalarizeInstruction(Instruction *Instr, VPValue *Def, VPUser &Operands, |
520 | const VPIteration &Instance, bool IfPredicateInstr, |
521 | VPTransformState &State); |
522 | |
523 | /// Widen an integer or floating-point induction variable \p IV. If \p Trunc |
524 | /// is provided, the integer induction variable will first be truncated to |
525 | /// the corresponding type. |
526 | void widenIntOrFpInduction(PHINode *IV, Value *Start, TruncInst *Trunc, |
527 | VPValue *Def, VPValue *CastDef, |
528 | VPTransformState &State); |
529 | |
530 | /// Construct the vector value of a scalarized value \p V one lane at a time. |
531 | void packScalarIntoVectorValue(VPValue *Def, const VPIteration &Instance, |
532 | VPTransformState &State); |
533 | |
534 | /// Try to vectorize interleaved access group \p Group with the base address |
535 | /// given in \p Addr, optionally masking the vector operations if \p |
536 | /// BlockInMask is non-null. Use \p State to translate given VPValues to IR |
537 | /// values in the vectorized loop. |
538 | void vectorizeInterleaveGroup(const InterleaveGroup<Instruction> *Group, |
539 | ArrayRef<VPValue *> VPDefs, |
540 | VPTransformState &State, VPValue *Addr, |
541 | ArrayRef<VPValue *> StoredValues, |
542 | VPValue *BlockInMask = nullptr); |
543 | |
544 | /// Vectorize Load and Store instructions with the base address given in \p |
545 | /// Addr, optionally masking the vector operations if \p BlockInMask is |
546 | /// non-null. Use \p State to translate given VPValues to IR values in the |
547 | /// vectorized loop. |
548 | void vectorizeMemoryInstruction(Instruction *Instr, VPTransformState &State, |
549 | VPValue *Def, VPValue *Addr, |
550 | VPValue *StoredValue, VPValue *BlockInMask); |
551 | |
552 | /// Set the debug location in the builder using the debug location in |
553 | /// the instruction. |
554 | void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr); |
555 | |
556 | /// Fix the non-induction PHIs in the OrigPHIsToFix vector. |
557 | void fixNonInductionPHIs(VPTransformState &State); |
558 | |
559 | /// Create a broadcast instruction. This method generates a broadcast |
560 | /// instruction (shuffle) for loop invariant values and for the induction |
561 | /// value. If this is the induction variable then we extend it to N, N+1, ... |
562 | /// this is needed because each iteration in the loop corresponds to a SIMD |
563 | /// element. |
564 | virtual Value *getBroadcastInstrs(Value *V); |
565 | |
566 | protected: |
567 | friend class LoopVectorizationPlanner; |
568 | |
569 | /// A small list of PHINodes. |
570 | using PhiVector = SmallVector<PHINode *, 4>; |
571 | |
572 | /// A type for scalarized values in the new loop. Each value from the |
573 | /// original loop, when scalarized, is represented by UF x VF scalar values |
574 | /// in the new unrolled loop, where UF is the unroll factor and VF is the |
575 | /// vectorization factor. |
576 | using ScalarParts = SmallVector<SmallVector<Value *, 4>, 2>; |
577 | |
578 | /// Set up the values of the IVs correctly when exiting the vector loop. |
579 | void fixupIVUsers(PHINode *OrigPhi, const InductionDescriptor &II, |
580 | Value *CountRoundDown, Value *EndValue, |
581 | BasicBlock *MiddleBlock); |
582 | |
583 | /// Create a new induction variable inside L. |
584 | PHINode *createInductionVariable(Loop *L, Value *Start, Value *End, |
585 | Value *Step, Instruction *DL); |
586 | |
587 | /// Handle all cross-iteration phis in the header. |
588 | void fixCrossIterationPHIs(VPTransformState &State); |
589 | |
590 | /// Fix a first-order recurrence. This is the second phase of vectorizing |
591 | /// this phi node. |
592 | void fixFirstOrderRecurrence(PHINode *Phi, VPTransformState &State); |
593 | |
594 | /// Fix a reduction cross-iteration phi. This is the second phase of |
595 | /// vectorizing this phi node. |
596 | void fixReduction(PHINode *Phi, VPTransformState &State); |
597 | |
598 | /// Clear NSW/NUW flags from reduction instructions if necessary. |
599 | void clearReductionWrapFlags(RecurrenceDescriptor &RdxDesc, |
600 | VPTransformState &State); |
601 | |
602 | /// Fixup the LCSSA phi nodes in the unique exit block. This simply |
603 | /// means we need to add the appropriate incoming value from the middle |
604 | /// block as exiting edges from the scalar epilogue loop (if present) are |
605 | /// already in place, and we exit the vector loop exclusively to the middle |
606 | /// block. |
607 | void fixLCSSAPHIs(VPTransformState &State); |
608 | |
609 | /// Iteratively sink the scalarized operands of a predicated instruction into |
610 | /// the block that was created for it. |
611 | void sinkScalarOperands(Instruction *PredInst); |
612 | |
613 | /// Shrinks vector element sizes to the smallest bitwidth they can be legally |
614 | /// represented as. |
615 | void truncateToMinimalBitwidths(VPTransformState &State); |
616 | |
617 | /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...) |
618 | /// to each vector element of Val. The sequence starts at StartIndex. |
619 | /// \p Opcode is relevant for FP induction variable. |
620 | virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step, |
621 | Instruction::BinaryOps Opcode = |
622 | Instruction::BinaryOpsEnd); |
623 | |
624 | /// Compute scalar induction steps. \p ScalarIV is the scalar induction |
625 | /// variable on which to base the steps, \p Step is the size of the step, and |
626 | /// \p EntryVal is the value from the original loop that maps to the steps. |
627 | /// Note that \p EntryVal doesn't have to be an induction variable - it |
628 | /// can also be a truncate instruction. |
629 | void buildScalarSteps(Value *ScalarIV, Value *Step, Instruction *EntryVal, |
630 | const InductionDescriptor &ID, VPValue *Def, |
631 | VPValue *CastDef, VPTransformState &State); |
632 | |
633 | /// Create a vector induction phi node based on an existing scalar one. \p |
634 | /// EntryVal is the value from the original loop that maps to the vector phi |
635 | /// node, and \p Step is the loop-invariant step. If \p EntryVal is a |
636 | /// truncate instruction, instead of widening the original IV, we widen a |
637 | /// version of the IV truncated to \p EntryVal's type. |
638 | void createVectorIntOrFpInductionPHI(const InductionDescriptor &II, |
639 | Value *Step, Value *Start, |
640 | Instruction *EntryVal, VPValue *Def, |
641 | VPValue *CastDef, |
642 | VPTransformState &State); |
643 | |
644 | /// Returns true if an instruction \p I should be scalarized instead of |
645 | /// vectorized for the chosen vectorization factor. |
646 | bool shouldScalarizeInstruction(Instruction *I) const; |
647 | |
648 | /// Returns true if we should generate a scalar version of \p IV. |
649 | bool needsScalarInduction(Instruction *IV) const; |
650 | |
651 | /// If there is a cast involved in the induction variable \p ID, which should |
652 | /// be ignored in the vectorized loop body, this function records the |
653 | /// VectorLoopValue of the respective Phi also as the VectorLoopValue of the |
654 | /// cast. We had already proved that the casted Phi is equal to the uncasted |
655 | /// Phi in the vectorized loop (under a runtime guard), and therefore |
656 | /// there is no need to vectorize the cast - the same value can be used in the |
657 | /// vector loop for both the Phi and the cast. |
658 | /// If \p VectorLoopValue is a scalarized value, \p Lane is also specified, |
659 | /// Otherwise, \p VectorLoopValue is a widened/vectorized value. |
660 | /// |
661 | /// \p EntryVal is the value from the original loop that maps to the vector |
662 | /// phi node and is used to distinguish what is the IV currently being |
663 | /// processed - original one (if \p EntryVal is a phi corresponding to the |
664 | /// original IV) or the "newly-created" one based on the proof mentioned above |
665 | /// (see also buildScalarSteps() and createVectorIntOrFPInductionPHI()). In the |
666 | /// latter case \p EntryVal is a TruncInst and we must not record anything for |
667 | /// that IV, but it's error-prone to expect callers of this routine to care |
668 | /// about that, hence this explicit parameter. |
669 | void recordVectorLoopValueForInductionCast( |
670 | const InductionDescriptor &ID, const Instruction *EntryVal, |
671 | Value *VectorLoopValue, VPValue *CastDef, VPTransformState &State, |
672 | unsigned Part, unsigned Lane = UINT_MAX(2147483647 *2U +1U)); |
673 | |
674 | /// Generate a shuffle sequence that will reverse the vector Vec. |
675 | virtual Value *reverseVector(Value *Vec); |
676 | |
677 | /// Returns (and creates if needed) the original loop trip count. |
678 | Value *getOrCreateTripCount(Loop *NewLoop); |
679 | |
680 | /// Returns (and creates if needed) the trip count of the widened loop. |
681 | Value *getOrCreateVectorTripCount(Loop *NewLoop); |
682 | |
683 | /// Returns a bitcasted value to the requested vector type. |
684 | /// Also handles bitcasts of vector<float> <-> vector<pointer> types. |
685 | Value *createBitOrPointerCast(Value *V, VectorType *DstVTy, |
686 | const DataLayout &DL); |
687 | |
688 | /// Emit a bypass check to see if the vector trip count is zero, including if |
689 | /// it overflows. |
690 | void emitMinimumIterationCountCheck(Loop *L, BasicBlock *Bypass); |
691 | |
692 | /// Emit a bypass check to see if all of the SCEV assumptions we've |
693 | /// had to make are correct. Returns the block containing the checks or |
694 | /// nullptr if no checks have been added. |
695 | BasicBlock *emitSCEVChecks(Loop *L, BasicBlock *Bypass); |
696 | |
697 | /// Emit bypass checks to check any memory assumptions we may have made. |
698 | /// Returns the block containing the checks or nullptr if no checks have been |
699 | /// added. |
700 | BasicBlock *emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass); |
701 | |
702 | /// Compute the transformed value of Index at offset StartValue using step |
703 | /// StepValue. |
704 | /// For integer induction, returns StartValue + Index * StepValue. |
705 | /// For pointer induction, returns StartValue[Index * StepValue]. |
706 | /// FIXME: The newly created binary instructions should contain nsw/nuw |
707 | /// flags, which can be found from the original scalar operations. |
708 | Value *emitTransformedIndex(IRBuilder<> &B, Value *Index, ScalarEvolution *SE, |
709 | const DataLayout &DL, |
710 | const InductionDescriptor &ID) const; |
711 | |
712 | /// Emit basic blocks (prefixed with \p Prefix) for the iteration check, |
713 | /// vector loop preheader, middle block and scalar preheader. Also |
714 | /// allocate a loop object for the new vector loop and return it. |
715 | Loop *createVectorLoopSkeleton(StringRef Prefix); |
716 | |
717 | /// Create new phi nodes for the induction variables to resume iteration count |
718 | /// in the scalar epilogue, from where the vectorized loop left off (given by |
719 | /// \p VectorTripCount). |
720 | /// In cases where the loop skeleton is more complicated (eg. epilogue |
721 | /// vectorization) and the resume values can come from an additional bypass |
722 | /// block, the \p AdditionalBypass pair provides information about the bypass |
723 | /// block and the end value on the edge from bypass to this loop. |
724 | void createInductionResumeValues( |
725 | Loop *L, Value *VectorTripCount, |
726 | std::pair<BasicBlock *, Value *> AdditionalBypass = {nullptr, nullptr}); |
727 | |
728 | /// Complete the loop skeleton by adding debug MDs, creating appropriate |
729 | /// conditional branches in the middle block, preparing the builder and |
730 | /// running the verifier. Take in the vector loop \p L as argument, and return |
731 | /// the preheader of the completed vector loop. |
732 | BasicBlock *completeLoopSkeleton(Loop *L, MDNode *OrigLoopID); |
733 | |
734 | /// Add additional metadata to \p To that was not present on \p Orig. |
735 | /// |
736 | /// Currently this is used to add the noalias annotations based on the |
737 | /// inserted memchecks. Use this for instructions that are *cloned* into the |
738 | /// vector loop. |
739 | void addNewMetadata(Instruction *To, const Instruction *Orig); |
740 | |
741 | /// Add metadata from one instruction to another. |
742 | /// |
743 | /// This includes both the original MDs from \p From and additional ones (\see |
744 | /// addNewMetadata). Use this for *newly created* instructions in the vector |
745 | /// loop. |
746 | void addMetadata(Instruction *To, Instruction *From); |
747 | |
748 | /// Similar to the previous function but it adds the metadata to a |
749 | /// vector of instructions. |
750 | void addMetadata(ArrayRef<Value *> To, Instruction *From); |
751 | |
752 | /// Allow subclasses to override and print debug traces before/after vplan |
753 | /// execution, when trace information is requested. |
754 | virtual void printDebugTracesAtStart(){}; |
755 | virtual void printDebugTracesAtEnd(){}; |
756 | |
757 | /// The original loop. |
758 | Loop *OrigLoop; |
759 | |
760 | /// A wrapper around ScalarEvolution used to add runtime SCEV checks. Applies |
761 | /// dynamic knowledge to simplify SCEV expressions and converts them to a |
762 | /// more usable form. |
763 | PredicatedScalarEvolution &PSE; |
764 | |
765 | /// Loop Info. |
766 | LoopInfo *LI; |
767 | |
768 | /// Dominator Tree. |
769 | DominatorTree *DT; |
770 | |
771 | /// Alias Analysis. |
772 | AAResults *AA; |
773 | |
774 | /// Target Library Info. |
775 | const TargetLibraryInfo *TLI; |
776 | |
777 | /// Target Transform Info. |
778 | const TargetTransformInfo *TTI; |
779 | |
780 | /// Assumption Cache. |
781 | AssumptionCache *AC; |
782 | |
783 | /// Interface to emit optimization remarks. |
784 | OptimizationRemarkEmitter *ORE; |
785 | |
786 | /// LoopVersioning. It's only set up (non-null) if memchecks were |
787 | /// used. |
788 | /// |
789 | /// This is currently only used to add no-alias metadata based on the |
790 | /// memchecks. The actually versioning is performed manually. |
791 | std::unique_ptr<LoopVersioning> LVer; |
792 | |
793 | /// The vectorization SIMD factor to use. Each vector will have this many |
794 | /// vector elements. |
795 | ElementCount VF; |
796 | |
797 | /// The vectorization unroll factor to use. Each scalar is vectorized to this |
798 | /// many different vector instructions. |
799 | unsigned UF; |
800 | |
801 | /// The builder that we use |
802 | IRBuilder<> Builder; |
803 | |
804 | // --- Vectorization state --- |
805 | |
806 | /// The vector-loop preheader. |
807 | BasicBlock *LoopVectorPreHeader; |
808 | |
809 | /// The scalar-loop preheader. |
810 | BasicBlock *LoopScalarPreHeader; |
811 | |
812 | /// Middle Block between the vector and the scalar. |
813 | BasicBlock *LoopMiddleBlock; |
814 | |
815 | /// The (unique) ExitBlock of the scalar loop. Note that |
816 | /// there can be multiple exiting edges reaching this block. |
817 | BasicBlock *LoopExitBlock; |
818 | |
819 | /// The vector loop body. |
820 | BasicBlock *LoopVectorBody; |
821 | |
822 | /// The scalar loop body. |
823 | BasicBlock *LoopScalarBody; |
824 | |
825 | /// A list of all bypass blocks. The first block is the entry of the loop. |
826 | SmallVector<BasicBlock *, 4> LoopBypassBlocks; |
827 | |
828 | /// The new Induction variable which was added to the new block. |
829 | PHINode *Induction = nullptr; |
830 | |
831 | /// The induction variable of the old basic block. |
832 | PHINode *OldInduction = nullptr; |
833 | |
834 | /// Store instructions that were predicated. |
835 | SmallVector<Instruction *, 4> PredicatedInstructions; |
836 | |
837 | /// Trip count of the original loop. |
838 | Value *TripCount = nullptr; |
839 | |
840 | /// Trip count of the widened loop (TripCount - TripCount % (VF*UF)) |
841 | Value *VectorTripCount = nullptr; |
842 | |
843 | /// The legality analysis. |
844 | LoopVectorizationLegality *Legal; |
845 | |
846 | /// The profitablity analysis. |
847 | LoopVectorizationCostModel *Cost; |
848 | |
849 | // Record whether runtime checks are added. |
850 | bool AddedSafetyChecks = false; |
851 | |
852 | // Holds the end values for each induction variable. We save the end values |
853 | // so we can later fix-up the external users of the induction variables. |
854 | DenseMap<PHINode *, Value *> IVEndValues; |
855 | |
856 | // Vector of original scalar PHIs whose corresponding widened PHIs need to be |
857 | // fixed up at the end of vector code generation. |
858 | SmallVector<PHINode *, 8> OrigPHIsToFix; |
859 | |
860 | /// BFI and PSI are used to check for profile guided size optimizations. |
861 | BlockFrequencyInfo *BFI; |
862 | ProfileSummaryInfo *PSI; |
863 | |
864 | // Whether this loop should be optimized for size based on profile guided size |
865 | // optimizatios. |
866 | bool OptForSizeBasedOnProfile; |
867 | |
868 | /// Structure to hold information about generated runtime checks, responsible |
869 | /// for cleaning the checks, if vectorization turns out unprofitable. |
870 | GeneratedRTChecks &RTChecks; |
871 | }; |
872 | |
873 | class InnerLoopUnroller : public InnerLoopVectorizer { |
874 | public: |
875 | InnerLoopUnroller(Loop *OrigLoop, PredicatedScalarEvolution &PSE, |
876 | LoopInfo *LI, DominatorTree *DT, |
877 | const TargetLibraryInfo *TLI, |
878 | const TargetTransformInfo *TTI, AssumptionCache *AC, |
879 | OptimizationRemarkEmitter *ORE, unsigned UnrollFactor, |
880 | LoopVectorizationLegality *LVL, |
881 | LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI, |
882 | ProfileSummaryInfo *PSI, GeneratedRTChecks &Check) |
883 | : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, |
884 | ElementCount::getFixed(1), UnrollFactor, LVL, CM, |
885 | BFI, PSI, Check) {} |
886 | |
887 | private: |
888 | Value *getBroadcastInstrs(Value *V) override; |
889 | Value *getStepVector(Value *Val, int StartIdx, Value *Step, |
890 | Instruction::BinaryOps Opcode = |
891 | Instruction::BinaryOpsEnd) override; |
892 | Value *reverseVector(Value *Vec) override; |
893 | }; |
894 | |
895 | /// Encapsulate information regarding vectorization of a loop and its epilogue. |
896 | /// This information is meant to be updated and used across two stages of |
897 | /// epilogue vectorization. |
898 | struct EpilogueLoopVectorizationInfo { |
899 | ElementCount MainLoopVF = ElementCount::getFixed(0); |
900 | unsigned MainLoopUF = 0; |
901 | ElementCount EpilogueVF = ElementCount::getFixed(0); |
902 | unsigned EpilogueUF = 0; |
903 | BasicBlock *MainLoopIterationCountCheck = nullptr; |
904 | BasicBlock *EpilogueIterationCountCheck = nullptr; |
905 | BasicBlock *SCEVSafetyCheck = nullptr; |
906 | BasicBlock *MemSafetyCheck = nullptr; |
907 | Value *TripCount = nullptr; |
908 | Value *VectorTripCount = nullptr; |
909 | |
910 | EpilogueLoopVectorizationInfo(unsigned MVF, unsigned MUF, unsigned EVF, |
911 | unsigned EUF) |
912 | : MainLoopVF(ElementCount::getFixed(MVF)), MainLoopUF(MUF), |
913 | EpilogueVF(ElementCount::getFixed(EVF)), EpilogueUF(EUF) { |
914 | assert(EUF == 1 &&((EUF == 1 && "A high UF for the epilogue loop is likely not beneficial." ) ? static_cast<void> (0) : __assert_fail ("EUF == 1 && \"A high UF for the epilogue loop is likely not beneficial.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 915, __PRETTY_FUNCTION__)) |
915 | "A high UF for the epilogue loop is likely not beneficial.")((EUF == 1 && "A high UF for the epilogue loop is likely not beneficial." ) ? static_cast<void> (0) : __assert_fail ("EUF == 1 && \"A high UF for the epilogue loop is likely not beneficial.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 915, __PRETTY_FUNCTION__)); |
916 | } |
917 | }; |
918 | |
919 | /// An extension of the inner loop vectorizer that creates a skeleton for a |
920 | /// vectorized loop that has its epilogue (residual) also vectorized. |
921 | /// The idea is to run the vplan on a given loop twice, firstly to setup the |
922 | /// skeleton and vectorize the main loop, and secondly to complete the skeleton |
923 | /// from the first step and vectorize the epilogue. This is achieved by |
924 | /// deriving two concrete strategy classes from this base class and invoking |
925 | /// them in succession from the loop vectorizer planner. |
926 | class InnerLoopAndEpilogueVectorizer : public InnerLoopVectorizer { |
927 | public: |
928 | InnerLoopAndEpilogueVectorizer( |
929 | Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI, |
930 | DominatorTree *DT, const TargetLibraryInfo *TLI, |
931 | const TargetTransformInfo *TTI, AssumptionCache *AC, |
932 | OptimizationRemarkEmitter *ORE, EpilogueLoopVectorizationInfo &EPI, |
933 | LoopVectorizationLegality *LVL, llvm::LoopVectorizationCostModel *CM, |
934 | BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, |
935 | GeneratedRTChecks &Checks) |
936 | : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, |
937 | EPI.MainLoopVF, EPI.MainLoopUF, LVL, CM, BFI, PSI, |
938 | Checks), |
939 | EPI(EPI) {} |
940 | |
941 | // Override this function to handle the more complex control flow around the |
942 | // three loops. |
943 | BasicBlock *createVectorizedLoopSkeleton() final override { |
944 | return createEpilogueVectorizedLoopSkeleton(); |
945 | } |
946 | |
947 | /// The interface for creating a vectorized skeleton using one of two |
948 | /// different strategies, each corresponding to one execution of the vplan |
949 | /// as described above. |
950 | virtual BasicBlock *createEpilogueVectorizedLoopSkeleton() = 0; |
951 | |
952 | /// Holds and updates state information required to vectorize the main loop |
953 | /// and its epilogue in two separate passes. This setup helps us avoid |
954 | /// regenerating and recomputing runtime safety checks. It also helps us to |
955 | /// shorten the iteration-count-check path length for the cases where the |
956 | /// iteration count of the loop is so small that the main vector loop is |
957 | /// completely skipped. |
958 | EpilogueLoopVectorizationInfo &EPI; |
959 | }; |
960 | |
961 | /// A specialized derived class of inner loop vectorizer that performs |
962 | /// vectorization of *main* loops in the process of vectorizing loops and their |
963 | /// epilogues. |
964 | class EpilogueVectorizerMainLoop : public InnerLoopAndEpilogueVectorizer { |
965 | public: |
966 | EpilogueVectorizerMainLoop( |
967 | Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI, |
968 | DominatorTree *DT, const TargetLibraryInfo *TLI, |
969 | const TargetTransformInfo *TTI, AssumptionCache *AC, |
970 | OptimizationRemarkEmitter *ORE, EpilogueLoopVectorizationInfo &EPI, |
971 | LoopVectorizationLegality *LVL, llvm::LoopVectorizationCostModel *CM, |
972 | BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, |
973 | GeneratedRTChecks &Check) |
974 | : InnerLoopAndEpilogueVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, |
975 | EPI, LVL, CM, BFI, PSI, Check) {} |
976 | /// Implements the interface for creating a vectorized skeleton using the |
977 | /// *main loop* strategy (ie the first pass of vplan execution). |
978 | BasicBlock *createEpilogueVectorizedLoopSkeleton() final override; |
979 | |
980 | protected: |
981 | /// Emits an iteration count bypass check once for the main loop (when \p |
982 | /// ForEpilogue is false) and once for the epilogue loop (when \p |
983 | /// ForEpilogue is true). |
984 | BasicBlock *emitMinimumIterationCountCheck(Loop *L, BasicBlock *Bypass, |
985 | bool ForEpilogue); |
986 | void printDebugTracesAtStart() override; |
987 | void printDebugTracesAtEnd() override; |
988 | }; |
989 | |
990 | // A specialized derived class of inner loop vectorizer that performs |
991 | // vectorization of *epilogue* loops in the process of vectorizing loops and |
992 | // their epilogues. |
993 | class EpilogueVectorizerEpilogueLoop : public InnerLoopAndEpilogueVectorizer { |
994 | public: |
995 | EpilogueVectorizerEpilogueLoop( |
996 | Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI, |
997 | DominatorTree *DT, const TargetLibraryInfo *TLI, |
998 | const TargetTransformInfo *TTI, AssumptionCache *AC, |
999 | OptimizationRemarkEmitter *ORE, EpilogueLoopVectorizationInfo &EPI, |
1000 | LoopVectorizationLegality *LVL, llvm::LoopVectorizationCostModel *CM, |
1001 | BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, |
1002 | GeneratedRTChecks &Checks) |
1003 | : InnerLoopAndEpilogueVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, |
1004 | EPI, LVL, CM, BFI, PSI, Checks) {} |
1005 | /// Implements the interface for creating a vectorized skeleton using the |
1006 | /// *epilogue loop* strategy (ie the second pass of vplan execution). |
1007 | BasicBlock *createEpilogueVectorizedLoopSkeleton() final override; |
1008 | |
1009 | protected: |
1010 | /// Emits an iteration count bypass check after the main vector loop has |
1011 | /// finished to see if there are any iterations left to execute by either |
1012 | /// the vector epilogue or the scalar epilogue. |
1013 | BasicBlock *emitMinimumVectorEpilogueIterCountCheck(Loop *L, |
1014 | BasicBlock *Bypass, |
1015 | BasicBlock *Insert); |
1016 | void printDebugTracesAtStart() override; |
1017 | void printDebugTracesAtEnd() override; |
1018 | }; |
1019 | } // end namespace llvm |
1020 | |
1021 | /// Look for a meaningful debug location on the instruction or it's |
1022 | /// operands. |
1023 | static Instruction *getDebugLocFromInstOrOperands(Instruction *I) { |
1024 | if (!I) |
1025 | return I; |
1026 | |
1027 | DebugLoc Empty; |
1028 | if (I->getDebugLoc() != Empty) |
1029 | return I; |
1030 | |
1031 | for (Use &Op : I->operands()) { |
1032 | if (Instruction *OpInst = dyn_cast<Instruction>(Op)) |
1033 | if (OpInst->getDebugLoc() != Empty) |
1034 | return OpInst; |
1035 | } |
1036 | |
1037 | return I; |
1038 | } |
1039 | |
1040 | void InnerLoopVectorizer::setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) { |
1041 | if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr)) { |
1042 | const DILocation *DIL = Inst->getDebugLoc(); |
1043 | if (DIL && Inst->getFunction()->isDebugInfoForProfiling() && |
1044 | !isa<DbgInfoIntrinsic>(Inst)) { |
1045 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1045, __PRETTY_FUNCTION__)); |
1046 | auto NewDIL = |
1047 | DIL->cloneByMultiplyingDuplicationFactor(UF * VF.getKnownMinValue()); |
1048 | if (NewDIL) |
1049 | B.SetCurrentDebugLocation(NewDIL.getValue()); |
1050 | else |
1051 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Failed to create new discriminator: " << DIL->getFilename() << " Line: " << DIL ->getLine(); } } while (false) |
1052 | << "Failed to create new discriminator: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Failed to create new discriminator: " << DIL->getFilename() << " Line: " << DIL ->getLine(); } } while (false) |
1053 | << DIL->getFilename() << " Line: " << DIL->getLine())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Failed to create new discriminator: " << DIL->getFilename() << " Line: " << DIL ->getLine(); } } while (false); |
1054 | } |
1055 | else |
1056 | B.SetCurrentDebugLocation(DIL); |
1057 | } else |
1058 | B.SetCurrentDebugLocation(DebugLoc()); |
1059 | } |
1060 | |
1061 | /// Write a record \p DebugMsg about vectorization failure to the debug |
1062 | /// output stream. If \p I is passed, it is an instruction that prevents |
1063 | /// vectorization. |
1064 | #ifndef NDEBUG |
1065 | static void debugVectorizationFailure(const StringRef DebugMsg, |
1066 | Instruction *I) { |
1067 | dbgs() << "LV: Not vectorizing: " << DebugMsg; |
1068 | if (I != nullptr) |
1069 | dbgs() << " " << *I; |
1070 | else |
1071 | dbgs() << '.'; |
1072 | dbgs() << '\n'; |
1073 | } |
1074 | #endif |
1075 | |
1076 | /// Create an analysis remark that explains why vectorization failed |
1077 | /// |
1078 | /// \p PassName is the name of the pass (e.g. can be AlwaysPrint). \p |
1079 | /// RemarkName is the identifier for the remark. If \p I is passed it is an |
1080 | /// instruction that prevents vectorization. Otherwise \p TheLoop is used for |
1081 | /// the location of the remark. \return the remark object that can be |
1082 | /// streamed to. |
1083 | static OptimizationRemarkAnalysis createLVAnalysis(const char *PassName, |
1084 | StringRef RemarkName, Loop *TheLoop, Instruction *I) { |
1085 | Value *CodeRegion = TheLoop->getHeader(); |
1086 | DebugLoc DL = TheLoop->getStartLoc(); |
1087 | |
1088 | if (I) { |
1089 | CodeRegion = I->getParent(); |
1090 | // If there is no debug location attached to the instruction, revert back to |
1091 | // using the loop's. |
1092 | if (I->getDebugLoc()) |
1093 | DL = I->getDebugLoc(); |
1094 | } |
1095 | |
1096 | OptimizationRemarkAnalysis R(PassName, RemarkName, DL, CodeRegion); |
1097 | R << "loop not vectorized: "; |
1098 | return R; |
1099 | } |
1100 | |
1101 | /// Return a value for Step multiplied by VF. |
1102 | static Value *createStepForVF(IRBuilder<> &B, Constant *Step, ElementCount VF) { |
1103 | assert(isa<ConstantInt>(Step) && "Expected an integer step")((isa<ConstantInt>(Step) && "Expected an integer step" ) ? static_cast<void> (0) : __assert_fail ("isa<ConstantInt>(Step) && \"Expected an integer step\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1103, __PRETTY_FUNCTION__)); |
1104 | Constant *StepVal = ConstantInt::get( |
1105 | Step->getType(), |
1106 | cast<ConstantInt>(Step)->getSExtValue() * VF.getKnownMinValue()); |
1107 | return VF.isScalable() ? B.CreateVScale(StepVal) : StepVal; |
1108 | } |
1109 | |
1110 | namespace llvm { |
1111 | |
1112 | /// Return the runtime value for VF. |
1113 | Value *getRuntimeVF(IRBuilder<> &B, Type *Ty, ElementCount VF) { |
1114 | Constant *EC = ConstantInt::get(Ty, VF.getKnownMinValue()); |
1115 | return VF.isScalable() ? B.CreateVScale(EC) : EC; |
1116 | } |
1117 | |
1118 | void reportVectorizationFailure(const StringRef DebugMsg, |
1119 | const StringRef OREMsg, const StringRef ORETag, |
1120 | OptimizationRemarkEmitter *ORE, Loop *TheLoop, Instruction *I) { |
1121 | LLVM_DEBUG(debugVectorizationFailure(DebugMsg, I))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { debugVectorizationFailure(DebugMsg, I); } } while (false); |
1122 | LoopVectorizeHints Hints(TheLoop, true /* doesn't matter */, *ORE); |
1123 | ORE->emit(createLVAnalysis(Hints.vectorizeAnalysisPassName(), |
1124 | ORETag, TheLoop, I) << OREMsg); |
1125 | } |
1126 | |
1127 | } // end namespace llvm |
1128 | |
1129 | #ifndef NDEBUG |
1130 | /// \return string containing a file name and a line # for the given loop. |
1131 | static std::string getDebugLocString(const Loop *L) { |
1132 | std::string Result; |
1133 | if (L) { |
1134 | raw_string_ostream OS(Result); |
1135 | if (const DebugLoc LoopDbgLoc = L->getStartLoc()) |
1136 | LoopDbgLoc.print(OS); |
1137 | else |
1138 | // Just print the module name. |
1139 | OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier(); |
1140 | OS.flush(); |
1141 | } |
1142 | return Result; |
1143 | } |
1144 | #endif |
1145 | |
1146 | void InnerLoopVectorizer::addNewMetadata(Instruction *To, |
1147 | const Instruction *Orig) { |
1148 | // If the loop was versioned with memchecks, add the corresponding no-alias |
1149 | // metadata. |
1150 | if (LVer && (isa<LoadInst>(Orig) || isa<StoreInst>(Orig))) |
1151 | LVer->annotateInstWithNoAlias(To, Orig); |
1152 | } |
1153 | |
1154 | void InnerLoopVectorizer::addMetadata(Instruction *To, |
1155 | Instruction *From) { |
1156 | propagateMetadata(To, From); |
1157 | addNewMetadata(To, From); |
1158 | } |
1159 | |
1160 | void InnerLoopVectorizer::addMetadata(ArrayRef<Value *> To, |
1161 | Instruction *From) { |
1162 | for (Value *V : To) { |
1163 | if (Instruction *I = dyn_cast<Instruction>(V)) |
1164 | addMetadata(I, From); |
1165 | } |
1166 | } |
1167 | |
1168 | namespace llvm { |
1169 | |
1170 | // Loop vectorization cost-model hints how the scalar epilogue loop should be |
1171 | // lowered. |
1172 | enum ScalarEpilogueLowering { |
1173 | |
1174 | // The default: allowing scalar epilogues. |
1175 | CM_ScalarEpilogueAllowed, |
1176 | |
1177 | // Vectorization with OptForSize: don't allow epilogues. |
1178 | CM_ScalarEpilogueNotAllowedOptSize, |
1179 | |
1180 | // A special case of vectorisation with OptForSize: loops with a very small |
1181 | // trip count are considered for vectorization under OptForSize, thereby |
1182 | // making sure the cost of their loop body is dominant, free of runtime |
1183 | // guards and scalar iteration overheads. |
1184 | CM_ScalarEpilogueNotAllowedLowTripLoop, |
1185 | |
1186 | // Loop hint predicate indicating an epilogue is undesired. |
1187 | CM_ScalarEpilogueNotNeededUsePredicate, |
1188 | |
1189 | // Directive indicating we must either tail fold or not vectorize |
1190 | CM_ScalarEpilogueNotAllowedUsePredicate |
1191 | }; |
1192 | |
1193 | /// LoopVectorizationCostModel - estimates the expected speedups due to |
1194 | /// vectorization. |
1195 | /// In many cases vectorization is not profitable. This can happen because of |
1196 | /// a number of reasons. In this class we mainly attempt to predict the |
1197 | /// expected speedup/slowdowns due to the supported instruction set. We use the |
1198 | /// TargetTransformInfo to query the different backends for the cost of |
1199 | /// different operations. |
1200 | class LoopVectorizationCostModel { |
1201 | public: |
1202 | LoopVectorizationCostModel(ScalarEpilogueLowering SEL, Loop *L, |
1203 | PredicatedScalarEvolution &PSE, LoopInfo *LI, |
1204 | LoopVectorizationLegality *Legal, |
1205 | const TargetTransformInfo &TTI, |
1206 | const TargetLibraryInfo *TLI, DemandedBits *DB, |
1207 | AssumptionCache *AC, |
1208 | OptimizationRemarkEmitter *ORE, const Function *F, |
1209 | const LoopVectorizeHints *Hints, |
1210 | InterleavedAccessInfo &IAI) |
1211 | : ScalarEpilogueStatus(SEL), TheLoop(L), PSE(PSE), LI(LI), Legal(Legal), |
1212 | TTI(TTI), TLI(TLI), DB(DB), AC(AC), ORE(ORE), TheFunction(F), |
1213 | Hints(Hints), InterleaveInfo(IAI) {} |
1214 | |
1215 | /// \return An upper bound for the vectorization factor, or None if |
1216 | /// vectorization and interleaving should be avoided up front. |
1217 | Optional<ElementCount> computeMaxVF(ElementCount UserVF, unsigned UserIC); |
1218 | |
1219 | /// \return True if runtime checks are required for vectorization, and false |
1220 | /// otherwise. |
1221 | bool runtimeChecksRequired(); |
1222 | |
1223 | /// \return The most profitable vectorization factor and the cost of that VF. |
1224 | /// This method checks every power of two up to MaxVF. If UserVF is not ZERO |
1225 | /// then this vectorization factor will be selected if vectorization is |
1226 | /// possible. |
1227 | VectorizationFactor selectVectorizationFactor(ElementCount MaxVF); |
1228 | VectorizationFactor |
1229 | selectEpilogueVectorizationFactor(const ElementCount MaxVF, |
1230 | const LoopVectorizationPlanner &LVP); |
1231 | |
1232 | /// Setup cost-based decisions for user vectorization factor. |
1233 | void selectUserVectorizationFactor(ElementCount UserVF) { |
1234 | collectUniformsAndScalars(UserVF); |
1235 | collectInstsToScalarize(UserVF); |
1236 | } |
1237 | |
1238 | /// \return The size (in bits) of the smallest and widest types in the code |
1239 | /// that needs to be vectorized. We ignore values that remain scalar such as |
1240 | /// 64 bit loop indices. |
1241 | std::pair<unsigned, unsigned> getSmallestAndWidestTypes(); |
1242 | |
1243 | /// \return The desired interleave count. |
1244 | /// If interleave count has been specified by metadata it will be returned. |
1245 | /// Otherwise, the interleave count is computed and returned. VF and LoopCost |
1246 | /// are the selected vectorization factor and the cost of the selected VF. |
1247 | unsigned selectInterleaveCount(ElementCount VF, unsigned LoopCost); |
1248 | |
1249 | /// Memory access instruction may be vectorized in more than one way. |
1250 | /// Form of instruction after vectorization depends on cost. |
1251 | /// This function takes cost-based decisions for Load/Store instructions |
1252 | /// and collects them in a map. This decisions map is used for building |
1253 | /// the lists of loop-uniform and loop-scalar instructions. |
1254 | /// The calculated cost is saved with widening decision in order to |
1255 | /// avoid redundant calculations. |
1256 | void setCostBasedWideningDecision(ElementCount VF); |
1257 | |
1258 | /// A struct that represents some properties of the register usage |
1259 | /// of a loop. |
1260 | struct RegisterUsage { |
1261 | /// Holds the number of loop invariant values that are used in the loop. |
1262 | /// The key is ClassID of target-provided register class. |
1263 | SmallMapVector<unsigned, unsigned, 4> LoopInvariantRegs; |
1264 | /// Holds the maximum number of concurrent live intervals in the loop. |
1265 | /// The key is ClassID of target-provided register class. |
1266 | SmallMapVector<unsigned, unsigned, 4> MaxLocalUsers; |
1267 | }; |
1268 | |
1269 | /// \return Returns information about the register usages of the loop for the |
1270 | /// given vectorization factors. |
1271 | SmallVector<RegisterUsage, 8> |
1272 | calculateRegisterUsage(ArrayRef<ElementCount> VFs); |
1273 | |
1274 | /// Collect values we want to ignore in the cost model. |
1275 | void collectValuesToIgnore(); |
1276 | |
1277 | /// Split reductions into those that happen in the loop, and those that happen |
1278 | /// outside. In loop reductions are collected into InLoopReductionChains. |
1279 | void collectInLoopReductions(); |
1280 | |
1281 | /// \returns The smallest bitwidth each instruction can be represented with. |
1282 | /// The vector equivalents of these instructions should be truncated to this |
1283 | /// type. |
1284 | const MapVector<Instruction *, uint64_t> &getMinimalBitwidths() const { |
1285 | return MinBWs; |
1286 | } |
1287 | |
1288 | /// \returns True if it is more profitable to scalarize instruction \p I for |
1289 | /// vectorization factor \p VF. |
1290 | bool isProfitableToScalarize(Instruction *I, ElementCount VF) const { |
1291 | assert(VF.isVector() &&((VF.isVector() && "Profitable to scalarize relevant only for VF > 1." ) ? static_cast<void> (0) : __assert_fail ("VF.isVector() && \"Profitable to scalarize relevant only for VF > 1.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1292, __PRETTY_FUNCTION__)) |
1292 | "Profitable to scalarize relevant only for VF > 1.")((VF.isVector() && "Profitable to scalarize relevant only for VF > 1." ) ? static_cast<void> (0) : __assert_fail ("VF.isVector() && \"Profitable to scalarize relevant only for VF > 1.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1292, __PRETTY_FUNCTION__)); |
1293 | |
1294 | // Cost model is not run in the VPlan-native path - return conservative |
1295 | // result until this changes. |
1296 | if (EnableVPlanNativePath) |
1297 | return false; |
1298 | |
1299 | auto Scalars = InstsToScalarize.find(VF); |
1300 | assert(Scalars != InstsToScalarize.end() &&((Scalars != InstsToScalarize.end() && "VF not yet analyzed for scalarization profitability" ) ? static_cast<void> (0) : __assert_fail ("Scalars != InstsToScalarize.end() && \"VF not yet analyzed for scalarization profitability\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1301, __PRETTY_FUNCTION__)) |
1301 | "VF not yet analyzed for scalarization profitability")((Scalars != InstsToScalarize.end() && "VF not yet analyzed for scalarization profitability" ) ? static_cast<void> (0) : __assert_fail ("Scalars != InstsToScalarize.end() && \"VF not yet analyzed for scalarization profitability\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1301, __PRETTY_FUNCTION__)); |
1302 | return Scalars->second.find(I) != Scalars->second.end(); |
1303 | } |
1304 | |
1305 | /// Returns true if \p I is known to be uniform after vectorization. |
1306 | bool isUniformAfterVectorization(Instruction *I, ElementCount VF) const { |
1307 | if (VF.isScalar()) |
1308 | return true; |
1309 | |
1310 | // Cost model is not run in the VPlan-native path - return conservative |
1311 | // result until this changes. |
1312 | if (EnableVPlanNativePath) |
1313 | return false; |
1314 | |
1315 | auto UniformsPerVF = Uniforms.find(VF); |
1316 | assert(UniformsPerVF != Uniforms.end() &&((UniformsPerVF != Uniforms.end() && "VF not yet analyzed for uniformity" ) ? static_cast<void> (0) : __assert_fail ("UniformsPerVF != Uniforms.end() && \"VF not yet analyzed for uniformity\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1317, __PRETTY_FUNCTION__)) |
1317 | "VF not yet analyzed for uniformity")((UniformsPerVF != Uniforms.end() && "VF not yet analyzed for uniformity" ) ? static_cast<void> (0) : __assert_fail ("UniformsPerVF != Uniforms.end() && \"VF not yet analyzed for uniformity\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1317, __PRETTY_FUNCTION__)); |
1318 | return UniformsPerVF->second.count(I); |
1319 | } |
1320 | |
1321 | /// Returns true if \p I is known to be scalar after vectorization. |
1322 | bool isScalarAfterVectorization(Instruction *I, ElementCount VF) const { |
1323 | if (VF.isScalar()) |
1324 | return true; |
1325 | |
1326 | // Cost model is not run in the VPlan-native path - return conservative |
1327 | // result until this changes. |
1328 | if (EnableVPlanNativePath) |
1329 | return false; |
1330 | |
1331 | auto ScalarsPerVF = Scalars.find(VF); |
1332 | assert(ScalarsPerVF != Scalars.end() &&((ScalarsPerVF != Scalars.end() && "Scalar values are not calculated for VF" ) ? static_cast<void> (0) : __assert_fail ("ScalarsPerVF != Scalars.end() && \"Scalar values are not calculated for VF\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1333, __PRETTY_FUNCTION__)) |
1333 | "Scalar values are not calculated for VF")((ScalarsPerVF != Scalars.end() && "Scalar values are not calculated for VF" ) ? static_cast<void> (0) : __assert_fail ("ScalarsPerVF != Scalars.end() && \"Scalar values are not calculated for VF\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1333, __PRETTY_FUNCTION__)); |
1334 | return ScalarsPerVF->second.count(I); |
1335 | } |
1336 | |
1337 | /// \returns True if instruction \p I can be truncated to a smaller bitwidth |
1338 | /// for vectorization factor \p VF. |
1339 | bool canTruncateToMinimalBitwidth(Instruction *I, ElementCount VF) const { |
1340 | return VF.isVector() && MinBWs.find(I) != MinBWs.end() && |
1341 | !isProfitableToScalarize(I, VF) && |
1342 | !isScalarAfterVectorization(I, VF); |
1343 | } |
1344 | |
1345 | /// Decision that was taken during cost calculation for memory instruction. |
1346 | enum InstWidening { |
1347 | CM_Unknown, |
1348 | CM_Widen, // For consecutive accesses with stride +1. |
1349 | CM_Widen_Reverse, // For consecutive accesses with stride -1. |
1350 | CM_Interleave, |
1351 | CM_GatherScatter, |
1352 | CM_Scalarize |
1353 | }; |
1354 | |
1355 | /// Save vectorization decision \p W and \p Cost taken by the cost model for |
1356 | /// instruction \p I and vector width \p VF. |
1357 | void setWideningDecision(Instruction *I, ElementCount VF, InstWidening W, |
1358 | InstructionCost Cost) { |
1359 | assert(VF.isVector() && "Expected VF >=2")((VF.isVector() && "Expected VF >=2") ? static_cast <void> (0) : __assert_fail ("VF.isVector() && \"Expected VF >=2\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1359, __PRETTY_FUNCTION__)); |
1360 | WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost); |
1361 | } |
1362 | |
1363 | /// Save vectorization decision \p W and \p Cost taken by the cost model for |
1364 | /// interleaving group \p Grp and vector width \p VF. |
1365 | void setWideningDecision(const InterleaveGroup<Instruction> *Grp, |
1366 | ElementCount VF, InstWidening W, |
1367 | InstructionCost Cost) { |
1368 | assert(VF.isVector() && "Expected VF >=2")((VF.isVector() && "Expected VF >=2") ? static_cast <void> (0) : __assert_fail ("VF.isVector() && \"Expected VF >=2\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1368, __PRETTY_FUNCTION__)); |
1369 | /// Broadcast this decicion to all instructions inside the group. |
1370 | /// But the cost will be assigned to one instruction only. |
1371 | for (unsigned i = 0; i < Grp->getFactor(); ++i) { |
1372 | if (auto *I = Grp->getMember(i)) { |
1373 | if (Grp->getInsertPos() == I) |
1374 | WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost); |
1375 | else |
1376 | WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, 0); |
1377 | } |
1378 | } |
1379 | } |
1380 | |
1381 | /// Return the cost model decision for the given instruction \p I and vector |
1382 | /// width \p VF. Return CM_Unknown if this instruction did not pass |
1383 | /// through the cost modeling. |
1384 | InstWidening getWideningDecision(Instruction *I, ElementCount VF) { |
1385 | assert(VF.isVector() && "Expected VF to be a vector VF")((VF.isVector() && "Expected VF to be a vector VF") ? static_cast<void> (0) : __assert_fail ("VF.isVector() && \"Expected VF to be a vector VF\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1385, __PRETTY_FUNCTION__)); |
1386 | // Cost model is not run in the VPlan-native path - return conservative |
1387 | // result until this changes. |
1388 | if (EnableVPlanNativePath) |
1389 | return CM_GatherScatter; |
1390 | |
1391 | std::pair<Instruction *, ElementCount> InstOnVF = std::make_pair(I, VF); |
1392 | auto Itr = WideningDecisions.find(InstOnVF); |
1393 | if (Itr == WideningDecisions.end()) |
1394 | return CM_Unknown; |
1395 | return Itr->second.first; |
1396 | } |
1397 | |
1398 | /// Return the vectorization cost for the given instruction \p I and vector |
1399 | /// width \p VF. |
1400 | InstructionCost getWideningCost(Instruction *I, ElementCount VF) { |
1401 | assert(VF.isVector() && "Expected VF >=2")((VF.isVector() && "Expected VF >=2") ? static_cast <void> (0) : __assert_fail ("VF.isVector() && \"Expected VF >=2\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1401, __PRETTY_FUNCTION__)); |
1402 | std::pair<Instruction *, ElementCount> InstOnVF = std::make_pair(I, VF); |
1403 | assert(WideningDecisions.find(InstOnVF) != WideningDecisions.end() &&((WideningDecisions.find(InstOnVF) != WideningDecisions.end() && "The cost is not calculated") ? static_cast<void > (0) : __assert_fail ("WideningDecisions.find(InstOnVF) != WideningDecisions.end() && \"The cost is not calculated\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1404, __PRETTY_FUNCTION__)) |
1404 | "The cost is not calculated")((WideningDecisions.find(InstOnVF) != WideningDecisions.end() && "The cost is not calculated") ? static_cast<void > (0) : __assert_fail ("WideningDecisions.find(InstOnVF) != WideningDecisions.end() && \"The cost is not calculated\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1404, __PRETTY_FUNCTION__)); |
1405 | return WideningDecisions[InstOnVF].second; |
1406 | } |
1407 | |
1408 | /// Return True if instruction \p I is an optimizable truncate whose operand |
1409 | /// is an induction variable. Such a truncate will be removed by adding a new |
1410 | /// induction variable with the destination type. |
1411 | bool isOptimizableIVTruncate(Instruction *I, ElementCount VF) { |
1412 | // If the instruction is not a truncate, return false. |
1413 | auto *Trunc = dyn_cast<TruncInst>(I); |
1414 | if (!Trunc) |
1415 | return false; |
1416 | |
1417 | // Get the source and destination types of the truncate. |
1418 | Type *SrcTy = ToVectorTy(cast<CastInst>(I)->getSrcTy(), VF); |
1419 | Type *DestTy = ToVectorTy(cast<CastInst>(I)->getDestTy(), VF); |
1420 | |
1421 | // If the truncate is free for the given types, return false. Replacing a |
1422 | // free truncate with an induction variable would add an induction variable |
1423 | // update instruction to each iteration of the loop. We exclude from this |
1424 | // check the primary induction variable since it will need an update |
1425 | // instruction regardless. |
1426 | Value *Op = Trunc->getOperand(0); |
1427 | if (Op != Legal->getPrimaryInduction() && TTI.isTruncateFree(SrcTy, DestTy)) |
1428 | return false; |
1429 | |
1430 | // If the truncated value is not an induction variable, return false. |
1431 | return Legal->isInductionPhi(Op); |
1432 | } |
1433 | |
1434 | /// Collects the instructions to scalarize for each predicated instruction in |
1435 | /// the loop. |
1436 | void collectInstsToScalarize(ElementCount VF); |
1437 | |
1438 | /// Collect Uniform and Scalar values for the given \p VF. |
1439 | /// The sets depend on CM decision for Load/Store instructions |
1440 | /// that may be vectorized as interleave, gather-scatter or scalarized. |
1441 | void collectUniformsAndScalars(ElementCount VF) { |
1442 | // Do the analysis once. |
1443 | if (VF.isScalar() || Uniforms.find(VF) != Uniforms.end()) |
1444 | return; |
1445 | setCostBasedWideningDecision(VF); |
1446 | collectLoopUniforms(VF); |
1447 | collectLoopScalars(VF); |
1448 | } |
1449 | |
1450 | /// Returns true if the target machine supports masked store operation |
1451 | /// for the given \p DataType and kind of access to \p Ptr. |
1452 | bool isLegalMaskedStore(Type *DataType, Value *Ptr, Align Alignment) { |
1453 | return Legal->isConsecutivePtr(Ptr) && |
1454 | TTI.isLegalMaskedStore(DataType, Alignment); |
1455 | } |
1456 | |
1457 | /// Returns true if the target machine supports masked load operation |
1458 | /// for the given \p DataType and kind of access to \p Ptr. |
1459 | bool isLegalMaskedLoad(Type *DataType, Value *Ptr, Align Alignment) { |
1460 | return Legal->isConsecutivePtr(Ptr) && |
1461 | TTI.isLegalMaskedLoad(DataType, Alignment); |
1462 | } |
1463 | |
1464 | /// Returns true if the target machine supports masked scatter operation |
1465 | /// for the given \p DataType. |
1466 | bool isLegalMaskedScatter(Type *DataType, Align Alignment) { |
1467 | return TTI.isLegalMaskedScatter(DataType, Alignment); |
1468 | } |
1469 | |
1470 | /// Returns true if the target machine supports masked gather operation |
1471 | /// for the given \p DataType. |
1472 | bool isLegalMaskedGather(Type *DataType, Align Alignment) { |
1473 | return TTI.isLegalMaskedGather(DataType, Alignment); |
1474 | } |
1475 | |
1476 | /// Returns true if the target machine can represent \p V as a masked gather |
1477 | /// or scatter operation. |
1478 | bool isLegalGatherOrScatter(Value *V) { |
1479 | bool LI = isa<LoadInst>(V); |
1480 | bool SI = isa<StoreInst>(V); |
1481 | if (!LI && !SI) |
1482 | return false; |
1483 | auto *Ty = getMemInstValueType(V); |
1484 | Align Align = getLoadStoreAlignment(V); |
1485 | return (LI && isLegalMaskedGather(Ty, Align)) || |
1486 | (SI && isLegalMaskedScatter(Ty, Align)); |
1487 | } |
1488 | |
1489 | /// Returns true if the target machine supports all of the reduction |
1490 | /// variables found for the given VF. |
1491 | bool canVectorizeReductions(ElementCount VF) { |
1492 | return (all_of(Legal->getReductionVars(), [&](auto &Reduction) -> bool { |
1493 | RecurrenceDescriptor RdxDesc = Reduction.second; |
1494 | return TTI.isLegalToVectorizeReduction(RdxDesc, VF); |
1495 | })); |
1496 | } |
1497 | |
1498 | /// Returns true if \p I is an instruction that will be scalarized with |
1499 | /// predication. Such instructions include conditional stores and |
1500 | /// instructions that may divide by zero. |
1501 | /// If a non-zero VF has been calculated, we check if I will be scalarized |
1502 | /// predication for that VF. |
1503 | bool isScalarWithPredication(Instruction *I, |
1504 | ElementCount VF = ElementCount::getFixed(1)); |
1505 | |
1506 | // Returns true if \p I is an instruction that will be predicated either |
1507 | // through scalar predication or masked load/store or masked gather/scatter. |
1508 | // Superset of instructions that return true for isScalarWithPredication. |
1509 | bool isPredicatedInst(Instruction *I) { |
1510 | if (!blockNeedsPredication(I->getParent())) |
1511 | return false; |
1512 | // Loads and stores that need some form of masked operation are predicated |
1513 | // instructions. |
1514 | if (isa<LoadInst>(I) || isa<StoreInst>(I)) |
1515 | return Legal->isMaskRequired(I); |
1516 | return isScalarWithPredication(I); |
1517 | } |
1518 | |
1519 | /// Returns true if \p I is a memory instruction with consecutive memory |
1520 | /// access that can be widened. |
1521 | bool |
1522 | memoryInstructionCanBeWidened(Instruction *I, |
1523 | ElementCount VF = ElementCount::getFixed(1)); |
1524 | |
1525 | /// Returns true if \p I is a memory instruction in an interleaved-group |
1526 | /// of memory accesses that can be vectorized with wide vector loads/stores |
1527 | /// and shuffles. |
1528 | bool |
1529 | interleavedAccessCanBeWidened(Instruction *I, |
1530 | ElementCount VF = ElementCount::getFixed(1)); |
1531 | |
1532 | /// Check if \p Instr belongs to any interleaved access group. |
1533 | bool isAccessInterleaved(Instruction *Instr) { |
1534 | return InterleaveInfo.isInterleaved(Instr); |
1535 | } |
1536 | |
1537 | /// Get the interleaved access group that \p Instr belongs to. |
1538 | const InterleaveGroup<Instruction> * |
1539 | getInterleavedAccessGroup(Instruction *Instr) { |
1540 | return InterleaveInfo.getInterleaveGroup(Instr); |
1541 | } |
1542 | |
1543 | /// Returns true if we're required to use a scalar epilogue for at least |
1544 | /// the final iteration of the original loop. |
1545 | bool requiresScalarEpilogue() const { |
1546 | if (!isScalarEpilogueAllowed()) |
1547 | return false; |
1548 | // If we might exit from anywhere but the latch, must run the exiting |
1549 | // iteration in scalar form. |
1550 | if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) |
1551 | return true; |
1552 | return InterleaveInfo.requiresScalarEpilogue(); |
1553 | } |
1554 | |
1555 | /// Returns true if a scalar epilogue is not allowed due to optsize or a |
1556 | /// loop hint annotation. |
1557 | bool isScalarEpilogueAllowed() const { |
1558 | return ScalarEpilogueStatus == CM_ScalarEpilogueAllowed; |
1559 | } |
1560 | |
1561 | /// Returns true if all loop blocks should be masked to fold tail loop. |
1562 | bool foldTailByMasking() const { return FoldTailByMasking; } |
1563 | |
1564 | bool blockNeedsPredication(BasicBlock *BB) { |
1565 | return foldTailByMasking() || Legal->blockNeedsPredication(BB); |
1566 | } |
1567 | |
1568 | /// A SmallMapVector to store the InLoop reduction op chains, mapping phi |
1569 | /// nodes to the chain of instructions representing the reductions. Uses a |
1570 | /// MapVector to ensure deterministic iteration order. |
1571 | using ReductionChainMap = |
1572 | SmallMapVector<PHINode *, SmallVector<Instruction *, 4>, 4>; |
1573 | |
1574 | /// Return the chain of instructions representing an inloop reduction. |
1575 | const ReductionChainMap &getInLoopReductionChains() const { |
1576 | return InLoopReductionChains; |
1577 | } |
1578 | |
1579 | /// Returns true if the Phi is part of an inloop reduction. |
1580 | bool isInLoopReduction(PHINode *Phi) const { |
1581 | return InLoopReductionChains.count(Phi); |
1582 | } |
1583 | |
1584 | /// Estimate cost of an intrinsic call instruction CI if it were vectorized |
1585 | /// with factor VF. Return the cost of the instruction, including |
1586 | /// scalarization overhead if it's needed. |
1587 | InstructionCost getVectorIntrinsicCost(CallInst *CI, ElementCount VF); |
1588 | |
1589 | /// Estimate cost of a call instruction CI if it were vectorized with factor |
1590 | /// VF. Return the cost of the instruction, including scalarization overhead |
1591 | /// if it's needed. The flag NeedToScalarize shows if the call needs to be |
1592 | /// scalarized - |
1593 | /// i.e. either vector version isn't available, or is too expensive. |
1594 | InstructionCost getVectorCallCost(CallInst *CI, ElementCount VF, |
1595 | bool &NeedToScalarize); |
1596 | |
1597 | /// Invalidates decisions already taken by the cost model. |
1598 | void invalidateCostModelingDecisions() { |
1599 | WideningDecisions.clear(); |
1600 | Uniforms.clear(); |
1601 | Scalars.clear(); |
1602 | } |
1603 | |
1604 | private: |
1605 | unsigned NumPredStores = 0; |
1606 | |
1607 | /// \return An upper bound for the vectorization factor, a power-of-2 larger |
1608 | /// than zero. One is returned if vectorization should best be avoided due |
1609 | /// to cost. |
1610 | ElementCount computeFeasibleMaxVF(unsigned ConstTripCount, |
1611 | ElementCount UserVF); |
1612 | |
1613 | /// The vectorization cost is a combination of the cost itself and a boolean |
1614 | /// indicating whether any of the contributing operations will actually |
1615 | /// operate on |
1616 | /// vector values after type legalization in the backend. If this latter value |
1617 | /// is |
1618 | /// false, then all operations will be scalarized (i.e. no vectorization has |
1619 | /// actually taken place). |
1620 | using VectorizationCostTy = std::pair<InstructionCost, bool>; |
1621 | |
1622 | /// Returns the expected execution cost. The unit of the cost does |
1623 | /// not matter because we use the 'cost' units to compare different |
1624 | /// vector widths. The cost that is returned is *not* normalized by |
1625 | /// the factor width. |
1626 | VectorizationCostTy expectedCost(ElementCount VF); |
1627 | |
1628 | /// Returns the execution time cost of an instruction for a given vector |
1629 | /// width. Vector width of one means scalar. |
1630 | VectorizationCostTy getInstructionCost(Instruction *I, ElementCount VF); |
1631 | |
1632 | /// The cost-computation logic from getInstructionCost which provides |
1633 | /// the vector type as an output parameter. |
1634 | InstructionCost getInstructionCost(Instruction *I, ElementCount VF, |
1635 | Type *&VectorTy); |
1636 | |
1637 | /// Return the cost of instructions in an inloop reduction pattern, if I is |
1638 | /// part of that pattern. |
1639 | InstructionCost getReductionPatternCost(Instruction *I, ElementCount VF, |
1640 | Type *VectorTy, |
1641 | TTI::TargetCostKind CostKind); |
1642 | |
1643 | /// Calculate vectorization cost of memory instruction \p I. |
1644 | InstructionCost getMemoryInstructionCost(Instruction *I, ElementCount VF); |
1645 | |
1646 | /// The cost computation for scalarized memory instruction. |
1647 | InstructionCost getMemInstScalarizationCost(Instruction *I, ElementCount VF); |
1648 | |
1649 | /// The cost computation for interleaving group of memory instructions. |
1650 | InstructionCost getInterleaveGroupCost(Instruction *I, ElementCount VF); |
1651 | |
1652 | /// The cost computation for Gather/Scatter instruction. |
1653 | InstructionCost getGatherScatterCost(Instruction *I, ElementCount VF); |
1654 | |
1655 | /// The cost computation for widening instruction \p I with consecutive |
1656 | /// memory access. |
1657 | InstructionCost getConsecutiveMemOpCost(Instruction *I, ElementCount VF); |
1658 | |
1659 | /// The cost calculation for Load/Store instruction \p I with uniform pointer - |
1660 | /// Load: scalar load + broadcast. |
1661 | /// Store: scalar store + (loop invariant value stored? 0 : extract of last |
1662 | /// element) |
1663 | InstructionCost getUniformMemOpCost(Instruction *I, ElementCount VF); |
1664 | |
1665 | /// Estimate the overhead of scalarizing an instruction. This is a |
1666 | /// convenience wrapper for the type-based getScalarizationOverhead API. |
1667 | InstructionCost getScalarizationOverhead(Instruction *I, ElementCount VF); |
1668 | |
1669 | /// Returns whether the instruction is a load or store and will be a emitted |
1670 | /// as a vector operation. |
1671 | bool isConsecutiveLoadOrStore(Instruction *I); |
1672 | |
1673 | /// Returns true if an artificially high cost for emulated masked memrefs |
1674 | /// should be used. |
1675 | bool useEmulatedMaskMemRefHack(Instruction *I); |
1676 | |
1677 | /// Map of scalar integer values to the smallest bitwidth they can be legally |
1678 | /// represented as. The vector equivalents of these values should be truncated |
1679 | /// to this type. |
1680 | MapVector<Instruction *, uint64_t> MinBWs; |
1681 | |
1682 | /// A type representing the costs for instructions if they were to be |
1683 | /// scalarized rather than vectorized. The entries are Instruction-Cost |
1684 | /// pairs. |
1685 | using ScalarCostsTy = DenseMap<Instruction *, InstructionCost>; |
1686 | |
1687 | /// A set containing all BasicBlocks that are known to present after |
1688 | /// vectorization as a predicated block. |
1689 | SmallPtrSet<BasicBlock *, 4> PredicatedBBsAfterVectorization; |
1690 | |
1691 | /// Records whether it is allowed to have the original scalar loop execute at |
1692 | /// least once. This may be needed as a fallback loop in case runtime |
1693 | /// aliasing/dependence checks fail, or to handle the tail/remainder |
1694 | /// iterations when the trip count is unknown or doesn't divide by the VF, |
1695 | /// or as a peel-loop to handle gaps in interleave-groups. |
1696 | /// Under optsize and when the trip count is very small we don't allow any |
1697 | /// iterations to execute in the scalar loop. |
1698 | ScalarEpilogueLowering ScalarEpilogueStatus = CM_ScalarEpilogueAllowed; |
1699 | |
1700 | /// All blocks of loop are to be masked to fold tail of scalar iterations. |
1701 | bool FoldTailByMasking = false; |
1702 | |
1703 | /// A map holding scalar costs for different vectorization factors. The |
1704 | /// presence of a cost for an instruction in the mapping indicates that the |
1705 | /// instruction will be scalarized when vectorizing with the associated |
1706 | /// vectorization factor. The entries are VF-ScalarCostTy pairs. |
1707 | DenseMap<ElementCount, ScalarCostsTy> InstsToScalarize; |
1708 | |
1709 | /// Holds the instructions known to be uniform after vectorization. |
1710 | /// The data is collected per VF. |
1711 | DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> Uniforms; |
1712 | |
1713 | /// Holds the instructions known to be scalar after vectorization. |
1714 | /// The data is collected per VF. |
1715 | DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> Scalars; |
1716 | |
1717 | /// Holds the instructions (address computations) that are forced to be |
1718 | /// scalarized. |
1719 | DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> ForcedScalars; |
1720 | |
1721 | /// PHINodes of the reductions that should be expanded in-loop along with |
1722 | /// their associated chains of reduction operations, in program order from top |
1723 | /// (PHI) to bottom |
1724 | ReductionChainMap InLoopReductionChains; |
1725 | |
1726 | /// A Map of inloop reduction operations and their immediate chain operand. |
1727 | /// FIXME: This can be removed once reductions can be costed correctly in |
1728 | /// vplan. This was added to allow quick lookup to the inloop operations, |
1729 | /// without having to loop through InLoopReductionChains. |
1730 | DenseMap<Instruction *, Instruction *> InLoopReductionImmediateChains; |
1731 | |
1732 | /// Returns the expected difference in cost from scalarizing the expression |
1733 | /// feeding a predicated instruction \p PredInst. The instructions to |
1734 | /// scalarize and their scalar costs are collected in \p ScalarCosts. A |
1735 | /// non-negative return value implies the expression will be scalarized. |
1736 | /// Currently, only single-use chains are considered for scalarization. |
1737 | int computePredInstDiscount(Instruction *PredInst, ScalarCostsTy &ScalarCosts, |
1738 | ElementCount VF); |
1739 | |
1740 | /// Collect the instructions that are uniform after vectorization. An |
1741 | /// instruction is uniform if we represent it with a single scalar value in |
1742 | /// the vectorized loop corresponding to each vector iteration. Examples of |
1743 | /// uniform instructions include pointer operands of consecutive or |
1744 | /// interleaved memory accesses. Note that although uniformity implies an |
1745 | /// instruction will be scalar, the reverse is not true. In general, a |
1746 | /// scalarized instruction will be represented by VF scalar values in the |
1747 | /// vectorized loop, each corresponding to an iteration of the original |
1748 | /// scalar loop. |
1749 | void collectLoopUniforms(ElementCount VF); |
1750 | |
1751 | /// Collect the instructions that are scalar after vectorization. An |
1752 | /// instruction is scalar if it is known to be uniform or will be scalarized |
1753 | /// during vectorization. Non-uniform scalarized instructions will be |
1754 | /// represented by VF values in the vectorized loop, each corresponding to an |
1755 | /// iteration of the original scalar loop. |
1756 | void collectLoopScalars(ElementCount VF); |
1757 | |
1758 | /// Keeps cost model vectorization decision and cost for instructions. |
1759 | /// Right now it is used for memory instructions only. |
1760 | using DecisionList = DenseMap<std::pair<Instruction *, ElementCount>, |
1761 | std::pair<InstWidening, InstructionCost>>; |
1762 | |
1763 | DecisionList WideningDecisions; |
1764 | |
1765 | /// Returns true if \p V is expected to be vectorized and it needs to be |
1766 | /// extracted. |
1767 | bool needsExtract(Value *V, ElementCount VF) const { |
1768 | Instruction *I = dyn_cast<Instruction>(V); |
1769 | if (VF.isScalar() || !I || !TheLoop->contains(I) || |
1770 | TheLoop->isLoopInvariant(I)) |
1771 | return false; |
1772 | |
1773 | // Assume we can vectorize V (and hence we need extraction) if the |
1774 | // scalars are not computed yet. This can happen, because it is called |
1775 | // via getScalarizationOverhead from setCostBasedWideningDecision, before |
1776 | // the scalars are collected. That should be a safe assumption in most |
1777 | // cases, because we check if the operands have vectorizable types |
1778 | // beforehand in LoopVectorizationLegality. |
1779 | return Scalars.find(VF) == Scalars.end() || |
1780 | !isScalarAfterVectorization(I, VF); |
1781 | }; |
1782 | |
1783 | /// Returns a range containing only operands needing to be extracted. |
1784 | SmallVector<Value *, 4> filterExtractingOperands(Instruction::op_range Ops, |
1785 | ElementCount VF) { |
1786 | return SmallVector<Value *, 4>(make_filter_range( |
1787 | Ops, [this, VF](Value *V) { return this->needsExtract(V, VF); })); |
1788 | } |
1789 | |
1790 | /// Determines if we have the infrastructure to vectorize loop \p L and its |
1791 | /// epilogue, assuming the main loop is vectorized by \p VF. |
1792 | bool isCandidateForEpilogueVectorization(const Loop &L, |
1793 | const ElementCount VF) const; |
1794 | |
1795 | /// Returns true if epilogue vectorization is considered profitable, and |
1796 | /// false otherwise. |
1797 | /// \p VF is the vectorization factor chosen for the original loop. |
1798 | bool isEpilogueVectorizationProfitable(const ElementCount VF) const; |
1799 | |
1800 | public: |
1801 | /// The loop that we evaluate. |
1802 | Loop *TheLoop; |
1803 | |
1804 | /// Predicated scalar evolution analysis. |
1805 | PredicatedScalarEvolution &PSE; |
1806 | |
1807 | /// Loop Info analysis. |
1808 | LoopInfo *LI; |
1809 | |
1810 | /// Vectorization legality. |
1811 | LoopVectorizationLegality *Legal; |
1812 | |
1813 | /// Vector target information. |
1814 | const TargetTransformInfo &TTI; |
1815 | |
1816 | /// Target Library Info. |
1817 | const TargetLibraryInfo *TLI; |
1818 | |
1819 | /// Demanded bits analysis. |
1820 | DemandedBits *DB; |
1821 | |
1822 | /// Assumption cache. |
1823 | AssumptionCache *AC; |
1824 | |
1825 | /// Interface to emit optimization remarks. |
1826 | OptimizationRemarkEmitter *ORE; |
1827 | |
1828 | const Function *TheFunction; |
1829 | |
1830 | /// Loop Vectorize Hint. |
1831 | const LoopVectorizeHints *Hints; |
1832 | |
1833 | /// The interleave access information contains groups of interleaved accesses |
1834 | /// with the same stride and close to each other. |
1835 | InterleavedAccessInfo &InterleaveInfo; |
1836 | |
1837 | /// Values to ignore in the cost model. |
1838 | SmallPtrSet<const Value *, 16> ValuesToIgnore; |
1839 | |
1840 | /// Values to ignore in the cost model when VF > 1. |
1841 | SmallPtrSet<const Value *, 16> VecValuesToIgnore; |
1842 | |
1843 | /// Profitable vector factors. |
1844 | SmallVector<VectorizationFactor, 8> ProfitableVFs; |
1845 | }; |
1846 | } // end namespace llvm |
1847 | |
1848 | /// Helper struct to manage generating runtime checks for vectorization. |
1849 | /// |
1850 | /// The runtime checks are created up-front in temporary blocks to allow better |
1851 | /// estimating the cost and un-linked from the existing IR. After deciding to |
1852 | /// vectorize, the checks are moved back. If deciding not to vectorize, the |
1853 | /// temporary blocks are completely removed. |
1854 | class GeneratedRTChecks { |
1855 | /// Basic block which contains the generated SCEV checks, if any. |
1856 | BasicBlock *SCEVCheckBlock = nullptr; |
1857 | |
1858 | /// The value representing the result of the generated SCEV checks. If it is |
1859 | /// nullptr, either no SCEV checks have been generated or they have been used. |
1860 | Value *SCEVCheckCond = nullptr; |
1861 | |
1862 | /// Basic block which contains the generated memory runtime checks, if any. |
1863 | BasicBlock *MemCheckBlock = nullptr; |
1864 | |
1865 | /// The value representing the result of the generated memory runtime checks. |
1866 | /// If it is nullptr, either no memory runtime checks have been generated or |
1867 | /// they have been used. |
1868 | Instruction *MemRuntimeCheckCond = nullptr; |
1869 | |
1870 | DominatorTree *DT; |
1871 | LoopInfo *LI; |
1872 | |
1873 | SCEVExpander SCEVExp; |
1874 | SCEVExpander MemCheckExp; |
1875 | |
1876 | public: |
1877 | GeneratedRTChecks(ScalarEvolution &SE, DominatorTree *DT, LoopInfo *LI, |
1878 | const DataLayout &DL) |
1879 | : DT(DT), LI(LI), SCEVExp(SE, DL, "scev.check"), |
1880 | MemCheckExp(SE, DL, "scev.check") {} |
1881 | |
1882 | /// Generate runtime checks in SCEVCheckBlock and MemCheckBlock, so we can |
1883 | /// accurately estimate the cost of the runtime checks. The blocks are |
1884 | /// un-linked from the IR and is added back during vector code generation. If |
1885 | /// there is no vector code generation, the check blocks are removed |
1886 | /// completely. |
1887 | void Create(Loop *L, const LoopAccessInfo &LAI, |
1888 | const SCEVUnionPredicate &UnionPred) { |
1889 | |
1890 | BasicBlock *LoopHeader = L->getHeader(); |
1891 | BasicBlock *Preheader = L->getLoopPreheader(); |
1892 | |
1893 | // Use SplitBlock to create blocks for SCEV & memory runtime checks to |
1894 | // ensure the blocks are properly added to LoopInfo & DominatorTree. Those |
1895 | // may be used by SCEVExpander. The blocks will be un-linked from their |
1896 | // predecessors and removed from LI & DT at the end of the function. |
1897 | if (!UnionPred.isAlwaysTrue()) { |
1898 | SCEVCheckBlock = SplitBlock(Preheader, Preheader->getTerminator(), DT, LI, |
1899 | nullptr, "vector.scevcheck"); |
1900 | |
1901 | SCEVCheckCond = SCEVExp.expandCodeForPredicate( |
1902 | &UnionPred, SCEVCheckBlock->getTerminator()); |
1903 | } |
1904 | |
1905 | const auto &RtPtrChecking = *LAI.getRuntimePointerChecking(); |
1906 | if (RtPtrChecking.Need) { |
1907 | auto *Pred = SCEVCheckBlock ? SCEVCheckBlock : Preheader; |
1908 | MemCheckBlock = SplitBlock(Pred, Pred->getTerminator(), DT, LI, nullptr, |
1909 | "vector.memcheck"); |
1910 | |
1911 | std::tie(std::ignore, MemRuntimeCheckCond) = |
1912 | addRuntimeChecks(MemCheckBlock->getTerminator(), L, |
1913 | RtPtrChecking.getChecks(), MemCheckExp); |
1914 | assert(MemRuntimeCheckCond &&((MemRuntimeCheckCond && "no RT checks generated although RtPtrChecking " "claimed checks are required") ? static_cast<void> (0) : __assert_fail ("MemRuntimeCheckCond && \"no RT checks generated although RtPtrChecking \" \"claimed checks are required\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1916, __PRETTY_FUNCTION__)) |
1915 | "no RT checks generated although RtPtrChecking "((MemRuntimeCheckCond && "no RT checks generated although RtPtrChecking " "claimed checks are required") ? static_cast<void> (0) : __assert_fail ("MemRuntimeCheckCond && \"no RT checks generated although RtPtrChecking \" \"claimed checks are required\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1916, __PRETTY_FUNCTION__)) |
1916 | "claimed checks are required")((MemRuntimeCheckCond && "no RT checks generated although RtPtrChecking " "claimed checks are required") ? static_cast<void> (0) : __assert_fail ("MemRuntimeCheckCond && \"no RT checks generated although RtPtrChecking \" \"claimed checks are required\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1916, __PRETTY_FUNCTION__)); |
1917 | } |
1918 | |
1919 | if (!MemCheckBlock && !SCEVCheckBlock) |
1920 | return; |
1921 | |
1922 | // Unhook the temporary block with the checks, update various places |
1923 | // accordingly. |
1924 | if (SCEVCheckBlock) |
1925 | SCEVCheckBlock->replaceAllUsesWith(Preheader); |
1926 | if (MemCheckBlock) |
1927 | MemCheckBlock->replaceAllUsesWith(Preheader); |
1928 | |
1929 | if (SCEVCheckBlock) { |
1930 | SCEVCheckBlock->getTerminator()->moveBefore(Preheader->getTerminator()); |
1931 | new UnreachableInst(Preheader->getContext(), SCEVCheckBlock); |
1932 | Preheader->getTerminator()->eraseFromParent(); |
1933 | } |
1934 | if (MemCheckBlock) { |
1935 | MemCheckBlock->getTerminator()->moveBefore(Preheader->getTerminator()); |
1936 | new UnreachableInst(Preheader->getContext(), MemCheckBlock); |
1937 | Preheader->getTerminator()->eraseFromParent(); |
1938 | } |
1939 | |
1940 | DT->changeImmediateDominator(LoopHeader, Preheader); |
1941 | if (MemCheckBlock) { |
1942 | DT->eraseNode(MemCheckBlock); |
1943 | LI->removeBlock(MemCheckBlock); |
1944 | } |
1945 | if (SCEVCheckBlock) { |
1946 | DT->eraseNode(SCEVCheckBlock); |
1947 | LI->removeBlock(SCEVCheckBlock); |
1948 | } |
1949 | } |
1950 | |
1951 | /// Remove the created SCEV & memory runtime check blocks & instructions, if |
1952 | /// unused. |
1953 | ~GeneratedRTChecks() { |
1954 | SCEVExpanderCleaner SCEVCleaner(SCEVExp, *DT); |
1955 | SCEVExpanderCleaner MemCheckCleaner(MemCheckExp, *DT); |
1956 | if (!SCEVCheckCond) |
1957 | SCEVCleaner.markResultUsed(); |
1958 | |
1959 | if (!MemRuntimeCheckCond) |
1960 | MemCheckCleaner.markResultUsed(); |
1961 | |
1962 | if (MemRuntimeCheckCond) { |
1963 | auto &SE = *MemCheckExp.getSE(); |
1964 | // Memory runtime check generation creates compares that use expanded |
1965 | // values. Remove them before running the SCEVExpanderCleaners. |
1966 | for (auto &I : make_early_inc_range(reverse(*MemCheckBlock))) { |
1967 | if (MemCheckExp.isInsertedInstruction(&I)) |
1968 | continue; |
1969 | SE.forgetValue(&I); |
1970 | SE.eraseValueFromMap(&I); |
1971 | I.eraseFromParent(); |
1972 | } |
1973 | } |
1974 | MemCheckCleaner.cleanup(); |
1975 | SCEVCleaner.cleanup(); |
1976 | |
1977 | if (SCEVCheckCond) |
1978 | SCEVCheckBlock->eraseFromParent(); |
1979 | if (MemRuntimeCheckCond) |
1980 | MemCheckBlock->eraseFromParent(); |
1981 | } |
1982 | |
1983 | /// Adds the generated SCEVCheckBlock before \p LoopVectorPreHeader and |
1984 | /// adjusts the branches to branch to the vector preheader or \p Bypass, |
1985 | /// depending on the generated condition. |
1986 | BasicBlock *emitSCEVChecks(Loop *L, BasicBlock *Bypass, |
1987 | BasicBlock *LoopVectorPreHeader, |
1988 | BasicBlock *LoopExitBlock) { |
1989 | if (!SCEVCheckCond) |
1990 | return nullptr; |
1991 | if (auto *C = dyn_cast<ConstantInt>(SCEVCheckCond)) |
1992 | if (C->isZero()) |
1993 | return nullptr; |
1994 | |
1995 | auto *Pred = LoopVectorPreHeader->getSinglePredecessor(); |
1996 | |
1997 | BranchInst::Create(LoopVectorPreHeader, SCEVCheckBlock); |
1998 | // Create new preheader for vector loop. |
1999 | if (auto *PL = LI->getLoopFor(LoopVectorPreHeader)) |
2000 | PL->addBasicBlockToLoop(SCEVCheckBlock, *LI); |
2001 | |
2002 | SCEVCheckBlock->getTerminator()->eraseFromParent(); |
2003 | SCEVCheckBlock->moveBefore(LoopVectorPreHeader); |
2004 | Pred->getTerminator()->replaceSuccessorWith(LoopVectorPreHeader, |
2005 | SCEVCheckBlock); |
2006 | |
2007 | DT->addNewBlock(SCEVCheckBlock, Pred); |
2008 | DT->changeImmediateDominator(LoopVectorPreHeader, SCEVCheckBlock); |
2009 | |
2010 | ReplaceInstWithInst( |
2011 | SCEVCheckBlock->getTerminator(), |
2012 | BranchInst::Create(Bypass, LoopVectorPreHeader, SCEVCheckCond)); |
2013 | // Mark the check as used, to prevent it from being removed during cleanup. |
2014 | SCEVCheckCond = nullptr; |
2015 | return SCEVCheckBlock; |
2016 | } |
2017 | |
2018 | /// Adds the generated MemCheckBlock before \p LoopVectorPreHeader and adjusts |
2019 | /// the branches to branch to the vector preheader or \p Bypass, depending on |
2020 | /// the generated condition. |
2021 | BasicBlock *emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass, |
2022 | BasicBlock *LoopVectorPreHeader) { |
2023 | // Check if we generated code that checks in runtime if arrays overlap. |
2024 | if (!MemRuntimeCheckCond) |
2025 | return nullptr; |
2026 | |
2027 | auto *Pred = LoopVectorPreHeader->getSinglePredecessor(); |
2028 | Pred->getTerminator()->replaceSuccessorWith(LoopVectorPreHeader, |
2029 | MemCheckBlock); |
2030 | |
2031 | DT->addNewBlock(MemCheckBlock, Pred); |
2032 | DT->changeImmediateDominator(LoopVectorPreHeader, MemCheckBlock); |
2033 | MemCheckBlock->moveBefore(LoopVectorPreHeader); |
2034 | |
2035 | if (auto *PL = LI->getLoopFor(LoopVectorPreHeader)) |
2036 | PL->addBasicBlockToLoop(MemCheckBlock, *LI); |
2037 | |
2038 | ReplaceInstWithInst( |
2039 | MemCheckBlock->getTerminator(), |
2040 | BranchInst::Create(Bypass, LoopVectorPreHeader, MemRuntimeCheckCond)); |
2041 | MemCheckBlock->getTerminator()->setDebugLoc( |
2042 | Pred->getTerminator()->getDebugLoc()); |
2043 | |
2044 | // Mark the check as used, to prevent it from being removed during cleanup. |
2045 | MemRuntimeCheckCond = nullptr; |
2046 | return MemCheckBlock; |
2047 | } |
2048 | }; |
2049 | |
2050 | // Return true if \p OuterLp is an outer loop annotated with hints for explicit |
2051 | // vectorization. The loop needs to be annotated with #pragma omp simd |
2052 | // simdlen(#) or #pragma clang vectorize(enable) vectorize_width(#). If the |
2053 | // vector length information is not provided, vectorization is not considered |
2054 | // explicit. Interleave hints are not allowed either. These limitations will be |
2055 | // relaxed in the future. |
2056 | // Please, note that we are currently forced to abuse the pragma 'clang |
2057 | // vectorize' semantics. This pragma provides *auto-vectorization hints* |
2058 | // (i.e., LV must check that vectorization is legal) whereas pragma 'omp simd' |
2059 | // provides *explicit vectorization hints* (LV can bypass legal checks and |
2060 | // assume that vectorization is legal). However, both hints are implemented |
2061 | // using the same metadata (llvm.loop.vectorize, processed by |
2062 | // LoopVectorizeHints). This will be fixed in the future when the native IR |
2063 | // representation for pragma 'omp simd' is introduced. |
2064 | static bool isExplicitVecOuterLoop(Loop *OuterLp, |
2065 | OptimizationRemarkEmitter *ORE) { |
2066 | assert(!OuterLp->isInnermost() && "This is not an outer loop")((!OuterLp->isInnermost() && "This is not an outer loop" ) ? static_cast<void> (0) : __assert_fail ("!OuterLp->isInnermost() && \"This is not an outer loop\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2066, __PRETTY_FUNCTION__)); |
2067 | LoopVectorizeHints Hints(OuterLp, true /*DisableInterleaving*/, *ORE); |
2068 | |
2069 | // Only outer loops with an explicit vectorization hint are supported. |
2070 | // Unannotated outer loops are ignored. |
2071 | if (Hints.getForce() == LoopVectorizeHints::FK_Undefined) |
2072 | return false; |
2073 | |
2074 | Function *Fn = OuterLp->getHeader()->getParent(); |
2075 | if (!Hints.allowVectorization(Fn, OuterLp, |
2076 | true /*VectorizeOnlyWhenForced*/)) { |
2077 | LLVM_DEBUG(dbgs() << "LV: Loop hints prevent outer loop vectorization.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints prevent outer loop vectorization.\n" ; } } while (false); |
2078 | return false; |
2079 | } |
2080 | |
2081 | if (Hints.getInterleave() > 1) { |
2082 | // TODO: Interleave support is future work. |
2083 | LLVM_DEBUG(dbgs() << "LV: Not vectorizing: Interleave is not supported for "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not vectorizing: Interleave is not supported for " "outer loops.\n"; } } while (false) |
2084 | "outer loops.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not vectorizing: Interleave is not supported for " "outer loops.\n"; } } while (false); |
2085 | Hints.emitRemarkWithHints(); |
2086 | return false; |
2087 | } |
2088 | |
2089 | return true; |
2090 | } |
2091 | |
2092 | static void collectSupportedLoops(Loop &L, LoopInfo *LI, |
2093 | OptimizationRemarkEmitter *ORE, |
2094 | SmallVectorImpl<Loop *> &V) { |
2095 | // Collect inner loops and outer loops without irreducible control flow. For |
2096 | // now, only collect outer loops that have explicit vectorization hints. If we |
2097 | // are stress testing the VPlan H-CFG construction, we collect the outermost |
2098 | // loop of every loop nest. |
2099 | if (L.isInnermost() || VPlanBuildStressTest || |
2100 | (EnableVPlanNativePath && isExplicitVecOuterLoop(&L, ORE))) { |
2101 | LoopBlocksRPO RPOT(&L); |
2102 | RPOT.perform(LI); |
2103 | if (!containsIrreducibleCFG<const BasicBlock *>(RPOT, *LI)) { |
2104 | V.push_back(&L); |
2105 | // TODO: Collect inner loops inside marked outer loops in case |
2106 | // vectorization fails for the outer loop. Do not invoke |
2107 | // 'containsIrreducibleCFG' again for inner loops when the outer loop is |
2108 | // already known to be reducible. We can use an inherited attribute for |
2109 | // that. |
2110 | return; |
2111 | } |
2112 | } |
2113 | for (Loop *InnerL : L) |
2114 | collectSupportedLoops(*InnerL, LI, ORE, V); |
2115 | } |
2116 | |
2117 | namespace { |
2118 | |
2119 | /// The LoopVectorize Pass. |
2120 | struct LoopVectorize : public FunctionPass { |
2121 | /// Pass identification, replacement for typeid |
2122 | static char ID; |
2123 | |
2124 | LoopVectorizePass Impl; |
2125 | |
2126 | explicit LoopVectorize(bool InterleaveOnlyWhenForced = false, |
2127 | bool VectorizeOnlyWhenForced = false) |
2128 | : FunctionPass(ID), |
2129 | Impl({InterleaveOnlyWhenForced, VectorizeOnlyWhenForced}) { |
2130 | initializeLoopVectorizePass(*PassRegistry::getPassRegistry()); |
2131 | } |
2132 | |
2133 | bool runOnFunction(Function &F) override { |
2134 | if (skipFunction(F)) |
2135 | return false; |
2136 | |
2137 | auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); |
2138 | auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); |
2139 | auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); |
2140 | auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
2141 | auto *BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI(); |
2142 | auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); |
2143 | auto *TLI = TLIP ? &TLIP->getTLI(F) : nullptr; |
2144 | auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); |
2145 | auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); |
2146 | auto *LAA = &getAnalysis<LoopAccessLegacyAnalysis>(); |
2147 | auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits(); |
2148 | auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(); |
2149 | auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); |
2150 | |
2151 | std::function<const LoopAccessInfo &(Loop &)> GetLAA = |
2152 | [&](Loop &L) -> const LoopAccessInfo & { return LAA->getInfo(&L); }; |
2153 | |
2154 | return Impl.runImpl(F, *SE, *LI, *TTI, *DT, *BFI, TLI, *DB, *AA, *AC, |
2155 | GetLAA, *ORE, PSI).MadeAnyChange; |
2156 | } |
2157 | |
2158 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
2159 | AU.addRequired<AssumptionCacheTracker>(); |
2160 | AU.addRequired<BlockFrequencyInfoWrapperPass>(); |
2161 | AU.addRequired<DominatorTreeWrapperPass>(); |
2162 | AU.addRequired<LoopInfoWrapperPass>(); |
2163 | AU.addRequired<ScalarEvolutionWrapperPass>(); |
2164 | AU.addRequired<TargetTransformInfoWrapperPass>(); |
2165 | AU.addRequired<AAResultsWrapperPass>(); |
2166 | AU.addRequired<LoopAccessLegacyAnalysis>(); |
2167 | AU.addRequired<DemandedBitsWrapperPass>(); |
2168 | AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); |
2169 | AU.addRequired<InjectTLIMappingsLegacy>(); |
2170 | |
2171 | // We currently do not preserve loopinfo/dominator analyses with outer loop |
2172 | // vectorization. Until this is addressed, mark these analyses as preserved |
2173 | // only for non-VPlan-native path. |
2174 | // TODO: Preserve Loop and Dominator analyses for VPlan-native path. |
2175 | if (!EnableVPlanNativePath) { |
2176 | AU.addPreserved<LoopInfoWrapperPass>(); |
2177 | AU.addPreserved<DominatorTreeWrapperPass>(); |
2178 | } |
2179 | |
2180 | AU.addPreserved<BasicAAWrapperPass>(); |
2181 | AU.addPreserved<GlobalsAAWrapperPass>(); |
2182 | AU.addRequired<ProfileSummaryInfoWrapperPass>(); |
2183 | } |
2184 | }; |
2185 | |
2186 | } // end anonymous namespace |
2187 | |
2188 | //===----------------------------------------------------------------------===// |
2189 | // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and |
2190 | // LoopVectorizationCostModel and LoopVectorizationPlanner. |
2191 | //===----------------------------------------------------------------------===// |
2192 | |
2193 | Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) { |
2194 | // We need to place the broadcast of invariant variables outside the loop, |
2195 | // but only if it's proven safe to do so. Else, broadcast will be inside |
2196 | // vector loop body. |
2197 | Instruction *Instr = dyn_cast<Instruction>(V); |
2198 | bool SafeToHoist = OrigLoop->isLoopInvariant(V) && |
2199 | (!Instr || |
2200 | DT->dominates(Instr->getParent(), LoopVectorPreHeader)); |
2201 | // Place the code for broadcasting invariant variables in the new preheader. |
2202 | IRBuilder<>::InsertPointGuard Guard(Builder); |
2203 | if (SafeToHoist) |
2204 | Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); |
2205 | |
2206 | // Broadcast the scalar into all locations in the vector. |
2207 | Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast"); |
2208 | |
2209 | return Shuf; |
2210 | } |
2211 | |
2212 | void InnerLoopVectorizer::createVectorIntOrFpInductionPHI( |
2213 | const InductionDescriptor &II, Value *Step, Value *Start, |
2214 | Instruction *EntryVal, VPValue *Def, VPValue *CastDef, |
2215 | VPTransformState &State) { |
2216 | assert((isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal)) &&(((isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal )) && "Expected either an induction phi-node or a truncate of it!" ) ? static_cast<void> (0) : __assert_fail ("(isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal)) && \"Expected either an induction phi-node or a truncate of it!\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2217, __PRETTY_FUNCTION__)) |
2217 | "Expected either an induction phi-node or a truncate of it!")(((isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal )) && "Expected either an induction phi-node or a truncate of it!" ) ? static_cast<void> (0) : __assert_fail ("(isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal)) && \"Expected either an induction phi-node or a truncate of it!\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2217, __PRETTY_FUNCTION__)); |
2218 | |
2219 | // Construct the initial value of the vector IV in the vector loop preheader |
2220 | auto CurrIP = Builder.saveIP(); |
2221 | Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); |
2222 | if (isa<TruncInst>(EntryVal)) { |
2223 | assert(Start->getType()->isIntegerTy() &&((Start->getType()->isIntegerTy() && "Truncation requires an integer type" ) ? static_cast<void> (0) : __assert_fail ("Start->getType()->isIntegerTy() && \"Truncation requires an integer type\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2224, __PRETTY_FUNCTION__)) |
2224 | "Truncation requires an integer type")((Start->getType()->isIntegerTy() && "Truncation requires an integer type" ) ? static_cast<void> (0) : __assert_fail ("Start->getType()->isIntegerTy() && \"Truncation requires an integer type\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2224, __PRETTY_FUNCTION__)); |
2225 | auto *TruncType = cast<IntegerType>(EntryVal->getType()); |
2226 | Step = Builder.CreateTrunc(Step, TruncType); |
2227 | Start = Builder.CreateCast(Instruction::Trunc, Start, TruncType); |
2228 | } |
2229 | Value *SplatStart = Builder.CreateVectorSplat(VF, Start); |
2230 | Value *SteppedStart = |
2231 | getStepVector(SplatStart, 0, Step, II.getInductionOpcode()); |
2232 | |
2233 | // We create vector phi nodes for both integer and floating-point induction |
2234 | // variables. Here, we determine the kind of arithmetic we will perform. |
2235 | Instruction::BinaryOps AddOp; |
2236 | Instruction::BinaryOps MulOp; |
2237 | if (Step->getType()->isIntegerTy()) { |
2238 | AddOp = Instruction::Add; |
2239 | MulOp = Instruction::Mul; |
2240 | } else { |
2241 | AddOp = II.getInductionOpcode(); |
2242 | MulOp = Instruction::FMul; |
2243 | } |
2244 | |
2245 | // Multiply the vectorization factor by the step using integer or |
2246 | // floating-point arithmetic as appropriate. |
2247 | Value *ConstVF = |
2248 | getSignedIntOrFpConstant(Step->getType(), VF.getKnownMinValue()); |
2249 | Value *Mul = Builder.CreateBinOp(MulOp, Step, ConstVF); |
2250 | |
2251 | // Create a vector splat to use in the induction update. |
2252 | // |
2253 | // FIXME: If the step is non-constant, we create the vector splat with |
2254 | // IRBuilder. IRBuilder can constant-fold the multiply, but it doesn't |
2255 | // handle a constant vector splat. |
2256 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2256, __PRETTY_FUNCTION__)); |
2257 | Value *SplatVF = isa<Constant>(Mul) |
2258 | ? ConstantVector::getSplat(VF, cast<Constant>(Mul)) |
2259 | : Builder.CreateVectorSplat(VF, Mul); |
2260 | Builder.restoreIP(CurrIP); |
2261 | |
2262 | // We may need to add the step a number of times, depending on the unroll |
2263 | // factor. The last of those goes into the PHI. |
2264 | PHINode *VecInd = PHINode::Create(SteppedStart->getType(), 2, "vec.ind", |
2265 | &*LoopVectorBody->getFirstInsertionPt()); |
2266 | VecInd->setDebugLoc(EntryVal->getDebugLoc()); |
2267 | Instruction *LastInduction = VecInd; |
2268 | for (unsigned Part = 0; Part < UF; ++Part) { |
2269 | State.set(Def, LastInduction, Part); |
2270 | |
2271 | if (isa<TruncInst>(EntryVal)) |
2272 | addMetadata(LastInduction, EntryVal); |
2273 | recordVectorLoopValueForInductionCast(II, EntryVal, LastInduction, CastDef, |
2274 | State, Part); |
2275 | |
2276 | LastInduction = cast<Instruction>( |
2277 | Builder.CreateBinOp(AddOp, LastInduction, SplatVF, "step.add")); |
2278 | LastInduction->setDebugLoc(EntryVal->getDebugLoc()); |
2279 | } |
2280 | |
2281 | // Move the last step to the end of the latch block. This ensures consistent |
2282 | // placement of all induction updates. |
2283 | auto *LoopVectorLatch = LI->getLoopFor(LoopVectorBody)->getLoopLatch(); |
2284 | auto *Br = cast<BranchInst>(LoopVectorLatch->getTerminator()); |
2285 | auto *ICmp = cast<Instruction>(Br->getCondition()); |
2286 | LastInduction->moveBefore(ICmp); |
2287 | LastInduction->setName("vec.ind.next"); |
2288 | |
2289 | VecInd->addIncoming(SteppedStart, LoopVectorPreHeader); |
2290 | VecInd->addIncoming(LastInduction, LoopVectorLatch); |
2291 | } |
2292 | |
2293 | bool InnerLoopVectorizer::shouldScalarizeInstruction(Instruction *I) const { |
2294 | return Cost->isScalarAfterVectorization(I, VF) || |
2295 | Cost->isProfitableToScalarize(I, VF); |
2296 | } |
2297 | |
2298 | bool InnerLoopVectorizer::needsScalarInduction(Instruction *IV) const { |
2299 | if (shouldScalarizeInstruction(IV)) |
2300 | return true; |
2301 | auto isScalarInst = [&](User *U) -> bool { |
2302 | auto *I = cast<Instruction>(U); |
2303 | return (OrigLoop->contains(I) && shouldScalarizeInstruction(I)); |
2304 | }; |
2305 | return llvm::any_of(IV->users(), isScalarInst); |
2306 | } |
2307 | |
2308 | void InnerLoopVectorizer::recordVectorLoopValueForInductionCast( |
2309 | const InductionDescriptor &ID, const Instruction *EntryVal, |
2310 | Value *VectorLoopVal, VPValue *CastDef, VPTransformState &State, |
2311 | unsigned Part, unsigned Lane) { |
2312 | assert((isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal)) &&(((isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal )) && "Expected either an induction phi-node or a truncate of it!" ) ? static_cast<void> (0) : __assert_fail ("(isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal)) && \"Expected either an induction phi-node or a truncate of it!\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2313, __PRETTY_FUNCTION__)) |
2313 | "Expected either an induction phi-node or a truncate of it!")(((isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal )) && "Expected either an induction phi-node or a truncate of it!" ) ? static_cast<void> (0) : __assert_fail ("(isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal)) && \"Expected either an induction phi-node or a truncate of it!\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2313, __PRETTY_FUNCTION__)); |
2314 | |
2315 | // This induction variable is not the phi from the original loop but the |
2316 | // newly-created IV based on the proof that casted Phi is equal to the |
2317 | // uncasted Phi in the vectorized loop (under a runtime guard possibly). It |
2318 | // re-uses the same InductionDescriptor that original IV uses but we don't |
2319 | // have to do any recording in this case - that is done when original IV is |
2320 | // processed. |
2321 | if (isa<TruncInst>(EntryVal)) |
2322 | return; |
2323 | |
2324 | const SmallVectorImpl<Instruction *> &Casts = ID.getCastInsts(); |
2325 | if (Casts.empty()) |
2326 | return; |
2327 | // Only the first Cast instruction in the Casts vector is of interest. |
2328 | // The rest of the Casts (if exist) have no uses outside the |
2329 | // induction update chain itself. |
2330 | if (Lane < UINT_MAX(2147483647 *2U +1U)) |
2331 | State.set(CastDef, VectorLoopVal, VPIteration(Part, Lane)); |
2332 | else |
2333 | State.set(CastDef, VectorLoopVal, Part); |
2334 | } |
2335 | |
2336 | void InnerLoopVectorizer::widenIntOrFpInduction(PHINode *IV, Value *Start, |
2337 | TruncInst *Trunc, VPValue *Def, |
2338 | VPValue *CastDef, |
2339 | VPTransformState &State) { |
2340 | assert((IV->getType()->isIntegerTy() || IV != OldInduction) &&(((IV->getType()->isIntegerTy() || IV != OldInduction) && "Primary induction variable must have an integer type") ? static_cast <void> (0) : __assert_fail ("(IV->getType()->isIntegerTy() || IV != OldInduction) && \"Primary induction variable must have an integer type\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2341, __PRETTY_FUNCTION__)) |
2341 | "Primary induction variable must have an integer type")(((IV->getType()->isIntegerTy() || IV != OldInduction) && "Primary induction variable must have an integer type") ? static_cast <void> (0) : __assert_fail ("(IV->getType()->isIntegerTy() || IV != OldInduction) && \"Primary induction variable must have an integer type\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2341, __PRETTY_FUNCTION__)); |
2342 | |
2343 | auto II = Legal->getInductionVars().find(IV); |
2344 | assert(II != Legal->getInductionVars().end() && "IV is not an induction")((II != Legal->getInductionVars().end() && "IV is not an induction" ) ? static_cast<void> (0) : __assert_fail ("II != Legal->getInductionVars().end() && \"IV is not an induction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2344, __PRETTY_FUNCTION__)); |
2345 | |
2346 | auto ID = II->second; |
2347 | assert(IV->getType() == ID.getStartValue()->getType() && "Types must match")((IV->getType() == ID.getStartValue()->getType() && "Types must match") ? static_cast<void> (0) : __assert_fail ("IV->getType() == ID.getStartValue()->getType() && \"Types must match\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2347, __PRETTY_FUNCTION__)); |
2348 | |
2349 | // The value from the original loop to which we are mapping the new induction |
2350 | // variable. |
2351 | Instruction *EntryVal = Trunc ? cast<Instruction>(Trunc) : IV; |
2352 | |
2353 | auto &DL = OrigLoop->getHeader()->getModule()->getDataLayout(); |
2354 | |
2355 | // Generate code for the induction step. Note that induction steps are |
2356 | // required to be loop-invariant |
2357 | auto CreateStepValue = [&](const SCEV *Step) -> Value * { |
2358 | assert(PSE.getSE()->isLoopInvariant(Step, OrigLoop) &&((PSE.getSE()->isLoopInvariant(Step, OrigLoop) && "Induction step should be loop invariant" ) ? static_cast<void> (0) : __assert_fail ("PSE.getSE()->isLoopInvariant(Step, OrigLoop) && \"Induction step should be loop invariant\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2359, __PRETTY_FUNCTION__)) |
2359 | "Induction step should be loop invariant")((PSE.getSE()->isLoopInvariant(Step, OrigLoop) && "Induction step should be loop invariant" ) ? static_cast<void> (0) : __assert_fail ("PSE.getSE()->isLoopInvariant(Step, OrigLoop) && \"Induction step should be loop invariant\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2359, __PRETTY_FUNCTION__)); |
2360 | if (PSE.getSE()->isSCEVable(IV->getType())) { |
2361 | SCEVExpander Exp(*PSE.getSE(), DL, "induction"); |
2362 | return Exp.expandCodeFor(Step, Step->getType(), |
2363 | LoopVectorPreHeader->getTerminator()); |
2364 | } |
2365 | return cast<SCEVUnknown>(Step)->getValue(); |
2366 | }; |
2367 | |
2368 | // The scalar value to broadcast. This is derived from the canonical |
2369 | // induction variable. If a truncation type is given, truncate the canonical |
2370 | // induction variable and step. Otherwise, derive these values from the |
2371 | // induction descriptor. |
2372 | auto CreateScalarIV = [&](Value *&Step) -> Value * { |
2373 | Value *ScalarIV = Induction; |
2374 | if (IV != OldInduction) { |
2375 | ScalarIV = IV->getType()->isIntegerTy() |
2376 | ? Builder.CreateSExtOrTrunc(Induction, IV->getType()) |
2377 | : Builder.CreateCast(Instruction::SIToFP, Induction, |
2378 | IV->getType()); |
2379 | ScalarIV = emitTransformedIndex(Builder, ScalarIV, PSE.getSE(), DL, ID); |
2380 | ScalarIV->setName("offset.idx"); |
2381 | } |
2382 | if (Trunc) { |
2383 | auto *TruncType = cast<IntegerType>(Trunc->getType()); |
2384 | assert(Step->getType()->isIntegerTy() &&((Step->getType()->isIntegerTy() && "Truncation requires an integer step" ) ? static_cast<void> (0) : __assert_fail ("Step->getType()->isIntegerTy() && \"Truncation requires an integer step\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2385, __PRETTY_FUNCTION__)) |
2385 | "Truncation requires an integer step")((Step->getType()->isIntegerTy() && "Truncation requires an integer step" ) ? static_cast<void> (0) : __assert_fail ("Step->getType()->isIntegerTy() && \"Truncation requires an integer step\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2385, __PRETTY_FUNCTION__)); |
2386 | ScalarIV = Builder.CreateTrunc(ScalarIV, TruncType); |
2387 | Step = Builder.CreateTrunc(Step, TruncType); |
2388 | } |
2389 | return ScalarIV; |
2390 | }; |
2391 | |
2392 | // Create the vector values from the scalar IV, in the absence of creating a |
2393 | // vector IV. |
2394 | auto CreateSplatIV = [&](Value *ScalarIV, Value *Step) { |
2395 | Value *Broadcasted = getBroadcastInstrs(ScalarIV); |
2396 | for (unsigned Part = 0; Part < UF; ++Part) { |
2397 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2397, __PRETTY_FUNCTION__)); |
2398 | Value *EntryPart = |
2399 | getStepVector(Broadcasted, VF.getKnownMinValue() * Part, Step, |
2400 | ID.getInductionOpcode()); |
2401 | State.set(Def, EntryPart, Part); |
2402 | if (Trunc) |
2403 | addMetadata(EntryPart, Trunc); |
2404 | recordVectorLoopValueForInductionCast(ID, EntryVal, EntryPart, CastDef, |
2405 | State, Part); |
2406 | } |
2407 | }; |
2408 | |
2409 | // Fast-math-flags propagate from the original induction instruction. |
2410 | IRBuilder<>::FastMathFlagGuard FMFG(Builder); |
2411 | if (ID.getInductionBinOp() && isa<FPMathOperator>(ID.getInductionBinOp())) |
2412 | Builder.setFastMathFlags(ID.getInductionBinOp()->getFastMathFlags()); |
2413 | |
2414 | // Now do the actual transformations, and start with creating the step value. |
2415 | Value *Step = CreateStepValue(ID.getStep()); |
2416 | if (VF.isZero() || VF.isScalar()) { |
2417 | Value *ScalarIV = CreateScalarIV(Step); |
2418 | CreateSplatIV(ScalarIV, Step); |
2419 | return; |
2420 | } |
2421 | |
2422 | // Determine if we want a scalar version of the induction variable. This is |
2423 | // true if the induction variable itself is not widened, or if it has at |
2424 | // least one user in the loop that is not widened. |
2425 | auto NeedsScalarIV = needsScalarInduction(EntryVal); |
2426 | if (!NeedsScalarIV) { |
2427 | createVectorIntOrFpInductionPHI(ID, Step, Start, EntryVal, Def, CastDef, |
2428 | State); |
2429 | return; |
2430 | } |
2431 | |
2432 | // Try to create a new independent vector induction variable. If we can't |
2433 | // create the phi node, we will splat the scalar induction variable in each |
2434 | // loop iteration. |
2435 | if (!shouldScalarizeInstruction(EntryVal)) { |
2436 | createVectorIntOrFpInductionPHI(ID, Step, Start, EntryVal, Def, CastDef, |
2437 | State); |
2438 | Value *ScalarIV = CreateScalarIV(Step); |
2439 | // Create scalar steps that can be used by instructions we will later |
2440 | // scalarize. Note that the addition of the scalar steps will not increase |
2441 | // the number of instructions in the loop in the common case prior to |
2442 | // InstCombine. We will be trading one vector extract for each scalar step. |
2443 | buildScalarSteps(ScalarIV, Step, EntryVal, ID, Def, CastDef, State); |
2444 | return; |
2445 | } |
2446 | |
2447 | // All IV users are scalar instructions, so only emit a scalar IV, not a |
2448 | // vectorised IV. Except when we tail-fold, then the splat IV feeds the |
2449 | // predicate used by the masked loads/stores. |
2450 | Value *ScalarIV = CreateScalarIV(Step); |
2451 | if (!Cost->isScalarEpilogueAllowed()) |
2452 | CreateSplatIV(ScalarIV, Step); |
2453 | buildScalarSteps(ScalarIV, Step, EntryVal, ID, Def, CastDef, State); |
2454 | } |
2455 | |
2456 | Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx, Value *Step, |
2457 | Instruction::BinaryOps BinOp) { |
2458 | // Create and check the types. |
2459 | auto *ValVTy = cast<FixedVectorType>(Val->getType()); |
2460 | int VLen = ValVTy->getNumElements(); |
2461 | |
2462 | Type *STy = Val->getType()->getScalarType(); |
2463 | assert((STy->isIntegerTy() || STy->isFloatingPointTy()) &&(((STy->isIntegerTy() || STy->isFloatingPointTy()) && "Induction Step must be an integer or FP") ? static_cast< void> (0) : __assert_fail ("(STy->isIntegerTy() || STy->isFloatingPointTy()) && \"Induction Step must be an integer or FP\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2464, __PRETTY_FUNCTION__)) |
2464 | "Induction Step must be an integer or FP")(((STy->isIntegerTy() || STy->isFloatingPointTy()) && "Induction Step must be an integer or FP") ? static_cast< void> (0) : __assert_fail ("(STy->isIntegerTy() || STy->isFloatingPointTy()) && \"Induction Step must be an integer or FP\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2464, __PRETTY_FUNCTION__)); |
2465 | assert(Step->getType() == STy && "Step has wrong type")((Step->getType() == STy && "Step has wrong type") ? static_cast<void> (0) : __assert_fail ("Step->getType() == STy && \"Step has wrong type\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2465, __PRETTY_FUNCTION__)); |
2466 | |
2467 | SmallVector<Constant *, 8> Indices; |
2468 | |
2469 | if (STy->isIntegerTy()) { |
2470 | // Create a vector of consecutive numbers from zero to VF. |
2471 | for (int i = 0; i < VLen; ++i) |
2472 | Indices.push_back(ConstantInt::get(STy, StartIdx + i)); |
2473 | |
2474 | // Add the consecutive indices to the vector value. |
2475 | Constant *Cv = ConstantVector::get(Indices); |
2476 | assert(Cv->getType() == Val->getType() && "Invalid consecutive vec")((Cv->getType() == Val->getType() && "Invalid consecutive vec" ) ? static_cast<void> (0) : __assert_fail ("Cv->getType() == Val->getType() && \"Invalid consecutive vec\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2476, __PRETTY_FUNCTION__)); |
2477 | Step = Builder.CreateVectorSplat(VLen, Step); |
2478 | assert(Step->getType() == Val->getType() && "Invalid step vec")((Step->getType() == Val->getType() && "Invalid step vec" ) ? static_cast<void> (0) : __assert_fail ("Step->getType() == Val->getType() && \"Invalid step vec\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2478, __PRETTY_FUNCTION__)); |
2479 | // FIXME: The newly created binary instructions should contain nsw/nuw flags, |
2480 | // which can be found from the original scalar operations. |
2481 | Step = Builder.CreateMul(Cv, Step); |
2482 | return Builder.CreateAdd(Val, Step, "induction"); |
2483 | } |
2484 | |
2485 | // Floating point induction. |
2486 | assert((BinOp == Instruction::FAdd || BinOp == Instruction::FSub) &&(((BinOp == Instruction::FAdd || BinOp == Instruction::FSub) && "Binary Opcode should be specified for FP induction") ? static_cast <void> (0) : __assert_fail ("(BinOp == Instruction::FAdd || BinOp == Instruction::FSub) && \"Binary Opcode should be specified for FP induction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2487, __PRETTY_FUNCTION__)) |
2487 | "Binary Opcode should be specified for FP induction")(((BinOp == Instruction::FAdd || BinOp == Instruction::FSub) && "Binary Opcode should be specified for FP induction") ? static_cast <void> (0) : __assert_fail ("(BinOp == Instruction::FAdd || BinOp == Instruction::FSub) && \"Binary Opcode should be specified for FP induction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2487, __PRETTY_FUNCTION__)); |
2488 | // Create a vector of consecutive numbers from zero to VF. |
2489 | for (int i = 0; i < VLen; ++i) |
2490 | Indices.push_back(ConstantFP::get(STy, (double)(StartIdx + i))); |
2491 | |
2492 | // Add the consecutive indices to the vector value. |
2493 | // Floating-point operations inherit FMF via the builder's flags. |
2494 | Constant *Cv = ConstantVector::get(Indices); |
2495 | Step = Builder.CreateVectorSplat(VLen, Step); |
2496 | Value *MulOp = Builder.CreateFMul(Cv, Step); |
2497 | return Builder.CreateBinOp(BinOp, Val, MulOp, "induction"); |
2498 | } |
2499 | |
2500 | void InnerLoopVectorizer::buildScalarSteps(Value *ScalarIV, Value *Step, |
2501 | Instruction *EntryVal, |
2502 | const InductionDescriptor &ID, |
2503 | VPValue *Def, VPValue *CastDef, |
2504 | VPTransformState &State) { |
2505 | // We shouldn't have to build scalar steps if we aren't vectorizing. |
2506 | assert(VF.isVector() && "VF should be greater than one")((VF.isVector() && "VF should be greater than one") ? static_cast<void> (0) : __assert_fail ("VF.isVector() && \"VF should be greater than one\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2506, __PRETTY_FUNCTION__)); |
2507 | // Get the value type and ensure it and the step have the same integer type. |
2508 | Type *ScalarIVTy = ScalarIV->getType()->getScalarType(); |
2509 | assert(ScalarIVTy == Step->getType() &&((ScalarIVTy == Step->getType() && "Val and Step should have the same type" ) ? static_cast<void> (0) : __assert_fail ("ScalarIVTy == Step->getType() && \"Val and Step should have the same type\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2510, __PRETTY_FUNCTION__)) |
2510 | "Val and Step should have the same type")((ScalarIVTy == Step->getType() && "Val and Step should have the same type" ) ? static_cast<void> (0) : __assert_fail ("ScalarIVTy == Step->getType() && \"Val and Step should have the same type\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2510, __PRETTY_FUNCTION__)); |
2511 | |
2512 | // We build scalar steps for both integer and floating-point induction |
2513 | // variables. Here, we determine the kind of arithmetic we will perform. |
2514 | Instruction::BinaryOps AddOp; |
2515 | Instruction::BinaryOps MulOp; |
2516 | if (ScalarIVTy->isIntegerTy()) { |
2517 | AddOp = Instruction::Add; |
2518 | MulOp = Instruction::Mul; |
2519 | } else { |
2520 | AddOp = ID.getInductionOpcode(); |
2521 | MulOp = Instruction::FMul; |
2522 | } |
2523 | |
2524 | // Determine the number of scalars we need to generate for each unroll |
2525 | // iteration. If EntryVal is uniform, we only need to generate the first |
2526 | // lane. Otherwise, we generate all VF values. |
2527 | unsigned Lanes = |
2528 | Cost->isUniformAfterVectorization(cast<Instruction>(EntryVal), VF) |
2529 | ? 1 |
2530 | : VF.getKnownMinValue(); |
2531 | assert((!VF.isScalable() || Lanes == 1) &&(((!VF.isScalable() || Lanes == 1) && "Should never scalarize a scalable vector" ) ? static_cast<void> (0) : __assert_fail ("(!VF.isScalable() || Lanes == 1) && \"Should never scalarize a scalable vector\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2532, __PRETTY_FUNCTION__)) |
2532 | "Should never scalarize a scalable vector")(((!VF.isScalable() || Lanes == 1) && "Should never scalarize a scalable vector" ) ? static_cast<void> (0) : __assert_fail ("(!VF.isScalable() || Lanes == 1) && \"Should never scalarize a scalable vector\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2532, __PRETTY_FUNCTION__)); |
2533 | // Compute the scalar steps and save the results in State. |
2534 | for (unsigned Part = 0; Part < UF; ++Part) { |
2535 | for (unsigned Lane = 0; Lane < Lanes; ++Lane) { |
2536 | auto *IntStepTy = IntegerType::get(ScalarIVTy->getContext(), |
2537 | ScalarIVTy->getScalarSizeInBits()); |
2538 | Value *StartIdx = |
2539 | createStepForVF(Builder, ConstantInt::get(IntStepTy, Part), VF); |
2540 | if (ScalarIVTy->isFloatingPointTy()) |
2541 | StartIdx = Builder.CreateSIToFP(StartIdx, ScalarIVTy); |
2542 | StartIdx = Builder.CreateBinOp( |
2543 | AddOp, StartIdx, getSignedIntOrFpConstant(ScalarIVTy, Lane)); |
2544 | // The step returned by `createStepForVF` is a runtime-evaluated value |
2545 | // when VF is scalable. Otherwise, it should be folded into a Constant. |
2546 | assert((VF.isScalable() || isa<Constant>(StartIdx)) &&(((VF.isScalable() || isa<Constant>(StartIdx)) && "Expected StartIdx to be folded to a constant when VF is not " "scalable") ? static_cast<void> (0) : __assert_fail ("(VF.isScalable() || isa<Constant>(StartIdx)) && \"Expected StartIdx to be folded to a constant when VF is not \" \"scalable\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2548, __PRETTY_FUNCTION__)) |
2547 | "Expected StartIdx to be folded to a constant when VF is not "(((VF.isScalable() || isa<Constant>(StartIdx)) && "Expected StartIdx to be folded to a constant when VF is not " "scalable") ? static_cast<void> (0) : __assert_fail ("(VF.isScalable() || isa<Constant>(StartIdx)) && \"Expected StartIdx to be folded to a constant when VF is not \" \"scalable\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2548, __PRETTY_FUNCTION__)) |
2548 | "scalable")(((VF.isScalable() || isa<Constant>(StartIdx)) && "Expected StartIdx to be folded to a constant when VF is not " "scalable") ? static_cast<void> (0) : __assert_fail ("(VF.isScalable() || isa<Constant>(StartIdx)) && \"Expected StartIdx to be folded to a constant when VF is not \" \"scalable\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2548, __PRETTY_FUNCTION__)); |
2549 | auto *Mul = Builder.CreateBinOp(MulOp, StartIdx, Step); |
2550 | auto *Add = Builder.CreateBinOp(AddOp, ScalarIV, Mul); |
2551 | State.set(Def, Add, VPIteration(Part, Lane)); |
2552 | recordVectorLoopValueForInductionCast(ID, EntryVal, Add, CastDef, State, |
2553 | Part, Lane); |
2554 | } |
2555 | } |
2556 | } |
2557 | |
2558 | void InnerLoopVectorizer::packScalarIntoVectorValue(VPValue *Def, |
2559 | const VPIteration &Instance, |
2560 | VPTransformState &State) { |
2561 | Value *ScalarInst = State.get(Def, Instance); |
2562 | Value *VectorValue = State.get(Def, Instance.Part); |
2563 | VectorValue = Builder.CreateInsertElement( |
2564 | VectorValue, ScalarInst, |
2565 | Instance.Lane.getAsRuntimeExpr(State.Builder, VF)); |
2566 | State.set(Def, VectorValue, Instance.Part); |
2567 | } |
2568 | |
2569 | Value *InnerLoopVectorizer::reverseVector(Value *Vec) { |
2570 | assert(Vec->getType()->isVectorTy() && "Invalid type")((Vec->getType()->isVectorTy() && "Invalid type" ) ? static_cast<void> (0) : __assert_fail ("Vec->getType()->isVectorTy() && \"Invalid type\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2570, __PRETTY_FUNCTION__)); |
2571 | assert(!VF.isScalable() && "Cannot reverse scalable vectors")((!VF.isScalable() && "Cannot reverse scalable vectors" ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"Cannot reverse scalable vectors\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2571, __PRETTY_FUNCTION__)); |
2572 | SmallVector<int, 8> ShuffleMask; |
2573 | for (unsigned i = 0; i < VF.getKnownMinValue(); ++i) |
2574 | ShuffleMask.push_back(VF.getKnownMinValue() - i - 1); |
2575 | |
2576 | return Builder.CreateShuffleVector(Vec, ShuffleMask, "reverse"); |
2577 | } |
2578 | |
2579 | // Return whether we allow using masked interleave-groups (for dealing with |
2580 | // strided loads/stores that reside in predicated blocks, or for dealing |
2581 | // with gaps). |
2582 | static bool useMaskedInterleavedAccesses(const TargetTransformInfo &TTI) { |
2583 | // If an override option has been passed in for interleaved accesses, use it. |
2584 | if (EnableMaskedInterleavedMemAccesses.getNumOccurrences() > 0) |
2585 | return EnableMaskedInterleavedMemAccesses; |
2586 | |
2587 | return TTI.enableMaskedInterleavedAccessVectorization(); |
2588 | } |
2589 | |
2590 | // Try to vectorize the interleave group that \p Instr belongs to. |
2591 | // |
2592 | // E.g. Translate following interleaved load group (factor = 3): |
2593 | // for (i = 0; i < N; i+=3) { |
2594 | // R = Pic[i]; // Member of index 0 |
2595 | // G = Pic[i+1]; // Member of index 1 |
2596 | // B = Pic[i+2]; // Member of index 2 |
2597 | // ... // do something to R, G, B |
2598 | // } |
2599 | // To: |
2600 | // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B |
2601 | // %R.vec = shuffle %wide.vec, poison, <0, 3, 6, 9> ; R elements |
2602 | // %G.vec = shuffle %wide.vec, poison, <1, 4, 7, 10> ; G elements |
2603 | // %B.vec = shuffle %wide.vec, poison, <2, 5, 8, 11> ; B elements |
2604 | // |
2605 | // Or translate following interleaved store group (factor = 3): |
2606 | // for (i = 0; i < N; i+=3) { |
2607 | // ... do something to R, G, B |
2608 | // Pic[i] = R; // Member of index 0 |
2609 | // Pic[i+1] = G; // Member of index 1 |
2610 | // Pic[i+2] = B; // Member of index 2 |
2611 | // } |
2612 | // To: |
2613 | // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7> |
2614 | // %B_U.vec = shuffle %B.vec, poison, <0, 1, 2, 3, u, u, u, u> |
2615 | // %interleaved.vec = shuffle %R_G.vec, %B_U.vec, |
2616 | // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements |
2617 | // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B |
2618 | void InnerLoopVectorizer::vectorizeInterleaveGroup( |
2619 | const InterleaveGroup<Instruction> *Group, ArrayRef<VPValue *> VPDefs, |
2620 | VPTransformState &State, VPValue *Addr, ArrayRef<VPValue *> StoredValues, |
2621 | VPValue *BlockInMask) { |
2622 | Instruction *Instr = Group->getInsertPos(); |
2623 | const DataLayout &DL = Instr->getModule()->getDataLayout(); |
2624 | |
2625 | // Prepare for the vector type of the interleaved load/store. |
2626 | Type *ScalarTy = getMemInstValueType(Instr); |
2627 | unsigned InterleaveFactor = Group->getFactor(); |
2628 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2628, __PRETTY_FUNCTION__)); |
2629 | auto *VecTy = VectorType::get(ScalarTy, VF * InterleaveFactor); |
2630 | |
2631 | // Prepare for the new pointers. |
2632 | SmallVector<Value *, 2> AddrParts; |
2633 | unsigned Index = Group->getIndex(Instr); |
2634 | |
2635 | // TODO: extend the masked interleaved-group support to reversed access. |
2636 | assert((!BlockInMask || !Group->isReverse()) &&(((!BlockInMask || !Group->isReverse()) && "Reversed masked interleave-group not supported." ) ? static_cast<void> (0) : __assert_fail ("(!BlockInMask || !Group->isReverse()) && \"Reversed masked interleave-group not supported.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2637, __PRETTY_FUNCTION__)) |
2637 | "Reversed masked interleave-group not supported.")(((!BlockInMask || !Group->isReverse()) && "Reversed masked interleave-group not supported." ) ? static_cast<void> (0) : __assert_fail ("(!BlockInMask || !Group->isReverse()) && \"Reversed masked interleave-group not supported.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2637, __PRETTY_FUNCTION__)); |
2638 | |
2639 | // If the group is reverse, adjust the index to refer to the last vector lane |
2640 | // instead of the first. We adjust the index from the first vector lane, |
2641 | // rather than directly getting the pointer for lane VF - 1, because the |
2642 | // pointer operand of the interleaved access is supposed to be uniform. For |
2643 | // uniform instructions, we're only required to generate a value for the |
2644 | // first vector lane in each unroll iteration. |
2645 | assert(!VF.isScalable() &&((!VF.isScalable() && "scalable vector reverse operation is not implemented" ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vector reverse operation is not implemented\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2646, __PRETTY_FUNCTION__)) |
2646 | "scalable vector reverse operation is not implemented")((!VF.isScalable() && "scalable vector reverse operation is not implemented" ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vector reverse operation is not implemented\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2646, __PRETTY_FUNCTION__)); |
2647 | if (Group->isReverse()) |
2648 | Index += (VF.getKnownMinValue() - 1) * Group->getFactor(); |
2649 | |
2650 | for (unsigned Part = 0; Part < UF; Part++) { |
2651 | Value *AddrPart = State.get(Addr, VPIteration(Part, 0)); |
2652 | setDebugLocFromInst(Builder, AddrPart); |
2653 | |
2654 | // Notice current instruction could be any index. Need to adjust the address |
2655 | // to the member of index 0. |
2656 | // |
2657 | // E.g. a = A[i+1]; // Member of index 1 (Current instruction) |
2658 | // b = A[i]; // Member of index 0 |
2659 | // Current pointer is pointed to A[i+1], adjust it to A[i]. |
2660 | // |
2661 | // E.g. A[i+1] = a; // Member of index 1 |
2662 | // A[i] = b; // Member of index 0 |
2663 | // A[i+2] = c; // Member of index 2 (Current instruction) |
2664 | // Current pointer is pointed to A[i+2], adjust it to A[i]. |
2665 | |
2666 | bool InBounds = false; |
2667 | if (auto *gep = dyn_cast<GetElementPtrInst>(AddrPart->stripPointerCasts())) |
2668 | InBounds = gep->isInBounds(); |
2669 | AddrPart = Builder.CreateGEP(ScalarTy, AddrPart, Builder.getInt32(-Index)); |
2670 | cast<GetElementPtrInst>(AddrPart)->setIsInBounds(InBounds); |
2671 | |
2672 | // Cast to the vector pointer type. |
2673 | unsigned AddressSpace = AddrPart->getType()->getPointerAddressSpace(); |
2674 | Type *PtrTy = VecTy->getPointerTo(AddressSpace); |
2675 | AddrParts.push_back(Builder.CreateBitCast(AddrPart, PtrTy)); |
2676 | } |
2677 | |
2678 | setDebugLocFromInst(Builder, Instr); |
2679 | Value *PoisonVec = PoisonValue::get(VecTy); |
2680 | |
2681 | Value *MaskForGaps = nullptr; |
2682 | if (Group->requiresScalarEpilogue() && !Cost->isScalarEpilogueAllowed()) { |
2683 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2683, __PRETTY_FUNCTION__)); |
2684 | MaskForGaps = createBitMaskForGaps(Builder, VF.getKnownMinValue(), *Group); |
2685 | assert(MaskForGaps && "Mask for Gaps is required but it is null")((MaskForGaps && "Mask for Gaps is required but it is null" ) ? static_cast<void> (0) : __assert_fail ("MaskForGaps && \"Mask for Gaps is required but it is null\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2685, __PRETTY_FUNCTION__)); |
2686 | } |
2687 | |
2688 | // Vectorize the interleaved load group. |
2689 | if (isa<LoadInst>(Instr)) { |
2690 | // For each unroll part, create a wide load for the group. |
2691 | SmallVector<Value *, 2> NewLoads; |
2692 | for (unsigned Part = 0; Part < UF; Part++) { |
2693 | Instruction *NewLoad; |
2694 | if (BlockInMask || MaskForGaps) { |
2695 | assert(useMaskedInterleavedAccesses(*TTI) &&((useMaskedInterleavedAccesses(*TTI) && "masked interleaved groups are not allowed." ) ? static_cast<void> (0) : __assert_fail ("useMaskedInterleavedAccesses(*TTI) && \"masked interleaved groups are not allowed.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2696, __PRETTY_FUNCTION__)) |
2696 | "masked interleaved groups are not allowed.")((useMaskedInterleavedAccesses(*TTI) && "masked interleaved groups are not allowed." ) ? static_cast<void> (0) : __assert_fail ("useMaskedInterleavedAccesses(*TTI) && \"masked interleaved groups are not allowed.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2696, __PRETTY_FUNCTION__)); |
2697 | Value *GroupMask = MaskForGaps; |
2698 | if (BlockInMask) { |
2699 | Value *BlockInMaskPart = State.get(BlockInMask, Part); |
2700 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2700, __PRETTY_FUNCTION__)); |
2701 | Value *ShuffledMask = Builder.CreateShuffleVector( |
2702 | BlockInMaskPart, |
2703 | createReplicatedMask(InterleaveFactor, VF.getKnownMinValue()), |
2704 | "interleaved.mask"); |
2705 | GroupMask = MaskForGaps |
2706 | ? Builder.CreateBinOp(Instruction::And, ShuffledMask, |
2707 | MaskForGaps) |
2708 | : ShuffledMask; |
2709 | } |
2710 | NewLoad = |
2711 | Builder.CreateMaskedLoad(AddrParts[Part], Group->getAlign(), |
2712 | GroupMask, PoisonVec, "wide.masked.vec"); |
2713 | } |
2714 | else |
2715 | NewLoad = Builder.CreateAlignedLoad(VecTy, AddrParts[Part], |
2716 | Group->getAlign(), "wide.vec"); |
2717 | Group->addMetadata(NewLoad); |
2718 | NewLoads.push_back(NewLoad); |
2719 | } |
2720 | |
2721 | // For each member in the group, shuffle out the appropriate data from the |
2722 | // wide loads. |
2723 | unsigned J = 0; |
2724 | for (unsigned I = 0; I < InterleaveFactor; ++I) { |
2725 | Instruction *Member = Group->getMember(I); |
2726 | |
2727 | // Skip the gaps in the group. |
2728 | if (!Member) |
2729 | continue; |
2730 | |
2731 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2731, __PRETTY_FUNCTION__)); |
2732 | auto StrideMask = |
2733 | createStrideMask(I, InterleaveFactor, VF.getKnownMinValue()); |
2734 | for (unsigned Part = 0; Part < UF; Part++) { |
2735 | Value *StridedVec = Builder.CreateShuffleVector( |
2736 | NewLoads[Part], StrideMask, "strided.vec"); |
2737 | |
2738 | // If this member has different type, cast the result type. |
2739 | if (Member->getType() != ScalarTy) { |
2740 | assert(!VF.isScalable() && "VF is assumed to be non scalable.")((!VF.isScalable() && "VF is assumed to be non scalable." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"VF is assumed to be non scalable.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2740, __PRETTY_FUNCTION__)); |
2741 | VectorType *OtherVTy = VectorType::get(Member->getType(), VF); |
2742 | StridedVec = createBitOrPointerCast(StridedVec, OtherVTy, DL); |
2743 | } |
2744 | |
2745 | if (Group->isReverse()) |
2746 | StridedVec = reverseVector(StridedVec); |
2747 | |
2748 | State.set(VPDefs[J], StridedVec, Part); |
2749 | } |
2750 | ++J; |
2751 | } |
2752 | return; |
2753 | } |
2754 | |
2755 | // The sub vector type for current instruction. |
2756 | assert(!VF.isScalable() && "VF is assumed to be non scalable.")((!VF.isScalable() && "VF is assumed to be non scalable." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"VF is assumed to be non scalable.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2756, __PRETTY_FUNCTION__)); |
2757 | auto *SubVT = VectorType::get(ScalarTy, VF); |
2758 | |
2759 | // Vectorize the interleaved store group. |
2760 | for (unsigned Part = 0; Part < UF; Part++) { |
2761 | // Collect the stored vector from each member. |
2762 | SmallVector<Value *, 4> StoredVecs; |
2763 | for (unsigned i = 0; i < InterleaveFactor; i++) { |
2764 | // Interleaved store group doesn't allow a gap, so each index has a member |
2765 | assert(Group->getMember(i) && "Fail to get a member from an interleaved store group")((Group->getMember(i) && "Fail to get a member from an interleaved store group" ) ? static_cast<void> (0) : __assert_fail ("Group->getMember(i) && \"Fail to get a member from an interleaved store group\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2765, __PRETTY_FUNCTION__)); |
2766 | |
2767 | Value *StoredVec = State.get(StoredValues[i], Part); |
2768 | |
2769 | if (Group->isReverse()) |
2770 | StoredVec = reverseVector(StoredVec); |
2771 | |
2772 | // If this member has different type, cast it to a unified type. |
2773 | |
2774 | if (StoredVec->getType() != SubVT) |
2775 | StoredVec = createBitOrPointerCast(StoredVec, SubVT, DL); |
2776 | |
2777 | StoredVecs.push_back(StoredVec); |
2778 | } |
2779 | |
2780 | // Concatenate all vectors into a wide vector. |
2781 | Value *WideVec = concatenateVectors(Builder, StoredVecs); |
2782 | |
2783 | // Interleave the elements in the wide vector. |
2784 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2784, __PRETTY_FUNCTION__)); |
2785 | Value *IVec = Builder.CreateShuffleVector( |
2786 | WideVec, createInterleaveMask(VF.getKnownMinValue(), InterleaveFactor), |
2787 | "interleaved.vec"); |
2788 | |
2789 | Instruction *NewStoreInstr; |
2790 | if (BlockInMask) { |
2791 | Value *BlockInMaskPart = State.get(BlockInMask, Part); |
2792 | Value *ShuffledMask = Builder.CreateShuffleVector( |
2793 | BlockInMaskPart, |
2794 | createReplicatedMask(InterleaveFactor, VF.getKnownMinValue()), |
2795 | "interleaved.mask"); |
2796 | NewStoreInstr = Builder.CreateMaskedStore( |
2797 | IVec, AddrParts[Part], Group->getAlign(), ShuffledMask); |
2798 | } |
2799 | else |
2800 | NewStoreInstr = |
2801 | Builder.CreateAlignedStore(IVec, AddrParts[Part], Group->getAlign()); |
2802 | |
2803 | Group->addMetadata(NewStoreInstr); |
2804 | } |
2805 | } |
2806 | |
2807 | void InnerLoopVectorizer::vectorizeMemoryInstruction( |
2808 | Instruction *Instr, VPTransformState &State, VPValue *Def, VPValue *Addr, |
2809 | VPValue *StoredValue, VPValue *BlockInMask) { |
2810 | // Attempt to issue a wide load. |
2811 | LoadInst *LI = dyn_cast<LoadInst>(Instr); |
2812 | StoreInst *SI = dyn_cast<StoreInst>(Instr); |
2813 | |
2814 | assert((LI || SI) && "Invalid Load/Store instruction")(((LI || SI) && "Invalid Load/Store instruction") ? static_cast <void> (0) : __assert_fail ("(LI || SI) && \"Invalid Load/Store instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2814, __PRETTY_FUNCTION__)); |
2815 | assert((!SI || StoredValue) && "No stored value provided for widened store")(((!SI || StoredValue) && "No stored value provided for widened store" ) ? static_cast<void> (0) : __assert_fail ("(!SI || StoredValue) && \"No stored value provided for widened store\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2815, __PRETTY_FUNCTION__)); |
2816 | assert((!LI || !StoredValue) && "Stored value provided for widened load")(((!LI || !StoredValue) && "Stored value provided for widened load" ) ? static_cast<void> (0) : __assert_fail ("(!LI || !StoredValue) && \"Stored value provided for widened load\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2816, __PRETTY_FUNCTION__)); |
2817 | |
2818 | LoopVectorizationCostModel::InstWidening Decision = |
2819 | Cost->getWideningDecision(Instr, VF); |
2820 | assert((Decision == LoopVectorizationCostModel::CM_Widen ||(((Decision == LoopVectorizationCostModel::CM_Widen || Decision == LoopVectorizationCostModel::CM_Widen_Reverse || Decision == LoopVectorizationCostModel::CM_GatherScatter) && "CM decision is not to widen the memory instruction" ) ? static_cast<void> (0) : __assert_fail ("(Decision == LoopVectorizationCostModel::CM_Widen || Decision == LoopVectorizationCostModel::CM_Widen_Reverse || Decision == LoopVectorizationCostModel::CM_GatherScatter) && \"CM decision is not to widen the memory instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2823, __PRETTY_FUNCTION__)) |
2821 | Decision == LoopVectorizationCostModel::CM_Widen_Reverse ||(((Decision == LoopVectorizationCostModel::CM_Widen || Decision == LoopVectorizationCostModel::CM_Widen_Reverse || Decision == LoopVectorizationCostModel::CM_GatherScatter) && "CM decision is not to widen the memory instruction" ) ? static_cast<void> (0) : __assert_fail ("(Decision == LoopVectorizationCostModel::CM_Widen || Decision == LoopVectorizationCostModel::CM_Widen_Reverse || Decision == LoopVectorizationCostModel::CM_GatherScatter) && \"CM decision is not to widen the memory instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2823, __PRETTY_FUNCTION__)) |
2822 | Decision == LoopVectorizationCostModel::CM_GatherScatter) &&(((Decision == LoopVectorizationCostModel::CM_Widen || Decision == LoopVectorizationCostModel::CM_Widen_Reverse || Decision == LoopVectorizationCostModel::CM_GatherScatter) && "CM decision is not to widen the memory instruction" ) ? static_cast<void> (0) : __assert_fail ("(Decision == LoopVectorizationCostModel::CM_Widen || Decision == LoopVectorizationCostModel::CM_Widen_Reverse || Decision == LoopVectorizationCostModel::CM_GatherScatter) && \"CM decision is not to widen the memory instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2823, __PRETTY_FUNCTION__)) |
2823 | "CM decision is not to widen the memory instruction")(((Decision == LoopVectorizationCostModel::CM_Widen || Decision == LoopVectorizationCostModel::CM_Widen_Reverse || Decision == LoopVectorizationCostModel::CM_GatherScatter) && "CM decision is not to widen the memory instruction" ) ? static_cast<void> (0) : __assert_fail ("(Decision == LoopVectorizationCostModel::CM_Widen || Decision == LoopVectorizationCostModel::CM_Widen_Reverse || Decision == LoopVectorizationCostModel::CM_GatherScatter) && \"CM decision is not to widen the memory instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2823, __PRETTY_FUNCTION__)); |
2824 | |
2825 | Type *ScalarDataTy = getMemInstValueType(Instr); |
2826 | |
2827 | auto *DataTy = VectorType::get(ScalarDataTy, VF); |
2828 | const Align Alignment = getLoadStoreAlignment(Instr); |
2829 | |
2830 | // Determine if the pointer operand of the access is either consecutive or |
2831 | // reverse consecutive. |
2832 | bool Reverse = (Decision == LoopVectorizationCostModel::CM_Widen_Reverse); |
2833 | bool ConsecutiveStride = |
2834 | Reverse || (Decision == LoopVectorizationCostModel::CM_Widen); |
2835 | bool CreateGatherScatter = |
2836 | (Decision == LoopVectorizationCostModel::CM_GatherScatter); |
2837 | |
2838 | // Either Ptr feeds a vector load/store, or a vector GEP should feed a vector |
2839 | // gather/scatter. Otherwise Decision should have been to Scalarize. |
2840 | assert((ConsecutiveStride || CreateGatherScatter) &&(((ConsecutiveStride || CreateGatherScatter) && "The instruction should be scalarized" ) ? static_cast<void> (0) : __assert_fail ("(ConsecutiveStride || CreateGatherScatter) && \"The instruction should be scalarized\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2841, __PRETTY_FUNCTION__)) |
2841 | "The instruction should be scalarized")(((ConsecutiveStride || CreateGatherScatter) && "The instruction should be scalarized" ) ? static_cast<void> (0) : __assert_fail ("(ConsecutiveStride || CreateGatherScatter) && \"The instruction should be scalarized\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2841, __PRETTY_FUNCTION__)); |
2842 | (void)ConsecutiveStride; |
2843 | |
2844 | VectorParts BlockInMaskParts(UF); |
2845 | bool isMaskRequired = BlockInMask; |
2846 | if (isMaskRequired) |
2847 | for (unsigned Part = 0; Part < UF; ++Part) |
2848 | BlockInMaskParts[Part] = State.get(BlockInMask, Part); |
2849 | |
2850 | const auto CreateVecPtr = [&](unsigned Part, Value *Ptr) -> Value * { |
2851 | // Calculate the pointer for the specific unroll-part. |
2852 | GetElementPtrInst *PartPtr = nullptr; |
2853 | |
2854 | bool InBounds = false; |
2855 | if (auto *gep = dyn_cast<GetElementPtrInst>(Ptr->stripPointerCasts())) |
2856 | InBounds = gep->isInBounds(); |
2857 | |
2858 | if (Reverse) { |
2859 | assert(!VF.isScalable() &&((!VF.isScalable() && "Reversing vectors is not yet supported for scalable vectors." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"Reversing vectors is not yet supported for scalable vectors.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2860, __PRETTY_FUNCTION__)) |
2860 | "Reversing vectors is not yet supported for scalable vectors.")((!VF.isScalable() && "Reversing vectors is not yet supported for scalable vectors." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"Reversing vectors is not yet supported for scalable vectors.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2860, __PRETTY_FUNCTION__)); |
2861 | |
2862 | // If the address is consecutive but reversed, then the |
2863 | // wide store needs to start at the last vector element. |
2864 | PartPtr = cast<GetElementPtrInst>(Builder.CreateGEP( |
2865 | ScalarDataTy, Ptr, Builder.getInt32(-Part * VF.getKnownMinValue()))); |
2866 | PartPtr->setIsInBounds(InBounds); |
2867 | PartPtr = cast<GetElementPtrInst>(Builder.CreateGEP( |
2868 | ScalarDataTy, PartPtr, Builder.getInt32(1 - VF.getKnownMinValue()))); |
2869 | PartPtr->setIsInBounds(InBounds); |
2870 | if (isMaskRequired) // Reverse of a null all-one mask is a null mask. |
2871 | BlockInMaskParts[Part] = reverseVector(BlockInMaskParts[Part]); |
2872 | } else { |
2873 | Value *Increment = createStepForVF(Builder, Builder.getInt32(Part), VF); |
2874 | PartPtr = cast<GetElementPtrInst>( |
2875 | Builder.CreateGEP(ScalarDataTy, Ptr, Increment)); |
2876 | PartPtr->setIsInBounds(InBounds); |
2877 | } |
2878 | |
2879 | unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace(); |
2880 | return Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace)); |
2881 | }; |
2882 | |
2883 | // Handle Stores: |
2884 | if (SI) { |
2885 | setDebugLocFromInst(Builder, SI); |
2886 | |
2887 | for (unsigned Part = 0; Part < UF; ++Part) { |
2888 | Instruction *NewSI = nullptr; |
2889 | Value *StoredVal = State.get(StoredValue, Part); |
2890 | if (CreateGatherScatter) { |
2891 | Value *MaskPart = isMaskRequired ? BlockInMaskParts[Part] : nullptr; |
2892 | Value *VectorGep = State.get(Addr, Part); |
2893 | NewSI = Builder.CreateMaskedScatter(StoredVal, VectorGep, Alignment, |
2894 | MaskPart); |
2895 | } else { |
2896 | if (Reverse) { |
2897 | // If we store to reverse consecutive memory locations, then we need |
2898 | // to reverse the order of elements in the stored value. |
2899 | StoredVal = reverseVector(StoredVal); |
2900 | // We don't want to update the value in the map as it might be used in |
2901 | // another expression. So don't call resetVectorValue(StoredVal). |
2902 | } |
2903 | auto *VecPtr = CreateVecPtr(Part, State.get(Addr, VPIteration(0, 0))); |
2904 | if (isMaskRequired) |
2905 | NewSI = Builder.CreateMaskedStore(StoredVal, VecPtr, Alignment, |
2906 | BlockInMaskParts[Part]); |
2907 | else |
2908 | NewSI = Builder.CreateAlignedStore(StoredVal, VecPtr, Alignment); |
2909 | } |
2910 | addMetadata(NewSI, SI); |
2911 | } |
2912 | return; |
2913 | } |
2914 | |
2915 | // Handle loads. |
2916 | assert(LI && "Must have a load instruction")((LI && "Must have a load instruction") ? static_cast <void> (0) : __assert_fail ("LI && \"Must have a load instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2916, __PRETTY_FUNCTION__)); |
2917 | setDebugLocFromInst(Builder, LI); |
2918 | for (unsigned Part = 0; Part < UF; ++Part) { |
2919 | Value *NewLI; |
2920 | if (CreateGatherScatter) { |
2921 | Value *MaskPart = isMaskRequired ? BlockInMaskParts[Part] : nullptr; |
2922 | Value *VectorGep = State.get(Addr, Part); |
2923 | NewLI = Builder.CreateMaskedGather(VectorGep, Alignment, MaskPart, |
2924 | nullptr, "wide.masked.gather"); |
2925 | addMetadata(NewLI, LI); |
2926 | } else { |
2927 | auto *VecPtr = CreateVecPtr(Part, State.get(Addr, VPIteration(0, 0))); |
2928 | if (isMaskRequired) |
2929 | NewLI = Builder.CreateMaskedLoad( |
2930 | VecPtr, Alignment, BlockInMaskParts[Part], PoisonValue::get(DataTy), |
2931 | "wide.masked.load"); |
2932 | else |
2933 | NewLI = |
2934 | Builder.CreateAlignedLoad(DataTy, VecPtr, Alignment, "wide.load"); |
2935 | |
2936 | // Add metadata to the load, but setVectorValue to the reverse shuffle. |
2937 | addMetadata(NewLI, LI); |
2938 | if (Reverse) |
2939 | NewLI = reverseVector(NewLI); |
2940 | } |
2941 | |
2942 | State.set(Def, NewLI, Part); |
2943 | } |
2944 | } |
2945 | |
2946 | void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, VPValue *Def, |
2947 | VPUser &User, |
2948 | const VPIteration &Instance, |
2949 | bool IfPredicateInstr, |
2950 | VPTransformState &State) { |
2951 | assert(!Instr->getType()->isAggregateType() && "Can't handle vectors")((!Instr->getType()->isAggregateType() && "Can't handle vectors" ) ? static_cast<void> (0) : __assert_fail ("!Instr->getType()->isAggregateType() && \"Can't handle vectors\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2951, __PRETTY_FUNCTION__)); |
2952 | |
2953 | // llvm.experimental.noalias.scope.decl intrinsics must only be duplicated for |
2954 | // the first lane and part. |
2955 | if (isa<NoAliasScopeDeclInst>(Instr)) |
2956 | if (!Instance.isFirstIteration()) |
2957 | return; |
2958 | |
2959 | setDebugLocFromInst(Builder, Instr); |
2960 | |
2961 | // Does this instruction return a value ? |
2962 | bool IsVoidRetTy = Instr->getType()->isVoidTy(); |
2963 | |
2964 | Instruction *Cloned = Instr->clone(); |
2965 | if (!IsVoidRetTy) |
2966 | Cloned->setName(Instr->getName() + ".cloned"); |
2967 | |
2968 | State.Builder.SetInsertPoint(Builder.GetInsertBlock(), |
2969 | Builder.GetInsertPoint()); |
2970 | // Replace the operands of the cloned instructions with their scalar |
2971 | // equivalents in the new loop. |
2972 | for (unsigned op = 0, e = User.getNumOperands(); op != e; ++op) { |
2973 | auto *Operand = dyn_cast<Instruction>(Instr->getOperand(op)); |
2974 | auto InputInstance = Instance; |
2975 | if (!Operand || !OrigLoop->contains(Operand) || |
2976 | (Cost->isUniformAfterVectorization(Operand, State.VF))) |
2977 | InputInstance.Lane = VPLane::getFirstLane(); |
2978 | auto *NewOp = State.get(User.getOperand(op), InputInstance); |
2979 | Cloned->setOperand(op, NewOp); |
2980 | } |
2981 | addNewMetadata(Cloned, Instr); |
2982 | |
2983 | // Place the cloned scalar in the new loop. |
2984 | Builder.Insert(Cloned); |
2985 | |
2986 | State.set(Def, Cloned, Instance); |
2987 | |
2988 | // If we just cloned a new assumption, add it the assumption cache. |
2989 | if (auto *II = dyn_cast<IntrinsicInst>(Cloned)) |
2990 | if (II->getIntrinsicID() == Intrinsic::assume) |
2991 | AC->registerAssumption(II); |
2992 | |
2993 | // End if-block. |
2994 | if (IfPredicateInstr) |
2995 | PredicatedInstructions.push_back(Cloned); |
2996 | } |
2997 | |
2998 | PHINode *InnerLoopVectorizer::createInductionVariable(Loop *L, Value *Start, |
2999 | Value *End, Value *Step, |
3000 | Instruction *DL) { |
3001 | BasicBlock *Header = L->getHeader(); |
3002 | BasicBlock *Latch = L->getLoopLatch(); |
3003 | // As we're just creating this loop, it's possible no latch exists |
3004 | // yet. If so, use the header as this will be a single block loop. |
3005 | if (!Latch) |
3006 | Latch = Header; |
3007 | |
3008 | IRBuilder<> Builder(&*Header->getFirstInsertionPt()); |
3009 | Instruction *OldInst = getDebugLocFromInstOrOperands(OldInduction); |
3010 | setDebugLocFromInst(Builder, OldInst); |
3011 | auto *Induction = Builder.CreatePHI(Start->getType(), 2, "index"); |
3012 | |
3013 | Builder.SetInsertPoint(Latch->getTerminator()); |
3014 | setDebugLocFromInst(Builder, OldInst); |
3015 | |
3016 | // Create i+1 and fill the PHINode. |
3017 | Value *Next = Builder.CreateAdd(Induction, Step, "index.next"); |
3018 | Induction->addIncoming(Start, L->getLoopPreheader()); |
3019 | Induction->addIncoming(Next, Latch); |
3020 | // Create the compare. |
3021 | Value *ICmp = Builder.CreateICmpEQ(Next, End); |
3022 | Builder.CreateCondBr(ICmp, L->getUniqueExitBlock(), Header); |
3023 | |
3024 | // Now we have two terminators. Remove the old one from the block. |
3025 | Latch->getTerminator()->eraseFromParent(); |
3026 | |
3027 | return Induction; |
3028 | } |
3029 | |
3030 | Value *InnerLoopVectorizer::getOrCreateTripCount(Loop *L) { |
3031 | if (TripCount) |
3032 | return TripCount; |
3033 | |
3034 | assert(L && "Create Trip Count for null loop.")((L && "Create Trip Count for null loop.") ? static_cast <void> (0) : __assert_fail ("L && \"Create Trip Count for null loop.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3034, __PRETTY_FUNCTION__)); |
3035 | IRBuilder<> Builder(L->getLoopPreheader()->getTerminator()); |
3036 | // Find the loop boundaries. |
3037 | ScalarEvolution *SE = PSE.getSE(); |
3038 | const SCEV *BackedgeTakenCount = PSE.getBackedgeTakenCount(); |
3039 | assert(!isa<SCEVCouldNotCompute>(BackedgeTakenCount) &&((!isa<SCEVCouldNotCompute>(BackedgeTakenCount) && "Invalid loop count") ? static_cast<void> (0) : __assert_fail ("!isa<SCEVCouldNotCompute>(BackedgeTakenCount) && \"Invalid loop count\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3040, __PRETTY_FUNCTION__)) |
3040 | "Invalid loop count")((!isa<SCEVCouldNotCompute>(BackedgeTakenCount) && "Invalid loop count") ? static_cast<void> (0) : __assert_fail ("!isa<SCEVCouldNotCompute>(BackedgeTakenCount) && \"Invalid loop count\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3040, __PRETTY_FUNCTION__)); |
3041 | |
3042 | Type *IdxTy = Legal->getWidestInductionType(); |
3043 | assert(IdxTy && "No type for induction")((IdxTy && "No type for induction") ? static_cast< void> (0) : __assert_fail ("IdxTy && \"No type for induction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3043, __PRETTY_FUNCTION__)); |
3044 | |
3045 | // The exit count might have the type of i64 while the phi is i32. This can |
3046 | // happen if we have an induction variable that is sign extended before the |
3047 | // compare. The only way that we get a backedge taken count is that the |
3048 | // induction variable was signed and as such will not overflow. In such a case |
3049 | // truncation is legal. |
3050 | if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) > |
3051 | IdxTy->getPrimitiveSizeInBits()) |
3052 | BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, IdxTy); |
3053 | BackedgeTakenCount = SE->getNoopOrZeroExtend(BackedgeTakenCount, IdxTy); |
3054 | |
3055 | // Get the total trip count from the count by adding 1. |
3056 | const SCEV *ExitCount = SE->getAddExpr( |
3057 | BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType())); |
3058 | |
3059 | const DataLayout &DL = L->getHeader()->getModule()->getDataLayout(); |
3060 | |
3061 | // Expand the trip count and place the new instructions in the preheader. |
3062 | // Notice that the pre-header does not change, only the loop body. |
3063 | SCEVExpander Exp(*SE, DL, "induction"); |
3064 | |
3065 | // Count holds the overall loop count (N). |
3066 | TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(), |
3067 | L->getLoopPreheader()->getTerminator()); |
3068 | |
3069 | if (TripCount->getType()->isPointerTy()) |
3070 | TripCount = |
3071 | CastInst::CreatePointerCast(TripCount, IdxTy, "exitcount.ptrcnt.to.int", |
3072 | L->getLoopPreheader()->getTerminator()); |
3073 | |
3074 | return TripCount; |
3075 | } |
3076 | |
3077 | Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) { |
3078 | if (VectorTripCount) |
3079 | return VectorTripCount; |
3080 | |
3081 | Value *TC = getOrCreateTripCount(L); |
3082 | IRBuilder<> Builder(L->getLoopPreheader()->getTerminator()); |
3083 | |
3084 | Type *Ty = TC->getType(); |
3085 | // This is where we can make the step a runtime constant. |
3086 | Value *Step = createStepForVF(Builder, ConstantInt::get(Ty, UF), VF); |
3087 | |
3088 | // If the tail is to be folded by masking, round the number of iterations N |
3089 | // up to a multiple of Step instead of rounding down. This is done by first |
3090 | // adding Step-1 and then rounding down. Note that it's ok if this addition |
3091 | // overflows: the vector induction variable will eventually wrap to zero given |
3092 | // that it starts at zero and its Step is a power of two; the loop will then |
3093 | // exit, with the last early-exit vector comparison also producing all-true. |
3094 | if (Cost->foldTailByMasking()) { |
3095 | assert(isPowerOf2_32(VF.getKnownMinValue() * UF) &&((isPowerOf2_32(VF.getKnownMinValue() * UF) && "VF*UF must be a power of 2 when folding tail by masking" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(VF.getKnownMinValue() * UF) && \"VF*UF must be a power of 2 when folding tail by masking\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3096, __PRETTY_FUNCTION__)) |
3096 | "VF*UF must be a power of 2 when folding tail by masking")((isPowerOf2_32(VF.getKnownMinValue() * UF) && "VF*UF must be a power of 2 when folding tail by masking" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(VF.getKnownMinValue() * UF) && \"VF*UF must be a power of 2 when folding tail by masking\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3096, __PRETTY_FUNCTION__)); |
3097 | assert(!VF.isScalable() &&((!VF.isScalable() && "Tail folding not yet supported for scalable vectors" ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"Tail folding not yet supported for scalable vectors\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3098, __PRETTY_FUNCTION__)) |
3098 | "Tail folding not yet supported for scalable vectors")((!VF.isScalable() && "Tail folding not yet supported for scalable vectors" ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"Tail folding not yet supported for scalable vectors\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3098, __PRETTY_FUNCTION__)); |
3099 | TC = Builder.CreateAdd( |
3100 | TC, ConstantInt::get(Ty, VF.getKnownMinValue() * UF - 1), "n.rnd.up"); |
3101 | } |
3102 | |
3103 | // Now we need to generate the expression for the part of the loop that the |
3104 | // vectorized body will execute. This is equal to N - (N % Step) if scalar |
3105 | // iterations are not required for correctness, or N - Step, otherwise. Step |
3106 | // is equal to the vectorization factor (number of SIMD elements) times the |
3107 | // unroll factor (number of SIMD instructions). |
3108 | Value *R = Builder.CreateURem(TC, Step, "n.mod.vf"); |
3109 | |
3110 | // There are two cases where we need to ensure (at least) the last iteration |
3111 | // runs in the scalar remainder loop. Thus, if the step evenly divides |
3112 | // the trip count, we set the remainder to be equal to the step. If the step |
3113 | // does not evenly divide the trip count, no adjustment is necessary since |
3114 | // there will already be scalar iterations. Note that the minimum iterations |
3115 | // check ensures that N >= Step. The cases are: |
3116 | // 1) If there is a non-reversed interleaved group that may speculatively |
3117 | // access memory out-of-bounds. |
3118 | // 2) If any instruction may follow a conditionally taken exit. That is, if |
3119 | // the loop contains multiple exiting blocks, or a single exiting block |
3120 | // which is not the latch. |
3121 | if (VF.isVector() && Cost->requiresScalarEpilogue()) { |
3122 | auto *IsZero = Builder.CreateICmpEQ(R, ConstantInt::get(R->getType(), 0)); |
3123 | R = Builder.CreateSelect(IsZero, Step, R); |
3124 | } |
3125 | |
3126 | VectorTripCount = Builder.CreateSub(TC, R, "n.vec"); |
3127 | |
3128 | return VectorTripCount; |
3129 | } |
3130 | |
3131 | Value *InnerLoopVectorizer::createBitOrPointerCast(Value *V, VectorType *DstVTy, |
3132 | const DataLayout &DL) { |
3133 | // Verify that V is a vector type with same number of elements as DstVTy. |
3134 | auto *DstFVTy = cast<FixedVectorType>(DstVTy); |
3135 | unsigned VF = DstFVTy->getNumElements(); |
3136 | auto *SrcVecTy = cast<FixedVectorType>(V->getType()); |
3137 | assert((VF == SrcVecTy->getNumElements()) && "Vector dimensions do not match")(((VF == SrcVecTy->getNumElements()) && "Vector dimensions do not match" ) ? static_cast<void> (0) : __assert_fail ("(VF == SrcVecTy->getNumElements()) && \"Vector dimensions do not match\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3137, __PRETTY_FUNCTION__)); |
3138 | Type *SrcElemTy = SrcVecTy->getElementType(); |
3139 | Type *DstElemTy = DstFVTy->getElementType(); |
3140 | assert((DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy)) &&(((DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy )) && "Vector elements must have same size") ? static_cast <void> (0) : __assert_fail ("(DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy)) && \"Vector elements must have same size\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3141, __PRETTY_FUNCTION__)) |
3141 | "Vector elements must have same size")(((DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy )) && "Vector elements must have same size") ? static_cast <void> (0) : __assert_fail ("(DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy)) && \"Vector elements must have same size\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3141, __PRETTY_FUNCTION__)); |
3142 | |
3143 | // Do a direct cast if element types are castable. |
3144 | if (CastInst::isBitOrNoopPointerCastable(SrcElemTy, DstElemTy, DL)) { |
3145 | return Builder.CreateBitOrPointerCast(V, DstFVTy); |
3146 | } |
3147 | // V cannot be directly casted to desired vector type. |
3148 | // May happen when V is a floating point vector but DstVTy is a vector of |
3149 | // pointers or vice-versa. Handle this using a two-step bitcast using an |
3150 | // intermediate Integer type for the bitcast i.e. Ptr <-> Int <-> Float. |
3151 | assert((DstElemTy->isPointerTy() != SrcElemTy->isPointerTy()) &&(((DstElemTy->isPointerTy() != SrcElemTy->isPointerTy() ) && "Only one type should be a pointer type") ? static_cast <void> (0) : __assert_fail ("(DstElemTy->isPointerTy() != SrcElemTy->isPointerTy()) && \"Only one type should be a pointer type\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3152, __PRETTY_FUNCTION__)) |
3152 | "Only one type should be a pointer type")(((DstElemTy->isPointerTy() != SrcElemTy->isPointerTy() ) && "Only one type should be a pointer type") ? static_cast <void> (0) : __assert_fail ("(DstElemTy->isPointerTy() != SrcElemTy->isPointerTy()) && \"Only one type should be a pointer type\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3152, __PRETTY_FUNCTION__)); |
3153 | assert((DstElemTy->isFloatingPointTy() != SrcElemTy->isFloatingPointTy()) &&(((DstElemTy->isFloatingPointTy() != SrcElemTy->isFloatingPointTy ()) && "Only one type should be a floating point type" ) ? static_cast<void> (0) : __assert_fail ("(DstElemTy->isFloatingPointTy() != SrcElemTy->isFloatingPointTy()) && \"Only one type should be a floating point type\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3154, __PRETTY_FUNCTION__)) |
3154 | "Only one type should be a floating point type")(((DstElemTy->isFloatingPointTy() != SrcElemTy->isFloatingPointTy ()) && "Only one type should be a floating point type" ) ? static_cast<void> (0) : __assert_fail ("(DstElemTy->isFloatingPointTy() != SrcElemTy->isFloatingPointTy()) && \"Only one type should be a floating point type\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3154, __PRETTY_FUNCTION__)); |
3155 | Type *IntTy = |
3156 | IntegerType::getIntNTy(V->getContext(), DL.getTypeSizeInBits(SrcElemTy)); |
3157 | auto *VecIntTy = FixedVectorType::get(IntTy, VF); |
3158 | Value *CastVal = Builder.CreateBitOrPointerCast(V, VecIntTy); |
3159 | return Builder.CreateBitOrPointerCast(CastVal, DstFVTy); |
3160 | } |
3161 | |
3162 | void InnerLoopVectorizer::emitMinimumIterationCountCheck(Loop *L, |
3163 | BasicBlock *Bypass) { |
3164 | Value *Count = getOrCreateTripCount(L); |
3165 | // Reuse existing vector loop preheader for TC checks. |
3166 | // Note that new preheader block is generated for vector loop. |
3167 | BasicBlock *const TCCheckBlock = LoopVectorPreHeader; |
3168 | IRBuilder<> Builder(TCCheckBlock->getTerminator()); |
3169 | |
3170 | // Generate code to check if the loop's trip count is less than VF * UF, or |
3171 | // equal to it in case a scalar epilogue is required; this implies that the |
3172 | // vector trip count is zero. This check also covers the case where adding one |
3173 | // to the backedge-taken count overflowed leading to an incorrect trip count |
3174 | // of zero. In this case we will also jump to the scalar loop. |
3175 | auto P = Cost->requiresScalarEpilogue() ? ICmpInst::ICMP_ULE |
3176 | : ICmpInst::ICMP_ULT; |
3177 | |
3178 | // If tail is to be folded, vector loop takes care of all iterations. |
3179 | Value *CheckMinIters = Builder.getFalse(); |
3180 | if (!Cost->foldTailByMasking()) { |
3181 | Value *Step = |
3182 | createStepForVF(Builder, ConstantInt::get(Count->getType(), UF), VF); |
3183 | CheckMinIters = Builder.CreateICmp(P, Count, Step, "min.iters.check"); |
3184 | } |
3185 | // Create new preheader for vector loop. |
3186 | LoopVectorPreHeader = |
3187 | SplitBlock(TCCheckBlock, TCCheckBlock->getTerminator(), DT, LI, nullptr, |
3188 | "vector.ph"); |
3189 | |
3190 | assert(DT->properlyDominates(DT->getNode(TCCheckBlock),((DT->properlyDominates(DT->getNode(TCCheckBlock), DT-> getNode(Bypass)->getIDom()) && "TC check is expected to dominate Bypass" ) ? static_cast<void> (0) : __assert_fail ("DT->properlyDominates(DT->getNode(TCCheckBlock), DT->getNode(Bypass)->getIDom()) && \"TC check is expected to dominate Bypass\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3192, __PRETTY_FUNCTION__)) |
3191 | DT->getNode(Bypass)->getIDom()) &&((DT->properlyDominates(DT->getNode(TCCheckBlock), DT-> getNode(Bypass)->getIDom()) && "TC check is expected to dominate Bypass" ) ? static_cast<void> (0) : __assert_fail ("DT->properlyDominates(DT->getNode(TCCheckBlock), DT->getNode(Bypass)->getIDom()) && \"TC check is expected to dominate Bypass\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3192, __PRETTY_FUNCTION__)) |
3192 | "TC check is expected to dominate Bypass")((DT->properlyDominates(DT->getNode(TCCheckBlock), DT-> getNode(Bypass)->getIDom()) && "TC check is expected to dominate Bypass" ) ? static_cast<void> (0) : __assert_fail ("DT->properlyDominates(DT->getNode(TCCheckBlock), DT->getNode(Bypass)->getIDom()) && \"TC check is expected to dominate Bypass\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3192, __PRETTY_FUNCTION__)); |
3193 | |
3194 | // Update dominator for Bypass & LoopExit. |
3195 | DT->changeImmediateDominator(Bypass, TCCheckBlock); |
3196 | DT->changeImmediateDominator(LoopExitBlock, TCCheckBlock); |
3197 | |
3198 | ReplaceInstWithInst( |
3199 | TCCheckBlock->getTerminator(), |
3200 | BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters)); |
3201 | LoopBypassBlocks.push_back(TCCheckBlock); |
3202 | } |
3203 | |
3204 | BasicBlock *InnerLoopVectorizer::emitSCEVChecks(Loop *L, BasicBlock *Bypass) { |
3205 | |
3206 | BasicBlock *const SCEVCheckBlock = |
3207 | RTChecks.emitSCEVChecks(L, Bypass, LoopVectorPreHeader, LoopExitBlock); |
3208 | if (!SCEVCheckBlock) |
3209 | return nullptr; |
3210 | |
3211 | assert(!(SCEVCheckBlock->getParent()->hasOptSize() ||((!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints ::FK_Enabled)) && "Cannot SCEV check stride or overflow when optimizing for size" ) ? static_cast<void> (0) : __assert_fail ("!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && \"Cannot SCEV check stride or overflow when optimizing for size\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3214, __PRETTY_FUNCTION__)) |
3212 | (OptForSizeBasedOnProfile &&((!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints ::FK_Enabled)) && "Cannot SCEV check stride or overflow when optimizing for size" ) ? static_cast<void> (0) : __assert_fail ("!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && \"Cannot SCEV check stride or overflow when optimizing for size\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3214, __PRETTY_FUNCTION__)) |
3213 | Cost->Hints->getForce() != LoopVectorizeHints::FK_Enabled)) &&((!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints ::FK_Enabled)) && "Cannot SCEV check stride or overflow when optimizing for size" ) ? static_cast<void> (0) : __assert_fail ("!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && \"Cannot SCEV check stride or overflow when optimizing for size\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3214, __PRETTY_FUNCTION__)) |
3214 | "Cannot SCEV check stride or overflow when optimizing for size")((!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints ::FK_Enabled)) && "Cannot SCEV check stride or overflow when optimizing for size" ) ? static_cast<void> (0) : __assert_fail ("!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && \"Cannot SCEV check stride or overflow when optimizing for size\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3214, __PRETTY_FUNCTION__)); |
3215 | |
3216 | |
3217 | // Update dominator only if this is first RT check. |
3218 | if (LoopBypassBlocks.empty()) { |
3219 | DT->changeImmediateDominator(Bypass, SCEVCheckBlock); |
3220 | DT->changeImmediateDominator(LoopExitBlock, SCEVCheckBlock); |
3221 | } |
3222 | |
3223 | LoopBypassBlocks.push_back(SCEVCheckBlock); |
3224 | AddedSafetyChecks = true; |
3225 | return SCEVCheckBlock; |
3226 | } |
3227 | |
3228 | BasicBlock *InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L, |
3229 | BasicBlock *Bypass) { |
3230 | // VPlan-native path does not do any analysis for runtime checks currently. |
3231 | if (EnableVPlanNativePath) |
3232 | return nullptr; |
3233 | |
3234 | BasicBlock *const MemCheckBlock = |
3235 | RTChecks.emitMemRuntimeChecks(L, Bypass, LoopVectorPreHeader); |
3236 | |
3237 | // Check if we generated code that checks in runtime if arrays overlap. We put |
3238 | // the checks into a separate block to make the more common case of few |
3239 | // elements faster. |
3240 | if (!MemCheckBlock) |
3241 | return nullptr; |
3242 | |
3243 | if (MemCheckBlock->getParent()->hasOptSize() || OptForSizeBasedOnProfile) { |
3244 | assert(Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled &&((Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled && "Cannot emit memory checks when optimizing for size, unless forced " "to vectorize.") ? static_cast<void> (0) : __assert_fail ("Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled && \"Cannot emit memory checks when optimizing for size, unless forced \" \"to vectorize.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3246, __PRETTY_FUNCTION__)) |
3245 | "Cannot emit memory checks when optimizing for size, unless forced "((Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled && "Cannot emit memory checks when optimizing for size, unless forced " "to vectorize.") ? static_cast<void> (0) : __assert_fail ("Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled && \"Cannot emit memory checks when optimizing for size, unless forced \" \"to vectorize.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3246, __PRETTY_FUNCTION__)) |
3246 | "to vectorize.")((Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled && "Cannot emit memory checks when optimizing for size, unless forced " "to vectorize.") ? static_cast<void> (0) : __assert_fail ("Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled && \"Cannot emit memory checks when optimizing for size, unless forced \" \"to vectorize.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3246, __PRETTY_FUNCTION__)); |
3247 | ORE->emit([&]() { |
3248 | return OptimizationRemarkAnalysis(DEBUG_TYPE"loop-vectorize", "VectorizationCodeSize", |
3249 | L->getStartLoc(), L->getHeader()) |
3250 | << "Code-size may be reduced by not forcing " |
3251 | "vectorization, or by source-code modifications " |
3252 | "eliminating the need for runtime checks " |
3253 | "(e.g., adding 'restrict')."; |
3254 | }); |
3255 | } |
3256 | |
3257 | LoopBypassBlocks.push_back(MemCheckBlock); |
3258 | |
3259 | AddedSafetyChecks = true; |
3260 | |
3261 | // We currently don't use LoopVersioning for the actual loop cloning but we |
3262 | // still use it to add the noalias metadata. |
3263 | LVer = std::make_unique<LoopVersioning>( |
3264 | *Legal->getLAI(), |
3265 | Legal->getLAI()->getRuntimePointerChecking()->getChecks(), OrigLoop, LI, |
3266 | DT, PSE.getSE()); |
3267 | LVer->prepareNoAliasMetadata(); |
3268 | return MemCheckBlock; |
3269 | } |
3270 | |
3271 | Value *InnerLoopVectorizer::emitTransformedIndex( |
3272 | IRBuilder<> &B, Value *Index, ScalarEvolution *SE, const DataLayout &DL, |
3273 | const InductionDescriptor &ID) const { |
3274 | |
3275 | SCEVExpander Exp(*SE, DL, "induction"); |
3276 | auto Step = ID.getStep(); |
3277 | auto StartValue = ID.getStartValue(); |
3278 | assert(Index->getType() == Step->getType() &&((Index->getType() == Step->getType() && "Index type does not match StepValue type" ) ? static_cast<void> (0) : __assert_fail ("Index->getType() == Step->getType() && \"Index type does not match StepValue type\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3279, __PRETTY_FUNCTION__)) |
3279 | "Index type does not match StepValue type")((Index->getType() == Step->getType() && "Index type does not match StepValue type" ) ? static_cast<void> (0) : __assert_fail ("Index->getType() == Step->getType() && \"Index type does not match StepValue type\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3279, __PRETTY_FUNCTION__)); |
3280 | |
3281 | // Note: the IR at this point is broken. We cannot use SE to create any new |
3282 | // SCEV and then expand it, hoping that SCEV's simplification will give us |
3283 | // a more optimal code. Unfortunately, attempt of doing so on invalid IR may |
3284 | // lead to various SCEV crashes. So all we can do is to use builder and rely |
3285 | // on InstCombine for future simplifications. Here we handle some trivial |
3286 | // cases only. |
3287 | auto CreateAdd = [&B](Value *X, Value *Y) { |
3288 | assert(X->getType() == Y->getType() && "Types don't match!")((X->getType() == Y->getType() && "Types don't match!" ) ? static_cast<void> (0) : __assert_fail ("X->getType() == Y->getType() && \"Types don't match!\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3288, __PRETTY_FUNCTION__)); |
3289 | if (auto *CX = dyn_cast<ConstantInt>(X)) |
3290 | if (CX->isZero()) |
3291 | return Y; |
3292 | if (auto *CY = dyn_cast<ConstantInt>(Y)) |
3293 | if (CY->isZero()) |
3294 | return X; |
3295 | return B.CreateAdd(X, Y); |
3296 | }; |
3297 | |
3298 | auto CreateMul = [&B](Value *X, Value *Y) { |
3299 | assert(X->getType() == Y->getType() && "Types don't match!")((X->getType() == Y->getType() && "Types don't match!" ) ? static_cast<void> (0) : __assert_fail ("X->getType() == Y->getType() && \"Types don't match!\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3299, __PRETTY_FUNCTION__)); |
3300 | if (auto *CX = dyn_cast<ConstantInt>(X)) |
3301 | if (CX->isOne()) |
3302 | return Y; |
3303 | if (auto *CY = dyn_cast<ConstantInt>(Y)) |
3304 | if (CY->isOne()) |
3305 | return X; |
3306 | return B.CreateMul(X, Y); |
3307 | }; |
3308 | |
3309 | // Get a suitable insert point for SCEV expansion. For blocks in the vector |
3310 | // loop, choose the end of the vector loop header (=LoopVectorBody), because |
3311 | // the DomTree is not kept up-to-date for additional blocks generated in the |
3312 | // vector loop. By using the header as insertion point, we guarantee that the |
3313 | // expanded instructions dominate all their uses. |
3314 | auto GetInsertPoint = [this, &B]() { |
3315 | BasicBlock *InsertBB = B.GetInsertPoint()->getParent(); |
3316 | if (InsertBB != LoopVectorBody && |
3317 | LI->getLoopFor(LoopVectorBody) == LI->getLoopFor(InsertBB)) |
3318 | return LoopVectorBody->getTerminator(); |
3319 | return &*B.GetInsertPoint(); |
3320 | }; |
3321 | |
3322 | switch (ID.getKind()) { |
3323 | case InductionDescriptor::IK_IntInduction: { |
3324 | assert(Index->getType() == StartValue->getType() &&((Index->getType() == StartValue->getType() && "Index type does not match StartValue type" ) ? static_cast<void> (0) : __assert_fail ("Index->getType() == StartValue->getType() && \"Index type does not match StartValue type\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3325, __PRETTY_FUNCTION__)) |
3325 | "Index type does not match StartValue type")((Index->getType() == StartValue->getType() && "Index type does not match StartValue type" ) ? static_cast<void> (0) : __assert_fail ("Index->getType() == StartValue->getType() && \"Index type does not match StartValue type\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3325, __PRETTY_FUNCTION__)); |
3326 | if (ID.getConstIntStepValue() && ID.getConstIntStepValue()->isMinusOne()) |
3327 | return B.CreateSub(StartValue, Index); |
3328 | auto *Offset = CreateMul( |
3329 | Index, Exp.expandCodeFor(Step, Index->getType(), GetInsertPoint())); |
3330 | return CreateAdd(StartValue, Offset); |
3331 | } |
3332 | case InductionDescriptor::IK_PtrInduction: { |
3333 | assert(isa<SCEVConstant>(Step) &&((isa<SCEVConstant>(Step) && "Expected constant step for pointer induction" ) ? static_cast<void> (0) : __assert_fail ("isa<SCEVConstant>(Step) && \"Expected constant step for pointer induction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3334, __PRETTY_FUNCTION__)) |
3334 | "Expected constant step for pointer induction")((isa<SCEVConstant>(Step) && "Expected constant step for pointer induction" ) ? static_cast<void> (0) : __assert_fail ("isa<SCEVConstant>(Step) && \"Expected constant step for pointer induction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3334, __PRETTY_FUNCTION__)); |
3335 | return B.CreateGEP( |
3336 | StartValue->getType()->getPointerElementType(), StartValue, |
3337 | CreateMul(Index, |
3338 | Exp.expandCodeFor(Step, Index->getType(), GetInsertPoint()))); |
3339 | } |
3340 | case InductionDescriptor::IK_FpInduction: { |
3341 | assert(Step->getType()->isFloatingPointTy() && "Expected FP Step value")((Step->getType()->isFloatingPointTy() && "Expected FP Step value" ) ? static_cast<void> (0) : __assert_fail ("Step->getType()->isFloatingPointTy() && \"Expected FP Step value\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3341, __PRETTY_FUNCTION__)); |
3342 | auto InductionBinOp = ID.getInductionBinOp(); |
3343 | assert(InductionBinOp &&((InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction ::FSub) && "Original bin op should be defined for FP induction" ) ? static_cast<void> (0) : __assert_fail ("InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction::FSub) && \"Original bin op should be defined for FP induction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3346, __PRETTY_FUNCTION__)) |
3344 | (InductionBinOp->getOpcode() == Instruction::FAdd ||((InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction ::FSub) && "Original bin op should be defined for FP induction" ) ? static_cast<void> (0) : __assert_fail ("InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction::FSub) && \"Original bin op should be defined for FP induction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3346, __PRETTY_FUNCTION__)) |
3345 | InductionBinOp->getOpcode() == Instruction::FSub) &&((InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction ::FSub) && "Original bin op should be defined for FP induction" ) ? static_cast<void> (0) : __assert_fail ("InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction::FSub) && \"Original bin op should be defined for FP induction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3346, __PRETTY_FUNCTION__)) |
3346 | "Original bin op should be defined for FP induction")((InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction ::FSub) && "Original bin op should be defined for FP induction" ) ? static_cast<void> (0) : __assert_fail ("InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction::FSub) && \"Original bin op should be defined for FP induction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3346, __PRETTY_FUNCTION__)); |
3347 | |
3348 | Value *StepValue = cast<SCEVUnknown>(Step)->getValue(); |
3349 | Value *MulExp = B.CreateFMul(StepValue, Index); |
3350 | return B.CreateBinOp(InductionBinOp->getOpcode(), StartValue, MulExp, |
3351 | "induction"); |
3352 | } |
3353 | case InductionDescriptor::IK_NoInduction: |
3354 | return nullptr; |
3355 | } |
3356 | llvm_unreachable("invalid enum")::llvm::llvm_unreachable_internal("invalid enum", "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3356); |
3357 | } |
3358 | |
3359 | Loop *InnerLoopVectorizer::createVectorLoopSkeleton(StringRef Prefix) { |
3360 | LoopScalarBody = OrigLoop->getHeader(); |
3361 | LoopVectorPreHeader = OrigLoop->getLoopPreheader(); |
3362 | LoopExitBlock = OrigLoop->getUniqueExitBlock(); |
3363 | assert(LoopExitBlock && "Must have an exit block")((LoopExitBlock && "Must have an exit block") ? static_cast <void> (0) : __assert_fail ("LoopExitBlock && \"Must have an exit block\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3363, __PRETTY_FUNCTION__)); |
3364 | assert(LoopVectorPreHeader && "Invalid loop structure")((LoopVectorPreHeader && "Invalid loop structure") ? static_cast <void> (0) : __assert_fail ("LoopVectorPreHeader && \"Invalid loop structure\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3364, __PRETTY_FUNCTION__)); |
3365 | |
3366 | LoopMiddleBlock = |
3367 | SplitBlock(LoopVectorPreHeader, LoopVectorPreHeader->getTerminator(), DT, |
3368 | LI, nullptr, Twine(Prefix) + "middle.block"); |
3369 | LoopScalarPreHeader = |
3370 | SplitBlock(LoopMiddleBlock, LoopMiddleBlock->getTerminator(), DT, LI, |
3371 | nullptr, Twine(Prefix) + "scalar.ph"); |
3372 | |
3373 | // Set up branch from middle block to the exit and scalar preheader blocks. |
3374 | // completeLoopSkeleton will update the condition to use an iteration check, |
3375 | // if required to decide whether to execute the remainder. |
3376 | BranchInst *BrInst = |
3377 | BranchInst::Create(LoopExitBlock, LoopScalarPreHeader, Builder.getTrue()); |
3378 | auto *ScalarLatchTerm = OrigLoop->getLoopLatch()->getTerminator(); |
3379 | BrInst->setDebugLoc(ScalarLatchTerm->getDebugLoc()); |
3380 | ReplaceInstWithInst(LoopMiddleBlock->getTerminator(), BrInst); |
3381 | |
3382 | // We intentionally don't let SplitBlock to update LoopInfo since |
3383 | // LoopVectorBody should belong to another loop than LoopVectorPreHeader. |
3384 | // LoopVectorBody is explicitly added to the correct place few lines later. |
3385 | LoopVectorBody = |
3386 | SplitBlock(LoopVectorPreHeader, LoopVectorPreHeader->getTerminator(), DT, |
3387 | nullptr, nullptr, Twine(Prefix) + "vector.body"); |
3388 | |
3389 | // Update dominator for loop exit. |
3390 | DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock); |
3391 | |
3392 | // Create and register the new vector loop. |
3393 | Loop *Lp = LI->AllocateLoop(); |
3394 | Loop *ParentLoop = OrigLoop->getParentLoop(); |
3395 | |
3396 | // Insert the new loop into the loop nest and register the new basic blocks |
3397 | // before calling any utilities such as SCEV that require valid LoopInfo. |
3398 | if (ParentLoop) { |
3399 | ParentLoop->addChildLoop(Lp); |
3400 | } else { |
3401 | LI->addTopLevelLoop(Lp); |
3402 | } |
3403 | Lp->addBasicBlockToLoop(LoopVectorBody, *LI); |
3404 | return Lp; |
3405 | } |
3406 | |
3407 | void InnerLoopVectorizer::createInductionResumeValues( |
3408 | Loop *L, Value *VectorTripCount, |
3409 | std::pair<BasicBlock *, Value *> AdditionalBypass) { |
3410 | assert(VectorTripCount && L && "Expected valid arguments")((VectorTripCount && L && "Expected valid arguments" ) ? static_cast<void> (0) : __assert_fail ("VectorTripCount && L && \"Expected valid arguments\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3410, __PRETTY_FUNCTION__)); |
3411 | assert(((AdditionalBypass.first && AdditionalBypass.second) ||((((AdditionalBypass.first && AdditionalBypass.second ) || (!AdditionalBypass.first && !AdditionalBypass.second )) && "Inconsistent information about additional bypass." ) ? static_cast<void> (0) : __assert_fail ("((AdditionalBypass.first && AdditionalBypass.second) || (!AdditionalBypass.first && !AdditionalBypass.second)) && \"Inconsistent information about additional bypass.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3413, __PRETTY_FUNCTION__)) |
3412 | (!AdditionalBypass.first && !AdditionalBypass.second)) &&((((AdditionalBypass.first && AdditionalBypass.second ) || (!AdditionalBypass.first && !AdditionalBypass.second )) && "Inconsistent information about additional bypass." ) ? static_cast<void> (0) : __assert_fail ("((AdditionalBypass.first && AdditionalBypass.second) || (!AdditionalBypass.first && !AdditionalBypass.second)) && \"Inconsistent information about additional bypass.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3413, __PRETTY_FUNCTION__)) |
3413 | "Inconsistent information about additional bypass.")((((AdditionalBypass.first && AdditionalBypass.second ) || (!AdditionalBypass.first && !AdditionalBypass.second )) && "Inconsistent information about additional bypass." ) ? static_cast<void> (0) : __assert_fail ("((AdditionalBypass.first && AdditionalBypass.second) || (!AdditionalBypass.first && !AdditionalBypass.second)) && \"Inconsistent information about additional bypass.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3413, __PRETTY_FUNCTION__)); |
3414 | // We are going to resume the execution of the scalar loop. |
3415 | // Go over all of the induction variables that we found and fix the |
3416 | // PHIs that are left in the scalar version of the loop. |
3417 | // The starting values of PHI nodes depend on the counter of the last |
3418 | // iteration in the vectorized loop. |
3419 | // If we come from a bypass edge then we need to start from the original |
3420 | // start value. |
3421 | for (auto &InductionEntry : Legal->getInductionVars()) { |
3422 | PHINode *OrigPhi = InductionEntry.first; |
3423 | InductionDescriptor II = InductionEntry.second; |
3424 | |
3425 | // Create phi nodes to merge from the backedge-taken check block. |
3426 | PHINode *BCResumeVal = |
3427 | PHINode::Create(OrigPhi->getType(), 3, "bc.resume.val", |
3428 | LoopScalarPreHeader->getTerminator()); |
3429 | // Copy original phi DL over to the new one. |
3430 | BCResumeVal->setDebugLoc(OrigPhi->getDebugLoc()); |
3431 | Value *&EndValue = IVEndValues[OrigPhi]; |
3432 | Value *EndValueFromAdditionalBypass = AdditionalBypass.second; |
3433 | if (OrigPhi == OldInduction) { |
3434 | // We know what the end value is. |
3435 | EndValue = VectorTripCount; |
3436 | } else { |
3437 | IRBuilder<> B(L->getLoopPreheader()->getTerminator()); |
3438 | |
3439 | // Fast-math-flags propagate from the original induction instruction. |
3440 | if (II.getInductionBinOp() && isa<FPMathOperator>(II.getInductionBinOp())) |
3441 | B.setFastMathFlags(II.getInductionBinOp()->getFastMathFlags()); |
3442 | |
3443 | Type *StepType = II.getStep()->getType(); |
3444 | Instruction::CastOps CastOp = |
3445 | CastInst::getCastOpcode(VectorTripCount, true, StepType, true); |
3446 | Value *CRD = B.CreateCast(CastOp, VectorTripCount, StepType, "cast.crd"); |
3447 | const DataLayout &DL = LoopScalarBody->getModule()->getDataLayout(); |
3448 | EndValue = emitTransformedIndex(B, CRD, PSE.getSE(), DL, II); |
3449 | EndValue->setName("ind.end"); |
3450 | |
3451 | // Compute the end value for the additional bypass (if applicable). |
3452 | if (AdditionalBypass.first) { |
3453 | B.SetInsertPoint(&(*AdditionalBypass.first->getFirstInsertionPt())); |
3454 | CastOp = CastInst::getCastOpcode(AdditionalBypass.second, true, |
3455 | StepType, true); |
3456 | CRD = |
3457 | B.CreateCast(CastOp, AdditionalBypass.second, StepType, "cast.crd"); |
3458 | EndValueFromAdditionalBypass = |
3459 | emitTransformedIndex(B, CRD, PSE.getSE(), DL, II); |
3460 | EndValueFromAdditionalBypass->setName("ind.end"); |
3461 | } |
3462 | } |
3463 | // The new PHI merges the original incoming value, in case of a bypass, |
3464 | // or the value at the end of the vectorized loop. |
3465 | BCResumeVal->addIncoming(EndValue, LoopMiddleBlock); |
3466 | |
3467 | // Fix the scalar body counter (PHI node). |
3468 | // The old induction's phi node in the scalar body needs the truncated |
3469 | // value. |
3470 | for (BasicBlock *BB : LoopBypassBlocks) |
3471 | BCResumeVal->addIncoming(II.getStartValue(), BB); |
3472 | |
3473 | if (AdditionalBypass.first) |
3474 | BCResumeVal->setIncomingValueForBlock(AdditionalBypass.first, |
3475 | EndValueFromAdditionalBypass); |
3476 | |
3477 | OrigPhi->setIncomingValueForBlock(LoopScalarPreHeader, BCResumeVal); |
3478 | } |
3479 | } |
3480 | |
3481 | BasicBlock *InnerLoopVectorizer::completeLoopSkeleton(Loop *L, |
3482 | MDNode *OrigLoopID) { |
3483 | assert(L && "Expected valid loop.")((L && "Expected valid loop.") ? static_cast<void> (0) : __assert_fail ("L && \"Expected valid loop.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3483, __PRETTY_FUNCTION__)); |
3484 | |
3485 | // The trip counts should be cached by now. |
3486 | Value *Count = getOrCreateTripCount(L); |
3487 | Value *VectorTripCount = getOrCreateVectorTripCount(L); |
3488 | |
3489 | auto *ScalarLatchTerm = OrigLoop->getLoopLatch()->getTerminator(); |
3490 | |
3491 | // Add a check in the middle block to see if we have completed |
3492 | // all of the iterations in the first vector loop. |
3493 | // If (N - N%VF) == N, then we *don't* need to run the remainder. |
3494 | // If tail is to be folded, we know we don't need to run the remainder. |
3495 | if (!Cost->foldTailByMasking()) { |
3496 | Instruction *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, |
3497 | Count, VectorTripCount, "cmp.n", |
3498 | LoopMiddleBlock->getTerminator()); |
3499 | |
3500 | // Here we use the same DebugLoc as the scalar loop latch terminator instead |
3501 | // of the corresponding compare because they may have ended up with |
3502 | // different line numbers and we want to avoid awkward line stepping while |
3503 | // debugging. Eg. if the compare has got a line number inside the loop. |
3504 | CmpN->setDebugLoc(ScalarLatchTerm->getDebugLoc()); |
3505 | cast<BranchInst>(LoopMiddleBlock->getTerminator())->setCondition(CmpN); |
3506 | } |
3507 | |
3508 | // Get ready to start creating new instructions into the vectorized body. |
3509 | assert(LoopVectorPreHeader == L->getLoopPreheader() &&((LoopVectorPreHeader == L->getLoopPreheader() && "Inconsistent vector loop preheader" ) ? static_cast<void> (0) : __assert_fail ("LoopVectorPreHeader == L->getLoopPreheader() && \"Inconsistent vector loop preheader\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3510, __PRETTY_FUNCTION__)) |
3510 | "Inconsistent vector loop preheader")((LoopVectorPreHeader == L->getLoopPreheader() && "Inconsistent vector loop preheader" ) ? static_cast<void> (0) : __assert_fail ("LoopVectorPreHeader == L->getLoopPreheader() && \"Inconsistent vector loop preheader\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3510, __PRETTY_FUNCTION__)); |
3511 | Builder.SetInsertPoint(&*LoopVectorBody->getFirstInsertionPt()); |
3512 | |
3513 | Optional<MDNode *> VectorizedLoopID = |
3514 | makeFollowupLoopID(OrigLoopID, {LLVMLoopVectorizeFollowupAll, |
3515 | LLVMLoopVectorizeFollowupVectorized}); |
3516 | if (VectorizedLoopID.hasValue()) { |
3517 | L->setLoopID(VectorizedLoopID.getValue()); |
3518 | |
3519 | // Do not setAlreadyVectorized if loop attributes have been defined |
3520 | // explicitly. |
3521 | return LoopVectorPreHeader; |
3522 | } |
3523 | |
3524 | // Keep all loop hints from the original loop on the vector loop (we'll |
3525 | // replace the vectorizer-specific hints below). |
3526 | if (MDNode *LID = OrigLoop->getLoopID()) |
3527 | L->setLoopID(LID); |
3528 | |
3529 | LoopVectorizeHints Hints(L, true, *ORE); |
3530 | Hints.setAlreadyVectorized(); |
3531 | |
3532 | #ifdef EXPENSIVE_CHECKS |
3533 | assert(DT->verify(DominatorTree::VerificationLevel::Fast))((DT->verify(DominatorTree::VerificationLevel::Fast)) ? static_cast <void> (0) : __assert_fail ("DT->verify(DominatorTree::VerificationLevel::Fast)" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3533, __PRETTY_FUNCTION__)); |
3534 | LI->verify(*DT); |
3535 | #endif |
3536 | |
3537 | return LoopVectorPreHeader; |
3538 | } |
3539 | |
3540 | BasicBlock *InnerLoopVectorizer::createVectorizedLoopSkeleton() { |
3541 | /* |
3542 | In this function we generate a new loop. The new loop will contain |
3543 | the vectorized instructions while the old loop will continue to run the |
3544 | scalar remainder. |
3545 | |
3546 | [ ] <-- loop iteration number check. |
3547 | / | |
3548 | / v |
3549 | | [ ] <-- vector loop bypass (may consist of multiple blocks). |
3550 | | / | |
3551 | | / v |
3552 | || [ ] <-- vector pre header. |
3553 | |/ | |
3554 | | v |
3555 | | [ ] \ |
3556 | | [ ]_| <-- vector loop. |
3557 | | | |
3558 | | v |
3559 | | -[ ] <--- middle-block. |
3560 | | / | |
3561 | | / v |
3562 | -|- >[ ] <--- new preheader. |
3563 | | | |
3564 | | v |
3565 | | [ ] \ |
3566 | | [ ]_| <-- old scalar loop to handle remainder. |
3567 | \ | |
3568 | \ v |
3569 | >[ ] <-- exit block. |
3570 | ... |
3571 | */ |
3572 | |
3573 | // Get the metadata of the original loop before it gets modified. |
3574 | MDNode *OrigLoopID = OrigLoop->getLoopID(); |
3575 | |
3576 | // Create an empty vector loop, and prepare basic blocks for the runtime |
3577 | // checks. |
3578 | Loop *Lp = createVectorLoopSkeleton(""); |
3579 | |
3580 | // Now, compare the new count to zero. If it is zero skip the vector loop and |
3581 | // jump to the scalar loop. This check also covers the case where the |
3582 | // backedge-taken count is uint##_max: adding one to it will overflow leading |
3583 | // to an incorrect trip count of zero. In this (rare) case we will also jump |
3584 | // to the scalar loop. |
3585 | emitMinimumIterationCountCheck(Lp, LoopScalarPreHeader); |
3586 | |
3587 | // Generate the code to check any assumptions that we've made for SCEV |
3588 | // expressions. |
3589 | emitSCEVChecks(Lp, LoopScalarPreHeader); |
3590 | |
3591 | // Generate the code that checks in runtime if arrays overlap. We put the |
3592 | // checks into a separate block to make the more common case of few elements |
3593 | // faster. |
3594 | emitMemRuntimeChecks(Lp, LoopScalarPreHeader); |
3595 | |
3596 | // Some loops have a single integer induction variable, while other loops |
3597 | // don't. One example is c++ iterators that often have multiple pointer |
3598 | // induction variables. In the code below we also support a case where we |
3599 | // don't have a single induction variable. |
3600 | // |
3601 | // We try to obtain an induction variable from the original loop as hard |
3602 | // as possible. However if we don't find one that: |
3603 | // - is an integer |
3604 | // - counts from zero, stepping by one |
3605 | // - is the size of the widest induction variable type |
3606 | // then we create a new one. |
3607 | OldInduction = Legal->getPrimaryInduction(); |
3608 | Type *IdxTy = Legal->getWidestInductionType(); |
3609 | Value *StartIdx = ConstantInt::get(IdxTy, 0); |
3610 | // The loop step is equal to the vectorization factor (num of SIMD elements) |
3611 | // times the unroll factor (num of SIMD instructions). |
3612 | Builder.SetInsertPoint(&*Lp->getHeader()->getFirstInsertionPt()); |
3613 | Value *Step = createStepForVF(Builder, ConstantInt::get(IdxTy, UF), VF); |
3614 | Value *CountRoundDown = getOrCreateVectorTripCount(Lp); |
3615 | Induction = |
3616 | createInductionVariable(Lp, StartIdx, CountRoundDown, Step, |
3617 | getDebugLocFromInstOrOperands(OldInduction)); |
3618 | |
3619 | // Emit phis for the new starting index of the scalar loop. |
3620 | createInductionResumeValues(Lp, CountRoundDown); |
3621 | |
3622 | return completeLoopSkeleton(Lp, OrigLoopID); |
3623 | } |
3624 | |
3625 | // Fix up external users of the induction variable. At this point, we are |
3626 | // in LCSSA form, with all external PHIs that use the IV having one input value, |
3627 | // coming from the remainder loop. We need those PHIs to also have a correct |
3628 | // value for the IV when arriving directly from the middle block. |
3629 | void InnerLoopVectorizer::fixupIVUsers(PHINode *OrigPhi, |
3630 | const InductionDescriptor &II, |
3631 | Value *CountRoundDown, Value *EndValue, |
3632 | BasicBlock *MiddleBlock) { |
3633 | // There are two kinds of external IV usages - those that use the value |
3634 | // computed in the last iteration (the PHI) and those that use the penultimate |
3635 | // value (the value that feeds into the phi from the loop latch). |
3636 | // We allow both, but they, obviously, have different values. |
3637 | |
3638 | assert(OrigLoop->getUniqueExitBlock() && "Expected a single exit block")((OrigLoop->getUniqueExitBlock() && "Expected a single exit block" ) ? static_cast<void> (0) : __assert_fail ("OrigLoop->getUniqueExitBlock() && \"Expected a single exit block\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3638, __PRETTY_FUNCTION__)); |
3639 | |
3640 | DenseMap<Value *, Value *> MissingVals; |
3641 | |
3642 | // An external user of the last iteration's value should see the value that |
3643 | // the remainder loop uses to initialize its own IV. |
3644 | Value *PostInc = OrigPhi->getIncomingValueForBlock(OrigLoop->getLoopLatch()); |
3645 | for (User *U : PostInc->users()) { |
3646 | Instruction *UI = cast<Instruction>(U); |
3647 | if (!OrigLoop->contains(UI)) { |
3648 | assert(isa<PHINode>(UI) && "Expected LCSSA form")((isa<PHINode>(UI) && "Expected LCSSA form") ? static_cast <void> (0) : __assert_fail ("isa<PHINode>(UI) && \"Expected LCSSA form\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3648, __PRETTY_FUNCTION__)); |
3649 | MissingVals[UI] = EndValue; |
3650 | } |
3651 | } |
3652 | |
3653 | // An external user of the penultimate value need to see EndValue - Step. |
3654 | // The simplest way to get this is to recompute it from the constituent SCEVs, |
3655 | // that is Start + (Step * (CRD - 1)). |
3656 | for (User *U : OrigPhi->users()) { |
3657 | auto *UI = cast<Instruction>(U); |
3658 | if (!OrigLoop->contains(UI)) { |
3659 | const DataLayout &DL = |
3660 | OrigLoop->getHeader()->getModule()->getDataLayout(); |
3661 | assert(isa<PHINode>(UI) && "Expected LCSSA form")((isa<PHINode>(UI) && "Expected LCSSA form") ? static_cast <void> (0) : __assert_fail ("isa<PHINode>(UI) && \"Expected LCSSA form\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3661, __PRETTY_FUNCTION__)); |
3662 | |
3663 | IRBuilder<> B(MiddleBlock->getTerminator()); |
3664 | |
3665 | // Fast-math-flags propagate from the original induction instruction. |
3666 | if (II.getInductionBinOp() && isa<FPMathOperator>(II.getInductionBinOp())) |
3667 | B.setFastMathFlags(II.getInductionBinOp()->getFastMathFlags()); |
3668 | |
3669 | Value *CountMinusOne = B.CreateSub( |
3670 | CountRoundDown, ConstantInt::get(CountRoundDown->getType(), 1)); |
3671 | Value *CMO = |
3672 | !II.getStep()->getType()->isIntegerTy() |
3673 | ? B.CreateCast(Instruction::SIToFP, CountMinusOne, |
3674 | II.getStep()->getType()) |
3675 | : B.CreateSExtOrTrunc(CountMinusOne, II.getStep()->getType()); |
3676 | CMO->setName("cast.cmo"); |
3677 | Value *Escape = emitTransformedIndex(B, CMO, PSE.getSE(), DL, II); |
3678 | Escape->setName("ind.escape"); |
3679 | MissingVals[UI] = Escape; |
3680 | } |
3681 | } |
3682 | |
3683 | for (auto &I : MissingVals) { |
3684 | PHINode *PHI = cast<PHINode>(I.first); |
3685 | // One corner case we have to handle is two IVs "chasing" each-other, |
3686 | // that is %IV2 = phi [...], [ %IV1, %latch ] |
3687 | // In this case, if IV1 has an external use, we need to avoid adding both |
3688 | // "last value of IV1" and "penultimate value of IV2". So, verify that we |
3689 | // don't already have an incoming value for the middle block. |
3690 | if (PHI->getBasicBlockIndex(MiddleBlock) == -1) |
3691 | PHI->addIncoming(I.second, MiddleBlock); |
3692 | } |
3693 | } |
3694 | |
3695 | namespace { |
3696 | |
3697 | struct CSEDenseMapInfo { |
3698 | static bool canHandle(const Instruction *I) { |
3699 | return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || |
3700 | isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I); |
3701 | } |
3702 | |
3703 | static inline Instruction *getEmptyKey() { |
3704 | return DenseMapInfo<Instruction *>::getEmptyKey(); |
3705 | } |
3706 | |
3707 | static inline Instruction *getTombstoneKey() { |
3708 | return DenseMapInfo<Instruction *>::getTombstoneKey(); |
3709 | } |
3710 | |
3711 | static unsigned getHashValue(const Instruction *I) { |
3712 | assert(canHandle(I) && "Unknown instruction!")((canHandle(I) && "Unknown instruction!") ? static_cast <void> (0) : __assert_fail ("canHandle(I) && \"Unknown instruction!\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3712, __PRETTY_FUNCTION__)); |
3713 | return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(), |
3714 | I->value_op_end())); |
3715 | } |
3716 | |
3717 | static bool isEqual(const Instruction *LHS, const Instruction *RHS) { |
3718 | if (LHS == getEmptyKey() || RHS == getEmptyKey() || |
3719 | LHS == getTombstoneKey() || RHS == getTombstoneKey()) |
3720 | return LHS == RHS; |
3721 | return LHS->isIdenticalTo(RHS); |
3722 | } |
3723 | }; |
3724 | |
3725 | } // end anonymous namespace |
3726 | |
3727 | ///Perform cse of induction variable instructions. |
3728 | static void cse(BasicBlock *BB) { |
3729 | // Perform simple cse. |
3730 | SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap; |
3731 | for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) { |
3732 | Instruction *In = &*I++; |
3733 | |
3734 | if (!CSEDenseMapInfo::canHandle(In)) |
3735 | continue; |
3736 | |
3737 | // Check if we can replace this instruction with any of the |
3738 | // visited instructions. |
3739 | if (Instruction *V = CSEMap.lookup(In)) { |
3740 | In->replaceAllUsesWith(V); |
3741 | In->eraseFromParent(); |
3742 | continue; |
3743 | } |
3744 | |
3745 | CSEMap[In] = In; |
3746 | } |
3747 | } |
3748 | |
3749 | InstructionCost |
3750 | LoopVectorizationCostModel::getVectorCallCost(CallInst *CI, ElementCount VF, |
3751 | bool &NeedToScalarize) { |
3752 | Function *F = CI->getCalledFunction(); |
3753 | Type *ScalarRetTy = CI->getType(); |
3754 | SmallVector<Type *, 4> Tys, ScalarTys; |
3755 | for (auto &ArgOp : CI->arg_operands()) |
3756 | ScalarTys.push_back(ArgOp->getType()); |
3757 | |
3758 | // Estimate cost of scalarized vector call. The source operands are assumed |
3759 | // to be vectors, so we need to extract individual elements from there, |
3760 | // execute VF scalar calls, and then gather the result into the vector return |
3761 | // value. |
3762 | InstructionCost ScalarCallCost = |
3763 | TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys, TTI::TCK_RecipThroughput); |
3764 | if (VF.isScalar()) |
3765 | return ScalarCallCost; |
3766 | |
3767 | // Compute corresponding vector type for return value and arguments. |
3768 | Type *RetTy = ToVectorTy(ScalarRetTy, VF); |
3769 | for (Type *ScalarTy : ScalarTys) |
3770 | Tys.push_back(ToVectorTy(ScalarTy, VF)); |
3771 | |
3772 | // Compute costs of unpacking argument values for the scalar calls and |
3773 | // packing the return values to a vector. |
3774 | InstructionCost ScalarizationCost = getScalarizationOverhead(CI, VF); |
3775 | |
3776 | InstructionCost Cost = |
3777 | ScalarCallCost * VF.getKnownMinValue() + ScalarizationCost; |
3778 | |
3779 | // If we can't emit a vector call for this function, then the currently found |
3780 | // cost is the cost we need to return. |
3781 | NeedToScalarize = true; |
3782 | VFShape Shape = VFShape::get(*CI, VF, false /*HasGlobalPred*/); |
3783 | Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape); |
3784 | |
3785 | if (!TLI || CI->isNoBuiltin() || !VecFunc) |
3786 | return Cost; |
3787 | |
3788 | // If the corresponding vector cost is cheaper, return its cost. |
3789 | InstructionCost VectorCallCost = |
3790 | TTI.getCallInstrCost(nullptr, RetTy, Tys, TTI::TCK_RecipThroughput); |
3791 | if (VectorCallCost < Cost) { |
3792 | NeedToScalarize = false; |
3793 | Cost = VectorCallCost; |
3794 | } |
3795 | return Cost; |
3796 | } |
3797 | |
3798 | static Type *MaybeVectorizeType(Type *Elt, ElementCount VF) { |
3799 | if (VF.isScalar() || (!Elt->isIntOrPtrTy() && !Elt->isFloatingPointTy())) |
3800 | return Elt; |
3801 | return VectorType::get(Elt, VF); |
3802 | } |
3803 | |
3804 | InstructionCost |
3805 | LoopVectorizationCostModel::getVectorIntrinsicCost(CallInst *CI, |
3806 | ElementCount VF) { |
3807 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); |
3808 | assert(ID && "Expected intrinsic call!")((ID && "Expected intrinsic call!") ? static_cast< void> (0) : __assert_fail ("ID && \"Expected intrinsic call!\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3808, __PRETTY_FUNCTION__)); |
3809 | Type *RetTy = MaybeVectorizeType(CI->getType(), VF); |
3810 | FastMathFlags FMF; |
3811 | if (auto *FPMO = dyn_cast<FPMathOperator>(CI)) |
3812 | FMF = FPMO->getFastMathFlags(); |
3813 | |
3814 | SmallVector<const Value *> Arguments(CI->arg_begin(), CI->arg_end()); |
3815 | FunctionType *FTy = CI->getCalledFunction()->getFunctionType(); |
3816 | SmallVector<Type *> ParamTys; |
3817 | std::transform(FTy->param_begin(), FTy->param_end(), |
3818 | std::back_inserter(ParamTys), |
3819 | [&](Type *Ty) { return MaybeVectorizeType(Ty, VF); }); |
3820 | |
3821 | IntrinsicCostAttributes CostAttrs(ID, RetTy, Arguments, ParamTys, FMF, |
3822 | dyn_cast<IntrinsicInst>(CI)); |
3823 | return TTI.getIntrinsicInstrCost(CostAttrs, |
3824 | TargetTransformInfo::TCK_RecipThroughput); |
3825 | } |
3826 | |
3827 | static Type *smallestIntegerVectorType(Type *T1, Type *T2) { |
3828 | auto *I1 = cast<IntegerType>(cast<VectorType>(T1)->getElementType()); |
3829 | auto *I2 = cast<IntegerType>(cast<VectorType>(T2)->getElementType()); |
3830 | return I1->getBitWidth() < I2->getBitWidth() ? T1 : T2; |
3831 | } |
3832 | |
3833 | static Type *largestIntegerVectorType(Type *T1, Type *T2) { |
3834 | auto *I1 = cast<IntegerType>(cast<VectorType>(T1)->getElementType()); |
3835 | auto *I2 = cast<IntegerType>(cast<VectorType>(T2)->getElementType()); |
3836 | return I1->getBitWidth() > I2->getBitWidth() ? T1 : T2; |
3837 | } |
3838 | |
3839 | void InnerLoopVectorizer::truncateToMinimalBitwidths(VPTransformState &State) { |
3840 | // For every instruction `I` in MinBWs, truncate the operands, create a |
3841 | // truncated version of `I` and reextend its result. InstCombine runs |
3842 | // later and will remove any ext/trunc pairs. |
3843 | SmallPtrSet<Value *, 4> Erased; |
3844 | for (const auto &KV : Cost->getMinimalBitwidths()) { |
3845 | // If the value wasn't vectorized, we must maintain the original scalar |
3846 | // type. The absence of the value from State indicates that it |
3847 | // wasn't vectorized. |
3848 | VPValue *Def = State.Plan->getVPValue(KV.first); |
3849 | if (!State.hasAnyVectorValue(Def)) |
3850 | continue; |
3851 | for (unsigned Part = 0; Part < UF; ++Part) { |
3852 | Value *I = State.get(Def, Part); |
3853 | if (Erased.count(I) || I->use_empty() || !isa<Instruction>(I)) |
3854 | continue; |
3855 | Type *OriginalTy = I->getType(); |
3856 | Type *ScalarTruncatedTy = |
3857 | IntegerType::get(OriginalTy->getContext(), KV.second); |
3858 | auto *TruncatedTy = FixedVectorType::get( |
3859 | ScalarTruncatedTy, |
3860 | cast<FixedVectorType>(OriginalTy)->getNumElements()); |
3861 | if (TruncatedTy == OriginalTy) |
3862 | continue; |
3863 | |
3864 | IRBuilder<> B(cast<Instruction>(I)); |
3865 | auto ShrinkOperand = [&](Value *V) -> Value * { |
3866 | if (auto *ZI = dyn_cast<ZExtInst>(V)) |
3867 | if (ZI->getSrcTy() == TruncatedTy) |
3868 | return ZI->getOperand(0); |
3869 | return B.CreateZExtOrTrunc(V, TruncatedTy); |
3870 | }; |
3871 | |
3872 | // The actual instruction modification depends on the instruction type, |
3873 | // unfortunately. |
3874 | Value *NewI = nullptr; |
3875 | if (auto *BO = dyn_cast<BinaryOperator>(I)) { |
3876 | NewI = B.CreateBinOp(BO->getOpcode(), ShrinkOperand(BO->getOperand(0)), |
3877 | ShrinkOperand(BO->getOperand(1))); |
3878 | |
3879 | // Any wrapping introduced by shrinking this operation shouldn't be |
3880 | // considered undefined behavior. So, we can't unconditionally copy |
3881 | // arithmetic wrapping flags to NewI. |
3882 | cast<BinaryOperator>(NewI)->copyIRFlags(I, /*IncludeWrapFlags=*/false); |
3883 | } else if (auto *CI = dyn_cast<ICmpInst>(I)) { |
3884 | NewI = |
3885 | B.CreateICmp(CI->getPredicate(), ShrinkOperand(CI->getOperand(0)), |
3886 | ShrinkOperand(CI->getOperand(1))); |
3887 | } else if (auto *SI = dyn_cast<SelectInst>(I)) { |
3888 | NewI = B.CreateSelect(SI->getCondition(), |
3889 | ShrinkOperand(SI->getTrueValue()), |
3890 | ShrinkOperand(SI->getFalseValue())); |
3891 | } else if (auto *CI = dyn_cast<CastInst>(I)) { |
3892 | switch (CI->getOpcode()) { |
3893 | default: |
3894 | llvm_unreachable("Unhandled cast!")::llvm::llvm_unreachable_internal("Unhandled cast!", "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3894); |
3895 | case Instruction::Trunc: |
3896 | NewI = ShrinkOperand(CI->getOperand(0)); |
3897 | break; |
3898 | case Instruction::SExt: |
3899 | NewI = B.CreateSExtOrTrunc( |
3900 | CI->getOperand(0), |
3901 | smallestIntegerVectorType(OriginalTy, TruncatedTy)); |
3902 | break; |
3903 | case Instruction::ZExt: |
3904 | NewI = B.CreateZExtOrTrunc( |
3905 | CI->getOperand(0), |
3906 | smallestIntegerVectorType(OriginalTy, TruncatedTy)); |
3907 | break; |
3908 | } |
3909 | } else if (auto *SI = dyn_cast<ShuffleVectorInst>(I)) { |
3910 | auto Elements0 = cast<FixedVectorType>(SI->getOperand(0)->getType()) |
3911 | ->getNumElements(); |
3912 | auto *O0 = B.CreateZExtOrTrunc( |
3913 | SI->getOperand(0), |
3914 | FixedVectorType::get(ScalarTruncatedTy, Elements0)); |
3915 | auto Elements1 = cast<FixedVectorType>(SI->getOperand(1)->getType()) |
3916 | ->getNumElements(); |
3917 | auto *O1 = B.CreateZExtOrTrunc( |
3918 | SI->getOperand(1), |
3919 | FixedVectorType::get(ScalarTruncatedTy, Elements1)); |
3920 | |
3921 | NewI = B.CreateShuffleVector(O0, O1, SI->getShuffleMask()); |
3922 | } else if (isa<LoadInst>(I) || isa<PHINode>(I)) { |
3923 | // Don't do anything with the operands, just extend the result. |
3924 | continue; |
3925 | } else if (auto *IE = dyn_cast<InsertElementInst>(I)) { |
3926 | auto Elements = cast<FixedVectorType>(IE->getOperand(0)->getType()) |
3927 | ->getNumElements(); |
3928 | auto *O0 = B.CreateZExtOrTrunc( |
3929 | IE->getOperand(0), |
3930 | FixedVectorType::get(ScalarTruncatedTy, Elements)); |
3931 | auto *O1 = B.CreateZExtOrTrunc(IE->getOperand(1), ScalarTruncatedTy); |
3932 | NewI = B.CreateInsertElement(O0, O1, IE->getOperand(2)); |
3933 | } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) { |
3934 | auto Elements = cast<FixedVectorType>(EE->getOperand(0)->getType()) |
3935 | ->getNumElements(); |
3936 | auto *O0 = B.CreateZExtOrTrunc( |
3937 | EE->getOperand(0), |
3938 | FixedVectorType::get(ScalarTruncatedTy, Elements)); |
3939 | NewI = B.CreateExtractElement(O0, EE->getOperand(2)); |
3940 | } else { |
3941 | // If we don't know what to do, be conservative and don't do anything. |
3942 | continue; |
3943 | } |
3944 | |
3945 | // Lastly, extend the result. |
3946 | NewI->takeName(cast<Instruction>(I)); |
3947 | Value *Res = B.CreateZExtOrTrunc(NewI, OriginalTy); |
3948 | I->replaceAllUsesWith(Res); |
3949 | cast<Instruction>(I)->eraseFromParent(); |
3950 | Erased.insert(I); |
3951 | State.reset(Def, Res, Part); |
3952 | } |
3953 | } |
3954 | |
3955 | // We'll have created a bunch of ZExts that are now parentless. Clean up. |
3956 | for (const auto &KV : Cost->getMinimalBitwidths()) { |
3957 | // If the value wasn't vectorized, we must maintain the original scalar |
3958 | // type. The absence of the value from State indicates that it |
3959 | // wasn't vectorized. |
3960 | VPValue *Def = State.Plan->getVPValue(KV.first); |
3961 | if (!State.hasAnyVectorValue(Def)) |
3962 | continue; |
3963 | for (unsigned Part = 0; Part < UF; ++Part) { |
3964 | Value *I = State.get(Def, Part); |
3965 | ZExtInst *Inst = dyn_cast<ZExtInst>(I); |
3966 | if (Inst && Inst->use_empty()) { |
3967 | Value *NewI = Inst->getOperand(0); |
3968 | Inst->eraseFromParent(); |
3969 | State.reset(Def, NewI, Part); |
3970 | } |
3971 | } |
3972 | } |
3973 | } |
3974 | |
3975 | void InnerLoopVectorizer::fixVectorizedLoop(VPTransformState &State) { |
3976 | // Insert truncates and extends for any truncated instructions as hints to |
3977 | // InstCombine. |
3978 | if (VF.isVector()) |
3979 | truncateToMinimalBitwidths(State); |
3980 | |
3981 | // Fix widened non-induction PHIs by setting up the PHI operands. |
3982 | if (OrigPHIsToFix.size()) { |
3983 | assert(EnableVPlanNativePath &&((EnableVPlanNativePath && "Unexpected non-induction PHIs for fixup in non VPlan-native path" ) ? static_cast<void> (0) : __assert_fail ("EnableVPlanNativePath && \"Unexpected non-induction PHIs for fixup in non VPlan-native path\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3984, __PRETTY_FUNCTION__)) |
3984 | "Unexpected non-induction PHIs for fixup in non VPlan-native path")((EnableVPlanNativePath && "Unexpected non-induction PHIs for fixup in non VPlan-native path" ) ? static_cast<void> (0) : __assert_fail ("EnableVPlanNativePath && \"Unexpected non-induction PHIs for fixup in non VPlan-native path\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3984, __PRETTY_FUNCTION__)); |
3985 | fixNonInductionPHIs(State); |
3986 | } |
3987 | |
3988 | // At this point every instruction in the original loop is widened to a |
3989 | // vector form. Now we need to fix the recurrences in the loop. These PHI |
3990 | // nodes are currently empty because we did not want to introduce cycles. |
3991 | // This is the second stage of vectorizing recurrences. |
3992 | fixCrossIterationPHIs(State); |
3993 | |
3994 | // Forget the original basic block. |
3995 | PSE.getSE()->forgetLoop(OrigLoop); |
3996 | |
3997 | // Fix-up external users of the induction variables. |
3998 | for (auto &Entry : Legal->getInductionVars()) |
3999 | fixupIVUsers(Entry.first, Entry.second, |
4000 | getOrCreateVectorTripCount(LI->getLoopFor(LoopVectorBody)), |
4001 | IVEndValues[Entry.first], LoopMiddleBlock); |
4002 | |
4003 | fixLCSSAPHIs(State); |
4004 | for (Instruction *PI : PredicatedInstructions) |
4005 | sinkScalarOperands(&*PI); |
4006 | |
4007 | // Remove redundant induction instructions. |
4008 | cse(LoopVectorBody); |
4009 | |
4010 | // Set/update profile weights for the vector and remainder loops as original |
4011 | // loop iterations are now distributed among them. Note that original loop |
4012 | // represented by LoopScalarBody becomes remainder loop after vectorization. |
4013 | // |
4014 | // For cases like foldTailByMasking() and requiresScalarEpiloque() we may |
4015 | // end up getting slightly roughened result but that should be OK since |
4016 | // profile is not inherently precise anyway. Note also possible bypass of |
4017 | // vector code caused by legality checks is ignored, assigning all the weight |
4018 | // to the vector loop, optimistically. |
4019 | // |
4020 | // For scalable vectorization we can't know at compile time how many iterations |
4021 | // of the loop are handled in one vector iteration, so instead assume a pessimistic |
4022 | // vscale of '1'. |
4023 | setProfileInfoAfterUnrolling( |
4024 | LI->getLoopFor(LoopScalarBody), LI->getLoopFor(LoopVectorBody), |
4025 | LI->getLoopFor(LoopScalarBody), VF.getKnownMinValue() * UF); |
4026 | } |
4027 | |
4028 | void InnerLoopVectorizer::fixCrossIterationPHIs(VPTransformState &State) { |
4029 | // In order to support recurrences we need to be able to vectorize Phi nodes. |
4030 | // Phi nodes have cycles, so we need to vectorize them in two stages. This is |
4031 | // stage #2: We now need to fix the recurrences by adding incoming edges to |
4032 | // the currently empty PHI nodes. At this point every instruction in the |
4033 | // original loop is widened to a vector form so we can use them to construct |
4034 | // the incoming edges. |
4035 | for (PHINode &Phi : OrigLoop->getHeader()->phis()) { |
4036 | // Handle first-order recurrences and reductions that need to be fixed. |
4037 | if (Legal->isFirstOrderRecurrence(&Phi)) |
4038 | fixFirstOrderRecurrence(&Phi, State); |
4039 | else if (Legal->isReductionVariable(&Phi)) |
4040 | fixReduction(&Phi, State); |
4041 | } |
4042 | } |
4043 | |
4044 | void InnerLoopVectorizer::fixFirstOrderRecurrence(PHINode *Phi, |
4045 | VPTransformState &State) { |
4046 | // This is the second phase of vectorizing first-order recurrences. An |
4047 | // overview of the transformation is described below. Suppose we have the |
4048 | // following loop. |
4049 | // |
4050 | // for (int i = 0; i < n; ++i) |
4051 | // b[i] = a[i] - a[i - 1]; |
4052 | // |
4053 | // There is a first-order recurrence on "a". For this loop, the shorthand |
4054 | // scalar IR looks like: |
4055 | // |
4056 | // scalar.ph: |
4057 | // s_init = a[-1] |
4058 | // br scalar.body |
4059 | // |
4060 | // scalar.body: |
4061 | // i = phi [0, scalar.ph], [i+1, scalar.body] |
4062 | // s1 = phi [s_init, scalar.ph], [s2, scalar.body] |
4063 | // s2 = a[i] |
4064 | // b[i] = s2 - s1 |
4065 | // br cond, scalar.body, ... |
4066 | // |
4067 | // In this example, s1 is a recurrence because it's value depends on the |
4068 | // previous iteration. In the first phase of vectorization, we created a |
4069 | // temporary value for s1. We now complete the vectorization and produce the |
4070 | // shorthand vector IR shown below (for VF = 4, UF = 1). |
4071 | // |
4072 | // vector.ph: |
4073 | // v_init = vector(..., ..., ..., a[-1]) |
4074 | // br vector.body |
4075 | // |
4076 | // vector.body |
4077 | // i = phi [0, vector.ph], [i+4, vector.body] |
4078 | // v1 = phi [v_init, vector.ph], [v2, vector.body] |
4079 | // v2 = a[i, i+1, i+2, i+3]; |
4080 | // v3 = vector(v1(3), v2(0, 1, 2)) |
4081 | // b[i, i+1, i+2, i+3] = v2 - v3 |
4082 | // br cond, vector.body, middle.block |
4083 | // |
4084 | // middle.block: |
4085 | // x = v2(3) |
4086 | // br scalar.ph |
4087 | // |
4088 | // scalar.ph: |
4089 | // s_init = phi [x, middle.block], [a[-1], otherwise] |
4090 | // br scalar.body |
4091 | // |
4092 | // After execution completes the vector loop, we extract the next value of |
4093 | // the recurrence (x) to use as the initial value in the scalar loop. |
4094 | |
4095 | // Get the original loop preheader and single loop latch. |
4096 | auto *Preheader = OrigLoop->getLoopPreheader(); |
4097 | auto *Latch = OrigLoop->getLoopLatch(); |
4098 | |
4099 | // Get the initial and previous values of the scalar recurrence. |
4100 | auto *ScalarInit = Phi->getIncomingValueForBlock(Preheader); |
4101 | auto *Previous = Phi->getIncomingValueForBlock(Latch); |
4102 | |
4103 | // Create a vector from the initial value. |
4104 | auto *VectorInit = ScalarInit; |
4105 | if (VF.isVector()) { |
4106 | Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); |
4107 | assert(!VF.isScalable() && "VF is assumed to be non scalable.")((!VF.isScalable() && "VF is assumed to be non scalable." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"VF is assumed to be non scalable.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4107, __PRETTY_FUNCTION__)); |
4108 | VectorInit = Builder.CreateInsertElement( |
4109 | PoisonValue::get(VectorType::get(VectorInit->getType(), VF)), VectorInit, |
4110 | Builder.getInt32(VF.getKnownMinValue() - 1), "vector.recur.init"); |
4111 | } |
4112 | |
4113 | VPValue *PhiDef = State.Plan->getVPValue(Phi); |
4114 | VPValue *PreviousDef = State.Plan->getVPValue(Previous); |
4115 | // We constructed a temporary phi node in the first phase of vectorization. |
4116 | // This phi node will eventually be deleted. |
4117 | Builder.SetInsertPoint(cast<Instruction>(State.get(PhiDef, 0))); |
4118 | |
4119 | // Create a phi node for the new recurrence. The current value will either be |
4120 | // the initial value inserted into a vector or loop-varying vector value. |
4121 | auto *VecPhi = Builder.CreatePHI(VectorInit->getType(), 2, "vector.recur"); |
4122 | VecPhi->addIncoming(VectorInit, LoopVectorPreHeader); |
4123 | |
4124 | // Get the vectorized previous value of the last part UF - 1. It appears last |
4125 | // among all unrolled iterations, due to the order of their construction. |
4126 | Value *PreviousLastPart = State.get(PreviousDef, UF - 1); |
4127 | |
4128 | // Find and set the insertion point after the previous value if it is an |
4129 | // instruction. |
4130 | BasicBlock::iterator InsertPt; |
4131 | // Note that the previous value may have been constant-folded so it is not |
4132 | // guaranteed to be an instruction in the vector loop. |
4133 | // FIXME: Loop invariant values do not form recurrences. We should deal with |
4134 | // them earlier. |
4135 | if (LI->getLoopFor(LoopVectorBody)->isLoopInvariant(PreviousLastPart)) |
4136 | InsertPt = LoopVectorBody->getFirstInsertionPt(); |
4137 | else { |
4138 | Instruction *PreviousInst = cast<Instruction>(PreviousLastPart); |
4139 | if (isa<PHINode>(PreviousLastPart)) |
4140 | // If the previous value is a phi node, we should insert after all the phi |
4141 | // nodes in the block containing the PHI to avoid breaking basic block |
4142 | // verification. Note that the basic block may be different to |
4143 | // LoopVectorBody, in case we predicate the loop. |
4144 | InsertPt = PreviousInst->getParent()->getFirstInsertionPt(); |
4145 | else |
4146 | InsertPt = ++PreviousInst->getIterator(); |
4147 | } |
4148 | Builder.SetInsertPoint(&*InsertPt); |
4149 | |
4150 | // We will construct a vector for the recurrence by combining the values for |
4151 | // the current and previous iterations. This is the required shuffle mask. |
4152 | assert(!VF.isScalable())((!VF.isScalable()) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable()", "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4152, __PRETTY_FUNCTION__)); |
4153 | SmallVector<int, 8> ShuffleMask(VF.getKnownMinValue()); |
4154 | ShuffleMask[0] = VF.getKnownMinValue() - 1; |
4155 | for (unsigned I = 1; I < VF.getKnownMinValue(); ++I) |
4156 | ShuffleMask[I] = I + VF.getKnownMinValue() - 1; |
4157 | |
4158 | // The vector from which to take the initial value for the current iteration |
4159 | // (actual or unrolled). Initially, this is the vector phi node. |
4160 | Value *Incoming = VecPhi; |
4161 | |
4162 | // Shuffle the current and previous vector and update the vector parts. |
4163 | for (unsigned Part = 0; Part < UF; ++Part) { |
4164 | Value *PreviousPart = State.get(PreviousDef, Part); |
4165 | Value *PhiPart = State.get(PhiDef, Part); |
4166 | auto *Shuffle = |
4167 | VF.isVector() |
4168 | ? Builder.CreateShuffleVector(Incoming, PreviousPart, ShuffleMask) |
4169 | : Incoming; |
4170 | PhiPart->replaceAllUsesWith(Shuffle); |
4171 | cast<Instruction>(PhiPart)->eraseFromParent(); |
4172 | State.reset(PhiDef, Shuffle, Part); |
4173 | Incoming = PreviousPart; |
4174 | } |
4175 | |
4176 | // Fix the latch value of the new recurrence in the vector loop. |
4177 | VecPhi->addIncoming(Incoming, LI->getLoopFor(LoopVectorBody)->getLoopLatch()); |
4178 | |
4179 | // Extract the last vector element in the middle block. This will be the |
4180 | // initial value for the recurrence when jumping to the scalar loop. |
4181 | auto *ExtractForScalar = Incoming; |
4182 | if (VF.isVector()) { |
4183 | Builder.SetInsertPoint(LoopMiddleBlock->getTerminator()); |
4184 | ExtractForScalar = Builder.CreateExtractElement( |
4185 | ExtractForScalar, Builder.getInt32(VF.getKnownMinValue() - 1), |
4186 | "vector.recur.extract"); |
4187 | } |
4188 | // Extract the second last element in the middle block if the |
4189 | // Phi is used outside the loop. We need to extract the phi itself |
4190 | // and not the last element (the phi update in the current iteration). This |
4191 | // will be the value when jumping to the exit block from the LoopMiddleBlock, |
4192 | // when the scalar loop is not run at all. |
4193 | Value *ExtractForPhiUsedOutsideLoop = nullptr; |
4194 | if (VF.isVector()) |
4195 | ExtractForPhiUsedOutsideLoop = Builder.CreateExtractElement( |
4196 | Incoming, Builder.getInt32(VF.getKnownMinValue() - 2), |
4197 | "vector.recur.extract.for.phi"); |
4198 | // When loop is unrolled without vectorizing, initialize |
4199 | // ExtractForPhiUsedOutsideLoop with the value just prior to unrolled value of |
4200 | // `Incoming`. This is analogous to the vectorized case above: extracting the |
4201 | // second last element when VF > 1. |
4202 | else if (UF > 1) |
4203 | ExtractForPhiUsedOutsideLoop = State.get(PreviousDef, UF - 2); |
4204 | |
4205 | // Fix the initial value of the original recurrence in the scalar loop. |
4206 | Builder.SetInsertPoint(&*LoopScalarPreHeader->begin()); |
4207 | auto *Start = Builder.CreatePHI(Phi->getType(), 2, "scalar.recur.init"); |
4208 | for (auto *BB : predecessors(LoopScalarPreHeader)) { |
4209 | auto *Incoming = BB == LoopMiddleBlock ? ExtractForScalar : ScalarInit; |
4210 | Start->addIncoming(Incoming, BB); |
4211 | } |
4212 | |
4213 | Phi->setIncomingValueForBlock(LoopScalarPreHeader, Start); |
4214 | Phi->setName("scalar.recur"); |
4215 | |
4216 | // Finally, fix users of the recurrence outside the loop. The users will need |
4217 | // either the last value of the scalar recurrence or the last value of the |
4218 | // vector recurrence we extracted in the middle block. Since the loop is in |
4219 | // LCSSA form, we just need to find all the phi nodes for the original scalar |
4220 | // recurrence in the exit block, and then add an edge for the middle block. |
4221 | // Note that LCSSA does not imply single entry when the original scalar loop |
4222 | // had multiple exiting edges (as we always run the last iteration in the |
4223 | // scalar epilogue); in that case, the exiting path through middle will be |
4224 | // dynamically dead and the value picked for the phi doesn't matter. |
4225 | for (PHINode &LCSSAPhi : LoopExitBlock->phis()) |
4226 | if (any_of(LCSSAPhi.incoming_values(), |
4227 | [Phi](Value *V) { return V == Phi; })) |
4228 | LCSSAPhi.addIncoming(ExtractForPhiUsedOutsideLoop, LoopMiddleBlock); |
4229 | } |
4230 | |
4231 | void InnerLoopVectorizer::fixReduction(PHINode *Phi, VPTransformState &State) { |
4232 | // Get it's reduction variable descriptor. |
4233 | assert(Legal->isReductionVariable(Phi) &&((Legal->isReductionVariable(Phi) && "Unable to find the reduction variable" ) ? static_cast<void> (0) : __assert_fail ("Legal->isReductionVariable(Phi) && \"Unable to find the reduction variable\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4234, __PRETTY_FUNCTION__)) |
4234 | "Unable to find the reduction variable")((Legal->isReductionVariable(Phi) && "Unable to find the reduction variable" ) ? static_cast<void> (0) : __assert_fail ("Legal->isReductionVariable(Phi) && \"Unable to find the reduction variable\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4234, __PRETTY_FUNCTION__)); |
4235 | RecurrenceDescriptor RdxDesc = Legal->getReductionVars()[Phi]; |
4236 | |
4237 | RecurKind RK = RdxDesc.getRecurrenceKind(); |
4238 | TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue(); |
4239 | Instruction *LoopExitInst = RdxDesc.getLoopExitInstr(); |
4240 | setDebugLocFromInst(Builder, ReductionStartValue); |
4241 | bool IsInLoopReductionPhi = Cost->isInLoopReduction(Phi); |
4242 | |
4243 | VPValue *LoopExitInstDef = State.Plan->getVPValue(LoopExitInst); |
4244 | // This is the vector-clone of the value that leaves the loop. |
4245 | Type *VecTy = State.get(LoopExitInstDef, 0)->getType(); |
4246 | |
4247 | // Wrap flags are in general invalid after vectorization, clear them. |
4248 | clearReductionWrapFlags(RdxDesc, State); |
4249 | |
4250 | // Fix the vector-loop phi. |
4251 | |
4252 | // Reductions do not have to start at zero. They can start with |
4253 | // any loop invariant values. |
4254 | BasicBlock *Latch = OrigLoop->getLoopLatch(); |
4255 | Value *LoopVal = Phi->getIncomingValueForBlock(Latch); |
4256 | |
4257 | for (unsigned Part = 0; Part < UF; ++Part) { |
4258 | Value *VecRdxPhi = State.get(State.Plan->getVPValue(Phi), Part); |
4259 | Value *Val = State.get(State.Plan->getVPValue(LoopVal), Part); |
4260 | cast<PHINode>(VecRdxPhi) |
4261 | ->addIncoming(Val, LI->getLoopFor(LoopVectorBody)->getLoopLatch()); |
4262 | } |
4263 | |
4264 | // Before each round, move the insertion point right between |
4265 | // the PHIs and the values we are going to write. |
4266 | // This allows us to write both PHINodes and the extractelement |
4267 | // instructions. |
4268 | Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt()); |
4269 | |
4270 | setDebugLocFromInst(Builder, LoopExitInst); |
4271 | |
4272 | // If tail is folded by masking, the vector value to leave the loop should be |
4273 | // a Select choosing between the vectorized LoopExitInst and vectorized Phi, |
4274 | // instead of the former. For an inloop reduction the reduction will already |
4275 | // be predicated, and does not need to be handled here. |
4276 | if (Cost->foldTailByMasking() && !IsInLoopReductionPhi) { |
4277 | for (unsigned Part = 0; Part < UF; ++Part) { |
4278 | Value *VecLoopExitInst = State.get(LoopExitInstDef, Part); |
4279 | Value *Sel = nullptr; |
4280 | for (User *U : VecLoopExitInst->users()) { |
4281 | if (isa<SelectInst>(U)) { |
4282 | assert(!Sel && "Reduction exit feeding two selects")((!Sel && "Reduction exit feeding two selects") ? static_cast <void> (0) : __assert_fail ("!Sel && \"Reduction exit feeding two selects\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4282, __PRETTY_FUNCTION__)); |
4283 | Sel = U; |
4284 | } else |
4285 | assert(isa<PHINode>(U) && "Reduction exit must feed Phi's or select")((isa<PHINode>(U) && "Reduction exit must feed Phi's or select" ) ? static_cast<void> (0) : __assert_fail ("isa<PHINode>(U) && \"Reduction exit must feed Phi's or select\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4285, __PRETTY_FUNCTION__)); |
4286 | } |
4287 | assert(Sel && "Reduction exit feeds no select")((Sel && "Reduction exit feeds no select") ? static_cast <void> (0) : __assert_fail ("Sel && \"Reduction exit feeds no select\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4287, __PRETTY_FUNCTION__)); |
4288 | State.reset(LoopExitInstDef, Sel, Part); |
4289 | |
4290 | // If the target can create a predicated operator for the reduction at no |
4291 | // extra cost in the loop (for example a predicated vadd), it can be |
4292 | // cheaper for the select to remain in the loop than be sunk out of it, |
4293 | // and so use the select value for the phi instead of the old |
4294 | // LoopExitValue. |
4295 | RecurrenceDescriptor RdxDesc = Legal->getReductionVars()[Phi]; |
4296 | if (PreferPredicatedReductionSelect || |
4297 | TTI->preferPredicatedReductionSelect( |
4298 | RdxDesc.getOpcode(), Phi->getType(), |
4299 | TargetTransformInfo::ReductionFlags())) { |
4300 | auto *VecRdxPhi = |
4301 | cast<PHINode>(State.get(State.Plan->getVPValue(Phi), Part)); |
4302 | VecRdxPhi->setIncomingValueForBlock( |
4303 | LI->getLoopFor(LoopVectorBody)->getLoopLatch(), Sel); |
4304 | } |
4305 | } |
4306 | } |
4307 | |
4308 | // If the vector reduction can be performed in a smaller type, we truncate |
4309 | // then extend the loop exit value to enable InstCombine to evaluate the |
4310 | // entire expression in the smaller type. |
4311 | if (VF.isVector() && Phi->getType() != RdxDesc.getRecurrenceType()) { |
4312 | assert(!IsInLoopReductionPhi && "Unexpected truncated inloop reduction!")((!IsInLoopReductionPhi && "Unexpected truncated inloop reduction!" ) ? static_cast<void> (0) : __assert_fail ("!IsInLoopReductionPhi && \"Unexpected truncated inloop reduction!\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4312, __PRETTY_FUNCTION__)); |
4313 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4313, __PRETTY_FUNCTION__)); |
4314 | Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF); |
4315 | Builder.SetInsertPoint( |
4316 | LI->getLoopFor(LoopVectorBody)->getLoopLatch()->getTerminator()); |
4317 | VectorParts RdxParts(UF); |
4318 | for (unsigned Part = 0; Part < UF; ++Part) { |
4319 | RdxParts[Part] = State.get(LoopExitInstDef, Part); |
4320 | Value *Trunc = Builder.CreateTrunc(RdxParts[Part], RdxVecTy); |
4321 | Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy) |
4322 | : Builder.CreateZExt(Trunc, VecTy); |
4323 | for (Value::user_iterator UI = RdxParts[Part]->user_begin(); |
4324 | UI != RdxParts[Part]->user_end();) |
4325 | if (*UI != Trunc) { |
4326 | (*UI++)->replaceUsesOfWith(RdxParts[Part], Extnd); |
4327 | RdxParts[Part] = Extnd; |
4328 | } else { |
4329 | ++UI; |
4330 | } |
4331 | } |
4332 | Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt()); |
4333 | for (unsigned Part = 0; Part < UF; ++Part) { |
4334 | RdxParts[Part] = Builder.CreateTrunc(RdxParts[Part], RdxVecTy); |
4335 | State.reset(LoopExitInstDef, RdxParts[Part], Part); |
4336 | } |
4337 | } |
4338 | |
4339 | // Reduce all of the unrolled parts into a single vector. |
4340 | Value *ReducedPartRdx = State.get(LoopExitInstDef, 0); |
4341 | unsigned Op = RecurrenceDescriptor::getOpcode(RK); |
4342 | |
4343 | // The middle block terminator has already been assigned a DebugLoc here (the |
4344 | // OrigLoop's single latch terminator). We want the whole middle block to |
4345 | // appear to execute on this line because: (a) it is all compiler generated, |
4346 | // (b) these instructions are always executed after evaluating the latch |
4347 | // conditional branch, and (c) other passes may add new predecessors which |
4348 | // terminate on this line. This is the easiest way to ensure we don't |
4349 | // accidentally cause an extra step back into the loop while debugging. |
4350 | setDebugLocFromInst(Builder, LoopMiddleBlock->getTerminator()); |
4351 | { |
4352 | // Floating-point operations should have some FMF to enable the reduction. |
4353 | IRBuilderBase::FastMathFlagGuard FMFG(Builder); |
4354 | Builder.setFastMathFlags(RdxDesc.getFastMathFlags()); |
4355 | for (unsigned Part = 1; Part < UF; ++Part) { |
4356 | Value *RdxPart = State.get(LoopExitInstDef, Part); |
4357 | if (Op != Instruction::ICmp && Op != Instruction::FCmp) { |
4358 | ReducedPartRdx = Builder.CreateBinOp( |
4359 | (Instruction::BinaryOps)Op, RdxPart, ReducedPartRdx, "bin.rdx"); |
4360 | } else { |
4361 | ReducedPartRdx = createMinMaxOp(Builder, RK, ReducedPartRdx, RdxPart); |
4362 | } |
4363 | } |
4364 | } |
4365 | |
4366 | // Create the reduction after the loop. Note that inloop reductions create the |
4367 | // target reduction in the loop using a Reduction recipe. |
4368 | if (VF.isVector() && !IsInLoopReductionPhi) { |
4369 | ReducedPartRdx = |
4370 | createTargetReduction(Builder, TTI, RdxDesc, ReducedPartRdx); |
4371 | // If the reduction can be performed in a smaller type, we need to extend |
4372 | // the reduction to the wider type before we branch to the original loop. |
4373 | if (Phi->getType() != RdxDesc.getRecurrenceType()) |
4374 | ReducedPartRdx = |
4375 | RdxDesc.isSigned() |
4376 | ? Builder.CreateSExt(ReducedPartRdx, Phi->getType()) |
4377 | : Builder.CreateZExt(ReducedPartRdx, Phi->getType()); |
4378 | } |
4379 | |
4380 | // Create a phi node that merges control-flow from the backedge-taken check |
4381 | // block and the middle block. |
4382 | PHINode *BCBlockPhi = PHINode::Create(Phi->getType(), 2, "bc.merge.rdx", |
4383 | LoopScalarPreHeader->getTerminator()); |
4384 | for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) |
4385 | BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[I]); |
4386 | BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock); |
4387 | |
4388 | // Now, we need to fix the users of the reduction variable |
4389 | // inside and outside of the scalar remainder loop. |
4390 | |
4391 | // We know that the loop is in LCSSA form. We need to update the PHI nodes |
4392 | // in the exit blocks. See comment on analogous loop in |
4393 | // fixFirstOrderRecurrence for a more complete explaination of the logic. |
4394 | for (PHINode &LCSSAPhi : LoopExitBlock->phis()) |
4395 | if (any_of(LCSSAPhi.incoming_values(), |
4396 | [LoopExitInst](Value *V) { return V == LoopExitInst; })) |
4397 | LCSSAPhi.addIncoming(ReducedPartRdx, LoopMiddleBlock); |
4398 | |
4399 | // Fix the scalar loop reduction variable with the incoming reduction sum |
4400 | // from the vector body and from the backedge value. |
4401 | int IncomingEdgeBlockIdx = |
4402 | Phi->getBasicBlockIndex(OrigLoop->getLoopLatch()); |
4403 | assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index")((IncomingEdgeBlockIdx >= 0 && "Invalid block index" ) ? static_cast<void> (0) : __assert_fail ("IncomingEdgeBlockIdx >= 0 && \"Invalid block index\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4403, __PRETTY_FUNCTION__)); |
4404 | // Pick the other block. |
4405 | int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); |
4406 | Phi->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi); |
4407 | Phi->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst); |
4408 | } |
4409 | |
4410 | void InnerLoopVectorizer::clearReductionWrapFlags(RecurrenceDescriptor &RdxDesc, |
4411 | VPTransformState &State) { |
4412 | RecurKind RK = RdxDesc.getRecurrenceKind(); |
4413 | if (RK != RecurKind::Add && RK != RecurKind::Mul) |
4414 | return; |
4415 | |
4416 | Instruction *LoopExitInstr = RdxDesc.getLoopExitInstr(); |
4417 | assert(LoopExitInstr && "null loop exit instruction")((LoopExitInstr && "null loop exit instruction") ? static_cast <void> (0) : __assert_fail ("LoopExitInstr && \"null loop exit instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4417, __PRETTY_FUNCTION__)); |
4418 | SmallVector<Instruction *, 8> Worklist; |
4419 | SmallPtrSet<Instruction *, 8> Visited; |
4420 | Worklist.push_back(LoopExitInstr); |
4421 | Visited.insert(LoopExitInstr); |
4422 | |
4423 | while (!Worklist.empty()) { |
4424 | Instruction *Cur = Worklist.pop_back_val(); |
4425 | if (isa<OverflowingBinaryOperator>(Cur)) |
4426 | for (unsigned Part = 0; Part < UF; ++Part) { |
4427 | Value *V = State.get(State.Plan->getVPValue(Cur), Part); |
4428 | cast<Instruction>(V)->dropPoisonGeneratingFlags(); |
4429 | } |
4430 | |
4431 | for (User *U : Cur->users()) { |
4432 | Instruction *UI = cast<Instruction>(U); |
4433 | if ((Cur != LoopExitInstr || OrigLoop->contains(UI->getParent())) && |
4434 | Visited.insert(UI).second) |
4435 | Worklist.push_back(UI); |
4436 | } |
4437 | } |
4438 | } |
4439 | |
4440 | void InnerLoopVectorizer::fixLCSSAPHIs(VPTransformState &State) { |
4441 | for (PHINode &LCSSAPhi : LoopExitBlock->phis()) { |
4442 | if (LCSSAPhi.getBasicBlockIndex(LoopMiddleBlock) != -1) |
4443 | // Some phis were already hand updated by the reduction and recurrence |
4444 | // code above, leave them alone. |
4445 | continue; |
4446 | |
4447 | auto *IncomingValue = LCSSAPhi.getIncomingValue(0); |
4448 | // Non-instruction incoming values will have only one value. |
4449 | |
4450 | VPLane Lane = VPLane::getFirstLane(); |
4451 | if (isa<Instruction>(IncomingValue) && |
4452 | !Cost->isUniformAfterVectorization(cast<Instruction>(IncomingValue), |
4453 | VF)) |
4454 | Lane = VPLane::getLastLaneForVF(VF); |
4455 | |
4456 | // Can be a loop invariant incoming value or the last scalar value to be |
4457 | // extracted from the vectorized loop. |
4458 | Builder.SetInsertPoint(LoopMiddleBlock->getTerminator()); |
4459 | Value *lastIncomingValue = |
4460 | OrigLoop->isLoopInvariant(IncomingValue) |
4461 | ? IncomingValue |
4462 | : State.get(State.Plan->getVPValue(IncomingValue), |
4463 | VPIteration(UF - 1, Lane)); |
4464 | LCSSAPhi.addIncoming(lastIncomingValue, LoopMiddleBlock); |
4465 | } |
4466 | } |
4467 | |
4468 | void InnerLoopVectorizer::sinkScalarOperands(Instruction *PredInst) { |
4469 | // The basic block and loop containing the predicated instruction. |
4470 | auto *PredBB = PredInst->getParent(); |
4471 | auto *VectorLoop = LI->getLoopFor(PredBB); |
4472 | |
4473 | // Initialize a worklist with the operands of the predicated instruction. |
4474 | SetVector<Value *> Worklist(PredInst->op_begin(), PredInst->op_end()); |
4475 | |
4476 | // Holds instructions that we need to analyze again. An instruction may be |
4477 | // reanalyzed if we don't yet know if we can sink it or not. |
4478 | SmallVector<Instruction *, 8> InstsToReanalyze; |
4479 | |
4480 | // Returns true if a given use occurs in the predicated block. Phi nodes use |
4481 | // their operands in their corresponding predecessor blocks. |
4482 | auto isBlockOfUsePredicated = [&](Use &U) -> bool { |
4483 | auto *I = cast<Instruction>(U.getUser()); |
4484 | BasicBlock *BB = I->getParent(); |
4485 | if (auto *Phi = dyn_cast<PHINode>(I)) |
4486 | BB = Phi->getIncomingBlock( |
4487 | PHINode::getIncomingValueNumForOperand(U.getOperandNo())); |
4488 | return BB == PredBB; |
4489 | }; |
4490 | |
4491 | // Iteratively sink the scalarized operands of the predicated instruction |
4492 | // into the block we created for it. When an instruction is sunk, it's |
4493 | // operands are then added to the worklist. The algorithm ends after one pass |
4494 | // through the worklist doesn't sink a single instruction. |
4495 | bool Changed; |
4496 | do { |
4497 | // Add the instructions that need to be reanalyzed to the worklist, and |
4498 | // reset the changed indicator. |
4499 | Worklist.insert(InstsToReanalyze.begin(), InstsToReanalyze.end()); |
4500 | InstsToReanalyze.clear(); |
4501 | Changed = false; |
4502 | |
4503 | while (!Worklist.empty()) { |
4504 | auto *I = dyn_cast<Instruction>(Worklist.pop_back_val()); |
4505 | |
4506 | // We can't sink an instruction if it is a phi node, is already in the |
4507 | // predicated block, is not in the loop, or may have side effects. |
4508 | if (!I || isa<PHINode>(I) || I->getParent() == PredBB || |
4509 | !VectorLoop->contains(I) || I->mayHaveSideEffects()) |
4510 | continue; |
4511 | |
4512 | // It's legal to sink the instruction if all its uses occur in the |
4513 | // predicated block. Otherwise, there's nothing to do yet, and we may |
4514 | // need to reanalyze the instruction. |
4515 | if (!llvm::all_of(I->uses(), isBlockOfUsePredicated)) { |
4516 | InstsToReanalyze.push_back(I); |
4517 | continue; |
4518 | } |
4519 | |
4520 | // Move the instruction to the beginning of the predicated block, and add |
4521 | // it's operands to the worklist. |
4522 | I->moveBefore(&*PredBB->getFirstInsertionPt()); |
4523 | Worklist.insert(I->op_begin(), I->op_end()); |
4524 | |
4525 | // The sinking may have enabled other instructions to be sunk, so we will |
4526 | // need to iterate. |
4527 | Changed = true; |
4528 | } |
4529 | } while (Changed); |
4530 | } |
4531 | |
4532 | void InnerLoopVectorizer::fixNonInductionPHIs(VPTransformState &State) { |
4533 | for (PHINode *OrigPhi : OrigPHIsToFix) { |
4534 | VPWidenPHIRecipe *VPPhi = |
4535 | cast<VPWidenPHIRecipe>(State.Plan->getVPValue(OrigPhi)); |
4536 | PHINode *NewPhi = cast<PHINode>(State.get(VPPhi, 0)); |
4537 | // Make sure the builder has a valid insert point. |
4538 | Builder.SetInsertPoint(NewPhi); |
4539 | for (unsigned i = 0; i < VPPhi->getNumOperands(); ++i) { |
4540 | VPValue *Inc = VPPhi->getIncomingValue(i); |
4541 | VPBasicBlock *VPBB = VPPhi->getIncomingBlock(i); |
4542 | NewPhi->addIncoming(State.get(Inc, 0), State.CFG.VPBB2IRBB[VPBB]); |
4543 | } |
4544 | } |
4545 | } |
4546 | |
4547 | void InnerLoopVectorizer::widenGEP(GetElementPtrInst *GEP, VPValue *VPDef, |
4548 | VPUser &Operands, unsigned UF, |
4549 | ElementCount VF, bool IsPtrLoopInvariant, |
4550 | SmallBitVector &IsIndexLoopInvariant, |
4551 | VPTransformState &State) { |
4552 | // Construct a vector GEP by widening the operands of the scalar GEP as |
4553 | // necessary. We mark the vector GEP 'inbounds' if appropriate. A GEP |
4554 | // results in a vector of pointers when at least one operand of the GEP |
4555 | // is vector-typed. Thus, to keep the representation compact, we only use |
4556 | // vector-typed operands for loop-varying values. |
4557 | |
4558 | if (VF.isVector() && IsPtrLoopInvariant && IsIndexLoopInvariant.all()) { |
4559 | // If we are vectorizing, but the GEP has only loop-invariant operands, |
4560 | // the GEP we build (by only using vector-typed operands for |
4561 | // loop-varying values) would be a scalar pointer. Thus, to ensure we |
4562 | // produce a vector of pointers, we need to either arbitrarily pick an |
4563 | // operand to broadcast, or broadcast a clone of the original GEP. |
4564 | // Here, we broadcast a clone of the original. |
4565 | // |
4566 | // TODO: If at some point we decide to scalarize instructions having |
4567 | // loop-invariant operands, this special case will no longer be |
4568 | // required. We would add the scalarization decision to |
4569 | // collectLoopScalars() and teach getVectorValue() to broadcast |
4570 | // the lane-zero scalar value. |
4571 | auto *Clone = Builder.Insert(GEP->clone()); |
4572 | for (unsigned Part = 0; Part < UF; ++Part) { |
4573 | Value *EntryPart = Builder.CreateVectorSplat(VF, Clone); |
4574 | State.set(VPDef, EntryPart, Part); |
4575 | addMetadata(EntryPart, GEP); |
4576 | } |
4577 | } else { |
4578 | // If the GEP has at least one loop-varying operand, we are sure to |
4579 | // produce a vector of pointers. But if we are only unrolling, we want |
4580 | // to produce a scalar GEP for each unroll part. Thus, the GEP we |
4581 | // produce with the code below will be scalar (if VF == 1) or vector |
4582 | // (otherwise). Note that for the unroll-only case, we still maintain |
4583 | // values in the vector mapping with initVector, as we do for other |
4584 | // instructions. |
4585 | for (unsigned Part = 0; Part < UF; ++Part) { |
4586 | // The pointer operand of the new GEP. If it's loop-invariant, we |
4587 | // won't broadcast it. |
4588 | auto *Ptr = IsPtrLoopInvariant |
4589 | ? State.get(Operands.getOperand(0), VPIteration(0, 0)) |
4590 | : State.get(Operands.getOperand(0), Part); |
4591 | |
4592 | // Collect all the indices for the new GEP. If any index is |
4593 | // loop-invariant, we won't broadcast it. |
4594 | SmallVector<Value *, 4> Indices; |
4595 | for (unsigned I = 1, E = Operands.getNumOperands(); I < E; I++) { |
4596 | VPValue *Operand = Operands.getOperand(I); |
4597 | if (IsIndexLoopInvariant[I - 1]) |
4598 | Indices.push_back(State.get(Operand, VPIteration(0, 0))); |
4599 | else |
4600 | Indices.push_back(State.get(Operand, Part)); |
4601 | } |
4602 | |
4603 | // Create the new GEP. Note that this GEP may be a scalar if VF == 1, |
4604 | // but it should be a vector, otherwise. |
4605 | auto *NewGEP = |
4606 | GEP->isInBounds() |
4607 | ? Builder.CreateInBoundsGEP(GEP->getSourceElementType(), Ptr, |
4608 | Indices) |
4609 | : Builder.CreateGEP(GEP->getSourceElementType(), Ptr, Indices); |
4610 | assert((VF.isScalar() || NewGEP->getType()->isVectorTy()) &&(((VF.isScalar() || NewGEP->getType()->isVectorTy()) && "NewGEP is not a pointer vector") ? static_cast<void> ( 0) : __assert_fail ("(VF.isScalar() || NewGEP->getType()->isVectorTy()) && \"NewGEP is not a pointer vector\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4611, __PRETTY_FUNCTION__)) |
4611 | "NewGEP is not a pointer vector")(((VF.isScalar() || NewGEP->getType()->isVectorTy()) && "NewGEP is not a pointer vector") ? static_cast<void> ( 0) : __assert_fail ("(VF.isScalar() || NewGEP->getType()->isVectorTy()) && \"NewGEP is not a pointer vector\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4611, __PRETTY_FUNCTION__)); |
4612 | State.set(VPDef, NewGEP, Part); |
4613 | addMetadata(NewGEP, GEP); |
4614 | } |
4615 | } |
4616 | } |
4617 | |
4618 | void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN, |
4619 | RecurrenceDescriptor *RdxDesc, |
4620 | VPValue *StartVPV, VPValue *Def, |
4621 | VPTransformState &State) { |
4622 | PHINode *P = cast<PHINode>(PN); |
4623 | if (EnableVPlanNativePath) { |
4624 | // Currently we enter here in the VPlan-native path for non-induction |
4625 | // PHIs where all control flow is uniform. We simply widen these PHIs. |
4626 | // Create a vector phi with no operands - the vector phi operands will be |
4627 | // set at the end of vector code generation. |
4628 | Type *VecTy = (State.VF.isScalar()) |
4629 | ? PN->getType() |
4630 | : VectorType::get(PN->getType(), State.VF); |
4631 | Value *VecPhi = Builder.CreatePHI(VecTy, PN->getNumOperands(), "vec.phi"); |
4632 | State.set(Def, VecPhi, 0); |
4633 | OrigPHIsToFix.push_back(P); |
4634 | |
4635 | return; |
4636 | } |
4637 | |
4638 | assert(PN->getParent() == OrigLoop->getHeader() &&((PN->getParent() == OrigLoop->getHeader() && "Non-header phis should have been handled elsewhere" ) ? static_cast<void> (0) : __assert_fail ("PN->getParent() == OrigLoop->getHeader() && \"Non-header phis should have been handled elsewhere\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4639, __PRETTY_FUNCTION__)) |
4639 | "Non-header phis should have been handled elsewhere")((PN->getParent() == OrigLoop->getHeader() && "Non-header phis should have been handled elsewhere" ) ? static_cast<void> (0) : __assert_fail ("PN->getParent() == OrigLoop->getHeader() && \"Non-header phis should have been handled elsewhere\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4639, __PRETTY_FUNCTION__)); |
4640 | |
4641 | Value *StartV = StartVPV ? StartVPV->getLiveInIRValue() : nullptr; |
4642 | // In order to support recurrences we need to be able to vectorize Phi nodes. |
4643 | // Phi nodes have cycles, so we need to vectorize them in two stages. This is |
4644 | // stage #1: We create a new vector PHI node with no incoming edges. We'll use |
4645 | // this value when we vectorize all of the instructions that use the PHI. |
4646 | if (RdxDesc || Legal->isFirstOrderRecurrence(P)) { |
4647 | Value *Iden = nullptr; |
4648 | bool ScalarPHI = |
4649 | (State.VF.isScalar()) || Cost->isInLoopReduction(cast<PHINode>(PN)); |
4650 | Type *VecTy = |
4651 | ScalarPHI ? PN->getType() : VectorType::get(PN->getType(), State.VF); |
4652 | |
4653 | if (RdxDesc) { |
4654 | assert(Legal->isReductionVariable(P) && StartV &&((Legal->isReductionVariable(P) && StartV && "RdxDesc should only be set for reduction variables; in that case " "a StartV is also required") ? static_cast<void> (0) : __assert_fail ("Legal->isReductionVariable(P) && StartV && \"RdxDesc should only be set for reduction variables; in that case \" \"a StartV is also required\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4656, __PRETTY_FUNCTION__)) |
4655 | "RdxDesc should only be set for reduction variables; in that case "((Legal->isReductionVariable(P) && StartV && "RdxDesc should only be set for reduction variables; in that case " "a StartV is also required") ? static_cast<void> (0) : __assert_fail ("Legal->isReductionVariable(P) && StartV && \"RdxDesc should only be set for reduction variables; in that case \" \"a StartV is also required\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4656, __PRETTY_FUNCTION__)) |
4656 | "a StartV is also required")((Legal->isReductionVariable(P) && StartV && "RdxDesc should only be set for reduction variables; in that case " "a StartV is also required") ? static_cast<void> (0) : __assert_fail ("Legal->isReductionVariable(P) && StartV && \"RdxDesc should only be set for reduction variables; in that case \" \"a StartV is also required\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4656, __PRETTY_FUNCTION__)); |
4657 | RecurKind RK = RdxDesc->getRecurrenceKind(); |
4658 | if (RecurrenceDescriptor::isMinMaxRecurrenceKind(RK)) { |
4659 | // MinMax reduction have the start value as their identify. |
4660 | if (ScalarPHI) { |
4661 | Iden = StartV; |
4662 | } else { |
4663 | IRBuilderBase::InsertPointGuard IPBuilder(Builder); |
4664 | Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); |
4665 | StartV = Iden = |
4666 | Builder.CreateVectorSplat(State.VF, StartV, "minmax.ident"); |
4667 | } |
4668 | } else { |
4669 | Constant *IdenC = RecurrenceDescriptor::getRecurrenceIdentity( |
4670 | RK, VecTy->getScalarType()); |
4671 | Iden = IdenC; |
4672 | |
4673 | if (!ScalarPHI) { |
4674 | Iden = ConstantVector::getSplat(State.VF, IdenC); |
4675 | IRBuilderBase::InsertPointGuard IPBuilder(Builder); |
4676 | Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); |
4677 | Constant *Zero = Builder.getInt32(0); |
4678 | StartV = Builder.CreateInsertElement(Iden, StartV, Zero); |
4679 | } |
4680 | } |
4681 | } |
4682 | |
4683 | for (unsigned Part = 0; Part < State.UF; ++Part) { |
4684 | // This is phase one of vectorizing PHIs. |
4685 | Value *EntryPart = PHINode::Create( |
4686 | VecTy, 2, "vec.phi", &*LoopVectorBody->getFirstInsertionPt()); |
4687 | State.set(Def, EntryPart, Part); |
4688 | if (StartV) { |
4689 | // Make sure to add the reduction start value only to the |
4690 | // first unroll part. |
4691 | Value *StartVal = (Part == 0) ? StartV : Iden; |
4692 | cast<PHINode>(EntryPart)->addIncoming(StartVal, LoopVectorPreHeader); |
4693 | } |
4694 | } |
4695 | return; |
4696 | } |
4697 | |
4698 | assert(!Legal->isReductionVariable(P) &&((!Legal->isReductionVariable(P) && "reductions should be handled above" ) ? static_cast<void> (0) : __assert_fail ("!Legal->isReductionVariable(P) && \"reductions should be handled above\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4699, __PRETTY_FUNCTION__)) |
4699 | "reductions should be handled above")((!Legal->isReductionVariable(P) && "reductions should be handled above" ) ? static_cast<void> (0) : __assert_fail ("!Legal->isReductionVariable(P) && \"reductions should be handled above\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4699, __PRETTY_FUNCTION__)); |
4700 | |
4701 | setDebugLocFromInst(Builder, P); |
4702 | |
4703 | // This PHINode must be an induction variable. |
4704 | // Make sure that we know about it. |
4705 | assert(Legal->getInductionVars().count(P) && "Not an induction variable")((Legal->getInductionVars().count(P) && "Not an induction variable" ) ? static_cast<void> (0) : __assert_fail ("Legal->getInductionVars().count(P) && \"Not an induction variable\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4705, __PRETTY_FUNCTION__)); |
4706 | |
4707 | InductionDescriptor II = Legal->getInductionVars().lookup(P); |
4708 | const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout(); |
4709 | |
4710 | // FIXME: The newly created binary instructions should contain nsw/nuw flags, |
4711 | // which can be found from the original scalar operations. |
4712 | switch (II.getKind()) { |
4713 | case InductionDescriptor::IK_NoInduction: |
4714 | llvm_unreachable("Unknown induction")::llvm::llvm_unreachable_internal("Unknown induction", "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4714); |
4715 | case InductionDescriptor::IK_IntInduction: |
4716 | case InductionDescriptor::IK_FpInduction: |
4717 | llvm_unreachable("Integer/fp induction is handled elsewhere.")::llvm::llvm_unreachable_internal("Integer/fp induction is handled elsewhere." , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4717); |
4718 | case InductionDescriptor::IK_PtrInduction: { |
4719 | // Handle the pointer induction variable case. |
4720 | assert(P->getType()->isPointerTy() && "Unexpected type.")((P->getType()->isPointerTy() && "Unexpected type." ) ? static_cast<void> (0) : __assert_fail ("P->getType()->isPointerTy() && \"Unexpected type.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4720, __PRETTY_FUNCTION__)); |
4721 | |
4722 | if (Cost->isScalarAfterVectorization(P, State.VF)) { |
4723 | // This is the normalized GEP that starts counting at zero. |
4724 | Value *PtrInd = |
4725 | Builder.CreateSExtOrTrunc(Induction, II.getStep()->getType()); |
4726 | // Determine the number of scalars we need to generate for each unroll |
4727 | // iteration. If the instruction is uniform, we only need to generate the |
4728 | // first lane. Otherwise, we generate all VF values. |
4729 | unsigned Lanes = Cost->isUniformAfterVectorization(P, State.VF) |
4730 | ? 1 |
4731 | : State.VF.getKnownMinValue(); |
4732 | for (unsigned Part = 0; Part < UF; ++Part) { |
4733 | for (unsigned Lane = 0; Lane < Lanes; ++Lane) { |
4734 | Constant *Idx = ConstantInt::get( |
4735 | PtrInd->getType(), Lane + Part * State.VF.getKnownMinValue()); |
4736 | Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx); |
4737 | Value *SclrGep = |
4738 | emitTransformedIndex(Builder, GlobalIdx, PSE.getSE(), DL, II); |
4739 | SclrGep->setName("next.gep"); |
4740 | State.set(Def, SclrGep, VPIteration(Part, Lane)); |
4741 | } |
4742 | } |
4743 | return; |
4744 | } |
4745 | assert(isa<SCEVConstant>(II.getStep()) &&((isa<SCEVConstant>(II.getStep()) && "Induction step not a SCEV constant!" ) ? static_cast<void> (0) : __assert_fail ("isa<SCEVConstant>(II.getStep()) && \"Induction step not a SCEV constant!\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4746, __PRETTY_FUNCTION__)) |
4746 | "Induction step not a SCEV constant!")((isa<SCEVConstant>(II.getStep()) && "Induction step not a SCEV constant!" ) ? static_cast<void> (0) : __assert_fail ("isa<SCEVConstant>(II.getStep()) && \"Induction step not a SCEV constant!\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4746, __PRETTY_FUNCTION__)); |
4747 | Type *PhiType = II.getStep()->getType(); |
4748 | |
4749 | // Build a pointer phi |
4750 | Value *ScalarStartValue = II.getStartValue(); |
4751 | Type *ScStValueType = ScalarStartValue->getType(); |
4752 | PHINode *NewPointerPhi = |
4753 | PHINode::Create(ScStValueType, 2, "pointer.phi", Induction); |
4754 | NewPointerPhi->addIncoming(ScalarStartValue, LoopVectorPreHeader); |
4755 | |
4756 | // A pointer induction, performed by using a gep |
4757 | BasicBlock *LoopLatch = LI->getLoopFor(LoopVectorBody)->getLoopLatch(); |
4758 | Instruction *InductionLoc = LoopLatch->getTerminator(); |
4759 | const SCEV *ScalarStep = II.getStep(); |
4760 | SCEVExpander Exp(*PSE.getSE(), DL, "induction"); |
4761 | Value *ScalarStepValue = |
4762 | Exp.expandCodeFor(ScalarStep, PhiType, InductionLoc); |
4763 | Value *InductionGEP = GetElementPtrInst::Create( |
4764 | ScStValueType->getPointerElementType(), NewPointerPhi, |
4765 | Builder.CreateMul( |
4766 | ScalarStepValue, |
4767 | ConstantInt::get(PhiType, State.VF.getKnownMinValue() * State.UF)), |
4768 | "ptr.ind", InductionLoc); |
4769 | NewPointerPhi->addIncoming(InductionGEP, LoopLatch); |
4770 | |
4771 | // Create UF many actual address geps that use the pointer |
4772 | // phi as base and a vectorized version of the step value |
4773 | // (<step*0, ..., step*N>) as offset. |
4774 | for (unsigned Part = 0; Part < State.UF; ++Part) { |
4775 | SmallVector<Constant *, 8> Indices; |
4776 | // Create a vector of consecutive numbers from zero to VF. |
4777 | for (unsigned i = 0; i < State.VF.getKnownMinValue(); ++i) |
4778 | Indices.push_back( |
4779 | ConstantInt::get(PhiType, i + Part * State.VF.getKnownMinValue())); |
4780 | Constant *StartOffset = ConstantVector::get(Indices); |
4781 | |
4782 | Value *GEP = Builder.CreateGEP( |
4783 | ScStValueType->getPointerElementType(), NewPointerPhi, |
4784 | Builder.CreateMul(StartOffset, |
4785 | Builder.CreateVectorSplat( |
4786 | State.VF.getKnownMinValue(), ScalarStepValue), |
4787 | "vector.gep")); |
4788 | State.set(Def, GEP, Part); |
4789 | } |
4790 | } |
4791 | } |
4792 | } |
4793 | |
4794 | /// A helper function for checking whether an integer division-related |
4795 | /// instruction may divide by zero (in which case it must be predicated if |
4796 | /// executed conditionally in the scalar code). |
4797 | /// TODO: It may be worthwhile to generalize and check isKnownNonZero(). |
4798 | /// Non-zero divisors that are non compile-time constants will not be |
4799 | /// converted into multiplication, so we will still end up scalarizing |
4800 | /// the division, but can do so w/o predication. |
4801 | static bool mayDivideByZero(Instruction &I) { |
4802 | assert((I.getOpcode() == Instruction::UDiv ||(((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction ::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && "Unexpected instruction") ? static_cast<void> (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4806, __PRETTY_FUNCTION__)) |
4803 | I.getOpcode() == Instruction::SDiv ||(((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction ::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && "Unexpected instruction") ? static_cast<void> (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4806, __PRETTY_FUNCTION__)) |
4804 | I.getOpcode() == Instruction::URem ||(((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction ::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && "Unexpected instruction") ? static_cast<void> (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4806, __PRETTY_FUNCTION__)) |
4805 | I.getOpcode() == Instruction::SRem) &&(((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction ::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && "Unexpected instruction") ? static_cast<void> (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4806, __PRETTY_FUNCTION__)) |
4806 | "Unexpected instruction")(((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction ::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && "Unexpected instruction") ? static_cast<void> (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4806, __PRETTY_FUNCTION__)); |
4807 | Value *Divisor = I.getOperand(1); |
4808 | auto *CInt = dyn_cast<ConstantInt>(Divisor); |
4809 | return !CInt || CInt->isZero(); |
4810 | } |
4811 | |
4812 | void InnerLoopVectorizer::widenInstruction(Instruction &I, VPValue *Def, |
4813 | VPUser &User, |
4814 | VPTransformState &State) { |
4815 | switch (I.getOpcode()) { |
4816 | case Instruction::Call: |
4817 | case Instruction::Br: |
4818 | case Instruction::PHI: |
4819 | case Instruction::GetElementPtr: |
4820 | case Instruction::Select: |
4821 | llvm_unreachable("This instruction is handled by a different recipe.")::llvm::llvm_unreachable_internal("This instruction is handled by a different recipe." , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4821); |
4822 | case Instruction::UDiv: |
4823 | case Instruction::SDiv: |
4824 | case Instruction::SRem: |
4825 | case Instruction::URem: |
4826 | case Instruction::Add: |
4827 | case Instruction::FAdd: |
4828 | case Instruction::Sub: |
4829 | case Instruction::FSub: |
4830 | case Instruction::FNeg: |
4831 | case Instruction::Mul: |
4832 | case Instruction::FMul: |
4833 | case Instruction::FDiv: |
4834 | case Instruction::FRem: |
4835 | case Instruction::Shl: |
4836 | case Instruction::LShr: |
4837 | case Instruction::AShr: |
4838 | case Instruction::And: |
4839 | case Instruction::Or: |
4840 | case Instruction::Xor: { |
4841 | // Just widen unops and binops. |
4842 | setDebugLocFromInst(Builder, &I); |
4843 | |
4844 | for (unsigned Part = 0; Part < UF; ++Part) { |
4845 | SmallVector<Value *, 2> Ops; |
4846 | for (VPValue *VPOp : User.operands()) |
4847 | Ops.push_back(State.get(VPOp, Part)); |
4848 | |
4849 | Value *V = Builder.CreateNAryOp(I.getOpcode(), Ops); |
4850 | |
4851 | if (auto *VecOp = dyn_cast<Instruction>(V)) |
4852 | VecOp->copyIRFlags(&I); |
4853 | |
4854 | // Use this vector value for all users of the original instruction. |
4855 | State.set(Def, V, Part); |
4856 | addMetadata(V, &I); |
4857 | } |
4858 | |
4859 | break; |
4860 | } |
4861 | case Instruction::ICmp: |
4862 | case Instruction::FCmp: { |
4863 | // Widen compares. Generate vector compares. |
4864 | bool FCmp = (I.getOpcode() == Instruction::FCmp); |
4865 | auto *Cmp = cast<CmpInst>(&I); |
4866 | setDebugLocFromInst(Builder, Cmp); |
4867 | for (unsigned Part = 0; Part < UF; ++Part) { |
4868 | Value *A = State.get(User.getOperand(0), Part); |
4869 | Value *B = State.get(User.getOperand(1), Part); |
4870 | Value *C = nullptr; |
4871 | if (FCmp) { |
4872 | // Propagate fast math flags. |
4873 | IRBuilder<>::FastMathFlagGuard FMFG(Builder); |
4874 | Builder.setFastMathFlags(Cmp->getFastMathFlags()); |
4875 | C = Builder.CreateFCmp(Cmp->getPredicate(), A, B); |
4876 | } else { |
4877 | C = Builder.CreateICmp(Cmp->getPredicate(), A, B); |
4878 | } |
4879 | State.set(Def, C, Part); |
4880 | addMetadata(C, &I); |
4881 | } |
4882 | |
4883 | break; |
4884 | } |
4885 | |
4886 | case Instruction::ZExt: |
4887 | case Instruction::SExt: |
4888 | case Instruction::FPToUI: |
4889 | case Instruction::FPToSI: |
4890 | case Instruction::FPExt: |
4891 | case Instruction::PtrToInt: |
4892 | case Instruction::IntToPtr: |
4893 | case Instruction::SIToFP: |
4894 | case Instruction::UIToFP: |
4895 | case Instruction::Trunc: |
4896 | case Instruction::FPTrunc: |
4897 | case Instruction::BitCast: { |
4898 | auto *CI = cast<CastInst>(&I); |
4899 | setDebugLocFromInst(Builder, CI); |
4900 | |
4901 | /// Vectorize casts. |
4902 | Type *DestTy = |
4903 | (VF.isScalar()) ? CI->getType() : VectorType::get(CI->getType(), VF); |
4904 | |
4905 | for (unsigned Part = 0; Part < UF; ++Part) { |
4906 | Value *A = State.get(User.getOperand(0), Part); |
4907 | Value *Cast = Builder.CreateCast(CI->getOpcode(), A, DestTy); |
4908 | State.set(Def, Cast, Part); |
4909 | addMetadata(Cast, &I); |
4910 | } |
4911 | break; |
4912 | } |
4913 | default: |
4914 | // This instruction is not vectorized by simple widening. |
4915 | LLVM_DEBUG(dbgs() << "LV: Found an unhandled instruction: " << I)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found an unhandled instruction: " << I; } } while (false); |
4916 | llvm_unreachable("Unhandled instruction!")::llvm::llvm_unreachable_internal("Unhandled instruction!", "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4916); |
4917 | } // end of switch. |
4918 | } |
4919 | |
4920 | void InnerLoopVectorizer::widenCallInstruction(CallInst &I, VPValue *Def, |
4921 | VPUser &ArgOperands, |
4922 | VPTransformState &State) { |
4923 | assert(!isa<DbgInfoIntrinsic>(I) &&((!isa<DbgInfoIntrinsic>(I) && "DbgInfoIntrinsic should have been dropped during VPlan construction" ) ? static_cast<void> (0) : __assert_fail ("!isa<DbgInfoIntrinsic>(I) && \"DbgInfoIntrinsic should have been dropped during VPlan construction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4924, __PRETTY_FUNCTION__)) |
4924 | "DbgInfoIntrinsic should have been dropped during VPlan construction")((!isa<DbgInfoIntrinsic>(I) && "DbgInfoIntrinsic should have been dropped during VPlan construction" ) ? static_cast<void> (0) : __assert_fail ("!isa<DbgInfoIntrinsic>(I) && \"DbgInfoIntrinsic should have been dropped during VPlan construction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4924, __PRETTY_FUNCTION__)); |
4925 | setDebugLocFromInst(Builder, &I); |
4926 | |
4927 | Module *M = I.getParent()->getParent()->getParent(); |
4928 | auto *CI = cast<CallInst>(&I); |
4929 | |
4930 | SmallVector<Type *, 4> Tys; |
4931 | for (Value *ArgOperand : CI->arg_operands()) |
4932 | Tys.push_back(ToVectorTy(ArgOperand->getType(), VF.getKnownMinValue())); |
4933 | |
4934 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); |
4935 | |
4936 | // The flag shows whether we use Intrinsic or a usual Call for vectorized |
4937 | // version of the instruction. |
4938 | // Is it beneficial to perform intrinsic call compared to lib call? |
4939 | bool NeedToScalarize = false; |
4940 | InstructionCost CallCost = Cost->getVectorCallCost(CI, VF, NeedToScalarize); |
4941 | InstructionCost IntrinsicCost = ID ? Cost->getVectorIntrinsicCost(CI, VF) : 0; |
4942 | bool UseVectorIntrinsic = ID && IntrinsicCost <= CallCost; |
4943 | assert((UseVectorIntrinsic || !NeedToScalarize) &&(((UseVectorIntrinsic || !NeedToScalarize) && "Instruction should be scalarized elsewhere." ) ? static_cast<void> (0) : __assert_fail ("(UseVectorIntrinsic || !NeedToScalarize) && \"Instruction should be scalarized elsewhere.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4944, __PRETTY_FUNCTION__)) |
4944 | "Instruction should be scalarized elsewhere.")(((UseVectorIntrinsic || !NeedToScalarize) && "Instruction should be scalarized elsewhere." ) ? static_cast<void> (0) : __assert_fail ("(UseVectorIntrinsic || !NeedToScalarize) && \"Instruction should be scalarized elsewhere.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4944, __PRETTY_FUNCTION__)); |
4945 | assert(IntrinsicCost.isValid() && CallCost.isValid() &&((IntrinsicCost.isValid() && CallCost.isValid() && "Cannot have invalid costs while widening") ? static_cast< void> (0) : __assert_fail ("IntrinsicCost.isValid() && CallCost.isValid() && \"Cannot have invalid costs while widening\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4946, __PRETTY_FUNCTION__)) |
4946 | "Cannot have invalid costs while widening")((IntrinsicCost.isValid() && CallCost.isValid() && "Cannot have invalid costs while widening") ? static_cast< void> (0) : __assert_fail ("IntrinsicCost.isValid() && CallCost.isValid() && \"Cannot have invalid costs while widening\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4946, __PRETTY_FUNCTION__)); |
4947 | |
4948 | for (unsigned Part = 0; Part < UF; ++Part) { |
4949 | SmallVector<Value *, 4> Args; |
4950 | for (auto &I : enumerate(ArgOperands.operands())) { |
4951 | // Some intrinsics have a scalar argument - don't replace it with a |
4952 | // vector. |
4953 | Value *Arg; |
4954 | if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, I.index())) |
4955 | Arg = State.get(I.value(), Part); |
4956 | else |
4957 | Arg = State.get(I.value(), VPIteration(0, 0)); |
4958 | Args.push_back(Arg); |
4959 | } |
4960 | |
4961 | Function *VectorF; |
4962 | if (UseVectorIntrinsic) { |
4963 | // Use vector version of the intrinsic. |
4964 | Type *TysForDecl[] = {CI->getType()}; |
4965 | if (VF.isVector()) |
4966 | TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF); |
4967 | VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl); |
4968 | assert(VectorF && "Can't retrieve vector intrinsic.")((VectorF && "Can't retrieve vector intrinsic.") ? static_cast <void> (0) : __assert_fail ("VectorF && \"Can't retrieve vector intrinsic.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4968, __PRETTY_FUNCTION__)); |
4969 | } else { |
4970 | // Use vector version of the function call. |
4971 | const VFShape Shape = VFShape::get(*CI, VF, false /*HasGlobalPred*/); |
4972 | #ifndef NDEBUG |
4973 | assert(VFDatabase(*CI).getVectorizedFunction(Shape) != nullptr &&((VFDatabase(*CI).getVectorizedFunction(Shape) != nullptr && "Can't create vector function.") ? static_cast<void> ( 0) : __assert_fail ("VFDatabase(*CI).getVectorizedFunction(Shape) != nullptr && \"Can't create vector function.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4974, __PRETTY_FUNCTION__)) |
4974 | "Can't create vector function.")((VFDatabase(*CI).getVectorizedFunction(Shape) != nullptr && "Can't create vector function.") ? static_cast<void> ( 0) : __assert_fail ("VFDatabase(*CI).getVectorizedFunction(Shape) != nullptr && \"Can't create vector function.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4974, __PRETTY_FUNCTION__)); |
4975 | #endif |
4976 | VectorF = VFDatabase(*CI).getVectorizedFunction(Shape); |
4977 | } |
4978 | SmallVector<OperandBundleDef, 1> OpBundles; |
4979 | CI->getOperandBundlesAsDefs(OpBundles); |
4980 | CallInst *V = Builder.CreateCall(VectorF, Args, OpBundles); |
4981 | |
4982 | if (isa<FPMathOperator>(V)) |
4983 | V->copyFastMathFlags(CI); |
4984 | |
4985 | State.set(Def, V, Part); |
4986 | addMetadata(V, &I); |
4987 | } |
4988 | } |
4989 | |
4990 | void InnerLoopVectorizer::widenSelectInstruction(SelectInst &I, VPValue *VPDef, |
4991 | VPUser &Operands, |
4992 | bool InvariantCond, |
4993 | VPTransformState &State) { |
4994 | setDebugLocFromInst(Builder, &I); |
4995 | |
4996 | // The condition can be loop invariant but still defined inside the |
4997 | // loop. This means that we can't just use the original 'cond' value. |
4998 | // We have to take the 'vectorized' value and pick the first lane. |
4999 | // Instcombine will make this a no-op. |
5000 | auto *InvarCond = InvariantCond |
5001 | ? State.get(Operands.getOperand(0), VPIteration(0, 0)) |
5002 | : nullptr; |
5003 | |
5004 | for (unsigned Part = 0; Part < UF; ++Part) { |
5005 | Value *Cond = |
5006 | InvarCond ? InvarCond : State.get(Operands.getOperand(0), Part); |
5007 | Value *Op0 = State.get(Operands.getOperand(1), Part); |
5008 | Value *Op1 = State.get(Operands.getOperand(2), Part); |
5009 | Value *Sel = Builder.CreateSelect(Cond, Op0, Op1); |
5010 | State.set(VPDef, Sel, Part); |
5011 | addMetadata(Sel, &I); |
5012 | } |
5013 | } |
5014 | |
5015 | void LoopVectorizationCostModel::collectLoopScalars(ElementCount VF) { |
5016 | // We should not collect Scalars more than once per VF. Right now, this |
5017 | // function is called from collectUniformsAndScalars(), which already does |
5018 | // this check. Collecting Scalars for VF=1 does not make any sense. |
5019 | assert(VF.isVector() && Scalars.find(VF) == Scalars.end() &&((VF.isVector() && Scalars.find(VF) == Scalars.end() && "This function should not be visited twice for the same VF") ? static_cast<void> (0) : __assert_fail ("VF.isVector() && Scalars.find(VF) == Scalars.end() && \"This function should not be visited twice for the same VF\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5020, __PRETTY_FUNCTION__)) |
5020 | "This function should not be visited twice for the same VF")((VF.isVector() && Scalars.find(VF) == Scalars.end() && "This function should not be visited twice for the same VF") ? static_cast<void> (0) : __assert_fail ("VF.isVector() && Scalars.find(VF) == Scalars.end() && \"This function should not be visited twice for the same VF\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5020, __PRETTY_FUNCTION__)); |
5021 | |
5022 | SmallSetVector<Instruction *, 8> Worklist; |
5023 | |
5024 | // These sets are used to seed the analysis with pointers used by memory |
5025 | // accesses that will remain scalar. |
5026 | SmallSetVector<Instruction *, 8> ScalarPtrs; |
5027 | SmallPtrSet<Instruction *, 8> PossibleNonScalarPtrs; |
5028 | auto *Latch = TheLoop->getLoopLatch(); |
5029 | |
5030 | // A helper that returns true if the use of Ptr by MemAccess will be scalar. |
5031 | // The pointer operands of loads and stores will be scalar as long as the |
5032 | // memory access is not a gather or scatter operation. The value operand of a |
5033 | // store will remain scalar if the store is scalarized. |
5034 | auto isScalarUse = [&](Instruction *MemAccess, Value *Ptr) { |
5035 | InstWidening WideningDecision = getWideningDecision(MemAccess, VF); |
5036 | assert(WideningDecision != CM_Unknown &&((WideningDecision != CM_Unknown && "Widening decision should be ready at this moment" ) ? static_cast<void> (0) : __assert_fail ("WideningDecision != CM_Unknown && \"Widening decision should be ready at this moment\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5037, __PRETTY_FUNCTION__)) |
5037 | "Widening decision should be ready at this moment")((WideningDecision != CM_Unknown && "Widening decision should be ready at this moment" ) ? static_cast<void> (0) : __assert_fail ("WideningDecision != CM_Unknown && \"Widening decision should be ready at this moment\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5037, __PRETTY_FUNCTION__)); |
5038 | if (auto *Store = dyn_cast<StoreInst>(MemAccess)) |
5039 | if (Ptr == Store->getValueOperand()) |
5040 | return WideningDecision == CM_Scalarize; |
5041 | assert(Ptr == getLoadStorePointerOperand(MemAccess) &&((Ptr == getLoadStorePointerOperand(MemAccess) && "Ptr is neither a value or pointer operand" ) ? static_cast<void> (0) : __assert_fail ("Ptr == getLoadStorePointerOperand(MemAccess) && \"Ptr is neither a value or pointer operand\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5042, __PRETTY_FUNCTION__)) |
5042 | "Ptr is neither a value or pointer operand")((Ptr == getLoadStorePointerOperand(MemAccess) && "Ptr is neither a value or pointer operand" ) ? static_cast<void> (0) : __assert_fail ("Ptr == getLoadStorePointerOperand(MemAccess) && \"Ptr is neither a value or pointer operand\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5042, __PRETTY_FUNCTION__)); |
5043 | return WideningDecision != CM_GatherScatter; |
5044 | }; |
5045 | |
5046 | // A helper that returns true if the given value is a bitcast or |
5047 | // getelementptr instruction contained in the loop. |
5048 | auto isLoopVaryingBitCastOrGEP = [&](Value *V) { |
5049 | return ((isa<BitCastInst>(V) && V->getType()->isPointerTy()) || |
5050 | isa<GetElementPtrInst>(V)) && |
5051 | !TheLoop->isLoopInvariant(V); |
5052 | }; |
5053 | |
5054 | auto isScalarPtrInduction = [&](Instruction *MemAccess, Value *Ptr) { |
5055 | if (!isa<PHINode>(Ptr) || |
5056 | !Legal->getInductionVars().count(cast<PHINode>(Ptr))) |
5057 | return false; |
5058 | auto &Induction = Legal->getInductionVars()[cast<PHINode>(Ptr)]; |
5059 | if (Induction.getKind() != InductionDescriptor::IK_PtrInduction) |
5060 | return false; |
5061 | return isScalarUse(MemAccess, Ptr); |
5062 | }; |
5063 | |
5064 | // A helper that evaluates a memory access's use of a pointer. If the |
5065 | // pointer is actually the pointer induction of a loop, it is being |
5066 | // inserted into Worklist. If the use will be a scalar use, and the |
5067 | // pointer is only used by memory accesses, we place the pointer in |
5068 | // ScalarPtrs. Otherwise, the pointer is placed in PossibleNonScalarPtrs. |
5069 | auto evaluatePtrUse = [&](Instruction *MemAccess, Value *Ptr) { |
5070 | if (isScalarPtrInduction(MemAccess, Ptr)) { |
5071 | Worklist.insert(cast<Instruction>(Ptr)); |
5072 | Instruction *Update = cast<Instruction>( |
5073 | cast<PHINode>(Ptr)->getIncomingValueForBlock(Latch)); |
5074 | Worklist.insert(Update); |
5075 | LLVM_DEBUG(dbgs() << "LV: Found new scalar instruction: " << *Ptrdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found new scalar instruction: " << *Ptr << "\n"; } } while (false) |
5076 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found new scalar instruction: " << *Ptr << "\n"; } } while (false); |
5077 | LLVM_DEBUG(dbgs() << "LV: Found new scalar instruction: " << *Updatedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found new scalar instruction: " << *Update << "\n"; } } while (false) |
5078 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found new scalar instruction: " << *Update << "\n"; } } while (false); |
5079 | return; |
5080 | } |
5081 | // We only care about bitcast and getelementptr instructions contained in |
5082 | // the loop. |
5083 | if (!isLoopVaryingBitCastOrGEP(Ptr)) |
5084 | return; |
5085 | |
5086 | // If the pointer has already been identified as scalar (e.g., if it was |
5087 | // also identified as uniform), there's nothing to do. |
5088 | auto *I = cast<Instruction>(Ptr); |
5089 | if (Worklist.count(I)) |
5090 | return; |
5091 | |
5092 | // If the use of the pointer will be a scalar use, and all users of the |
5093 | // pointer are memory accesses, place the pointer in ScalarPtrs. Otherwise, |
5094 | // place the pointer in PossibleNonScalarPtrs. |
5095 | if (isScalarUse(MemAccess, Ptr) && llvm::all_of(I->users(), [&](User *U) { |
5096 | return isa<LoadInst>(U) || isa<StoreInst>(U); |
5097 | })) |
5098 | ScalarPtrs.insert(I); |
5099 | else |
5100 | PossibleNonScalarPtrs.insert(I); |
5101 | }; |
5102 | |
5103 | // We seed the scalars analysis with three classes of instructions: (1) |
5104 | // instructions marked uniform-after-vectorization and (2) bitcast, |
5105 | // getelementptr and (pointer) phi instructions used by memory accesses |
5106 | // requiring a scalar use. |
5107 | // |
5108 | // (1) Add to the worklist all instructions that have been identified as |
5109 | // uniform-after-vectorization. |
5110 | Worklist.insert(Uniforms[VF].begin(), Uniforms[VF].end()); |
5111 | |
5112 | // (2) Add to the worklist all bitcast and getelementptr instructions used by |
5113 | // memory accesses requiring a scalar use. The pointer operands of loads and |
5114 | // stores will be scalar as long as the memory accesses is not a gather or |
5115 | // scatter operation. The value operand of a store will remain scalar if the |
5116 | // store is scalarized. |
5117 | for (auto *BB : TheLoop->blocks()) |
5118 | for (auto &I : *BB) { |
5119 | if (auto *Load = dyn_cast<LoadInst>(&I)) { |
5120 | evaluatePtrUse(Load, Load->getPointerOperand()); |
5121 | } else if (auto *Store = dyn_cast<StoreInst>(&I)) { |
5122 | evaluatePtrUse(Store, Store->getPointerOperand()); |
5123 | evaluatePtrUse(Store, Store->getValueOperand()); |
5124 | } |
5125 | } |
5126 | for (auto *I : ScalarPtrs) |
5127 | if (!PossibleNonScalarPtrs.count(I)) { |
5128 | LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: " << *I << "\n"; } } while (false); |
5129 | Worklist.insert(I); |
5130 | } |
5131 | |
5132 | // Insert the forced scalars. |
5133 | // FIXME: Currently widenPHIInstruction() often creates a dead vector |
5134 | // induction variable when the PHI user is scalarized. |
5135 | auto ForcedScalar = ForcedScalars.find(VF); |
5136 | if (ForcedScalar != ForcedScalars.end()) |
5137 | for (auto *I : ForcedScalar->second) |
5138 | Worklist.insert(I); |
5139 | |
5140 | // Expand the worklist by looking through any bitcasts and getelementptr |
5141 | // instructions we've already identified as scalar. This is similar to the |
5142 | // expansion step in collectLoopUniforms(); however, here we're only |
5143 | // expanding to include additional bitcasts and getelementptr instructions. |
5144 | unsigned Idx = 0; |
5145 | while (Idx != Worklist.size()) { |
5146 | Instruction *Dst = Worklist[Idx++]; |
5147 | if (!isLoopVaryingBitCastOrGEP(Dst->getOperand(0))) |
5148 | continue; |
5149 | auto *Src = cast<Instruction>(Dst->getOperand(0)); |
5150 | if (llvm::all_of(Src->users(), [&](User *U) -> bool { |
5151 | auto *J = cast<Instruction>(U); |
5152 | return !TheLoop->contains(J) || Worklist.count(J) || |
5153 | ((isa<LoadInst>(J) || isa<StoreInst>(J)) && |
5154 | isScalarUse(J, Src)); |
5155 | })) { |
5156 | Worklist.insert(Src); |
5157 | LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *Src << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: " << *Src << "\n"; } } while (false); |
5158 | } |
5159 | } |
5160 | |
5161 | // An induction variable will remain scalar if all users of the induction |
5162 | // variable and induction variable update remain scalar. |
5163 | for (auto &Induction : Legal->getInductionVars()) { |
5164 | auto *Ind = Induction.first; |
5165 | auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch)); |
5166 | |
5167 | // If tail-folding is applied, the primary induction variable will be used |
5168 | // to feed a vector compare. |
5169 | if (Ind == Legal->getPrimaryInduction() && foldTailByMasking()) |
5170 | continue; |
5171 | |
5172 | // Determine if all users of the induction variable are scalar after |
5173 | // vectorization. |
5174 | auto ScalarInd = llvm::all_of(Ind->users(), [&](User *U) -> bool { |
5175 | auto *I = cast<Instruction>(U); |
5176 | return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I); |
5177 | }); |
5178 | if (!ScalarInd) |
5179 | continue; |
5180 | |
5181 | // Determine if all users of the induction variable update instruction are |
5182 | // scalar after vectorization. |
5183 | auto ScalarIndUpdate = |
5184 | llvm::all_of(IndUpdate->users(), [&](User *U) -> bool { |
5185 | auto *I = cast<Instruction>(U); |
5186 | return I == Ind || !TheLoop->contains(I) || Worklist.count(I); |
5187 | }); |
5188 | if (!ScalarIndUpdate) |
5189 | continue; |
5190 | |
5191 | // The induction variable and its update instruction will remain scalar. |
5192 | Worklist.insert(Ind); |
5193 | Worklist.insert(IndUpdate); |
5194 | LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *Ind << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: " << *Ind << "\n"; } } while (false); |
5195 | LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *IndUpdatedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: " << *IndUpdate << "\n"; } } while (false) |
5196 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: " << *IndUpdate << "\n"; } } while (false); |
5197 | } |
5198 | |
5199 | Scalars[VF].insert(Worklist.begin(), Worklist.end()); |
5200 | } |
5201 | |
5202 | bool LoopVectorizationCostModel::isScalarWithPredication(Instruction *I, |
5203 | ElementCount VF) { |
5204 | if (!blockNeedsPredication(I->getParent())) |
5205 | return false; |
5206 | switch(I->getOpcode()) { |
5207 | default: |
5208 | break; |
5209 | case Instruction::Load: |
5210 | case Instruction::Store: { |
5211 | if (!Legal->isMaskRequired(I)) |
5212 | return false; |
5213 | auto *Ptr = getLoadStorePointerOperand(I); |
5214 | auto *Ty = getMemInstValueType(I); |
5215 | // We have already decided how to vectorize this instruction, get that |
5216 | // result. |
5217 | if (VF.isVector()) { |
5218 | InstWidening WideningDecision = getWideningDecision(I, VF); |
5219 | assert(WideningDecision != CM_Unknown &&((WideningDecision != CM_Unknown && "Widening decision should be ready at this moment" ) ? static_cast<void> (0) : __assert_fail ("WideningDecision != CM_Unknown && \"Widening decision should be ready at this moment\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5220, __PRETTY_FUNCTION__)) |
5220 | "Widening decision should be ready at this moment")((WideningDecision != CM_Unknown && "Widening decision should be ready at this moment" ) ? static_cast<void> (0) : __assert_fail ("WideningDecision != CM_Unknown && \"Widening decision should be ready at this moment\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5220, __PRETTY_FUNCTION__)); |
5221 | return WideningDecision == CM_Scalarize; |
5222 | } |
5223 | const Align Alignment = getLoadStoreAlignment(I); |
5224 | return isa<LoadInst>(I) ? !(isLegalMaskedLoad(Ty, Ptr, Alignment) || |
5225 | isLegalMaskedGather(Ty, Alignment)) |
5226 | : !(isLegalMaskedStore(Ty, Ptr, Alignment) || |
5227 | isLegalMaskedScatter(Ty, Alignment)); |
5228 | } |
5229 | case Instruction::UDiv: |
5230 | case Instruction::SDiv: |
5231 | case Instruction::SRem: |
5232 | case Instruction::URem: |
5233 | return mayDivideByZero(*I); |
5234 | } |
5235 | return false; |
5236 | } |
5237 | |
5238 | bool LoopVectorizationCostModel::interleavedAccessCanBeWidened( |
5239 | Instruction *I, ElementCount VF) { |
5240 | assert(isAccessInterleaved(I) && "Expecting interleaved access.")((isAccessInterleaved(I) && "Expecting interleaved access." ) ? static_cast<void> (0) : __assert_fail ("isAccessInterleaved(I) && \"Expecting interleaved access.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5240, __PRETTY_FUNCTION__)); |
5241 | assert(getWideningDecision(I, VF) == CM_Unknown &&((getWideningDecision(I, VF) == CM_Unknown && "Decision should not be set yet." ) ? static_cast<void> (0) : __assert_fail ("getWideningDecision(I, VF) == CM_Unknown && \"Decision should not be set yet.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5242, __PRETTY_FUNCTION__)) |
5242 | "Decision should not be set yet.")((getWideningDecision(I, VF) == CM_Unknown && "Decision should not be set yet." ) ? static_cast<void> (0) : __assert_fail ("getWideningDecision(I, VF) == CM_Unknown && \"Decision should not be set yet.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5242, __PRETTY_FUNCTION__)); |
5243 | auto *Group = getInterleavedAccessGroup(I); |
5244 | assert(Group && "Must have a group.")((Group && "Must have a group.") ? static_cast<void > (0) : __assert_fail ("Group && \"Must have a group.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5244, __PRETTY_FUNCTION__)); |
5245 | |
5246 | // If the instruction's allocated size doesn't equal it's type size, it |
5247 | // requires padding and will be scalarized. |
5248 | auto &DL = I->getModule()->getDataLayout(); |
5249 | auto *ScalarTy = getMemInstValueType(I); |
5250 | if (hasIrregularType(ScalarTy, DL, VF)) |
5251 | return false; |
5252 | |
5253 | // Check if masking is required. |
5254 | // A Group may need masking for one of two reasons: it resides in a block that |
5255 | // needs predication, or it was decided to use masking to deal with gaps. |
5256 | bool PredicatedAccessRequiresMasking = |
5257 | Legal->blockNeedsPredication(I->getParent()) && Legal->isMaskRequired(I); |
5258 | bool AccessWithGapsRequiresMasking = |
5259 | Group->requiresScalarEpilogue() && !isScalarEpilogueAllowed(); |
5260 | if (!PredicatedAccessRequiresMasking && !AccessWithGapsRequiresMasking) |
5261 | return true; |
5262 | |
5263 | // If masked interleaving is required, we expect that the user/target had |
5264 | // enabled it, because otherwise it either wouldn't have been created or |
5265 | // it should have been invalidated by the CostModel. |
5266 | assert(useMaskedInterleavedAccesses(TTI) &&((useMaskedInterleavedAccesses(TTI) && "Masked interleave-groups for predicated accesses are not enabled." ) ? static_cast<void> (0) : __assert_fail ("useMaskedInterleavedAccesses(TTI) && \"Masked interleave-groups for predicated accesses are not enabled.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5267, __PRETTY_FUNCTION__)) |
5267 | "Masked interleave-groups for predicated accesses are not enabled.")((useMaskedInterleavedAccesses(TTI) && "Masked interleave-groups for predicated accesses are not enabled." ) ? static_cast<void> (0) : __assert_fail ("useMaskedInterleavedAccesses(TTI) && \"Masked interleave-groups for predicated accesses are not enabled.\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5267, __PRETTY_FUNCTION__)); |
5268 | |
5269 | auto *Ty = getMemInstValueType(I); |
5270 | const Align Alignment = getLoadStoreAlignment(I); |
5271 | return isa<LoadInst>(I) ? TTI.isLegalMaskedLoad(Ty, Alignment) |
5272 | : TTI.isLegalMaskedStore(Ty, Alignment); |
5273 | } |
5274 | |
5275 | bool LoopVectorizationCostModel::memoryInstructionCanBeWidened( |
5276 | Instruction *I, ElementCount VF) { |
5277 | // Get and ensure we have a valid memory instruction. |
5278 | LoadInst *LI = dyn_cast<LoadInst>(I); |
5279 | StoreInst *SI = dyn_cast<StoreInst>(I); |
5280 | assert((LI || SI) && "Invalid memory instruction")(((LI || SI) && "Invalid memory instruction") ? static_cast <void> (0) : __assert_fail ("(LI || SI) && \"Invalid memory instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5280, __PRETTY_FUNCTION__)); |
5281 | |
5282 | auto *Ptr = getLoadStorePointerOperand(I); |
5283 | |
5284 | // In order to be widened, the pointer should be consecutive, first of all. |
5285 | if (!Legal->isConsecutivePtr(Ptr)) |
5286 | return false; |
5287 | |
5288 | // If the instruction is a store located in a predicated block, it will be |
5289 | // scalarized. |
5290 | if (isScalarWithPredication(I)) |
5291 | return false; |
5292 | |
5293 | // If the instruction's allocated size doesn't equal it's type size, it |
5294 | // requires padding and will be scalarized. |
5295 | auto &DL = I->getModule()->getDataLayout(); |
5296 | auto *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType(); |
5297 | if (hasIrregularType(ScalarTy, DL, VF)) |
5298 | return false; |
5299 | |
5300 | return true; |
5301 | } |
5302 | |
5303 | void LoopVectorizationCostModel::collectLoopUniforms(ElementCount VF) { |
5304 | // We should not collect Uniforms more than once per VF. Right now, |
5305 | // this function is called from collectUniformsAndScalars(), which |
5306 | // already does this check. Collecting Uniforms for VF=1 does not make any |
5307 | // sense. |
5308 | |
5309 | assert(VF.isVector() && Uniforms.find(VF) == Uniforms.end() &&((VF.isVector() && Uniforms.find(VF) == Uniforms.end( ) && "This function should not be visited twice for the same VF" ) ? static_cast<void> (0) : __assert_fail ("VF.isVector() && Uniforms.find(VF) == Uniforms.end() && \"This function should not be visited twice for the same VF\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5310, __PRETTY_FUNCTION__)) |
5310 | "This function should not be visited twice for the same VF")((VF.isVector() && Uniforms.find(VF) == Uniforms.end( ) && "This function should not be visited twice for the same VF" ) ? static_cast<void> (0) : __assert_fail ("VF.isVector() && Uniforms.find(VF) == Uniforms.end() && \"This function should not be visited twice for the same VF\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5310, __PRETTY_FUNCTION__)); |
5311 | |
5312 | // Visit the list of Uniforms. If we'll not find any uniform value, we'll |
5313 | // not analyze again. Uniforms.count(VF) will return 1. |
5314 | Uniforms[VF].clear(); |
5315 | |
5316 | // We now know that the loop is vectorizable! |
5317 | // Collect instructions inside the loop that will remain uniform after |
5318 | // vectorization. |
5319 | |
5320 | // Global values, params and instructions outside of current loop are out of |
5321 | // scope. |
5322 | auto isOutOfScope = [&](Value *V) -> bool { |
5323 | Instruction *I = dyn_cast<Instruction>(V); |
5324 | return (!I || !TheLoop->contains(I)); |
5325 | }; |
5326 | |
5327 | SetVector<Instruction *> Worklist; |
5328 | BasicBlock *Latch = TheLoop->getLoopLatch(); |
5329 | |
5330 | // Instructions that are scalar with predication must not be considered |
5331 | // uniform after vectorization, because that would create an erroneous |
5332 | // replicating region where only a single instance out of VF should be formed. |
5333 | // TODO: optimize such seldom cases if found important, see PR40816. |
5334 | auto addToWorklistIfAllowed = [&](Instruction *I) -> void { |
5335 | if (isOutOfScope(I)) { |
5336 | LLVM_DEBUG(dbgs() << "LV: Found not uniform due to scope: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found not uniform due to scope: " << *I << "\n"; } } while (false) |
5337 | << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found not uniform due to scope: " << *I << "\n"; } } while (false); |
5338 | return; |
5339 | } |
5340 | if (isScalarWithPredication(I, VF)) { |
5341 | LLVM_DEBUG(dbgs() << "LV: Found not uniform being ScalarWithPredication: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found not uniform being ScalarWithPredication: " << *I << "\n"; } } while (false) |
5342 | << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found not uniform being ScalarWithPredication: " << *I << "\n"; } } while (false); |
5343 | return; |
5344 | } |
5345 | LLVM_DEBUG(dbgs() << "LV: Found uniform instruction: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found uniform instruction: " << *I << "\n"; } } while (false); |
5346 | Worklist.insert(I); |
5347 | }; |
5348 | |
5349 | // Start with the conditional branch. If the branch condition is an |
5350 | // instruction contained in the loop that is only used by the branch, it is |
5351 | // uniform. |
5352 | auto *Cmp = dyn_cast<Instruction>(Latch->getTerminator()->getOperand(0)); |
5353 | if (Cmp && TheLoop->contains(Cmp) && Cmp->hasOneUse()) |
5354 | addToWorklistIfAllowed(Cmp); |
5355 | |
5356 | auto isUniformDecision = [&](Instruction *I, ElementCount VF) { |
5357 | InstWidening WideningDecision = getWideningDecision(I, VF); |
5358 | assert(WideningDecision != CM_Unknown &&((WideningDecision != CM_Unknown && "Widening decision should be ready at this moment" ) ? static_cast<void> (0) : __assert_fail ("WideningDecision != CM_Unknown && \"Widening decision should be ready at this moment\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5359, __PRETTY_FUNCTION__)) |
5359 | "Widening decision should be ready at this moment")((WideningDecision != CM_Unknown && "Widening decision should be ready at this moment" ) ? static_cast<void> (0) : __assert_fail ("WideningDecision != CM_Unknown && \"Widening decision should be ready at this moment\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5359, __PRETTY_FUNCTION__)); |
5360 | |
5361 | // A uniform memory op is itself uniform. We exclude uniform stores |
5362 | // here as they demand the last lane, not the first one. |
5363 | if (isa<LoadInst>(I) && Legal->isUniformMemOp(*I)) { |
5364 | assert(WideningDecision == CM_Scalarize)((WideningDecision == CM_Scalarize) ? static_cast<void> (0) : __assert_fail ("WideningDecision == CM_Scalarize", "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5364, __PRETTY_FUNCTION__)); |
5365 | return true; |
5366 | } |
5367 | |
5368 | return (WideningDecision == CM_Widen || |
5369 | WideningDecision == CM_Widen_Reverse || |
5370 | WideningDecision == CM_Interleave); |
5371 | }; |
5372 | |
5373 | |
5374 | // Returns true if Ptr is the pointer operand of a memory access instruction |
5375 | // I, and I is known to not require scalarization. |
5376 | auto isVectorizedMemAccessUse = [&](Instruction *I, Value *Ptr) -> bool { |
5377 | return getLoadStorePointerOperand(I) == Ptr && isUniformDecision(I, VF); |
5378 | }; |
5379 | |
5380 | // Holds a list of values which are known to have at least one uniform use. |
5381 | // Note that there may be other uses which aren't uniform. A "uniform use" |
5382 | // here is something which only demands lane 0 of the unrolled iterations; |
5383 | // it does not imply that all lanes produce the same value (e.g. this is not |
5384 | // the usual meaning of uniform) |
5385 | SmallPtrSet<Value *, 8> HasUniformUse; |
5386 | |
5387 | // Scan the loop for instructions which are either a) known to have only |
5388 | // lane 0 demanded or b) are uses which demand only lane 0 of their operand. |
5389 | for (auto *BB : TheLoop->blocks()) |
5390 | for (auto &I : *BB) { |
5391 | // If there's no pointer operand, there's nothing to do. |
5392 | auto *Ptr = getLoadStorePointerOperand(&I); |
5393 | if (!Ptr) |
5394 | continue; |
5395 | |
5396 | // A uniform memory op is itself uniform. We exclude uniform stores |
5397 | // here as they demand the last lane, not the first one. |
5398 | if (isa<LoadInst>(I) && Legal->isUniformMemOp(I)) |
5399 | addToWorklistIfAllowed(&I); |
5400 | |
5401 | if (isUniformDecision(&I, VF)) { |
5402 | assert(isVectorizedMemAccessUse(&I, Ptr) && "consistency check")((isVectorizedMemAccessUse(&I, Ptr) && "consistency check" ) ? static_cast<void> (0) : __assert_fail ("isVectorizedMemAccessUse(&I, Ptr) && \"consistency check\"" , "/build/llvm-toolchain-snapshot-13~++20210307111131+ab67fd39fc14/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5402, __PRETTY_FUNCTION__)); |
5403 | HasUniformUse.insert(Ptr); |
5404 | } |
5405 | } |
5406 | |
5407 | // Add to the worklist any operands which have *only* uniform (e.g. lane 0 |
5408 | // demanding) users. Since loops are assumed to be in LCSSA form, this |
5409 | // disallows uses outside the loop as well. |
5410 | for (auto *V : HasUniformUse) { |
5411 | if (isOutOfScope(V)) |
5412 | continue; |
5413 | auto *I = cast<Instruction>(V); |
5414 | auto UsersAreMemAccesses = |
5415 | llvm::all_of(I->users(), [&](User *U) -> bool { |
5416 | return isVectorizedMemAccessUse(cast<Instruction>(U), V); |
5417 | }); |
5418 | if (UsersAreMemAccesses) |
5419 | addToWorklistIfAllowed(I); |
5420 | } |
5421 | |
5422 | // Expand Worklist in topological order: whenever a new instruction |
5423 | // is added , its users should be already inside Worklist. It ensures |
5424 | // a uniform instruction will only be used by uniform instructions. |
5425 | unsigned idx = 0; |
5426 | while (idx != Worklist.size()) { |
5427 | Instruction *I = Worklist[idx++]; |
5428 | |
5429 | for (auto OV : I->operand_values()) { |
5430 | // isOutOfScope operands cannot be uniform instructions. |
5431 | if (isOutOfScope(OV)) |
5432 | continue; |
5433 | // First order recurrence Phi's should typically be considered |
5434 | // non-uniform. |
5435 | auto *OP = dyn_cast<PHINode>(OV); |
5436 | if (OP && Legal->isFirstOrderRecurrence(OP)) |
5437 | continue; |
5438 | // If all the users of the operand are uniform, then add the |
5439 | // operand into the uniform worklist. |
5440 | auto *OI = cast<Instruction>(OV); |
5441 | if (llvm::all_of(OI->users(), [&](User *U) -> bool { |
5442 | auto *J = cast<Instruction>(U); |
5443 | return Worklist.count(J) || isVectorizedMemAccessUse(J, OI); |
5444 | })) |