File: | build/llvm-toolchain-snapshot-15~++20220319111418+752e9cdbb06a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp |
Warning: | line 2907, column 23 Called C++ object pointer is null |
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/SmallSet.h" | |||
74 | #include "llvm/ADT/SmallVector.h" | |||
75 | #include "llvm/ADT/Statistic.h" | |||
76 | #include "llvm/ADT/StringRef.h" | |||
77 | #include "llvm/ADT/Twine.h" | |||
78 | #include "llvm/ADT/iterator_range.h" | |||
79 | #include "llvm/Analysis/AssumptionCache.h" | |||
80 | #include "llvm/Analysis/BasicAliasAnalysis.h" | |||
81 | #include "llvm/Analysis/BlockFrequencyInfo.h" | |||
82 | #include "llvm/Analysis/CFG.h" | |||
83 | #include "llvm/Analysis/CodeMetrics.h" | |||
84 | #include "llvm/Analysis/DemandedBits.h" | |||
85 | #include "llvm/Analysis/GlobalsModRef.h" | |||
86 | #include "llvm/Analysis/LoopAccessAnalysis.h" | |||
87 | #include "llvm/Analysis/LoopAnalysisManager.h" | |||
88 | #include "llvm/Analysis/LoopInfo.h" | |||
89 | #include "llvm/Analysis/LoopIterator.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/Metadata.h" | |||
116 | #include "llvm/IR/Module.h" | |||
117 | #include "llvm/IR/Operator.h" | |||
118 | #include "llvm/IR/PatternMatch.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 <functional> | |||
147 | #include <iterator> | |||
148 | #include <limits> | |||
149 | #include <memory> | |||
150 | #include <string> | |||
151 | #include <tuple> | |||
152 | #include <utility> | |||
153 | ||||
154 | using namespace llvm; | |||
155 | ||||
156 | #define LV_NAME"loop-vectorize" "loop-vectorize" | |||
157 | #define DEBUG_TYPE"loop-vectorize" LV_NAME"loop-vectorize" | |||
158 | ||||
159 | #ifndef NDEBUG | |||
160 | const char VerboseDebug[] = DEBUG_TYPE"loop-vectorize" "-verbose"; | |||
161 | #endif | |||
162 | ||||
163 | /// @{ | |||
164 | /// Metadata attribute names | |||
165 | const char LLVMLoopVectorizeFollowupAll[] = "llvm.loop.vectorize.followup_all"; | |||
166 | const char LLVMLoopVectorizeFollowupVectorized[] = | |||
167 | "llvm.loop.vectorize.followup_vectorized"; | |||
168 | const char LLVMLoopVectorizeFollowupEpilogue[] = | |||
169 | "llvm.loop.vectorize.followup_epilogue"; | |||
170 | /// @} | |||
171 | ||||
172 | STATISTIC(LoopsVectorized, "Number of loops vectorized")static llvm::Statistic LoopsVectorized = {"loop-vectorize", "LoopsVectorized" , "Number of loops vectorized"}; | |||
173 | STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization")static llvm::Statistic LoopsAnalyzed = {"loop-vectorize", "LoopsAnalyzed" , "Number of loops analyzed for vectorization"}; | |||
174 | STATISTIC(LoopsEpilogueVectorized, "Number of epilogues vectorized")static llvm::Statistic LoopsEpilogueVectorized = {"loop-vectorize" , "LoopsEpilogueVectorized", "Number of epilogues vectorized" }; | |||
175 | ||||
176 | static cl::opt<bool> EnableEpilogueVectorization( | |||
177 | "enable-epilogue-vectorization", cl::init(true), cl::Hidden, | |||
178 | cl::desc("Enable vectorization of epilogue loops.")); | |||
179 | ||||
180 | static cl::opt<unsigned> EpilogueVectorizationForceVF( | |||
181 | "epilogue-vectorization-force-VF", cl::init(1), cl::Hidden, | |||
182 | cl::desc("When epilogue vectorization is enabled, and a value greater than " | |||
183 | "1 is specified, forces the given VF for all applicable epilogue " | |||
184 | "loops.")); | |||
185 | ||||
186 | static cl::opt<unsigned> EpilogueVectorizationMinVF( | |||
187 | "epilogue-vectorization-minimum-VF", cl::init(16), cl::Hidden, | |||
188 | cl::desc("Only loops with vectorization factor equal to or larger than " | |||
189 | "the specified value are considered for epilogue vectorization.")); | |||
190 | ||||
191 | /// Loops with a known constant trip count below this number are vectorized only | |||
192 | /// if no scalar iteration overheads are incurred. | |||
193 | static cl::opt<unsigned> TinyTripCountVectorThreshold( | |||
194 | "vectorizer-min-trip-count", cl::init(16), cl::Hidden, | |||
195 | cl::desc("Loops with a constant trip count that is smaller than this " | |||
196 | "value are vectorized only if no scalar iteration overheads " | |||
197 | "are incurred.")); | |||
198 | ||||
199 | static cl::opt<unsigned> PragmaVectorizeMemoryCheckThreshold( | |||
200 | "pragma-vectorize-memory-check-threshold", cl::init(128), cl::Hidden, | |||
201 | cl::desc("The maximum allowed number of runtime memory checks with a " | |||
202 | "vectorize(enable) pragma.")); | |||
203 | ||||
204 | // Option prefer-predicate-over-epilogue indicates that an epilogue is undesired, | |||
205 | // that predication is preferred, and this lists all options. I.e., the | |||
206 | // vectorizer will try to fold the tail-loop (epilogue) into the vector body | |||
207 | // and predicate the instructions accordingly. If tail-folding fails, there are | |||
208 | // different fallback strategies depending on these values: | |||
209 | namespace PreferPredicateTy { | |||
210 | enum Option { | |||
211 | ScalarEpilogue = 0, | |||
212 | PredicateElseScalarEpilogue, | |||
213 | PredicateOrDontVectorize | |||
214 | }; | |||
215 | } // namespace PreferPredicateTy | |||
216 | ||||
217 | static cl::opt<PreferPredicateTy::Option> PreferPredicateOverEpilogue( | |||
218 | "prefer-predicate-over-epilogue", | |||
219 | cl::init(PreferPredicateTy::ScalarEpilogue), | |||
220 | cl::Hidden, | |||
221 | cl::desc("Tail-folding and predication preferences over creating a scalar " | |||
222 | "epilogue loop."), | |||
223 | cl::values(clEnumValN(PreferPredicateTy::ScalarEpilogue,llvm::cl::OptionEnumValue { "scalar-epilogue", int(PreferPredicateTy ::ScalarEpilogue), "Don't tail-predicate loops, create scalar epilogue" } | |||
224 | "scalar-epilogue",llvm::cl::OptionEnumValue { "scalar-epilogue", int(PreferPredicateTy ::ScalarEpilogue), "Don't tail-predicate loops, create scalar epilogue" } | |||
225 | "Don't tail-predicate loops, create scalar epilogue")llvm::cl::OptionEnumValue { "scalar-epilogue", int(PreferPredicateTy ::ScalarEpilogue), "Don't tail-predicate loops, create scalar epilogue" }, | |||
226 | clEnumValN(PreferPredicateTy::PredicateElseScalarEpilogue,llvm::cl::OptionEnumValue { "predicate-else-scalar-epilogue", int(PreferPredicateTy::PredicateElseScalarEpilogue), "prefer tail-folding, create scalar epilogue if tail " "folding fails." } | |||
227 | "predicate-else-scalar-epilogue",llvm::cl::OptionEnumValue { "predicate-else-scalar-epilogue", int(PreferPredicateTy::PredicateElseScalarEpilogue), "prefer tail-folding, create scalar epilogue if tail " "folding fails." } | |||
228 | "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." } | |||
229 | "folding fails.")llvm::cl::OptionEnumValue { "predicate-else-scalar-epilogue", int(PreferPredicateTy::PredicateElseScalarEpilogue), "prefer tail-folding, create scalar epilogue if tail " "folding fails." }, | |||
230 | clEnumValN(PreferPredicateTy::PredicateOrDontVectorize,llvm::cl::OptionEnumValue { "predicate-dont-vectorize", int(PreferPredicateTy ::PredicateOrDontVectorize), "prefers tail-folding, don't attempt vectorization if " "tail-folding fails." } | |||
231 | "predicate-dont-vectorize",llvm::cl::OptionEnumValue { "predicate-dont-vectorize", int(PreferPredicateTy ::PredicateOrDontVectorize), "prefers tail-folding, don't attempt vectorization if " "tail-folding fails." } | |||
232 | "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." } | |||
233 | "tail-folding fails.")llvm::cl::OptionEnumValue { "predicate-dont-vectorize", int(PreferPredicateTy ::PredicateOrDontVectorize), "prefers tail-folding, don't attempt vectorization if " "tail-folding fails." })); | |||
234 | ||||
235 | static cl::opt<bool> MaximizeBandwidth( | |||
236 | "vectorizer-maximize-bandwidth", cl::init(false), cl::Hidden, | |||
237 | cl::desc("Maximize bandwidth when selecting vectorization factor which " | |||
238 | "will be determined by the smallest type in loop.")); | |||
239 | ||||
240 | static cl::opt<bool> EnableInterleavedMemAccesses( | |||
241 | "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden, | |||
242 | cl::desc("Enable vectorization on interleaved memory accesses in a loop")); | |||
243 | ||||
244 | /// An interleave-group may need masking if it resides in a block that needs | |||
245 | /// predication, or in order to mask away gaps. | |||
246 | static cl::opt<bool> EnableMaskedInterleavedMemAccesses( | |||
247 | "enable-masked-interleaved-mem-accesses", cl::init(false), cl::Hidden, | |||
248 | cl::desc("Enable vectorization on masked interleaved memory accesses in a loop")); | |||
249 | ||||
250 | static cl::opt<unsigned> TinyTripCountInterleaveThreshold( | |||
251 | "tiny-trip-count-interleave-threshold", cl::init(128), cl::Hidden, | |||
252 | cl::desc("We don't interleave loops with a estimated constant trip count " | |||
253 | "below this number")); | |||
254 | ||||
255 | static cl::opt<unsigned> ForceTargetNumScalarRegs( | |||
256 | "force-target-num-scalar-regs", cl::init(0), cl::Hidden, | |||
257 | cl::desc("A flag that overrides the target's number of scalar registers.")); | |||
258 | ||||
259 | static cl::opt<unsigned> ForceTargetNumVectorRegs( | |||
260 | "force-target-num-vector-regs", cl::init(0), cl::Hidden, | |||
261 | cl::desc("A flag that overrides the target's number of vector registers.")); | |||
262 | ||||
263 | static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor( | |||
264 | "force-target-max-scalar-interleave", cl::init(0), cl::Hidden, | |||
265 | cl::desc("A flag that overrides the target's max interleave factor for " | |||
266 | "scalar loops.")); | |||
267 | ||||
268 | static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor( | |||
269 | "force-target-max-vector-interleave", cl::init(0), cl::Hidden, | |||
270 | cl::desc("A flag that overrides the target's max interleave factor for " | |||
271 | "vectorized loops.")); | |||
272 | ||||
273 | static cl::opt<unsigned> ForceTargetInstructionCost( | |||
274 | "force-target-instruction-cost", cl::init(0), cl::Hidden, | |||
275 | cl::desc("A flag that overrides the target's expected cost for " | |||
276 | "an instruction to a single constant value. Mostly " | |||
277 | "useful for getting consistent testing.")); | |||
278 | ||||
279 | static cl::opt<bool> ForceTargetSupportsScalableVectors( | |||
280 | "force-target-supports-scalable-vectors", cl::init(false), cl::Hidden, | |||
281 | cl::desc( | |||
282 | "Pretend that scalable vectors are supported, even if the target does " | |||
283 | "not support them. This flag should only be used for testing.")); | |||
284 | ||||
285 | static cl::opt<unsigned> SmallLoopCost( | |||
286 | "small-loop-cost", cl::init(20), cl::Hidden, | |||
287 | cl::desc( | |||
288 | "The cost of a loop that is considered 'small' by the interleaver.")); | |||
289 | ||||
290 | static cl::opt<bool> LoopVectorizeWithBlockFrequency( | |||
291 | "loop-vectorize-with-block-frequency", cl::init(true), cl::Hidden, | |||
292 | cl::desc("Enable the use of the block frequency analysis to access PGO " | |||
293 | "heuristics minimizing code growth in cold regions and being more " | |||
294 | "aggressive in hot regions.")); | |||
295 | ||||
296 | // Runtime interleave loops for load/store throughput. | |||
297 | static cl::opt<bool> EnableLoadStoreRuntimeInterleave( | |||
298 | "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden, | |||
299 | cl::desc( | |||
300 | "Enable runtime interleaving until load/store ports are saturated")); | |||
301 | ||||
302 | /// Interleave small loops with scalar reductions. | |||
303 | static cl::opt<bool> InterleaveSmallLoopScalarReduction( | |||
304 | "interleave-small-loop-scalar-reduction", cl::init(false), cl::Hidden, | |||
305 | cl::desc("Enable interleaving for loops with small iteration counts that " | |||
306 | "contain scalar reductions to expose ILP.")); | |||
307 | ||||
308 | /// The number of stores in a loop that are allowed to need predication. | |||
309 | static cl::opt<unsigned> NumberOfStoresToPredicate( | |||
310 | "vectorize-num-stores-pred", cl::init(1), cl::Hidden, | |||
311 | cl::desc("Max number of stores to be predicated behind an if.")); | |||
312 | ||||
313 | static cl::opt<bool> EnableIndVarRegisterHeur( | |||
314 | "enable-ind-var-reg-heur", cl::init(true), cl::Hidden, | |||
315 | cl::desc("Count the induction variable only once when interleaving")); | |||
316 | ||||
317 | static cl::opt<bool> EnableCondStoresVectorization( | |||
318 | "enable-cond-stores-vec", cl::init(true), cl::Hidden, | |||
319 | cl::desc("Enable if predication of stores during vectorization.")); | |||
320 | ||||
321 | static cl::opt<unsigned> MaxNestedScalarReductionIC( | |||
322 | "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden, | |||
323 | cl::desc("The maximum interleave count to use when interleaving a scalar " | |||
324 | "reduction in a nested loop.")); | |||
325 | ||||
326 | static cl::opt<bool> | |||
327 | PreferInLoopReductions("prefer-inloop-reductions", cl::init(false), | |||
328 | cl::Hidden, | |||
329 | cl::desc("Prefer in-loop vector reductions, " | |||
330 | "overriding the targets preference.")); | |||
331 | ||||
332 | static cl::opt<bool> ForceOrderedReductions( | |||
333 | "force-ordered-reductions", cl::init(false), cl::Hidden, | |||
334 | cl::desc("Enable the vectorisation of loops with in-order (strict) " | |||
335 | "FP reductions")); | |||
336 | ||||
337 | static cl::opt<bool> PreferPredicatedReductionSelect( | |||
338 | "prefer-predicated-reduction-select", cl::init(false), cl::Hidden, | |||
339 | cl::desc( | |||
340 | "Prefer predicating a reduction operation over an after loop select.")); | |||
341 | ||||
342 | cl::opt<bool> EnableVPlanNativePath( | |||
343 | "enable-vplan-native-path", cl::init(false), cl::Hidden, | |||
344 | cl::desc("Enable VPlan-native vectorization path with " | |||
345 | "support for outer loop vectorization.")); | |||
346 | ||||
347 | // FIXME: Remove this switch once we have divergence analysis. Currently we | |||
348 | // assume divergent non-backedge branches when this switch is true. | |||
349 | cl::opt<bool> EnableVPlanPredication( | |||
350 | "enable-vplan-predication", cl::init(false), cl::Hidden, | |||
351 | cl::desc("Enable VPlan-native vectorization path predicator with " | |||
352 | "support for outer loop vectorization.")); | |||
353 | ||||
354 | // This flag enables the stress testing of the VPlan H-CFG construction in the | |||
355 | // VPlan-native vectorization path. It must be used in conjuction with | |||
356 | // -enable-vplan-native-path. -vplan-verify-hcfg can also be used to enable the | |||
357 | // verification of the H-CFGs built. | |||
358 | static cl::opt<bool> VPlanBuildStressTest( | |||
359 | "vplan-build-stress-test", cl::init(false), cl::Hidden, | |||
360 | cl::desc( | |||
361 | "Build VPlan for every supported loop nest in the function and bail " | |||
362 | "out right after the build (stress test the VPlan H-CFG construction " | |||
363 | "in the VPlan-native vectorization path).")); | |||
364 | ||||
365 | cl::opt<bool> llvm::EnableLoopInterleaving( | |||
366 | "interleave-loops", cl::init(true), cl::Hidden, | |||
367 | cl::desc("Enable loop interleaving in Loop vectorization passes")); | |||
368 | cl::opt<bool> llvm::EnableLoopVectorization( | |||
369 | "vectorize-loops", cl::init(true), cl::Hidden, | |||
370 | cl::desc("Run the Loop vectorization passes")); | |||
371 | ||||
372 | cl::opt<bool> PrintVPlansInDotFormat( | |||
373 | "vplan-print-in-dot-format", cl::init(false), cl::Hidden, | |||
374 | cl::desc("Use dot format instead of plain text when dumping VPlans")); | |||
375 | ||||
376 | /// A helper function that returns true if the given type is irregular. The | |||
377 | /// type is irregular if its allocated size doesn't equal the store size of an | |||
378 | /// element of the corresponding vector type. | |||
379 | static bool hasIrregularType(Type *Ty, const DataLayout &DL) { | |||
380 | // Determine if an array of N elements of type Ty is "bitcast compatible" | |||
381 | // with a <N x Ty> vector. | |||
382 | // This is only true if there is no padding between the array elements. | |||
383 | return DL.getTypeAllocSizeInBits(Ty) != DL.getTypeSizeInBits(Ty); | |||
384 | } | |||
385 | ||||
386 | /// A helper function that returns the reciprocal of the block probability of | |||
387 | /// predicated blocks. If we return X, we are assuming the predicated block | |||
388 | /// will execute once for every X iterations of the loop header. | |||
389 | /// | |||
390 | /// TODO: We should use actual block probability here, if available. Currently, | |||
391 | /// we always assume predicated blocks have a 50% chance of executing. | |||
392 | static unsigned getReciprocalPredBlockProb() { return 2; } | |||
393 | ||||
394 | /// A helper function that returns an integer or floating-point constant with | |||
395 | /// value C. | |||
396 | static Constant *getSignedIntOrFpConstant(Type *Ty, int64_t C) { | |||
397 | return Ty->isIntegerTy() ? ConstantInt::getSigned(Ty, C) | |||
398 | : ConstantFP::get(Ty, C); | |||
399 | } | |||
400 | ||||
401 | /// Returns "best known" trip count for the specified loop \p L as defined by | |||
402 | /// the following procedure: | |||
403 | /// 1) Returns exact trip count if it is known. | |||
404 | /// 2) Returns expected trip count according to profile data if any. | |||
405 | /// 3) Returns upper bound estimate if it is known. | |||
406 | /// 4) Returns None if all of the above failed. | |||
407 | static Optional<unsigned> getSmallBestKnownTC(ScalarEvolution &SE, Loop *L) { | |||
408 | // Check if exact trip count is known. | |||
409 | if (unsigned ExpectedTC = SE.getSmallConstantTripCount(L)) | |||
410 | return ExpectedTC; | |||
411 | ||||
412 | // Check if there is an expected trip count available from profile data. | |||
413 | if (LoopVectorizeWithBlockFrequency) | |||
414 | if (auto EstimatedTC = getLoopEstimatedTripCount(L)) | |||
415 | return EstimatedTC; | |||
416 | ||||
417 | // Check if upper bound estimate is known. | |||
418 | if (unsigned ExpectedTC = SE.getSmallConstantMaxTripCount(L)) | |||
419 | return ExpectedTC; | |||
420 | ||||
421 | return None; | |||
422 | } | |||
423 | ||||
424 | // Forward declare GeneratedRTChecks. | |||
425 | class GeneratedRTChecks; | |||
426 | ||||
427 | namespace llvm { | |||
428 | ||||
429 | AnalysisKey ShouldRunExtraVectorPasses::Key; | |||
430 | ||||
431 | /// InnerLoopVectorizer vectorizes loops which contain only one basic | |||
432 | /// block to a specified vectorization factor (VF). | |||
433 | /// This class performs the widening of scalars into vectors, or multiple | |||
434 | /// scalars. This class also implements the following features: | |||
435 | /// * It inserts an epilogue loop for handling loops that don't have iteration | |||
436 | /// counts that are known to be a multiple of the vectorization factor. | |||
437 | /// * It handles the code generation for reduction variables. | |||
438 | /// * Scalarization (implementation using scalars) of un-vectorizable | |||
439 | /// instructions. | |||
440 | /// InnerLoopVectorizer does not perform any vectorization-legality | |||
441 | /// checks, and relies on the caller to check for the different legality | |||
442 | /// aspects. The InnerLoopVectorizer relies on the | |||
443 | /// LoopVectorizationLegality class to provide information about the induction | |||
444 | /// and reduction variables that were found to a given vectorization factor. | |||
445 | class InnerLoopVectorizer { | |||
446 | public: | |||
447 | InnerLoopVectorizer(Loop *OrigLoop, PredicatedScalarEvolution &PSE, | |||
448 | LoopInfo *LI, DominatorTree *DT, | |||
449 | const TargetLibraryInfo *TLI, | |||
450 | const TargetTransformInfo *TTI, AssumptionCache *AC, | |||
451 | OptimizationRemarkEmitter *ORE, ElementCount VecWidth, | |||
452 | unsigned UnrollFactor, LoopVectorizationLegality *LVL, | |||
453 | LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI, | |||
454 | ProfileSummaryInfo *PSI, GeneratedRTChecks &RTChecks) | |||
455 | : OrigLoop(OrigLoop), PSE(PSE), LI(LI), DT(DT), TLI(TLI), TTI(TTI), | |||
456 | AC(AC), ORE(ORE), VF(VecWidth), UF(UnrollFactor), | |||
457 | Builder(PSE.getSE()->getContext()), Legal(LVL), Cost(CM), BFI(BFI), | |||
458 | PSI(PSI), RTChecks(RTChecks) { | |||
459 | // Query this against the original loop and save it here because the profile | |||
460 | // of the original loop header may change as the transformation happens. | |||
461 | OptForSizeBasedOnProfile = llvm::shouldOptimizeForSize( | |||
462 | OrigLoop->getHeader(), PSI, BFI, PGSOQueryType::IRPass); | |||
463 | } | |||
464 | ||||
465 | virtual ~InnerLoopVectorizer() = default; | |||
466 | ||||
467 | /// Create a new empty loop that will contain vectorized instructions later | |||
468 | /// on, while the old loop will be used as the scalar remainder. Control flow | |||
469 | /// is generated around the vectorized (and scalar epilogue) loops consisting | |||
470 | /// of various checks and bypasses. Return the pre-header block of the new | |||
471 | /// loop and the start value for the canonical induction, if it is != 0. The | |||
472 | /// latter is the case when vectorizing the epilogue loop. In the case of | |||
473 | /// epilogue vectorization, this function is overriden to handle the more | |||
474 | /// complex control flow around the loops. | |||
475 | virtual std::pair<BasicBlock *, Value *> createVectorizedLoopSkeleton(); | |||
476 | ||||
477 | /// Widen a single call instruction within the innermost loop. | |||
478 | void widenCallInstruction(CallInst &I, VPValue *Def, VPUser &ArgOperands, | |||
479 | VPTransformState &State); | |||
480 | ||||
481 | /// Fix the vectorized code, taking care of header phi's, live-outs, and more. | |||
482 | void fixVectorizedLoop(VPTransformState &State); | |||
483 | ||||
484 | // Return true if any runtime check is added. | |||
485 | bool areSafetyChecksAdded() { return AddedSafetyChecks; } | |||
486 | ||||
487 | /// A type for vectorized values in the new loop. Each value from the | |||
488 | /// original loop, when vectorized, is represented by UF vector values in the | |||
489 | /// new unrolled loop, where UF is the unroll factor. | |||
490 | using VectorParts = SmallVector<Value *, 2>; | |||
491 | ||||
492 | /// Vectorize a single first-order recurrence or pointer induction PHINode in | |||
493 | /// a block. This method handles the induction variable canonicalization. It | |||
494 | /// supports both VF = 1 for unrolled loops and arbitrary length vectors. | |||
495 | void widenPHIInstruction(Instruction *PN, VPWidenPHIRecipe *PhiR, | |||
496 | VPTransformState &State); | |||
497 | ||||
498 | /// A helper function to scalarize a single Instruction in the innermost loop. | |||
499 | /// Generates a sequence of scalar instances for each lane between \p MinLane | |||
500 | /// and \p MaxLane, times each part between \p MinPart and \p MaxPart, | |||
501 | /// inclusive. Uses the VPValue operands from \p RepRecipe instead of \p | |||
502 | /// Instr's operands. | |||
503 | void scalarizeInstruction(Instruction *Instr, VPReplicateRecipe *RepRecipe, | |||
504 | const VPIteration &Instance, bool IfPredicateInstr, | |||
505 | VPTransformState &State); | |||
506 | ||||
507 | /// Construct the vector value of a scalarized value \p V one lane at a time. | |||
508 | void packScalarIntoVectorValue(VPValue *Def, const VPIteration &Instance, | |||
509 | VPTransformState &State); | |||
510 | ||||
511 | /// Try to vectorize interleaved access group \p Group with the base address | |||
512 | /// given in \p Addr, optionally masking the vector operations if \p | |||
513 | /// BlockInMask is non-null. Use \p State to translate given VPValues to IR | |||
514 | /// values in the vectorized loop. | |||
515 | void vectorizeInterleaveGroup(const InterleaveGroup<Instruction> *Group, | |||
516 | ArrayRef<VPValue *> VPDefs, | |||
517 | VPTransformState &State, VPValue *Addr, | |||
518 | ArrayRef<VPValue *> StoredValues, | |||
519 | VPValue *BlockInMask = nullptr); | |||
520 | ||||
521 | /// Set the debug location in the builder \p Ptr using the debug location in | |||
522 | /// \p V. If \p Ptr is None then it uses the class member's Builder. | |||
523 | void setDebugLocFromInst(const Value *V, | |||
524 | Optional<IRBuilderBase *> CustomBuilder = None); | |||
525 | ||||
526 | /// Fix the non-induction PHIs in the OrigPHIsToFix vector. | |||
527 | void fixNonInductionPHIs(VPTransformState &State); | |||
528 | ||||
529 | /// Returns true if the reordering of FP operations is not allowed, but we are | |||
530 | /// able to vectorize with strict in-order reductions for the given RdxDesc. | |||
531 | bool useOrderedReductions(const RecurrenceDescriptor &RdxDesc); | |||
532 | ||||
533 | /// Create a broadcast instruction. This method generates a broadcast | |||
534 | /// instruction (shuffle) for loop invariant values and for the induction | |||
535 | /// value. If this is the induction variable then we extend it to N, N+1, ... | |||
536 | /// this is needed because each iteration in the loop corresponds to a SIMD | |||
537 | /// element. | |||
538 | virtual Value *getBroadcastInstrs(Value *V); | |||
539 | ||||
540 | /// Add metadata from one instruction to another. | |||
541 | /// | |||
542 | /// This includes both the original MDs from \p From and additional ones (\see | |||
543 | /// addNewMetadata). Use this for *newly created* instructions in the vector | |||
544 | /// loop. | |||
545 | void addMetadata(Instruction *To, Instruction *From); | |||
546 | ||||
547 | /// Similar to the previous function but it adds the metadata to a | |||
548 | /// vector of instructions. | |||
549 | void addMetadata(ArrayRef<Value *> To, Instruction *From); | |||
550 | ||||
551 | // Returns the resume value (bc.merge.rdx) for a reduction as | |||
552 | // generated by fixReduction. | |||
553 | PHINode *getReductionResumeValue(const RecurrenceDescriptor &RdxDesc); | |||
554 | ||||
555 | protected: | |||
556 | friend class LoopVectorizationPlanner; | |||
557 | ||||
558 | /// A small list of PHINodes. | |||
559 | using PhiVector = SmallVector<PHINode *, 4>; | |||
560 | ||||
561 | /// A type for scalarized values in the new loop. Each value from the | |||
562 | /// original loop, when scalarized, is represented by UF x VF scalar values | |||
563 | /// in the new unrolled loop, where UF is the unroll factor and VF is the | |||
564 | /// vectorization factor. | |||
565 | using ScalarParts = SmallVector<SmallVector<Value *, 4>, 2>; | |||
566 | ||||
567 | /// Set up the values of the IVs correctly when exiting the vector loop. | |||
568 | void fixupIVUsers(PHINode *OrigPhi, const InductionDescriptor &II, | |||
569 | Value *CountRoundDown, Value *EndValue, | |||
570 | BasicBlock *MiddleBlock, BasicBlock *VectorHeader); | |||
571 | ||||
572 | /// Introduce a conditional branch (on true, condition to be set later) at the | |||
573 | /// end of the header=latch connecting it to itself (across the backedge) and | |||
574 | /// to the exit block of \p L. | |||
575 | void createHeaderBranch(Loop *L); | |||
576 | ||||
577 | /// Handle all cross-iteration phis in the header. | |||
578 | void fixCrossIterationPHIs(VPTransformState &State); | |||
579 | ||||
580 | /// Create the exit value of first order recurrences in the middle block and | |||
581 | /// update their users. | |||
582 | void fixFirstOrderRecurrence(VPFirstOrderRecurrencePHIRecipe *PhiR, | |||
583 | VPTransformState &State); | |||
584 | ||||
585 | /// Create code for the loop exit value of the reduction. | |||
586 | void fixReduction(VPReductionPHIRecipe *Phi, VPTransformState &State); | |||
587 | ||||
588 | /// Clear NSW/NUW flags from reduction instructions if necessary. | |||
589 | void clearReductionWrapFlags(const RecurrenceDescriptor &RdxDesc, | |||
590 | VPTransformState &State); | |||
591 | ||||
592 | /// Fixup the LCSSA phi nodes in the unique exit block. This simply | |||
593 | /// means we need to add the appropriate incoming value from the middle | |||
594 | /// block as exiting edges from the scalar epilogue loop (if present) are | |||
595 | /// already in place, and we exit the vector loop exclusively to the middle | |||
596 | /// block. | |||
597 | void fixLCSSAPHIs(VPTransformState &State); | |||
598 | ||||
599 | /// Iteratively sink the scalarized operands of a predicated instruction into | |||
600 | /// the block that was created for it. | |||
601 | void sinkScalarOperands(Instruction *PredInst); | |||
602 | ||||
603 | /// Shrinks vector element sizes to the smallest bitwidth they can be legally | |||
604 | /// represented as. | |||
605 | void truncateToMinimalBitwidths(VPTransformState &State); | |||
606 | ||||
607 | /// Returns (and creates if needed) the original loop trip count. | |||
608 | Value *getOrCreateTripCount(Loop *NewLoop); | |||
609 | ||||
610 | /// Returns (and creates if needed) the trip count of the widened loop. | |||
611 | Value *getOrCreateVectorTripCount(Loop *NewLoop); | |||
612 | ||||
613 | /// Returns a bitcasted value to the requested vector type. | |||
614 | /// Also handles bitcasts of vector<float> <-> vector<pointer> types. | |||
615 | Value *createBitOrPointerCast(Value *V, VectorType *DstVTy, | |||
616 | const DataLayout &DL); | |||
617 | ||||
618 | /// Emit a bypass check to see if the vector trip count is zero, including if | |||
619 | /// it overflows. | |||
620 | void emitMinimumIterationCountCheck(Loop *L, BasicBlock *Bypass); | |||
621 | ||||
622 | /// Emit a bypass check to see if all of the SCEV assumptions we've | |||
623 | /// had to make are correct. Returns the block containing the checks or | |||
624 | /// nullptr if no checks have been added. | |||
625 | BasicBlock *emitSCEVChecks(BasicBlock *Bypass); | |||
626 | ||||
627 | /// Emit bypass checks to check any memory assumptions we may have made. | |||
628 | /// Returns the block containing the checks or nullptr if no checks have been | |||
629 | /// added. | |||
630 | BasicBlock *emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass); | |||
631 | ||||
632 | /// Emit basic blocks (prefixed with \p Prefix) for the iteration check, | |||
633 | /// vector loop preheader, middle block and scalar preheader. Also | |||
634 | /// allocate a loop object for the new vector loop and return it. | |||
635 | Loop *createVectorLoopSkeleton(StringRef Prefix); | |||
636 | ||||
637 | /// Create new phi nodes for the induction variables to resume iteration count | |||
638 | /// in the scalar epilogue, from where the vectorized loop left off. | |||
639 | /// In cases where the loop skeleton is more complicated (eg. epilogue | |||
640 | /// vectorization) and the resume values can come from an additional bypass | |||
641 | /// block, the \p AdditionalBypass pair provides information about the bypass | |||
642 | /// block and the end value on the edge from bypass to this loop. | |||
643 | void createInductionResumeValues( | |||
644 | Loop *L, | |||
645 | std::pair<BasicBlock *, Value *> AdditionalBypass = {nullptr, nullptr}); | |||
646 | ||||
647 | /// Complete the loop skeleton by adding debug MDs, creating appropriate | |||
648 | /// conditional branches in the middle block, preparing the builder and | |||
649 | /// running the verifier. Take in the vector loop \p L as argument, and return | |||
650 | /// the preheader of the completed vector loop. | |||
651 | BasicBlock *completeLoopSkeleton(Loop *L, MDNode *OrigLoopID); | |||
652 | ||||
653 | /// Add additional metadata to \p To that was not present on \p Orig. | |||
654 | /// | |||
655 | /// Currently this is used to add the noalias annotations based on the | |||
656 | /// inserted memchecks. Use this for instructions that are *cloned* into the | |||
657 | /// vector loop. | |||
658 | void addNewMetadata(Instruction *To, const Instruction *Orig); | |||
659 | ||||
660 | /// Collect poison-generating recipes that may generate a poison value that is | |||
661 | /// used after vectorization, even when their operands are not poison. Those | |||
662 | /// recipes meet the following conditions: | |||
663 | /// * Contribute to the address computation of a recipe generating a widen | |||
664 | /// memory load/store (VPWidenMemoryInstructionRecipe or | |||
665 | /// VPInterleaveRecipe). | |||
666 | /// * Such a widen memory load/store has at least one underlying Instruction | |||
667 | /// that is in a basic block that needs predication and after vectorization | |||
668 | /// the generated instruction won't be predicated. | |||
669 | void collectPoisonGeneratingRecipes(VPTransformState &State); | |||
670 | ||||
671 | /// Allow subclasses to override and print debug traces before/after vplan | |||
672 | /// execution, when trace information is requested. | |||
673 | virtual void printDebugTracesAtStart(){}; | |||
674 | virtual void printDebugTracesAtEnd(){}; | |||
675 | ||||
676 | /// The original loop. | |||
677 | Loop *OrigLoop; | |||
678 | ||||
679 | /// A wrapper around ScalarEvolution used to add runtime SCEV checks. Applies | |||
680 | /// dynamic knowledge to simplify SCEV expressions and converts them to a | |||
681 | /// more usable form. | |||
682 | PredicatedScalarEvolution &PSE; | |||
683 | ||||
684 | /// Loop Info. | |||
685 | LoopInfo *LI; | |||
686 | ||||
687 | /// Dominator Tree. | |||
688 | DominatorTree *DT; | |||
689 | ||||
690 | /// Alias Analysis. | |||
691 | AAResults *AA; | |||
692 | ||||
693 | /// Target Library Info. | |||
694 | const TargetLibraryInfo *TLI; | |||
695 | ||||
696 | /// Target Transform Info. | |||
697 | const TargetTransformInfo *TTI; | |||
698 | ||||
699 | /// Assumption Cache. | |||
700 | AssumptionCache *AC; | |||
701 | ||||
702 | /// Interface to emit optimization remarks. | |||
703 | OptimizationRemarkEmitter *ORE; | |||
704 | ||||
705 | /// LoopVersioning. It's only set up (non-null) if memchecks were | |||
706 | /// used. | |||
707 | /// | |||
708 | /// This is currently only used to add no-alias metadata based on the | |||
709 | /// memchecks. The actually versioning is performed manually. | |||
710 | std::unique_ptr<LoopVersioning> LVer; | |||
711 | ||||
712 | /// The vectorization SIMD factor to use. Each vector will have this many | |||
713 | /// vector elements. | |||
714 | ElementCount VF; | |||
715 | ||||
716 | /// The vectorization unroll factor to use. Each scalar is vectorized to this | |||
717 | /// many different vector instructions. | |||
718 | unsigned UF; | |||
719 | ||||
720 | /// The builder that we use | |||
721 | IRBuilder<> Builder; | |||
722 | ||||
723 | // --- Vectorization state --- | |||
724 | ||||
725 | /// The vector-loop preheader. | |||
726 | BasicBlock *LoopVectorPreHeader; | |||
727 | ||||
728 | /// The scalar-loop preheader. | |||
729 | BasicBlock *LoopScalarPreHeader; | |||
730 | ||||
731 | /// Middle Block between the vector and the scalar. | |||
732 | BasicBlock *LoopMiddleBlock; | |||
733 | ||||
734 | /// The unique ExitBlock of the scalar loop if one exists. Note that | |||
735 | /// there can be multiple exiting edges reaching this block. | |||
736 | BasicBlock *LoopExitBlock; | |||
737 | ||||
738 | /// The scalar loop body. | |||
739 | BasicBlock *LoopScalarBody; | |||
740 | ||||
741 | /// A list of all bypass blocks. The first block is the entry of the loop. | |||
742 | SmallVector<BasicBlock *, 4> LoopBypassBlocks; | |||
743 | ||||
744 | /// Store instructions that were predicated. | |||
745 | SmallVector<Instruction *, 4> PredicatedInstructions; | |||
746 | ||||
747 | /// Trip count of the original loop. | |||
748 | Value *TripCount = nullptr; | |||
749 | ||||
750 | /// Trip count of the widened loop (TripCount - TripCount % (VF*UF)) | |||
751 | Value *VectorTripCount = nullptr; | |||
752 | ||||
753 | /// The legality analysis. | |||
754 | LoopVectorizationLegality *Legal; | |||
755 | ||||
756 | /// The profitablity analysis. | |||
757 | LoopVectorizationCostModel *Cost; | |||
758 | ||||
759 | // Record whether runtime checks are added. | |||
760 | bool AddedSafetyChecks = false; | |||
761 | ||||
762 | // Holds the end values for each induction variable. We save the end values | |||
763 | // so we can later fix-up the external users of the induction variables. | |||
764 | DenseMap<PHINode *, Value *> IVEndValues; | |||
765 | ||||
766 | // Vector of original scalar PHIs whose corresponding widened PHIs need to be | |||
767 | // fixed up at the end of vector code generation. | |||
768 | SmallVector<PHINode *, 8> OrigPHIsToFix; | |||
769 | ||||
770 | /// BFI and PSI are used to check for profile guided size optimizations. | |||
771 | BlockFrequencyInfo *BFI; | |||
772 | ProfileSummaryInfo *PSI; | |||
773 | ||||
774 | // Whether this loop should be optimized for size based on profile guided size | |||
775 | // optimizatios. | |||
776 | bool OptForSizeBasedOnProfile; | |||
777 | ||||
778 | /// Structure to hold information about generated runtime checks, responsible | |||
779 | /// for cleaning the checks, if vectorization turns out unprofitable. | |||
780 | GeneratedRTChecks &RTChecks; | |||
781 | ||||
782 | // Holds the resume values for reductions in the loops, used to set the | |||
783 | // correct start value of reduction PHIs when vectorizing the epilogue. | |||
784 | SmallMapVector<const RecurrenceDescriptor *, PHINode *, 4> | |||
785 | ReductionResumeValues; | |||
786 | }; | |||
787 | ||||
788 | class InnerLoopUnroller : public InnerLoopVectorizer { | |||
789 | public: | |||
790 | InnerLoopUnroller(Loop *OrigLoop, PredicatedScalarEvolution &PSE, | |||
791 | LoopInfo *LI, DominatorTree *DT, | |||
792 | const TargetLibraryInfo *TLI, | |||
793 | const TargetTransformInfo *TTI, AssumptionCache *AC, | |||
794 | OptimizationRemarkEmitter *ORE, unsigned UnrollFactor, | |||
795 | LoopVectorizationLegality *LVL, | |||
796 | LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI, | |||
797 | ProfileSummaryInfo *PSI, GeneratedRTChecks &Check) | |||
798 | : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, | |||
799 | ElementCount::getFixed(1), UnrollFactor, LVL, CM, | |||
800 | BFI, PSI, Check) {} | |||
801 | ||||
802 | private: | |||
803 | Value *getBroadcastInstrs(Value *V) override; | |||
804 | }; | |||
805 | ||||
806 | /// Encapsulate information regarding vectorization of a loop and its epilogue. | |||
807 | /// This information is meant to be updated and used across two stages of | |||
808 | /// epilogue vectorization. | |||
809 | struct EpilogueLoopVectorizationInfo { | |||
810 | ElementCount MainLoopVF = ElementCount::getFixed(0); | |||
811 | unsigned MainLoopUF = 0; | |||
812 | ElementCount EpilogueVF = ElementCount::getFixed(0); | |||
813 | unsigned EpilogueUF = 0; | |||
814 | BasicBlock *MainLoopIterationCountCheck = nullptr; | |||
815 | BasicBlock *EpilogueIterationCountCheck = nullptr; | |||
816 | BasicBlock *SCEVSafetyCheck = nullptr; | |||
817 | BasicBlock *MemSafetyCheck = nullptr; | |||
818 | Value *TripCount = nullptr; | |||
819 | Value *VectorTripCount = nullptr; | |||
820 | ||||
821 | EpilogueLoopVectorizationInfo(ElementCount MVF, unsigned MUF, | |||
822 | ElementCount EVF, unsigned EUF) | |||
823 | : MainLoopVF(MVF), MainLoopUF(MUF), EpilogueVF(EVF), EpilogueUF(EUF) { | |||
824 | assert(EUF == 1 &&(static_cast <bool> (EUF == 1 && "A high UF for the epilogue loop is likely not beneficial." ) ? void (0) : __assert_fail ("EUF == 1 && \"A high UF for the epilogue loop is likely not beneficial.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 825, __extension__ __PRETTY_FUNCTION__)) | |||
825 | "A high UF for the epilogue loop is likely not beneficial.")(static_cast <bool> (EUF == 1 && "A high UF for the epilogue loop is likely not beneficial." ) ? void (0) : __assert_fail ("EUF == 1 && \"A high UF for the epilogue loop is likely not beneficial.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 825, __extension__ __PRETTY_FUNCTION__)); | |||
826 | } | |||
827 | }; | |||
828 | ||||
829 | /// An extension of the inner loop vectorizer that creates a skeleton for a | |||
830 | /// vectorized loop that has its epilogue (residual) also vectorized. | |||
831 | /// The idea is to run the vplan on a given loop twice, firstly to setup the | |||
832 | /// skeleton and vectorize the main loop, and secondly to complete the skeleton | |||
833 | /// from the first step and vectorize the epilogue. This is achieved by | |||
834 | /// deriving two concrete strategy classes from this base class and invoking | |||
835 | /// them in succession from the loop vectorizer planner. | |||
836 | class InnerLoopAndEpilogueVectorizer : public InnerLoopVectorizer { | |||
837 | public: | |||
838 | InnerLoopAndEpilogueVectorizer( | |||
839 | Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI, | |||
840 | DominatorTree *DT, const TargetLibraryInfo *TLI, | |||
841 | const TargetTransformInfo *TTI, AssumptionCache *AC, | |||
842 | OptimizationRemarkEmitter *ORE, EpilogueLoopVectorizationInfo &EPI, | |||
843 | LoopVectorizationLegality *LVL, llvm::LoopVectorizationCostModel *CM, | |||
844 | BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, | |||
845 | GeneratedRTChecks &Checks) | |||
846 | : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, | |||
847 | EPI.MainLoopVF, EPI.MainLoopUF, LVL, CM, BFI, PSI, | |||
848 | Checks), | |||
849 | EPI(EPI) {} | |||
850 | ||||
851 | // Override this function to handle the more complex control flow around the | |||
852 | // three loops. | |||
853 | std::pair<BasicBlock *, Value *> | |||
854 | createVectorizedLoopSkeleton() final override { | |||
855 | return createEpilogueVectorizedLoopSkeleton(); | |||
856 | } | |||
857 | ||||
858 | /// The interface for creating a vectorized skeleton using one of two | |||
859 | /// different strategies, each corresponding to one execution of the vplan | |||
860 | /// as described above. | |||
861 | virtual std::pair<BasicBlock *, Value *> | |||
862 | createEpilogueVectorizedLoopSkeleton() = 0; | |||
863 | ||||
864 | /// Holds and updates state information required to vectorize the main loop | |||
865 | /// and its epilogue in two separate passes. This setup helps us avoid | |||
866 | /// regenerating and recomputing runtime safety checks. It also helps us to | |||
867 | /// shorten the iteration-count-check path length for the cases where the | |||
868 | /// iteration count of the loop is so small that the main vector loop is | |||
869 | /// completely skipped. | |||
870 | EpilogueLoopVectorizationInfo &EPI; | |||
871 | }; | |||
872 | ||||
873 | /// A specialized derived class of inner loop vectorizer that performs | |||
874 | /// vectorization of *main* loops in the process of vectorizing loops and their | |||
875 | /// epilogues. | |||
876 | class EpilogueVectorizerMainLoop : public InnerLoopAndEpilogueVectorizer { | |||
877 | public: | |||
878 | EpilogueVectorizerMainLoop( | |||
879 | Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI, | |||
880 | DominatorTree *DT, const TargetLibraryInfo *TLI, | |||
881 | const TargetTransformInfo *TTI, AssumptionCache *AC, | |||
882 | OptimizationRemarkEmitter *ORE, EpilogueLoopVectorizationInfo &EPI, | |||
883 | LoopVectorizationLegality *LVL, llvm::LoopVectorizationCostModel *CM, | |||
884 | BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, | |||
885 | GeneratedRTChecks &Check) | |||
886 | : InnerLoopAndEpilogueVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, | |||
887 | EPI, LVL, CM, BFI, PSI, Check) {} | |||
888 | /// Implements the interface for creating a vectorized skeleton using the | |||
889 | /// *main loop* strategy (ie the first pass of vplan execution). | |||
890 | std::pair<BasicBlock *, Value *> | |||
891 | createEpilogueVectorizedLoopSkeleton() final override; | |||
892 | ||||
893 | protected: | |||
894 | /// Emits an iteration count bypass check once for the main loop (when \p | |||
895 | /// ForEpilogue is false) and once for the epilogue loop (when \p | |||
896 | /// ForEpilogue is true). | |||
897 | BasicBlock *emitMinimumIterationCountCheck(Loop *L, BasicBlock *Bypass, | |||
898 | bool ForEpilogue); | |||
899 | void printDebugTracesAtStart() override; | |||
900 | void printDebugTracesAtEnd() override; | |||
901 | }; | |||
902 | ||||
903 | // A specialized derived class of inner loop vectorizer that performs | |||
904 | // vectorization of *epilogue* loops in the process of vectorizing loops and | |||
905 | // their epilogues. | |||
906 | class EpilogueVectorizerEpilogueLoop : public InnerLoopAndEpilogueVectorizer { | |||
907 | public: | |||
908 | EpilogueVectorizerEpilogueLoop( | |||
909 | Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI, | |||
910 | DominatorTree *DT, const TargetLibraryInfo *TLI, | |||
911 | const TargetTransformInfo *TTI, AssumptionCache *AC, | |||
912 | OptimizationRemarkEmitter *ORE, EpilogueLoopVectorizationInfo &EPI, | |||
913 | LoopVectorizationLegality *LVL, llvm::LoopVectorizationCostModel *CM, | |||
914 | BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, | |||
915 | GeneratedRTChecks &Checks) | |||
916 | : InnerLoopAndEpilogueVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, | |||
917 | EPI, LVL, CM, BFI, PSI, Checks) {} | |||
918 | /// Implements the interface for creating a vectorized skeleton using the | |||
919 | /// *epilogue loop* strategy (ie the second pass of vplan execution). | |||
920 | std::pair<BasicBlock *, Value *> | |||
921 | createEpilogueVectorizedLoopSkeleton() final override; | |||
922 | ||||
923 | protected: | |||
924 | /// Emits an iteration count bypass check after the main vector loop has | |||
925 | /// finished to see if there are any iterations left to execute by either | |||
926 | /// the vector epilogue or the scalar epilogue. | |||
927 | BasicBlock *emitMinimumVectorEpilogueIterCountCheck( | |||
928 | BasicBlock *Bypass, | |||
929 | BasicBlock *Insert); | |||
930 | void printDebugTracesAtStart() override; | |||
931 | void printDebugTracesAtEnd() override; | |||
932 | }; | |||
933 | } // end namespace llvm | |||
934 | ||||
935 | /// Look for a meaningful debug location on the instruction or it's | |||
936 | /// operands. | |||
937 | static Instruction *getDebugLocFromInstOrOperands(Instruction *I) { | |||
938 | if (!I) | |||
939 | return I; | |||
940 | ||||
941 | DebugLoc Empty; | |||
942 | if (I->getDebugLoc() != Empty) | |||
943 | return I; | |||
944 | ||||
945 | for (Use &Op : I->operands()) { | |||
946 | if (Instruction *OpInst = dyn_cast<Instruction>(Op)) | |||
947 | if (OpInst->getDebugLoc() != Empty) | |||
948 | return OpInst; | |||
949 | } | |||
950 | ||||
951 | return I; | |||
952 | } | |||
953 | ||||
954 | void InnerLoopVectorizer::setDebugLocFromInst( | |||
955 | const Value *V, Optional<IRBuilderBase *> CustomBuilder) { | |||
956 | IRBuilderBase *B = (CustomBuilder == None) ? &Builder : *CustomBuilder; | |||
957 | if (const Instruction *Inst = dyn_cast_or_null<Instruction>(V)) { | |||
958 | const DILocation *DIL = Inst->getDebugLoc(); | |||
959 | ||||
960 | // When a FSDiscriminator is enabled, we don't need to add the multiply | |||
961 | // factors to the discriminators. | |||
962 | if (DIL && Inst->getFunction()->isDebugInfoForProfiling() && | |||
963 | !isa<DbgInfoIntrinsic>(Inst) && !EnableFSDiscriminator) { | |||
964 | // FIXME: For scalable vectors, assume vscale=1. | |||
965 | auto NewDIL = | |||
966 | DIL->cloneByMultiplyingDuplicationFactor(UF * VF.getKnownMinValue()); | |||
967 | if (NewDIL) | |||
968 | B->SetCurrentDebugLocation(NewDIL.getValue()); | |||
969 | else | |||
970 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Failed to create new discriminator: " << DIL->getFilename() << " Line: " << DIL ->getLine(); } } while (false) | |||
971 | << "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) | |||
972 | << 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); | |||
973 | } else | |||
974 | B->SetCurrentDebugLocation(DIL); | |||
975 | } else | |||
976 | B->SetCurrentDebugLocation(DebugLoc()); | |||
977 | } | |||
978 | ||||
979 | /// Write a \p DebugMsg about vectorization to the debug output stream. If \p I | |||
980 | /// is passed, the message relates to that particular instruction. | |||
981 | #ifndef NDEBUG | |||
982 | static void debugVectorizationMessage(const StringRef Prefix, | |||
983 | const StringRef DebugMsg, | |||
984 | Instruction *I) { | |||
985 | dbgs() << "LV: " << Prefix << DebugMsg; | |||
986 | if (I != nullptr) | |||
987 | dbgs() << " " << *I; | |||
988 | else | |||
989 | dbgs() << '.'; | |||
990 | dbgs() << '\n'; | |||
991 | } | |||
992 | #endif | |||
993 | ||||
994 | /// Create an analysis remark that explains why vectorization failed | |||
995 | /// | |||
996 | /// \p PassName is the name of the pass (e.g. can be AlwaysPrint). \p | |||
997 | /// RemarkName is the identifier for the remark. If \p I is passed it is an | |||
998 | /// instruction that prevents vectorization. Otherwise \p TheLoop is used for | |||
999 | /// the location of the remark. \return the remark object that can be | |||
1000 | /// streamed to. | |||
1001 | static OptimizationRemarkAnalysis createLVAnalysis(const char *PassName, | |||
1002 | StringRef RemarkName, Loop *TheLoop, Instruction *I) { | |||
1003 | Value *CodeRegion = TheLoop->getHeader(); | |||
1004 | DebugLoc DL = TheLoop->getStartLoc(); | |||
1005 | ||||
1006 | if (I) { | |||
1007 | CodeRegion = I->getParent(); | |||
1008 | // If there is no debug location attached to the instruction, revert back to | |||
1009 | // using the loop's. | |||
1010 | if (I->getDebugLoc()) | |||
1011 | DL = I->getDebugLoc(); | |||
1012 | } | |||
1013 | ||||
1014 | return OptimizationRemarkAnalysis(PassName, RemarkName, DL, CodeRegion); | |||
1015 | } | |||
1016 | ||||
1017 | namespace llvm { | |||
1018 | ||||
1019 | /// Return a value for Step multiplied by VF. | |||
1020 | Value *createStepForVF(IRBuilderBase &B, Type *Ty, ElementCount VF, | |||
1021 | int64_t Step) { | |||
1022 | assert(Ty->isIntegerTy() && "Expected an integer step")(static_cast <bool> (Ty->isIntegerTy() && "Expected an integer step" ) ? void (0) : __assert_fail ("Ty->isIntegerTy() && \"Expected an integer step\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1022, __extension__ __PRETTY_FUNCTION__)); | |||
1023 | Constant *StepVal = ConstantInt::get(Ty, Step * VF.getKnownMinValue()); | |||
1024 | return VF.isScalable() ? B.CreateVScale(StepVal) : StepVal; | |||
1025 | } | |||
1026 | ||||
1027 | /// Return the runtime value for VF. | |||
1028 | Value *getRuntimeVF(IRBuilderBase &B, Type *Ty, ElementCount VF) { | |||
1029 | Constant *EC = ConstantInt::get(Ty, VF.getKnownMinValue()); | |||
1030 | return VF.isScalable() ? B.CreateVScale(EC) : EC; | |||
1031 | } | |||
1032 | ||||
1033 | static Value *getRuntimeVFAsFloat(IRBuilderBase &B, Type *FTy, | |||
1034 | ElementCount VF) { | |||
1035 | assert(FTy->isFloatingPointTy() && "Expected floating point type!")(static_cast <bool> (FTy->isFloatingPointTy() && "Expected floating point type!") ? void (0) : __assert_fail ( "FTy->isFloatingPointTy() && \"Expected floating point type!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1035, __extension__ __PRETTY_FUNCTION__)); | |||
1036 | Type *IntTy = IntegerType::get(FTy->getContext(), FTy->getScalarSizeInBits()); | |||
1037 | Value *RuntimeVF = getRuntimeVF(B, IntTy, VF); | |||
1038 | return B.CreateUIToFP(RuntimeVF, FTy); | |||
1039 | } | |||
1040 | ||||
1041 | void reportVectorizationFailure(const StringRef DebugMsg, | |||
1042 | const StringRef OREMsg, const StringRef ORETag, | |||
1043 | OptimizationRemarkEmitter *ORE, Loop *TheLoop, | |||
1044 | Instruction *I) { | |||
1045 | LLVM_DEBUG(debugVectorizationMessage("Not vectorizing: ", DebugMsg, I))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { debugVectorizationMessage("Not vectorizing: " , DebugMsg, I); } } while (false); | |||
1046 | LoopVectorizeHints Hints(TheLoop, true /* doesn't matter */, *ORE); | |||
1047 | ORE->emit( | |||
1048 | createLVAnalysis(Hints.vectorizeAnalysisPassName(), ORETag, TheLoop, I) | |||
1049 | << "loop not vectorized: " << OREMsg); | |||
1050 | } | |||
1051 | ||||
1052 | void reportVectorizationInfo(const StringRef Msg, const StringRef ORETag, | |||
1053 | OptimizationRemarkEmitter *ORE, Loop *TheLoop, | |||
1054 | Instruction *I) { | |||
1055 | LLVM_DEBUG(debugVectorizationMessage("", Msg, I))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { debugVectorizationMessage("", Msg, I); } } while (false); | |||
1056 | LoopVectorizeHints Hints(TheLoop, true /* doesn't matter */, *ORE); | |||
1057 | ORE->emit( | |||
1058 | createLVAnalysis(Hints.vectorizeAnalysisPassName(), ORETag, TheLoop, I) | |||
1059 | << Msg); | |||
1060 | } | |||
1061 | ||||
1062 | } // end namespace llvm | |||
1063 | ||||
1064 | #ifndef NDEBUG | |||
1065 | /// \return string containing a file name and a line # for the given loop. | |||
1066 | static std::string getDebugLocString(const Loop *L) { | |||
1067 | std::string Result; | |||
1068 | if (L) { | |||
1069 | raw_string_ostream OS(Result); | |||
1070 | if (const DebugLoc LoopDbgLoc = L->getStartLoc()) | |||
1071 | LoopDbgLoc.print(OS); | |||
1072 | else | |||
1073 | // Just print the module name. | |||
1074 | OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier(); | |||
1075 | OS.flush(); | |||
1076 | } | |||
1077 | return Result; | |||
1078 | } | |||
1079 | #endif | |||
1080 | ||||
1081 | void InnerLoopVectorizer::addNewMetadata(Instruction *To, | |||
1082 | const Instruction *Orig) { | |||
1083 | // If the loop was versioned with memchecks, add the corresponding no-alias | |||
1084 | // metadata. | |||
1085 | if (LVer && (isa<LoadInst>(Orig) || isa<StoreInst>(Orig))) | |||
1086 | LVer->annotateInstWithNoAlias(To, Orig); | |||
1087 | } | |||
1088 | ||||
1089 | void InnerLoopVectorizer::collectPoisonGeneratingRecipes( | |||
1090 | VPTransformState &State) { | |||
1091 | ||||
1092 | // Collect recipes in the backward slice of `Root` that may generate a poison | |||
1093 | // value that is used after vectorization. | |||
1094 | SmallPtrSet<VPRecipeBase *, 16> Visited; | |||
1095 | auto collectPoisonGeneratingInstrsInBackwardSlice([&](VPRecipeBase *Root) { | |||
1096 | SmallVector<VPRecipeBase *, 16> Worklist; | |||
1097 | Worklist.push_back(Root); | |||
1098 | ||||
1099 | // Traverse the backward slice of Root through its use-def chain. | |||
1100 | while (!Worklist.empty()) { | |||
1101 | VPRecipeBase *CurRec = Worklist.back(); | |||
1102 | Worklist.pop_back(); | |||
1103 | ||||
1104 | if (!Visited.insert(CurRec).second) | |||
1105 | continue; | |||
1106 | ||||
1107 | // Prune search if we find another recipe generating a widen memory | |||
1108 | // instruction. Widen memory instructions involved in address computation | |||
1109 | // will lead to gather/scatter instructions, which don't need to be | |||
1110 | // handled. | |||
1111 | if (isa<VPWidenMemoryInstructionRecipe>(CurRec) || | |||
1112 | isa<VPInterleaveRecipe>(CurRec) || | |||
1113 | isa<VPScalarIVStepsRecipe>(CurRec) || | |||
1114 | isa<VPCanonicalIVPHIRecipe>(CurRec)) | |||
1115 | continue; | |||
1116 | ||||
1117 | // This recipe contributes to the address computation of a widen | |||
1118 | // load/store. Collect recipe if its underlying instruction has | |||
1119 | // poison-generating flags. | |||
1120 | Instruction *Instr = CurRec->getUnderlyingInstr(); | |||
1121 | if (Instr && Instr->hasPoisonGeneratingFlags()) | |||
1122 | State.MayGeneratePoisonRecipes.insert(CurRec); | |||
1123 | ||||
1124 | // Add new definitions to the worklist. | |||
1125 | for (VPValue *operand : CurRec->operands()) | |||
1126 | if (VPDef *OpDef = operand->getDef()) | |||
1127 | Worklist.push_back(cast<VPRecipeBase>(OpDef)); | |||
1128 | } | |||
1129 | }); | |||
1130 | ||||
1131 | // Traverse all the recipes in the VPlan and collect the poison-generating | |||
1132 | // recipes in the backward slice starting at the address of a VPWidenRecipe or | |||
1133 | // VPInterleaveRecipe. | |||
1134 | auto Iter = depth_first( | |||
1135 | VPBlockRecursiveTraversalWrapper<VPBlockBase *>(State.Plan->getEntry())); | |||
1136 | for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(Iter)) { | |||
1137 | for (VPRecipeBase &Recipe : *VPBB) { | |||
1138 | if (auto *WidenRec = dyn_cast<VPWidenMemoryInstructionRecipe>(&Recipe)) { | |||
1139 | Instruction *UnderlyingInstr = WidenRec->getUnderlyingInstr(); | |||
1140 | VPDef *AddrDef = WidenRec->getAddr()->getDef(); | |||
1141 | if (AddrDef && WidenRec->isConsecutive() && UnderlyingInstr && | |||
1142 | Legal->blockNeedsPredication(UnderlyingInstr->getParent())) | |||
1143 | collectPoisonGeneratingInstrsInBackwardSlice( | |||
1144 | cast<VPRecipeBase>(AddrDef)); | |||
1145 | } else if (auto *InterleaveRec = dyn_cast<VPInterleaveRecipe>(&Recipe)) { | |||
1146 | VPDef *AddrDef = InterleaveRec->getAddr()->getDef(); | |||
1147 | if (AddrDef) { | |||
1148 | // Check if any member of the interleave group needs predication. | |||
1149 | const InterleaveGroup<Instruction> *InterGroup = | |||
1150 | InterleaveRec->getInterleaveGroup(); | |||
1151 | bool NeedPredication = false; | |||
1152 | for (int I = 0, NumMembers = InterGroup->getNumMembers(); | |||
1153 | I < NumMembers; ++I) { | |||
1154 | Instruction *Member = InterGroup->getMember(I); | |||
1155 | if (Member) | |||
1156 | NeedPredication |= | |||
1157 | Legal->blockNeedsPredication(Member->getParent()); | |||
1158 | } | |||
1159 | ||||
1160 | if (NeedPredication) | |||
1161 | collectPoisonGeneratingInstrsInBackwardSlice( | |||
1162 | cast<VPRecipeBase>(AddrDef)); | |||
1163 | } | |||
1164 | } | |||
1165 | } | |||
1166 | } | |||
1167 | } | |||
1168 | ||||
1169 | void InnerLoopVectorizer::addMetadata(Instruction *To, | |||
1170 | Instruction *From) { | |||
1171 | propagateMetadata(To, From); | |||
1172 | addNewMetadata(To, From); | |||
1173 | } | |||
1174 | ||||
1175 | void InnerLoopVectorizer::addMetadata(ArrayRef<Value *> To, | |||
1176 | Instruction *From) { | |||
1177 | for (Value *V : To) { | |||
1178 | if (Instruction *I = dyn_cast<Instruction>(V)) | |||
1179 | addMetadata(I, From); | |||
1180 | } | |||
1181 | } | |||
1182 | ||||
1183 | PHINode *InnerLoopVectorizer::getReductionResumeValue( | |||
1184 | const RecurrenceDescriptor &RdxDesc) { | |||
1185 | auto It = ReductionResumeValues.find(&RdxDesc); | |||
1186 | assert(It != ReductionResumeValues.end() &&(static_cast <bool> (It != ReductionResumeValues.end() && "Expected to find a resume value for the reduction.") ? void (0) : __assert_fail ("It != ReductionResumeValues.end() && \"Expected to find a resume value for the reduction.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1187, __extension__ __PRETTY_FUNCTION__)) | |||
1187 | "Expected to find a resume value for the reduction.")(static_cast <bool> (It != ReductionResumeValues.end() && "Expected to find a resume value for the reduction.") ? void (0) : __assert_fail ("It != ReductionResumeValues.end() && \"Expected to find a resume value for the reduction.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1187, __extension__ __PRETTY_FUNCTION__)); | |||
1188 | return It->second; | |||
1189 | } | |||
1190 | ||||
1191 | namespace llvm { | |||
1192 | ||||
1193 | // Loop vectorization cost-model hints how the scalar epilogue loop should be | |||
1194 | // lowered. | |||
1195 | enum ScalarEpilogueLowering { | |||
1196 | ||||
1197 | // The default: allowing scalar epilogues. | |||
1198 | CM_ScalarEpilogueAllowed, | |||
1199 | ||||
1200 | // Vectorization with OptForSize: don't allow epilogues. | |||
1201 | CM_ScalarEpilogueNotAllowedOptSize, | |||
1202 | ||||
1203 | // A special case of vectorisation with OptForSize: loops with a very small | |||
1204 | // trip count are considered for vectorization under OptForSize, thereby | |||
1205 | // making sure the cost of their loop body is dominant, free of runtime | |||
1206 | // guards and scalar iteration overheads. | |||
1207 | CM_ScalarEpilogueNotAllowedLowTripLoop, | |||
1208 | ||||
1209 | // Loop hint predicate indicating an epilogue is undesired. | |||
1210 | CM_ScalarEpilogueNotNeededUsePredicate, | |||
1211 | ||||
1212 | // Directive indicating we must either tail fold or not vectorize | |||
1213 | CM_ScalarEpilogueNotAllowedUsePredicate | |||
1214 | }; | |||
1215 | ||||
1216 | /// ElementCountComparator creates a total ordering for ElementCount | |||
1217 | /// for the purposes of using it in a set structure. | |||
1218 | struct ElementCountComparator { | |||
1219 | bool operator()(const ElementCount &LHS, const ElementCount &RHS) const { | |||
1220 | return std::make_tuple(LHS.isScalable(), LHS.getKnownMinValue()) < | |||
1221 | std::make_tuple(RHS.isScalable(), RHS.getKnownMinValue()); | |||
1222 | } | |||
1223 | }; | |||
1224 | using ElementCountSet = SmallSet<ElementCount, 16, ElementCountComparator>; | |||
1225 | ||||
1226 | /// LoopVectorizationCostModel - estimates the expected speedups due to | |||
1227 | /// vectorization. | |||
1228 | /// In many cases vectorization is not profitable. This can happen because of | |||
1229 | /// a number of reasons. In this class we mainly attempt to predict the | |||
1230 | /// expected speedup/slowdowns due to the supported instruction set. We use the | |||
1231 | /// TargetTransformInfo to query the different backends for the cost of | |||
1232 | /// different operations. | |||
1233 | class LoopVectorizationCostModel { | |||
1234 | public: | |||
1235 | LoopVectorizationCostModel(ScalarEpilogueLowering SEL, Loop *L, | |||
1236 | PredicatedScalarEvolution &PSE, LoopInfo *LI, | |||
1237 | LoopVectorizationLegality *Legal, | |||
1238 | const TargetTransformInfo &TTI, | |||
1239 | const TargetLibraryInfo *TLI, DemandedBits *DB, | |||
1240 | AssumptionCache *AC, | |||
1241 | OptimizationRemarkEmitter *ORE, const Function *F, | |||
1242 | const LoopVectorizeHints *Hints, | |||
1243 | InterleavedAccessInfo &IAI) | |||
1244 | : ScalarEpilogueStatus(SEL), TheLoop(L), PSE(PSE), LI(LI), Legal(Legal), | |||
1245 | TTI(TTI), TLI(TLI), DB(DB), AC(AC), ORE(ORE), TheFunction(F), | |||
1246 | Hints(Hints), InterleaveInfo(IAI) {} | |||
1247 | ||||
1248 | /// \return An upper bound for the vectorization factors (both fixed and | |||
1249 | /// scalable). If the factors are 0, vectorization and interleaving should be | |||
1250 | /// avoided up front. | |||
1251 | FixedScalableVFPair computeMaxVF(ElementCount UserVF, unsigned UserIC); | |||
1252 | ||||
1253 | /// \return True if runtime checks are required for vectorization, and false | |||
1254 | /// otherwise. | |||
1255 | bool runtimeChecksRequired(); | |||
1256 | ||||
1257 | /// \return The most profitable vectorization factor and the cost of that VF. | |||
1258 | /// This method checks every VF in \p CandidateVFs. If UserVF is not ZERO | |||
1259 | /// then this vectorization factor will be selected if vectorization is | |||
1260 | /// possible. | |||
1261 | VectorizationFactor | |||
1262 | selectVectorizationFactor(const ElementCountSet &CandidateVFs); | |||
1263 | ||||
1264 | VectorizationFactor | |||
1265 | selectEpilogueVectorizationFactor(const ElementCount MaxVF, | |||
1266 | const LoopVectorizationPlanner &LVP); | |||
1267 | ||||
1268 | /// Setup cost-based decisions for user vectorization factor. | |||
1269 | /// \return true if the UserVF is a feasible VF to be chosen. | |||
1270 | bool selectUserVectorizationFactor(ElementCount UserVF) { | |||
1271 | collectUniformsAndScalars(UserVF); | |||
1272 | collectInstsToScalarize(UserVF); | |||
1273 | return expectedCost(UserVF).first.isValid(); | |||
1274 | } | |||
1275 | ||||
1276 | /// \return The size (in bits) of the smallest and widest types in the code | |||
1277 | /// that needs to be vectorized. We ignore values that remain scalar such as | |||
1278 | /// 64 bit loop indices. | |||
1279 | std::pair<unsigned, unsigned> getSmallestAndWidestTypes(); | |||
1280 | ||||
1281 | /// \return The desired interleave count. | |||
1282 | /// If interleave count has been specified by metadata it will be returned. | |||
1283 | /// Otherwise, the interleave count is computed and returned. VF and LoopCost | |||
1284 | /// are the selected vectorization factor and the cost of the selected VF. | |||
1285 | unsigned selectInterleaveCount(ElementCount VF, unsigned LoopCost); | |||
1286 | ||||
1287 | /// Memory access instruction may be vectorized in more than one way. | |||
1288 | /// Form of instruction after vectorization depends on cost. | |||
1289 | /// This function takes cost-based decisions for Load/Store instructions | |||
1290 | /// and collects them in a map. This decisions map is used for building | |||
1291 | /// the lists of loop-uniform and loop-scalar instructions. | |||
1292 | /// The calculated cost is saved with widening decision in order to | |||
1293 | /// avoid redundant calculations. | |||
1294 | void setCostBasedWideningDecision(ElementCount VF); | |||
1295 | ||||
1296 | /// A struct that represents some properties of the register usage | |||
1297 | /// of a loop. | |||
1298 | struct RegisterUsage { | |||
1299 | /// Holds the number of loop invariant values that are used in the loop. | |||
1300 | /// The key is ClassID of target-provided register class. | |||
1301 | SmallMapVector<unsigned, unsigned, 4> LoopInvariantRegs; | |||
1302 | /// Holds the maximum number of concurrent live intervals in the loop. | |||
1303 | /// The key is ClassID of target-provided register class. | |||
1304 | SmallMapVector<unsigned, unsigned, 4> MaxLocalUsers; | |||
1305 | }; | |||
1306 | ||||
1307 | /// \return Returns information about the register usages of the loop for the | |||
1308 | /// given vectorization factors. | |||
1309 | SmallVector<RegisterUsage, 8> | |||
1310 | calculateRegisterUsage(ArrayRef<ElementCount> VFs); | |||
1311 | ||||
1312 | /// Collect values we want to ignore in the cost model. | |||
1313 | void collectValuesToIgnore(); | |||
1314 | ||||
1315 | /// Collect all element types in the loop for which widening is needed. | |||
1316 | void collectElementTypesForWidening(); | |||
1317 | ||||
1318 | /// Split reductions into those that happen in the loop, and those that happen | |||
1319 | /// outside. In loop reductions are collected into InLoopReductionChains. | |||
1320 | void collectInLoopReductions(); | |||
1321 | ||||
1322 | /// Returns true if we should use strict in-order reductions for the given | |||
1323 | /// RdxDesc. This is true if the -enable-strict-reductions flag is passed, | |||
1324 | /// the IsOrdered flag of RdxDesc is set and we do not allow reordering | |||
1325 | /// of FP operations. | |||
1326 | bool useOrderedReductions(const RecurrenceDescriptor &RdxDesc) { | |||
1327 | return !Hints->allowReordering() && RdxDesc.isOrdered(); | |||
1328 | } | |||
1329 | ||||
1330 | /// \returns The smallest bitwidth each instruction can be represented with. | |||
1331 | /// The vector equivalents of these instructions should be truncated to this | |||
1332 | /// type. | |||
1333 | const MapVector<Instruction *, uint64_t> &getMinimalBitwidths() const { | |||
1334 | return MinBWs; | |||
1335 | } | |||
1336 | ||||
1337 | /// \returns True if it is more profitable to scalarize instruction \p I for | |||
1338 | /// vectorization factor \p VF. | |||
1339 | bool isProfitableToScalarize(Instruction *I, ElementCount VF) const { | |||
1340 | assert(VF.isVector() &&(static_cast <bool> (VF.isVector() && "Profitable to scalarize relevant only for VF > 1." ) ? void (0) : __assert_fail ("VF.isVector() && \"Profitable to scalarize relevant only for VF > 1.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1341, __extension__ __PRETTY_FUNCTION__)) | |||
1341 | "Profitable to scalarize relevant only for VF > 1.")(static_cast <bool> (VF.isVector() && "Profitable to scalarize relevant only for VF > 1." ) ? void (0) : __assert_fail ("VF.isVector() && \"Profitable to scalarize relevant only for VF > 1.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1341, __extension__ __PRETTY_FUNCTION__)); | |||
1342 | ||||
1343 | // Cost model is not run in the VPlan-native path - return conservative | |||
1344 | // result until this changes. | |||
1345 | if (EnableVPlanNativePath) | |||
1346 | return false; | |||
1347 | ||||
1348 | auto Scalars = InstsToScalarize.find(VF); | |||
1349 | assert(Scalars != InstsToScalarize.end() &&(static_cast <bool> (Scalars != InstsToScalarize.end() && "VF not yet analyzed for scalarization profitability") ? void (0) : __assert_fail ("Scalars != InstsToScalarize.end() && \"VF not yet analyzed for scalarization profitability\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1350, __extension__ __PRETTY_FUNCTION__)) | |||
1350 | "VF not yet analyzed for scalarization profitability")(static_cast <bool> (Scalars != InstsToScalarize.end() && "VF not yet analyzed for scalarization profitability") ? void (0) : __assert_fail ("Scalars != InstsToScalarize.end() && \"VF not yet analyzed for scalarization profitability\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1350, __extension__ __PRETTY_FUNCTION__)); | |||
1351 | return Scalars->second.find(I) != Scalars->second.end(); | |||
1352 | } | |||
1353 | ||||
1354 | /// Returns true if \p I is known to be uniform after vectorization. | |||
1355 | bool isUniformAfterVectorization(Instruction *I, ElementCount VF) const { | |||
1356 | if (VF.isScalar()) | |||
1357 | return true; | |||
1358 | ||||
1359 | // Cost model is not run in the VPlan-native path - return conservative | |||
1360 | // result until this changes. | |||
1361 | if (EnableVPlanNativePath) | |||
1362 | return false; | |||
1363 | ||||
1364 | auto UniformsPerVF = Uniforms.find(VF); | |||
1365 | assert(UniformsPerVF != Uniforms.end() &&(static_cast <bool> (UniformsPerVF != Uniforms.end() && "VF not yet analyzed for uniformity") ? void (0) : __assert_fail ("UniformsPerVF != Uniforms.end() && \"VF not yet analyzed for uniformity\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1366, __extension__ __PRETTY_FUNCTION__)) | |||
1366 | "VF not yet analyzed for uniformity")(static_cast <bool> (UniformsPerVF != Uniforms.end() && "VF not yet analyzed for uniformity") ? void (0) : __assert_fail ("UniformsPerVF != Uniforms.end() && \"VF not yet analyzed for uniformity\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1366, __extension__ __PRETTY_FUNCTION__)); | |||
1367 | return UniformsPerVF->second.count(I); | |||
1368 | } | |||
1369 | ||||
1370 | /// Returns true if \p I is known to be scalar after vectorization. | |||
1371 | bool isScalarAfterVectorization(Instruction *I, ElementCount VF) const { | |||
1372 | if (VF.isScalar()) | |||
1373 | return true; | |||
1374 | ||||
1375 | // Cost model is not run in the VPlan-native path - return conservative | |||
1376 | // result until this changes. | |||
1377 | if (EnableVPlanNativePath) | |||
1378 | return false; | |||
1379 | ||||
1380 | auto ScalarsPerVF = Scalars.find(VF); | |||
1381 | assert(ScalarsPerVF != Scalars.end() &&(static_cast <bool> (ScalarsPerVF != Scalars.end() && "Scalar values are not calculated for VF") ? void (0) : __assert_fail ("ScalarsPerVF != Scalars.end() && \"Scalar values are not calculated for VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1382, __extension__ __PRETTY_FUNCTION__)) | |||
1382 | "Scalar values are not calculated for VF")(static_cast <bool> (ScalarsPerVF != Scalars.end() && "Scalar values are not calculated for VF") ? void (0) : __assert_fail ("ScalarsPerVF != Scalars.end() && \"Scalar values are not calculated for VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1382, __extension__ __PRETTY_FUNCTION__)); | |||
1383 | return ScalarsPerVF->second.count(I); | |||
1384 | } | |||
1385 | ||||
1386 | /// \returns True if instruction \p I can be truncated to a smaller bitwidth | |||
1387 | /// for vectorization factor \p VF. | |||
1388 | bool canTruncateToMinimalBitwidth(Instruction *I, ElementCount VF) const { | |||
1389 | return VF.isVector() && MinBWs.find(I) != MinBWs.end() && | |||
1390 | !isProfitableToScalarize(I, VF) && | |||
1391 | !isScalarAfterVectorization(I, VF); | |||
1392 | } | |||
1393 | ||||
1394 | /// Decision that was taken during cost calculation for memory instruction. | |||
1395 | enum InstWidening { | |||
1396 | CM_Unknown, | |||
1397 | CM_Widen, // For consecutive accesses with stride +1. | |||
1398 | CM_Widen_Reverse, // For consecutive accesses with stride -1. | |||
1399 | CM_Interleave, | |||
1400 | CM_GatherScatter, | |||
1401 | CM_Scalarize | |||
1402 | }; | |||
1403 | ||||
1404 | /// Save vectorization decision \p W and \p Cost taken by the cost model for | |||
1405 | /// instruction \p I and vector width \p VF. | |||
1406 | void setWideningDecision(Instruction *I, ElementCount VF, InstWidening W, | |||
1407 | InstructionCost Cost) { | |||
1408 | assert(VF.isVector() && "Expected VF >=2")(static_cast <bool> (VF.isVector() && "Expected VF >=2" ) ? void (0) : __assert_fail ("VF.isVector() && \"Expected VF >=2\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1408, __extension__ __PRETTY_FUNCTION__)); | |||
1409 | WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost); | |||
1410 | } | |||
1411 | ||||
1412 | /// Save vectorization decision \p W and \p Cost taken by the cost model for | |||
1413 | /// interleaving group \p Grp and vector width \p VF. | |||
1414 | void setWideningDecision(const InterleaveGroup<Instruction> *Grp, | |||
1415 | ElementCount VF, InstWidening W, | |||
1416 | InstructionCost Cost) { | |||
1417 | assert(VF.isVector() && "Expected VF >=2")(static_cast <bool> (VF.isVector() && "Expected VF >=2" ) ? void (0) : __assert_fail ("VF.isVector() && \"Expected VF >=2\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1417, __extension__ __PRETTY_FUNCTION__)); | |||
1418 | /// Broadcast this decicion to all instructions inside the group. | |||
1419 | /// But the cost will be assigned to one instruction only. | |||
1420 | for (unsigned i = 0; i < Grp->getFactor(); ++i) { | |||
1421 | if (auto *I = Grp->getMember(i)) { | |||
1422 | if (Grp->getInsertPos() == I) | |||
1423 | WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost); | |||
1424 | else | |||
1425 | WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, 0); | |||
1426 | } | |||
1427 | } | |||
1428 | } | |||
1429 | ||||
1430 | /// Return the cost model decision for the given instruction \p I and vector | |||
1431 | /// width \p VF. Return CM_Unknown if this instruction did not pass | |||
1432 | /// through the cost modeling. | |||
1433 | InstWidening getWideningDecision(Instruction *I, ElementCount VF) const { | |||
1434 | assert(VF.isVector() && "Expected VF to be a vector VF")(static_cast <bool> (VF.isVector() && "Expected VF to be a vector VF" ) ? void (0) : __assert_fail ("VF.isVector() && \"Expected VF to be a vector VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1434, __extension__ __PRETTY_FUNCTION__)); | |||
1435 | // Cost model is not run in the VPlan-native path - return conservative | |||
1436 | // result until this changes. | |||
1437 | if (EnableVPlanNativePath) | |||
1438 | return CM_GatherScatter; | |||
1439 | ||||
1440 | std::pair<Instruction *, ElementCount> InstOnVF = std::make_pair(I, VF); | |||
1441 | auto Itr = WideningDecisions.find(InstOnVF); | |||
1442 | if (Itr == WideningDecisions.end()) | |||
1443 | return CM_Unknown; | |||
1444 | return Itr->second.first; | |||
1445 | } | |||
1446 | ||||
1447 | /// Return the vectorization cost for the given instruction \p I and vector | |||
1448 | /// width \p VF. | |||
1449 | InstructionCost getWideningCost(Instruction *I, ElementCount VF) { | |||
1450 | assert(VF.isVector() && "Expected VF >=2")(static_cast <bool> (VF.isVector() && "Expected VF >=2" ) ? void (0) : __assert_fail ("VF.isVector() && \"Expected VF >=2\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1450, __extension__ __PRETTY_FUNCTION__)); | |||
1451 | std::pair<Instruction *, ElementCount> InstOnVF = std::make_pair(I, VF); | |||
1452 | assert(WideningDecisions.find(InstOnVF) != WideningDecisions.end() &&(static_cast <bool> (WideningDecisions.find(InstOnVF) != WideningDecisions.end() && "The cost is not calculated" ) ? void (0) : __assert_fail ("WideningDecisions.find(InstOnVF) != WideningDecisions.end() && \"The cost is not calculated\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1453, __extension__ __PRETTY_FUNCTION__)) | |||
1453 | "The cost is not calculated")(static_cast <bool> (WideningDecisions.find(InstOnVF) != WideningDecisions.end() && "The cost is not calculated" ) ? void (0) : __assert_fail ("WideningDecisions.find(InstOnVF) != WideningDecisions.end() && \"The cost is not calculated\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1453, __extension__ __PRETTY_FUNCTION__)); | |||
1454 | return WideningDecisions[InstOnVF].second; | |||
1455 | } | |||
1456 | ||||
1457 | /// Return True if instruction \p I is an optimizable truncate whose operand | |||
1458 | /// is an induction variable. Such a truncate will be removed by adding a new | |||
1459 | /// induction variable with the destination type. | |||
1460 | bool isOptimizableIVTruncate(Instruction *I, ElementCount VF) { | |||
1461 | // If the instruction is not a truncate, return false. | |||
1462 | auto *Trunc = dyn_cast<TruncInst>(I); | |||
1463 | if (!Trunc) | |||
1464 | return false; | |||
1465 | ||||
1466 | // Get the source and destination types of the truncate. | |||
1467 | Type *SrcTy = ToVectorTy(cast<CastInst>(I)->getSrcTy(), VF); | |||
1468 | Type *DestTy = ToVectorTy(cast<CastInst>(I)->getDestTy(), VF); | |||
1469 | ||||
1470 | // If the truncate is free for the given types, return false. Replacing a | |||
1471 | // free truncate with an induction variable would add an induction variable | |||
1472 | // update instruction to each iteration of the loop. We exclude from this | |||
1473 | // check the primary induction variable since it will need an update | |||
1474 | // instruction regardless. | |||
1475 | Value *Op = Trunc->getOperand(0); | |||
1476 | if (Op != Legal->getPrimaryInduction() && TTI.isTruncateFree(SrcTy, DestTy)) | |||
1477 | return false; | |||
1478 | ||||
1479 | // If the truncated value is not an induction variable, return false. | |||
1480 | return Legal->isInductionPhi(Op); | |||
1481 | } | |||
1482 | ||||
1483 | /// Collects the instructions to scalarize for each predicated instruction in | |||
1484 | /// the loop. | |||
1485 | void collectInstsToScalarize(ElementCount VF); | |||
1486 | ||||
1487 | /// Collect Uniform and Scalar values for the given \p VF. | |||
1488 | /// The sets depend on CM decision for Load/Store instructions | |||
1489 | /// that may be vectorized as interleave, gather-scatter or scalarized. | |||
1490 | void collectUniformsAndScalars(ElementCount VF) { | |||
1491 | // Do the analysis once. | |||
1492 | if (VF.isScalar() || Uniforms.find(VF) != Uniforms.end()) | |||
1493 | return; | |||
1494 | setCostBasedWideningDecision(VF); | |||
1495 | collectLoopUniforms(VF); | |||
1496 | collectLoopScalars(VF); | |||
1497 | } | |||
1498 | ||||
1499 | /// Returns true if the target machine supports masked store operation | |||
1500 | /// for the given \p DataType and kind of access to \p Ptr. | |||
1501 | bool isLegalMaskedStore(Type *DataType, Value *Ptr, Align Alignment) const { | |||
1502 | return Legal->isConsecutivePtr(DataType, Ptr) && | |||
1503 | TTI.isLegalMaskedStore(DataType, Alignment); | |||
1504 | } | |||
1505 | ||||
1506 | /// Returns true if the target machine supports masked load operation | |||
1507 | /// for the given \p DataType and kind of access to \p Ptr. | |||
1508 | bool isLegalMaskedLoad(Type *DataType, Value *Ptr, Align Alignment) const { | |||
1509 | return Legal->isConsecutivePtr(DataType, Ptr) && | |||
1510 | TTI.isLegalMaskedLoad(DataType, Alignment); | |||
1511 | } | |||
1512 | ||||
1513 | /// Returns true if the target machine can represent \p V as a masked gather | |||
1514 | /// or scatter operation. | |||
1515 | bool isLegalGatherOrScatter(Value *V, | |||
1516 | ElementCount VF = ElementCount::getFixed(1)) { | |||
1517 | bool LI = isa<LoadInst>(V); | |||
1518 | bool SI = isa<StoreInst>(V); | |||
1519 | if (!LI && !SI) | |||
1520 | return false; | |||
1521 | auto *Ty = getLoadStoreType(V); | |||
1522 | Align Align = getLoadStoreAlignment(V); | |||
1523 | if (VF.isVector()) | |||
1524 | Ty = VectorType::get(Ty, VF); | |||
1525 | return (LI && TTI.isLegalMaskedGather(Ty, Align)) || | |||
1526 | (SI && TTI.isLegalMaskedScatter(Ty, Align)); | |||
1527 | } | |||
1528 | ||||
1529 | /// Returns true if the target machine supports all of the reduction | |||
1530 | /// variables found for the given VF. | |||
1531 | bool canVectorizeReductions(ElementCount VF) const { | |||
1532 | return (all_of(Legal->getReductionVars(), [&](auto &Reduction) -> bool { | |||
1533 | const RecurrenceDescriptor &RdxDesc = Reduction.second; | |||
1534 | return TTI.isLegalToVectorizeReduction(RdxDesc, VF); | |||
1535 | })); | |||
1536 | } | |||
1537 | ||||
1538 | /// Returns true if \p I is an instruction that will be scalarized with | |||
1539 | /// predication when vectorizing \p I with vectorization factor \p VF. Such | |||
1540 | /// instructions include conditional stores and instructions that may divide | |||
1541 | /// by zero. | |||
1542 | bool isScalarWithPredication(Instruction *I, ElementCount VF) const; | |||
1543 | ||||
1544 | // Returns true if \p I is an instruction that will be predicated either | |||
1545 | // through scalar predication or masked load/store or masked gather/scatter. | |||
1546 | // \p VF is the vectorization factor that will be used to vectorize \p I. | |||
1547 | // Superset of instructions that return true for isScalarWithPredication. | |||
1548 | bool isPredicatedInst(Instruction *I, ElementCount VF, | |||
1549 | bool IsKnownUniform = false) { | |||
1550 | // When we know the load is uniform and the original scalar loop was not | |||
1551 | // predicated we don't need to mark it as a predicated instruction. Any | |||
1552 | // vectorised blocks created when tail-folding are something artificial we | |||
1553 | // have introduced and we know there is always at least one active lane. | |||
1554 | // That's why we call Legal->blockNeedsPredication here because it doesn't | |||
1555 | // query tail-folding. | |||
1556 | if (IsKnownUniform && isa<LoadInst>(I) && | |||
1557 | !Legal->blockNeedsPredication(I->getParent())) | |||
1558 | return false; | |||
1559 | if (!blockNeedsPredicationForAnyReason(I->getParent())) | |||
1560 | return false; | |||
1561 | // Loads and stores that need some form of masked operation are predicated | |||
1562 | // instructions. | |||
1563 | if (isa<LoadInst>(I) || isa<StoreInst>(I)) | |||
1564 | return Legal->isMaskRequired(I); | |||
1565 | return isScalarWithPredication(I, VF); | |||
1566 | } | |||
1567 | ||||
1568 | /// Returns true if \p I is a memory instruction with consecutive memory | |||
1569 | /// access that can be widened. | |||
1570 | bool | |||
1571 | memoryInstructionCanBeWidened(Instruction *I, | |||
1572 | ElementCount VF = ElementCount::getFixed(1)); | |||
1573 | ||||
1574 | /// Returns true if \p I is a memory instruction in an interleaved-group | |||
1575 | /// of memory accesses that can be vectorized with wide vector loads/stores | |||
1576 | /// and shuffles. | |||
1577 | bool | |||
1578 | interleavedAccessCanBeWidened(Instruction *I, | |||
1579 | ElementCount VF = ElementCount::getFixed(1)); | |||
1580 | ||||
1581 | /// Check if \p Instr belongs to any interleaved access group. | |||
1582 | bool isAccessInterleaved(Instruction *Instr) { | |||
1583 | return InterleaveInfo.isInterleaved(Instr); | |||
1584 | } | |||
1585 | ||||
1586 | /// Get the interleaved access group that \p Instr belongs to. | |||
1587 | const InterleaveGroup<Instruction> * | |||
1588 | getInterleavedAccessGroup(Instruction *Instr) { | |||
1589 | return InterleaveInfo.getInterleaveGroup(Instr); | |||
1590 | } | |||
1591 | ||||
1592 | /// Returns true if we're required to use a scalar epilogue for at least | |||
1593 | /// the final iteration of the original loop. | |||
1594 | bool requiresScalarEpilogue(ElementCount VF) const { | |||
1595 | if (!isScalarEpilogueAllowed()) | |||
1596 | return false; | |||
1597 | // If we might exit from anywhere but the latch, must run the exiting | |||
1598 | // iteration in scalar form. | |||
1599 | if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) | |||
1600 | return true; | |||
1601 | return VF.isVector() && InterleaveInfo.requiresScalarEpilogue(); | |||
1602 | } | |||
1603 | ||||
1604 | /// Returns true if a scalar epilogue is not allowed due to optsize or a | |||
1605 | /// loop hint annotation. | |||
1606 | bool isScalarEpilogueAllowed() const { | |||
1607 | return ScalarEpilogueStatus == CM_ScalarEpilogueAllowed; | |||
1608 | } | |||
1609 | ||||
1610 | /// Returns true if all loop blocks should be masked to fold tail loop. | |||
1611 | bool foldTailByMasking() const { return FoldTailByMasking; } | |||
1612 | ||||
1613 | /// Returns true if the instructions in this block requires predication | |||
1614 | /// for any reason, e.g. because tail folding now requires a predicate | |||
1615 | /// or because the block in the original loop was predicated. | |||
1616 | bool blockNeedsPredicationForAnyReason(BasicBlock *BB) const { | |||
1617 | return foldTailByMasking() || Legal->blockNeedsPredication(BB); | |||
1618 | } | |||
1619 | ||||
1620 | /// A SmallMapVector to store the InLoop reduction op chains, mapping phi | |||
1621 | /// nodes to the chain of instructions representing the reductions. Uses a | |||
1622 | /// MapVector to ensure deterministic iteration order. | |||
1623 | using ReductionChainMap = | |||
1624 | SmallMapVector<PHINode *, SmallVector<Instruction *, 4>, 4>; | |||
1625 | ||||
1626 | /// Return the chain of instructions representing an inloop reduction. | |||
1627 | const ReductionChainMap &getInLoopReductionChains() const { | |||
1628 | return InLoopReductionChains; | |||
1629 | } | |||
1630 | ||||
1631 | /// Returns true if the Phi is part of an inloop reduction. | |||
1632 | bool isInLoopReduction(PHINode *Phi) const { | |||
1633 | return InLoopReductionChains.count(Phi); | |||
1634 | } | |||
1635 | ||||
1636 | /// Estimate cost of an intrinsic call instruction CI if it were vectorized | |||
1637 | /// with factor VF. Return the cost of the instruction, including | |||
1638 | /// scalarization overhead if it's needed. | |||
1639 | InstructionCost getVectorIntrinsicCost(CallInst *CI, ElementCount VF) const; | |||
1640 | ||||
1641 | /// Estimate cost of a call instruction CI if it were vectorized with factor | |||
1642 | /// VF. Return the cost of the instruction, including scalarization overhead | |||
1643 | /// if it's needed. The flag NeedToScalarize shows if the call needs to be | |||
1644 | /// scalarized - | |||
1645 | /// i.e. either vector version isn't available, or is too expensive. | |||
1646 | InstructionCost getVectorCallCost(CallInst *CI, ElementCount VF, | |||
1647 | bool &NeedToScalarize) const; | |||
1648 | ||||
1649 | /// Returns true if the per-lane cost of VectorizationFactor A is lower than | |||
1650 | /// that of B. | |||
1651 | bool isMoreProfitable(const VectorizationFactor &A, | |||
1652 | const VectorizationFactor &B) const; | |||
1653 | ||||
1654 | /// Invalidates decisions already taken by the cost model. | |||
1655 | void invalidateCostModelingDecisions() { | |||
1656 | WideningDecisions.clear(); | |||
1657 | Uniforms.clear(); | |||
1658 | Scalars.clear(); | |||
1659 | } | |||
1660 | ||||
1661 | private: | |||
1662 | unsigned NumPredStores = 0; | |||
1663 | ||||
1664 | /// Convenience function that returns the value of vscale_range iff | |||
1665 | /// vscale_range.min == vscale_range.max or otherwise returns the value | |||
1666 | /// returned by the corresponding TLI method. | |||
1667 | Optional<unsigned> getVScaleForTuning() const; | |||
1668 | ||||
1669 | /// \return An upper bound for the vectorization factors for both | |||
1670 | /// fixed and scalable vectorization, where the minimum-known number of | |||
1671 | /// elements is a power-of-2 larger than zero. If scalable vectorization is | |||
1672 | /// disabled or unsupported, then the scalable part will be equal to | |||
1673 | /// ElementCount::getScalable(0). | |||
1674 | FixedScalableVFPair computeFeasibleMaxVF(unsigned ConstTripCount, | |||
1675 | ElementCount UserVF, | |||
1676 | bool FoldTailByMasking); | |||
1677 | ||||
1678 | /// \return the maximized element count based on the targets vector | |||
1679 | /// registers and the loop trip-count, but limited to a maximum safe VF. | |||
1680 | /// This is a helper function of computeFeasibleMaxVF. | |||
1681 | /// FIXME: MaxSafeVF is currently passed by reference to avoid some obscure | |||
1682 | /// issue that occurred on one of the buildbots which cannot be reproduced | |||
1683 | /// without having access to the properietary compiler (see comments on | |||
1684 | /// D98509). The issue is currently under investigation and this workaround | |||
1685 | /// will be removed as soon as possible. | |||
1686 | ElementCount getMaximizedVFForTarget(unsigned ConstTripCount, | |||
1687 | unsigned SmallestType, | |||
1688 | unsigned WidestType, | |||
1689 | const ElementCount &MaxSafeVF, | |||
1690 | bool FoldTailByMasking); | |||
1691 | ||||
1692 | /// \return the maximum legal scalable VF, based on the safe max number | |||
1693 | /// of elements. | |||
1694 | ElementCount getMaxLegalScalableVF(unsigned MaxSafeElements); | |||
1695 | ||||
1696 | /// The vectorization cost is a combination of the cost itself and a boolean | |||
1697 | /// indicating whether any of the contributing operations will actually | |||
1698 | /// operate on vector values after type legalization in the backend. If this | |||
1699 | /// latter value is false, then all operations will be scalarized (i.e. no | |||
1700 | /// vectorization has actually taken place). | |||
1701 | using VectorizationCostTy = std::pair<InstructionCost, bool>; | |||
1702 | ||||
1703 | /// Returns the expected execution cost. The unit of the cost does | |||
1704 | /// not matter because we use the 'cost' units to compare different | |||
1705 | /// vector widths. The cost that is returned is *not* normalized by | |||
1706 | /// the factor width. If \p Invalid is not nullptr, this function | |||
1707 | /// will add a pair(Instruction*, ElementCount) to \p Invalid for | |||
1708 | /// each instruction that has an Invalid cost for the given VF. | |||
1709 | using InstructionVFPair = std::pair<Instruction *, ElementCount>; | |||
1710 | VectorizationCostTy | |||
1711 | expectedCost(ElementCount VF, | |||
1712 | SmallVectorImpl<InstructionVFPair> *Invalid = nullptr); | |||
1713 | ||||
1714 | /// Returns the execution time cost of an instruction for a given vector | |||
1715 | /// width. Vector width of one means scalar. | |||
1716 | VectorizationCostTy getInstructionCost(Instruction *I, ElementCount VF); | |||
1717 | ||||
1718 | /// The cost-computation logic from getInstructionCost which provides | |||
1719 | /// the vector type as an output parameter. | |||
1720 | InstructionCost getInstructionCost(Instruction *I, ElementCount VF, | |||
1721 | Type *&VectorTy); | |||
1722 | ||||
1723 | /// Return the cost of instructions in an inloop reduction pattern, if I is | |||
1724 | /// part of that pattern. | |||
1725 | Optional<InstructionCost> | |||
1726 | getReductionPatternCost(Instruction *I, ElementCount VF, Type *VectorTy, | |||
1727 | TTI::TargetCostKind CostKind); | |||
1728 | ||||
1729 | /// Calculate vectorization cost of memory instruction \p I. | |||
1730 | InstructionCost getMemoryInstructionCost(Instruction *I, ElementCount VF); | |||
1731 | ||||
1732 | /// The cost computation for scalarized memory instruction. | |||
1733 | InstructionCost getMemInstScalarizationCost(Instruction *I, ElementCount VF); | |||
1734 | ||||
1735 | /// The cost computation for interleaving group of memory instructions. | |||
1736 | InstructionCost getInterleaveGroupCost(Instruction *I, ElementCount VF); | |||
1737 | ||||
1738 | /// The cost computation for Gather/Scatter instruction. | |||
1739 | InstructionCost getGatherScatterCost(Instruction *I, ElementCount VF); | |||
1740 | ||||
1741 | /// The cost computation for widening instruction \p I with consecutive | |||
1742 | /// memory access. | |||
1743 | InstructionCost getConsecutiveMemOpCost(Instruction *I, ElementCount VF); | |||
1744 | ||||
1745 | /// The cost calculation for Load/Store instruction \p I with uniform pointer - | |||
1746 | /// Load: scalar load + broadcast. | |||
1747 | /// Store: scalar store + (loop invariant value stored? 0 : extract of last | |||
1748 | /// element) | |||
1749 | InstructionCost getUniformMemOpCost(Instruction *I, ElementCount VF); | |||
1750 | ||||
1751 | /// Estimate the overhead of scalarizing an instruction. This is a | |||
1752 | /// convenience wrapper for the type-based getScalarizationOverhead API. | |||
1753 | InstructionCost getScalarizationOverhead(Instruction *I, | |||
1754 | ElementCount VF) const; | |||
1755 | ||||
1756 | /// Returns whether the instruction is a load or store and will be a emitted | |||
1757 | /// as a vector operation. | |||
1758 | bool isConsecutiveLoadOrStore(Instruction *I); | |||
1759 | ||||
1760 | /// Returns true if an artificially high cost for emulated masked memrefs | |||
1761 | /// should be used. | |||
1762 | bool useEmulatedMaskMemRefHack(Instruction *I, ElementCount VF); | |||
1763 | ||||
1764 | /// Map of scalar integer values to the smallest bitwidth they can be legally | |||
1765 | /// represented as. The vector equivalents of these values should be truncated | |||
1766 | /// to this type. | |||
1767 | MapVector<Instruction *, uint64_t> MinBWs; | |||
1768 | ||||
1769 | /// A type representing the costs for instructions if they were to be | |||
1770 | /// scalarized rather than vectorized. The entries are Instruction-Cost | |||
1771 | /// pairs. | |||
1772 | using ScalarCostsTy = DenseMap<Instruction *, InstructionCost>; | |||
1773 | ||||
1774 | /// A set containing all BasicBlocks that are known to present after | |||
1775 | /// vectorization as a predicated block. | |||
1776 | SmallPtrSet<BasicBlock *, 4> PredicatedBBsAfterVectorization; | |||
1777 | ||||
1778 | /// Records whether it is allowed to have the original scalar loop execute at | |||
1779 | /// least once. This may be needed as a fallback loop in case runtime | |||
1780 | /// aliasing/dependence checks fail, or to handle the tail/remainder | |||
1781 | /// iterations when the trip count is unknown or doesn't divide by the VF, | |||
1782 | /// or as a peel-loop to handle gaps in interleave-groups. | |||
1783 | /// Under optsize and when the trip count is very small we don't allow any | |||
1784 | /// iterations to execute in the scalar loop. | |||
1785 | ScalarEpilogueLowering ScalarEpilogueStatus = CM_ScalarEpilogueAllowed; | |||
1786 | ||||
1787 | /// All blocks of loop are to be masked to fold tail of scalar iterations. | |||
1788 | bool FoldTailByMasking = false; | |||
1789 | ||||
1790 | /// A map holding scalar costs for different vectorization factors. The | |||
1791 | /// presence of a cost for an instruction in the mapping indicates that the | |||
1792 | /// instruction will be scalarized when vectorizing with the associated | |||
1793 | /// vectorization factor. The entries are VF-ScalarCostTy pairs. | |||
1794 | DenseMap<ElementCount, ScalarCostsTy> InstsToScalarize; | |||
1795 | ||||
1796 | /// Holds the instructions known to be uniform after vectorization. | |||
1797 | /// The data is collected per VF. | |||
1798 | DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> Uniforms; | |||
1799 | ||||
1800 | /// Holds the instructions known to be scalar after vectorization. | |||
1801 | /// The data is collected per VF. | |||
1802 | DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> Scalars; | |||
1803 | ||||
1804 | /// Holds the instructions (address computations) that are forced to be | |||
1805 | /// scalarized. | |||
1806 | DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> ForcedScalars; | |||
1807 | ||||
1808 | /// PHINodes of the reductions that should be expanded in-loop along with | |||
1809 | /// their associated chains of reduction operations, in program order from top | |||
1810 | /// (PHI) to bottom | |||
1811 | ReductionChainMap InLoopReductionChains; | |||
1812 | ||||
1813 | /// A Map of inloop reduction operations and their immediate chain operand. | |||
1814 | /// FIXME: This can be removed once reductions can be costed correctly in | |||
1815 | /// vplan. This was added to allow quick lookup to the inloop operations, | |||
1816 | /// without having to loop through InLoopReductionChains. | |||
1817 | DenseMap<Instruction *, Instruction *> InLoopReductionImmediateChains; | |||
1818 | ||||
1819 | /// Returns the expected difference in cost from scalarizing the expression | |||
1820 | /// feeding a predicated instruction \p PredInst. The instructions to | |||
1821 | /// scalarize and their scalar costs are collected in \p ScalarCosts. A | |||
1822 | /// non-negative return value implies the expression will be scalarized. | |||
1823 | /// Currently, only single-use chains are considered for scalarization. | |||
1824 | int computePredInstDiscount(Instruction *PredInst, ScalarCostsTy &ScalarCosts, | |||
1825 | ElementCount VF); | |||
1826 | ||||
1827 | /// Collect the instructions that are uniform after vectorization. An | |||
1828 | /// instruction is uniform if we represent it with a single scalar value in | |||
1829 | /// the vectorized loop corresponding to each vector iteration. Examples of | |||
1830 | /// uniform instructions include pointer operands of consecutive or | |||
1831 | /// interleaved memory accesses. Note that although uniformity implies an | |||
1832 | /// instruction will be scalar, the reverse is not true. In general, a | |||
1833 | /// scalarized instruction will be represented by VF scalar values in the | |||
1834 | /// vectorized loop, each corresponding to an iteration of the original | |||
1835 | /// scalar loop. | |||
1836 | void collectLoopUniforms(ElementCount VF); | |||
1837 | ||||
1838 | /// Collect the instructions that are scalar after vectorization. An | |||
1839 | /// instruction is scalar if it is known to be uniform or will be scalarized | |||
1840 | /// during vectorization. collectLoopScalars should only add non-uniform nodes | |||
1841 | /// to the list if they are used by a load/store instruction that is marked as | |||
1842 | /// CM_Scalarize. Non-uniform scalarized instructions will be represented by | |||
1843 | /// VF values in the vectorized loop, each corresponding to an iteration of | |||
1844 | /// the original scalar loop. | |||
1845 | void collectLoopScalars(ElementCount VF); | |||
1846 | ||||
1847 | /// Keeps cost model vectorization decision and cost for instructions. | |||
1848 | /// Right now it is used for memory instructions only. | |||
1849 | using DecisionList = DenseMap<std::pair<Instruction *, ElementCount>, | |||
1850 | std::pair<InstWidening, InstructionCost>>; | |||
1851 | ||||
1852 | DecisionList WideningDecisions; | |||
1853 | ||||
1854 | /// Returns true if \p V is expected to be vectorized and it needs to be | |||
1855 | /// extracted. | |||
1856 | bool needsExtract(Value *V, ElementCount VF) const { | |||
1857 | Instruction *I = dyn_cast<Instruction>(V); | |||
1858 | if (VF.isScalar() || !I || !TheLoop->contains(I) || | |||
1859 | TheLoop->isLoopInvariant(I)) | |||
1860 | return false; | |||
1861 | ||||
1862 | // Assume we can vectorize V (and hence we need extraction) if the | |||
1863 | // scalars are not computed yet. This can happen, because it is called | |||
1864 | // via getScalarizationOverhead from setCostBasedWideningDecision, before | |||
1865 | // the scalars are collected. That should be a safe assumption in most | |||
1866 | // cases, because we check if the operands have vectorizable types | |||
1867 | // beforehand in LoopVectorizationLegality. | |||
1868 | return Scalars.find(VF) == Scalars.end() || | |||
1869 | !isScalarAfterVectorization(I, VF); | |||
1870 | }; | |||
1871 | ||||
1872 | /// Returns a range containing only operands needing to be extracted. | |||
1873 | SmallVector<Value *, 4> filterExtractingOperands(Instruction::op_range Ops, | |||
1874 | ElementCount VF) const { | |||
1875 | return SmallVector<Value *, 4>(make_filter_range( | |||
1876 | Ops, [this, VF](Value *V) { return this->needsExtract(V, VF); })); | |||
1877 | } | |||
1878 | ||||
1879 | /// Determines if we have the infrastructure to vectorize loop \p L and its | |||
1880 | /// epilogue, assuming the main loop is vectorized by \p VF. | |||
1881 | bool isCandidateForEpilogueVectorization(const Loop &L, | |||
1882 | const ElementCount VF) const; | |||
1883 | ||||
1884 | /// Returns true if epilogue vectorization is considered profitable, and | |||
1885 | /// false otherwise. | |||
1886 | /// \p VF is the vectorization factor chosen for the original loop. | |||
1887 | bool isEpilogueVectorizationProfitable(const ElementCount VF) const; | |||
1888 | ||||
1889 | public: | |||
1890 | /// The loop that we evaluate. | |||
1891 | Loop *TheLoop; | |||
1892 | ||||
1893 | /// Predicated scalar evolution analysis. | |||
1894 | PredicatedScalarEvolution &PSE; | |||
1895 | ||||
1896 | /// Loop Info analysis. | |||
1897 | LoopInfo *LI; | |||
1898 | ||||
1899 | /// Vectorization legality. | |||
1900 | LoopVectorizationLegality *Legal; | |||
1901 | ||||
1902 | /// Vector target information. | |||
1903 | const TargetTransformInfo &TTI; | |||
1904 | ||||
1905 | /// Target Library Info. | |||
1906 | const TargetLibraryInfo *TLI; | |||
1907 | ||||
1908 | /// Demanded bits analysis. | |||
1909 | DemandedBits *DB; | |||
1910 | ||||
1911 | /// Assumption cache. | |||
1912 | AssumptionCache *AC; | |||
1913 | ||||
1914 | /// Interface to emit optimization remarks. | |||
1915 | OptimizationRemarkEmitter *ORE; | |||
1916 | ||||
1917 | const Function *TheFunction; | |||
1918 | ||||
1919 | /// Loop Vectorize Hint. | |||
1920 | const LoopVectorizeHints *Hints; | |||
1921 | ||||
1922 | /// The interleave access information contains groups of interleaved accesses | |||
1923 | /// with the same stride and close to each other. | |||
1924 | InterleavedAccessInfo &InterleaveInfo; | |||
1925 | ||||
1926 | /// Values to ignore in the cost model. | |||
1927 | SmallPtrSet<const Value *, 16> ValuesToIgnore; | |||
1928 | ||||
1929 | /// Values to ignore in the cost model when VF > 1. | |||
1930 | SmallPtrSet<const Value *, 16> VecValuesToIgnore; | |||
1931 | ||||
1932 | /// All element types found in the loop. | |||
1933 | SmallPtrSet<Type *, 16> ElementTypesInLoop; | |||
1934 | ||||
1935 | /// Profitable vector factors. | |||
1936 | SmallVector<VectorizationFactor, 8> ProfitableVFs; | |||
1937 | }; | |||
1938 | } // end namespace llvm | |||
1939 | ||||
1940 | /// Helper struct to manage generating runtime checks for vectorization. | |||
1941 | /// | |||
1942 | /// The runtime checks are created up-front in temporary blocks to allow better | |||
1943 | /// estimating the cost and un-linked from the existing IR. After deciding to | |||
1944 | /// vectorize, the checks are moved back. If deciding not to vectorize, the | |||
1945 | /// temporary blocks are completely removed. | |||
1946 | class GeneratedRTChecks { | |||
1947 | /// Basic block which contains the generated SCEV checks, if any. | |||
1948 | BasicBlock *SCEVCheckBlock = nullptr; | |||
1949 | ||||
1950 | /// The value representing the result of the generated SCEV checks. If it is | |||
1951 | /// nullptr, either no SCEV checks have been generated or they have been used. | |||
1952 | Value *SCEVCheckCond = nullptr; | |||
1953 | ||||
1954 | /// Basic block which contains the generated memory runtime checks, if any. | |||
1955 | BasicBlock *MemCheckBlock = nullptr; | |||
1956 | ||||
1957 | /// The value representing the result of the generated memory runtime checks. | |||
1958 | /// If it is nullptr, either no memory runtime checks have been generated or | |||
1959 | /// they have been used. | |||
1960 | Value *MemRuntimeCheckCond = nullptr; | |||
1961 | ||||
1962 | DominatorTree *DT; | |||
1963 | LoopInfo *LI; | |||
1964 | ||||
1965 | SCEVExpander SCEVExp; | |||
1966 | SCEVExpander MemCheckExp; | |||
1967 | ||||
1968 | public: | |||
1969 | GeneratedRTChecks(ScalarEvolution &SE, DominatorTree *DT, LoopInfo *LI, | |||
1970 | const DataLayout &DL) | |||
1971 | : DT(DT), LI(LI), SCEVExp(SE, DL, "scev.check"), | |||
1972 | MemCheckExp(SE, DL, "scev.check") {} | |||
1973 | ||||
1974 | /// Generate runtime checks in SCEVCheckBlock and MemCheckBlock, so we can | |||
1975 | /// accurately estimate the cost of the runtime checks. The blocks are | |||
1976 | /// un-linked from the IR and is added back during vector code generation. If | |||
1977 | /// there is no vector code generation, the check blocks are removed | |||
1978 | /// completely. | |||
1979 | void Create(Loop *L, const LoopAccessInfo &LAI, | |||
1980 | const SCEVPredicate &Pred) { | |||
1981 | ||||
1982 | BasicBlock *LoopHeader = L->getHeader(); | |||
1983 | BasicBlock *Preheader = L->getLoopPreheader(); | |||
1984 | ||||
1985 | // Use SplitBlock to create blocks for SCEV & memory runtime checks to | |||
1986 | // ensure the blocks are properly added to LoopInfo & DominatorTree. Those | |||
1987 | // may be used by SCEVExpander. The blocks will be un-linked from their | |||
1988 | // predecessors and removed from LI & DT at the end of the function. | |||
1989 | if (!Pred.isAlwaysTrue()) { | |||
1990 | SCEVCheckBlock = SplitBlock(Preheader, Preheader->getTerminator(), DT, LI, | |||
1991 | nullptr, "vector.scevcheck"); | |||
1992 | ||||
1993 | SCEVCheckCond = SCEVExp.expandCodeForPredicate( | |||
1994 | &Pred, SCEVCheckBlock->getTerminator()); | |||
1995 | } | |||
1996 | ||||
1997 | const auto &RtPtrChecking = *LAI.getRuntimePointerChecking(); | |||
1998 | if (RtPtrChecking.Need) { | |||
1999 | auto *Pred = SCEVCheckBlock ? SCEVCheckBlock : Preheader; | |||
2000 | MemCheckBlock = SplitBlock(Pred, Pred->getTerminator(), DT, LI, nullptr, | |||
2001 | "vector.memcheck"); | |||
2002 | ||||
2003 | MemRuntimeCheckCond = | |||
2004 | addRuntimeChecks(MemCheckBlock->getTerminator(), L, | |||
2005 | RtPtrChecking.getChecks(), MemCheckExp); | |||
2006 | assert(MemRuntimeCheckCond &&(static_cast <bool> (MemRuntimeCheckCond && "no RT checks generated although RtPtrChecking " "claimed checks are required") ? void (0) : __assert_fail ("MemRuntimeCheckCond && \"no RT checks generated although RtPtrChecking \" \"claimed checks are required\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2008, __extension__ __PRETTY_FUNCTION__)) | |||
2007 | "no RT checks generated although RtPtrChecking "(static_cast <bool> (MemRuntimeCheckCond && "no RT checks generated although RtPtrChecking " "claimed checks are required") ? void (0) : __assert_fail ("MemRuntimeCheckCond && \"no RT checks generated although RtPtrChecking \" \"claimed checks are required\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2008, __extension__ __PRETTY_FUNCTION__)) | |||
2008 | "claimed checks are required")(static_cast <bool> (MemRuntimeCheckCond && "no RT checks generated although RtPtrChecking " "claimed checks are required") ? void (0) : __assert_fail ("MemRuntimeCheckCond && \"no RT checks generated although RtPtrChecking \" \"claimed checks are required\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2008, __extension__ __PRETTY_FUNCTION__)); | |||
2009 | } | |||
2010 | ||||
2011 | if (!MemCheckBlock && !SCEVCheckBlock) | |||
2012 | return; | |||
2013 | ||||
2014 | // Unhook the temporary block with the checks, update various places | |||
2015 | // accordingly. | |||
2016 | if (SCEVCheckBlock) | |||
2017 | SCEVCheckBlock->replaceAllUsesWith(Preheader); | |||
2018 | if (MemCheckBlock) | |||
2019 | MemCheckBlock->replaceAllUsesWith(Preheader); | |||
2020 | ||||
2021 | if (SCEVCheckBlock) { | |||
2022 | SCEVCheckBlock->getTerminator()->moveBefore(Preheader->getTerminator()); | |||
2023 | new UnreachableInst(Preheader->getContext(), SCEVCheckBlock); | |||
2024 | Preheader->getTerminator()->eraseFromParent(); | |||
2025 | } | |||
2026 | if (MemCheckBlock) { | |||
2027 | MemCheckBlock->getTerminator()->moveBefore(Preheader->getTerminator()); | |||
2028 | new UnreachableInst(Preheader->getContext(), MemCheckBlock); | |||
2029 | Preheader->getTerminator()->eraseFromParent(); | |||
2030 | } | |||
2031 | ||||
2032 | DT->changeImmediateDominator(LoopHeader, Preheader); | |||
2033 | if (MemCheckBlock) { | |||
2034 | DT->eraseNode(MemCheckBlock); | |||
2035 | LI->removeBlock(MemCheckBlock); | |||
2036 | } | |||
2037 | if (SCEVCheckBlock) { | |||
2038 | DT->eraseNode(SCEVCheckBlock); | |||
2039 | LI->removeBlock(SCEVCheckBlock); | |||
2040 | } | |||
2041 | } | |||
2042 | ||||
2043 | /// Remove the created SCEV & memory runtime check blocks & instructions, if | |||
2044 | /// unused. | |||
2045 | ~GeneratedRTChecks() { | |||
2046 | SCEVExpanderCleaner SCEVCleaner(SCEVExp); | |||
2047 | SCEVExpanderCleaner MemCheckCleaner(MemCheckExp); | |||
2048 | if (!SCEVCheckCond) | |||
2049 | SCEVCleaner.markResultUsed(); | |||
2050 | ||||
2051 | if (!MemRuntimeCheckCond) | |||
2052 | MemCheckCleaner.markResultUsed(); | |||
2053 | ||||
2054 | if (MemRuntimeCheckCond) { | |||
2055 | auto &SE = *MemCheckExp.getSE(); | |||
2056 | // Memory runtime check generation creates compares that use expanded | |||
2057 | // values. Remove them before running the SCEVExpanderCleaners. | |||
2058 | for (auto &I : make_early_inc_range(reverse(*MemCheckBlock))) { | |||
2059 | if (MemCheckExp.isInsertedInstruction(&I)) | |||
2060 | continue; | |||
2061 | SE.forgetValue(&I); | |||
2062 | I.eraseFromParent(); | |||
2063 | } | |||
2064 | } | |||
2065 | MemCheckCleaner.cleanup(); | |||
2066 | SCEVCleaner.cleanup(); | |||
2067 | ||||
2068 | if (SCEVCheckCond) | |||
2069 | SCEVCheckBlock->eraseFromParent(); | |||
2070 | if (MemRuntimeCheckCond) | |||
2071 | MemCheckBlock->eraseFromParent(); | |||
2072 | } | |||
2073 | ||||
2074 | /// Adds the generated SCEVCheckBlock before \p LoopVectorPreHeader and | |||
2075 | /// adjusts the branches to branch to the vector preheader or \p Bypass, | |||
2076 | /// depending on the generated condition. | |||
2077 | BasicBlock *emitSCEVChecks(BasicBlock *Bypass, | |||
2078 | BasicBlock *LoopVectorPreHeader, | |||
2079 | BasicBlock *LoopExitBlock) { | |||
2080 | if (!SCEVCheckCond) | |||
2081 | return nullptr; | |||
2082 | if (auto *C = dyn_cast<ConstantInt>(SCEVCheckCond)) | |||
2083 | if (C->isZero()) | |||
2084 | return nullptr; | |||
2085 | ||||
2086 | auto *Pred = LoopVectorPreHeader->getSinglePredecessor(); | |||
2087 | ||||
2088 | BranchInst::Create(LoopVectorPreHeader, SCEVCheckBlock); | |||
2089 | // Create new preheader for vector loop. | |||
2090 | if (auto *PL = LI->getLoopFor(LoopVectorPreHeader)) | |||
2091 | PL->addBasicBlockToLoop(SCEVCheckBlock, *LI); | |||
2092 | ||||
2093 | SCEVCheckBlock->getTerminator()->eraseFromParent(); | |||
2094 | SCEVCheckBlock->moveBefore(LoopVectorPreHeader); | |||
2095 | Pred->getTerminator()->replaceSuccessorWith(LoopVectorPreHeader, | |||
2096 | SCEVCheckBlock); | |||
2097 | ||||
2098 | DT->addNewBlock(SCEVCheckBlock, Pred); | |||
2099 | DT->changeImmediateDominator(LoopVectorPreHeader, SCEVCheckBlock); | |||
2100 | ||||
2101 | ReplaceInstWithInst( | |||
2102 | SCEVCheckBlock->getTerminator(), | |||
2103 | BranchInst::Create(Bypass, LoopVectorPreHeader, SCEVCheckCond)); | |||
2104 | // Mark the check as used, to prevent it from being removed during cleanup. | |||
2105 | SCEVCheckCond = nullptr; | |||
2106 | return SCEVCheckBlock; | |||
2107 | } | |||
2108 | ||||
2109 | /// Adds the generated MemCheckBlock before \p LoopVectorPreHeader and adjusts | |||
2110 | /// the branches to branch to the vector preheader or \p Bypass, depending on | |||
2111 | /// the generated condition. | |||
2112 | BasicBlock *emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass, | |||
2113 | BasicBlock *LoopVectorPreHeader) { | |||
2114 | // Check if we generated code that checks in runtime if arrays overlap. | |||
2115 | if (!MemRuntimeCheckCond) | |||
2116 | return nullptr; | |||
2117 | ||||
2118 | auto *Pred = LoopVectorPreHeader->getSinglePredecessor(); | |||
2119 | Pred->getTerminator()->replaceSuccessorWith(LoopVectorPreHeader, | |||
2120 | MemCheckBlock); | |||
2121 | ||||
2122 | DT->addNewBlock(MemCheckBlock, Pred); | |||
2123 | DT->changeImmediateDominator(LoopVectorPreHeader, MemCheckBlock); | |||
2124 | MemCheckBlock->moveBefore(LoopVectorPreHeader); | |||
2125 | ||||
2126 | if (auto *PL = LI->getLoopFor(LoopVectorPreHeader)) | |||
2127 | PL->addBasicBlockToLoop(MemCheckBlock, *LI); | |||
2128 | ||||
2129 | ReplaceInstWithInst( | |||
2130 | MemCheckBlock->getTerminator(), | |||
2131 | BranchInst::Create(Bypass, LoopVectorPreHeader, MemRuntimeCheckCond)); | |||
2132 | MemCheckBlock->getTerminator()->setDebugLoc( | |||
2133 | Pred->getTerminator()->getDebugLoc()); | |||
2134 | ||||
2135 | // Mark the check as used, to prevent it from being removed during cleanup. | |||
2136 | MemRuntimeCheckCond = nullptr; | |||
2137 | return MemCheckBlock; | |||
2138 | } | |||
2139 | }; | |||
2140 | ||||
2141 | // Return true if \p OuterLp is an outer loop annotated with hints for explicit | |||
2142 | // vectorization. The loop needs to be annotated with #pragma omp simd | |||
2143 | // simdlen(#) or #pragma clang vectorize(enable) vectorize_width(#). If the | |||
2144 | // vector length information is not provided, vectorization is not considered | |||
2145 | // explicit. Interleave hints are not allowed either. These limitations will be | |||
2146 | // relaxed in the future. | |||
2147 | // Please, note that we are currently forced to abuse the pragma 'clang | |||
2148 | // vectorize' semantics. This pragma provides *auto-vectorization hints* | |||
2149 | // (i.e., LV must check that vectorization is legal) whereas pragma 'omp simd' | |||
2150 | // provides *explicit vectorization hints* (LV can bypass legal checks and | |||
2151 | // assume that vectorization is legal). However, both hints are implemented | |||
2152 | // using the same metadata (llvm.loop.vectorize, processed by | |||
2153 | // LoopVectorizeHints). This will be fixed in the future when the native IR | |||
2154 | // representation for pragma 'omp simd' is introduced. | |||
2155 | static bool isExplicitVecOuterLoop(Loop *OuterLp, | |||
2156 | OptimizationRemarkEmitter *ORE) { | |||
2157 | assert(!OuterLp->isInnermost() && "This is not an outer loop")(static_cast <bool> (!OuterLp->isInnermost() && "This is not an outer loop") ? void (0) : __assert_fail ("!OuterLp->isInnermost() && \"This is not an outer loop\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2157, __extension__ __PRETTY_FUNCTION__)); | |||
2158 | LoopVectorizeHints Hints(OuterLp, true /*DisableInterleaving*/, *ORE); | |||
2159 | ||||
2160 | // Only outer loops with an explicit vectorization hint are supported. | |||
2161 | // Unannotated outer loops are ignored. | |||
2162 | if (Hints.getForce() == LoopVectorizeHints::FK_Undefined) | |||
2163 | return false; | |||
2164 | ||||
2165 | Function *Fn = OuterLp->getHeader()->getParent(); | |||
2166 | if (!Hints.allowVectorization(Fn, OuterLp, | |||
2167 | true /*VectorizeOnlyWhenForced*/)) { | |||
2168 | 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); | |||
2169 | return false; | |||
2170 | } | |||
2171 | ||||
2172 | if (Hints.getInterleave() > 1) { | |||
2173 | // TODO: Interleave support is future work. | |||
2174 | 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) | |||
2175 | "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); | |||
2176 | Hints.emitRemarkWithHints(); | |||
2177 | return false; | |||
2178 | } | |||
2179 | ||||
2180 | return true; | |||
2181 | } | |||
2182 | ||||
2183 | static void collectSupportedLoops(Loop &L, LoopInfo *LI, | |||
2184 | OptimizationRemarkEmitter *ORE, | |||
2185 | SmallVectorImpl<Loop *> &V) { | |||
2186 | // Collect inner loops and outer loops without irreducible control flow. For | |||
2187 | // now, only collect outer loops that have explicit vectorization hints. If we | |||
2188 | // are stress testing the VPlan H-CFG construction, we collect the outermost | |||
2189 | // loop of every loop nest. | |||
2190 | if (L.isInnermost() || VPlanBuildStressTest || | |||
2191 | (EnableVPlanNativePath && isExplicitVecOuterLoop(&L, ORE))) { | |||
2192 | LoopBlocksRPO RPOT(&L); | |||
2193 | RPOT.perform(LI); | |||
2194 | if (!containsIrreducibleCFG<const BasicBlock *>(RPOT, *LI)) { | |||
2195 | V.push_back(&L); | |||
2196 | // TODO: Collect inner loops inside marked outer loops in case | |||
2197 | // vectorization fails for the outer loop. Do not invoke | |||
2198 | // 'containsIrreducibleCFG' again for inner loops when the outer loop is | |||
2199 | // already known to be reducible. We can use an inherited attribute for | |||
2200 | // that. | |||
2201 | return; | |||
2202 | } | |||
2203 | } | |||
2204 | for (Loop *InnerL : L) | |||
2205 | collectSupportedLoops(*InnerL, LI, ORE, V); | |||
2206 | } | |||
2207 | ||||
2208 | namespace { | |||
2209 | ||||
2210 | /// The LoopVectorize Pass. | |||
2211 | struct LoopVectorize : public FunctionPass { | |||
2212 | /// Pass identification, replacement for typeid | |||
2213 | static char ID; | |||
2214 | ||||
2215 | LoopVectorizePass Impl; | |||
2216 | ||||
2217 | explicit LoopVectorize(bool InterleaveOnlyWhenForced = false, | |||
2218 | bool VectorizeOnlyWhenForced = false) | |||
2219 | : FunctionPass(ID), | |||
2220 | Impl({InterleaveOnlyWhenForced, VectorizeOnlyWhenForced}) { | |||
2221 | initializeLoopVectorizePass(*PassRegistry::getPassRegistry()); | |||
2222 | } | |||
2223 | ||||
2224 | bool runOnFunction(Function &F) override { | |||
2225 | if (skipFunction(F)) | |||
2226 | return false; | |||
2227 | ||||
2228 | auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); | |||
2229 | auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); | |||
2230 | auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); | |||
2231 | auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); | |||
2232 | auto *BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI(); | |||
2233 | auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); | |||
2234 | auto *TLI = TLIP ? &TLIP->getTLI(F) : nullptr; | |||
2235 | auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); | |||
2236 | auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); | |||
2237 | auto *LAA = &getAnalysis<LoopAccessLegacyAnalysis>(); | |||
2238 | auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits(); | |||
2239 | auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(); | |||
2240 | auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); | |||
2241 | ||||
2242 | std::function<const LoopAccessInfo &(Loop &)> GetLAA = | |||
2243 | [&](Loop &L) -> const LoopAccessInfo & { return LAA->getInfo(&L); }; | |||
2244 | ||||
2245 | return Impl.runImpl(F, *SE, *LI, *TTI, *DT, *BFI, TLI, *DB, *AA, *AC, | |||
2246 | GetLAA, *ORE, PSI).MadeAnyChange; | |||
2247 | } | |||
2248 | ||||
2249 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
2250 | AU.addRequired<AssumptionCacheTracker>(); | |||
2251 | AU.addRequired<BlockFrequencyInfoWrapperPass>(); | |||
2252 | AU.addRequired<DominatorTreeWrapperPass>(); | |||
2253 | AU.addRequired<LoopInfoWrapperPass>(); | |||
2254 | AU.addRequired<ScalarEvolutionWrapperPass>(); | |||
2255 | AU.addRequired<TargetTransformInfoWrapperPass>(); | |||
2256 | AU.addRequired<AAResultsWrapperPass>(); | |||
2257 | AU.addRequired<LoopAccessLegacyAnalysis>(); | |||
2258 | AU.addRequired<DemandedBitsWrapperPass>(); | |||
2259 | AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); | |||
2260 | AU.addRequired<InjectTLIMappingsLegacy>(); | |||
2261 | ||||
2262 | // We currently do not preserve loopinfo/dominator analyses with outer loop | |||
2263 | // vectorization. Until this is addressed, mark these analyses as preserved | |||
2264 | // only for non-VPlan-native path. | |||
2265 | // TODO: Preserve Loop and Dominator analyses for VPlan-native path. | |||
2266 | if (!EnableVPlanNativePath) { | |||
2267 | AU.addPreserved<LoopInfoWrapperPass>(); | |||
2268 | AU.addPreserved<DominatorTreeWrapperPass>(); | |||
2269 | } | |||
2270 | ||||
2271 | AU.addPreserved<BasicAAWrapperPass>(); | |||
2272 | AU.addPreserved<GlobalsAAWrapperPass>(); | |||
2273 | AU.addRequired<ProfileSummaryInfoWrapperPass>(); | |||
2274 | } | |||
2275 | }; | |||
2276 | ||||
2277 | } // end anonymous namespace | |||
2278 | ||||
2279 | //===----------------------------------------------------------------------===// | |||
2280 | // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and | |||
2281 | // LoopVectorizationCostModel and LoopVectorizationPlanner. | |||
2282 | //===----------------------------------------------------------------------===// | |||
2283 | ||||
2284 | Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) { | |||
2285 | // We need to place the broadcast of invariant variables outside the loop, | |||
2286 | // but only if it's proven safe to do so. Else, broadcast will be inside | |||
2287 | // vector loop body. | |||
2288 | Instruction *Instr = dyn_cast<Instruction>(V); | |||
2289 | bool SafeToHoist = OrigLoop->isLoopInvariant(V) && | |||
2290 | (!Instr || | |||
2291 | DT->dominates(Instr->getParent(), LoopVectorPreHeader)); | |||
2292 | // Place the code for broadcasting invariant variables in the new preheader. | |||
2293 | IRBuilder<>::InsertPointGuard Guard(Builder); | |||
2294 | if (SafeToHoist) | |||
2295 | Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); | |||
2296 | ||||
2297 | // Broadcast the scalar into all locations in the vector. | |||
2298 | Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast"); | |||
2299 | ||||
2300 | return Shuf; | |||
2301 | } | |||
2302 | ||||
2303 | /// This function adds | |||
2304 | /// (StartIdx * Step, (StartIdx + 1) * Step, (StartIdx + 2) * Step, ...) | |||
2305 | /// to each vector element of Val. The sequence starts at StartIndex. | |||
2306 | /// \p Opcode is relevant for FP induction variable. | |||
2307 | static Value *getStepVector(Value *Val, Value *StartIdx, Value *Step, | |||
2308 | Instruction::BinaryOps BinOp, ElementCount VF, | |||
2309 | IRBuilderBase &Builder) { | |||
2310 | assert(VF.isVector() && "only vector VFs are supported")(static_cast <bool> (VF.isVector() && "only vector VFs are supported" ) ? void (0) : __assert_fail ("VF.isVector() && \"only vector VFs are supported\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2310, __extension__ __PRETTY_FUNCTION__)); | |||
2311 | ||||
2312 | // Create and check the types. | |||
2313 | auto *ValVTy = cast<VectorType>(Val->getType()); | |||
2314 | ElementCount VLen = ValVTy->getElementCount(); | |||
2315 | ||||
2316 | Type *STy = Val->getType()->getScalarType(); | |||
2317 | assert((STy->isIntegerTy() || STy->isFloatingPointTy()) &&(static_cast <bool> ((STy->isIntegerTy() || STy-> isFloatingPointTy()) && "Induction Step must be an integer or FP" ) ? void (0) : __assert_fail ("(STy->isIntegerTy() || STy->isFloatingPointTy()) && \"Induction Step must be an integer or FP\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2318, __extension__ __PRETTY_FUNCTION__)) | |||
2318 | "Induction Step must be an integer or FP")(static_cast <bool> ((STy->isIntegerTy() || STy-> isFloatingPointTy()) && "Induction Step must be an integer or FP" ) ? void (0) : __assert_fail ("(STy->isIntegerTy() || STy->isFloatingPointTy()) && \"Induction Step must be an integer or FP\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2318, __extension__ __PRETTY_FUNCTION__)); | |||
2319 | assert(Step->getType() == STy && "Step has wrong type")(static_cast <bool> (Step->getType() == STy && "Step has wrong type") ? void (0) : __assert_fail ("Step->getType() == STy && \"Step has wrong type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2319, __extension__ __PRETTY_FUNCTION__)); | |||
2320 | ||||
2321 | SmallVector<Constant *, 8> Indices; | |||
2322 | ||||
2323 | // Create a vector of consecutive numbers from zero to VF. | |||
2324 | VectorType *InitVecValVTy = ValVTy; | |||
2325 | if (STy->isFloatingPointTy()) { | |||
2326 | Type *InitVecValSTy = | |||
2327 | IntegerType::get(STy->getContext(), STy->getScalarSizeInBits()); | |||
2328 | InitVecValVTy = VectorType::get(InitVecValSTy, VLen); | |||
2329 | } | |||
2330 | Value *InitVec = Builder.CreateStepVector(InitVecValVTy); | |||
2331 | ||||
2332 | // Splat the StartIdx | |||
2333 | Value *StartIdxSplat = Builder.CreateVectorSplat(VLen, StartIdx); | |||
2334 | ||||
2335 | if (STy->isIntegerTy()) { | |||
2336 | InitVec = Builder.CreateAdd(InitVec, StartIdxSplat); | |||
2337 | Step = Builder.CreateVectorSplat(VLen, Step); | |||
2338 | assert(Step->getType() == Val->getType() && "Invalid step vec")(static_cast <bool> (Step->getType() == Val->getType () && "Invalid step vec") ? void (0) : __assert_fail ( "Step->getType() == Val->getType() && \"Invalid step vec\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2338, __extension__ __PRETTY_FUNCTION__)); | |||
2339 | // FIXME: The newly created binary instructions should contain nsw/nuw | |||
2340 | // flags, which can be found from the original scalar operations. | |||
2341 | Step = Builder.CreateMul(InitVec, Step); | |||
2342 | return Builder.CreateAdd(Val, Step, "induction"); | |||
2343 | } | |||
2344 | ||||
2345 | // Floating point induction. | |||
2346 | assert((BinOp == Instruction::FAdd || BinOp == Instruction::FSub) &&(static_cast <bool> ((BinOp == Instruction::FAdd || BinOp == Instruction::FSub) && "Binary Opcode should be specified for FP induction" ) ? void (0) : __assert_fail ("(BinOp == Instruction::FAdd || BinOp == Instruction::FSub) && \"Binary Opcode should be specified for FP induction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2347, __extension__ __PRETTY_FUNCTION__)) | |||
2347 | "Binary Opcode should be specified for FP induction")(static_cast <bool> ((BinOp == Instruction::FAdd || BinOp == Instruction::FSub) && "Binary Opcode should be specified for FP induction" ) ? void (0) : __assert_fail ("(BinOp == Instruction::FAdd || BinOp == Instruction::FSub) && \"Binary Opcode should be specified for FP induction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2347, __extension__ __PRETTY_FUNCTION__)); | |||
2348 | InitVec = Builder.CreateUIToFP(InitVec, ValVTy); | |||
2349 | InitVec = Builder.CreateFAdd(InitVec, StartIdxSplat); | |||
2350 | ||||
2351 | Step = Builder.CreateVectorSplat(VLen, Step); | |||
2352 | Value *MulOp = Builder.CreateFMul(InitVec, Step); | |||
2353 | return Builder.CreateBinOp(BinOp, Val, MulOp, "induction"); | |||
2354 | } | |||
2355 | ||||
2356 | /// Compute scalar induction steps. \p ScalarIV is the scalar induction | |||
2357 | /// variable on which to base the steps, \p Step is the size of the step. | |||
2358 | static void buildScalarSteps(Value *ScalarIV, Value *Step, | |||
2359 | const InductionDescriptor &ID, VPValue *Def, | |||
2360 | VPTransformState &State) { | |||
2361 | IRBuilderBase &Builder = State.Builder; | |||
2362 | // We shouldn't have to build scalar steps if we aren't vectorizing. | |||
2363 | assert(State.VF.isVector() && "VF should be greater than one")(static_cast <bool> (State.VF.isVector() && "VF should be greater than one" ) ? void (0) : __assert_fail ("State.VF.isVector() && \"VF should be greater than one\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2363, __extension__ __PRETTY_FUNCTION__)); | |||
2364 | // Get the value type and ensure it and the step have the same integer type. | |||
2365 | Type *ScalarIVTy = ScalarIV->getType()->getScalarType(); | |||
2366 | assert(ScalarIVTy == Step->getType() &&(static_cast <bool> (ScalarIVTy == Step->getType() && "Val and Step should have the same type") ? void (0) : __assert_fail ("ScalarIVTy == Step->getType() && \"Val and Step should have the same type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2367, __extension__ __PRETTY_FUNCTION__)) | |||
2367 | "Val and Step should have the same type")(static_cast <bool> (ScalarIVTy == Step->getType() && "Val and Step should have the same type") ? void (0) : __assert_fail ("ScalarIVTy == Step->getType() && \"Val and Step should have the same type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2367, __extension__ __PRETTY_FUNCTION__)); | |||
2368 | ||||
2369 | // We build scalar steps for both integer and floating-point induction | |||
2370 | // variables. Here, we determine the kind of arithmetic we will perform. | |||
2371 | Instruction::BinaryOps AddOp; | |||
2372 | Instruction::BinaryOps MulOp; | |||
2373 | if (ScalarIVTy->isIntegerTy()) { | |||
2374 | AddOp = Instruction::Add; | |||
2375 | MulOp = Instruction::Mul; | |||
2376 | } else { | |||
2377 | AddOp = ID.getInductionOpcode(); | |||
2378 | MulOp = Instruction::FMul; | |||
2379 | } | |||
2380 | ||||
2381 | // Determine the number of scalars we need to generate for each unroll | |||
2382 | // iteration. | |||
2383 | bool FirstLaneOnly = vputils::onlyFirstLaneUsed(Def); | |||
2384 | unsigned Lanes = FirstLaneOnly ? 1 : State.VF.getKnownMinValue(); | |||
2385 | // Compute the scalar steps and save the results in State. | |||
2386 | Type *IntStepTy = IntegerType::get(ScalarIVTy->getContext(), | |||
2387 | ScalarIVTy->getScalarSizeInBits()); | |||
2388 | Type *VecIVTy = nullptr; | |||
2389 | Value *UnitStepVec = nullptr, *SplatStep = nullptr, *SplatIV = nullptr; | |||
2390 | if (!FirstLaneOnly && State.VF.isScalable()) { | |||
2391 | VecIVTy = VectorType::get(ScalarIVTy, State.VF); | |||
2392 | UnitStepVec = | |||
2393 | Builder.CreateStepVector(VectorType::get(IntStepTy, State.VF)); | |||
2394 | SplatStep = Builder.CreateVectorSplat(State.VF, Step); | |||
2395 | SplatIV = Builder.CreateVectorSplat(State.VF, ScalarIV); | |||
2396 | } | |||
2397 | ||||
2398 | for (unsigned Part = 0; Part < State.UF; ++Part) { | |||
2399 | Value *StartIdx0 = createStepForVF(Builder, IntStepTy, State.VF, Part); | |||
2400 | ||||
2401 | if (!FirstLaneOnly && State.VF.isScalable()) { | |||
2402 | auto *SplatStartIdx = Builder.CreateVectorSplat(State.VF, StartIdx0); | |||
2403 | auto *InitVec = Builder.CreateAdd(SplatStartIdx, UnitStepVec); | |||
2404 | if (ScalarIVTy->isFloatingPointTy()) | |||
2405 | InitVec = Builder.CreateSIToFP(InitVec, VecIVTy); | |||
2406 | auto *Mul = Builder.CreateBinOp(MulOp, InitVec, SplatStep); | |||
2407 | auto *Add = Builder.CreateBinOp(AddOp, SplatIV, Mul); | |||
2408 | State.set(Def, Add, Part); | |||
2409 | // It's useful to record the lane values too for the known minimum number | |||
2410 | // of elements so we do those below. This improves the code quality when | |||
2411 | // trying to extract the first element, for example. | |||
2412 | } | |||
2413 | ||||
2414 | if (ScalarIVTy->isFloatingPointTy()) | |||
2415 | StartIdx0 = Builder.CreateSIToFP(StartIdx0, ScalarIVTy); | |||
2416 | ||||
2417 | for (unsigned Lane = 0; Lane < Lanes; ++Lane) { | |||
2418 | Value *StartIdx = Builder.CreateBinOp( | |||
2419 | AddOp, StartIdx0, getSignedIntOrFpConstant(ScalarIVTy, Lane)); | |||
2420 | // The step returned by `createStepForVF` is a runtime-evaluated value | |||
2421 | // when VF is scalable. Otherwise, it should be folded into a Constant. | |||
2422 | assert((State.VF.isScalable() || isa<Constant>(StartIdx)) &&(static_cast <bool> ((State.VF.isScalable() || isa<Constant >(StartIdx)) && "Expected StartIdx to be folded to a constant when VF is not " "scalable") ? void (0) : __assert_fail ("(State.VF.isScalable() || isa<Constant>(StartIdx)) && \"Expected StartIdx to be folded to a constant when VF is not \" \"scalable\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2424, __extension__ __PRETTY_FUNCTION__)) | |||
2423 | "Expected StartIdx to be folded to a constant when VF is not "(static_cast <bool> ((State.VF.isScalable() || isa<Constant >(StartIdx)) && "Expected StartIdx to be folded to a constant when VF is not " "scalable") ? void (0) : __assert_fail ("(State.VF.isScalable() || isa<Constant>(StartIdx)) && \"Expected StartIdx to be folded to a constant when VF is not \" \"scalable\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2424, __extension__ __PRETTY_FUNCTION__)) | |||
2424 | "scalable")(static_cast <bool> ((State.VF.isScalable() || isa<Constant >(StartIdx)) && "Expected StartIdx to be folded to a constant when VF is not " "scalable") ? void (0) : __assert_fail ("(State.VF.isScalable() || isa<Constant>(StartIdx)) && \"Expected StartIdx to be folded to a constant when VF is not \" \"scalable\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2424, __extension__ __PRETTY_FUNCTION__)); | |||
2425 | auto *Mul = Builder.CreateBinOp(MulOp, StartIdx, Step); | |||
2426 | auto *Add = Builder.CreateBinOp(AddOp, ScalarIV, Mul); | |||
2427 | State.set(Def, Add, VPIteration(Part, Lane)); | |||
2428 | } | |||
2429 | } | |||
2430 | } | |||
2431 | ||||
2432 | // Generate code for the induction step. Note that induction steps are | |||
2433 | // required to be loop-invariant | |||
2434 | static Value *CreateStepValue(const SCEV *Step, ScalarEvolution &SE, | |||
2435 | Instruction *InsertBefore, | |||
2436 | Loop *OrigLoop = nullptr) { | |||
2437 | const DataLayout &DL = SE.getDataLayout(); | |||
2438 | assert((!OrigLoop || SE.isLoopInvariant(Step, OrigLoop)) &&(static_cast <bool> ((!OrigLoop || SE.isLoopInvariant(Step , OrigLoop)) && "Induction step should be loop invariant" ) ? void (0) : __assert_fail ("(!OrigLoop || SE.isLoopInvariant(Step, OrigLoop)) && \"Induction step should be loop invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2439, __extension__ __PRETTY_FUNCTION__)) | |||
2439 | "Induction step should be loop invariant")(static_cast <bool> ((!OrigLoop || SE.isLoopInvariant(Step , OrigLoop)) && "Induction step should be loop invariant" ) ? void (0) : __assert_fail ("(!OrigLoop || SE.isLoopInvariant(Step, OrigLoop)) && \"Induction step should be loop invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2439, __extension__ __PRETTY_FUNCTION__)); | |||
2440 | if (auto *E = dyn_cast<SCEVUnknown>(Step)) | |||
2441 | return E->getValue(); | |||
2442 | ||||
2443 | SCEVExpander Exp(SE, DL, "induction"); | |||
2444 | return Exp.expandCodeFor(Step, Step->getType(), InsertBefore); | |||
2445 | } | |||
2446 | ||||
2447 | /// Compute the transformed value of Index at offset StartValue using step | |||
2448 | /// StepValue. | |||
2449 | /// For integer induction, returns StartValue + Index * StepValue. | |||
2450 | /// For pointer induction, returns StartValue[Index * StepValue]. | |||
2451 | /// FIXME: The newly created binary instructions should contain nsw/nuw | |||
2452 | /// flags, which can be found from the original scalar operations. | |||
2453 | static Value *emitTransformedIndex(IRBuilderBase &B, Value *Index, | |||
2454 | Value *StartValue, Value *Step, | |||
2455 | const InductionDescriptor &ID) { | |||
2456 | assert(Index->getType()->getScalarType() == Step->getType() &&(static_cast <bool> (Index->getType()->getScalarType () == Step->getType() && "Index scalar type does not match StepValue type" ) ? void (0) : __assert_fail ("Index->getType()->getScalarType() == Step->getType() && \"Index scalar type does not match StepValue type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2457, __extension__ __PRETTY_FUNCTION__)) | |||
2457 | "Index scalar type does not match StepValue type")(static_cast <bool> (Index->getType()->getScalarType () == Step->getType() && "Index scalar type does not match StepValue type" ) ? void (0) : __assert_fail ("Index->getType()->getScalarType() == Step->getType() && \"Index scalar type does not match StepValue type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2457, __extension__ __PRETTY_FUNCTION__)); | |||
2458 | ||||
2459 | // Note: the IR at this point is broken. We cannot use SE to create any new | |||
2460 | // SCEV and then expand it, hoping that SCEV's simplification will give us | |||
2461 | // a more optimal code. Unfortunately, attempt of doing so on invalid IR may | |||
2462 | // lead to various SCEV crashes. So all we can do is to use builder and rely | |||
2463 | // on InstCombine for future simplifications. Here we handle some trivial | |||
2464 | // cases only. | |||
2465 | auto CreateAdd = [&B](Value *X, Value *Y) { | |||
2466 | assert(X->getType() == Y->getType() && "Types don't match!")(static_cast <bool> (X->getType() == Y->getType() && "Types don't match!") ? void (0) : __assert_fail ( "X->getType() == Y->getType() && \"Types don't match!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2466, __extension__ __PRETTY_FUNCTION__)); | |||
2467 | if (auto *CX = dyn_cast<ConstantInt>(X)) | |||
2468 | if (CX->isZero()) | |||
2469 | return Y; | |||
2470 | if (auto *CY = dyn_cast<ConstantInt>(Y)) | |||
2471 | if (CY->isZero()) | |||
2472 | return X; | |||
2473 | return B.CreateAdd(X, Y); | |||
2474 | }; | |||
2475 | ||||
2476 | // We allow X to be a vector type, in which case Y will potentially be | |||
2477 | // splatted into a vector with the same element count. | |||
2478 | auto CreateMul = [&B](Value *X, Value *Y) { | |||
2479 | assert(X->getType()->getScalarType() == Y->getType() &&(static_cast <bool> (X->getType()->getScalarType( ) == Y->getType() && "Types don't match!") ? void ( 0) : __assert_fail ("X->getType()->getScalarType() == Y->getType() && \"Types don't match!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2480, __extension__ __PRETTY_FUNCTION__)) | |||
2480 | "Types don't match!")(static_cast <bool> (X->getType()->getScalarType( ) == Y->getType() && "Types don't match!") ? void ( 0) : __assert_fail ("X->getType()->getScalarType() == Y->getType() && \"Types don't match!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2480, __extension__ __PRETTY_FUNCTION__)); | |||
2481 | if (auto *CX = dyn_cast<ConstantInt>(X)) | |||
2482 | if (CX->isOne()) | |||
2483 | return Y; | |||
2484 | if (auto *CY = dyn_cast<ConstantInt>(Y)) | |||
2485 | if (CY->isOne()) | |||
2486 | return X; | |||
2487 | VectorType *XVTy = dyn_cast<VectorType>(X->getType()); | |||
2488 | if (XVTy && !isa<VectorType>(Y->getType())) | |||
2489 | Y = B.CreateVectorSplat(XVTy->getElementCount(), Y); | |||
2490 | return B.CreateMul(X, Y); | |||
2491 | }; | |||
2492 | ||||
2493 | switch (ID.getKind()) { | |||
2494 | case InductionDescriptor::IK_IntInduction: { | |||
2495 | assert(!isa<VectorType>(Index->getType()) &&(static_cast <bool> (!isa<VectorType>(Index->getType ()) && "Vector indices not supported for integer inductions yet" ) ? void (0) : __assert_fail ("!isa<VectorType>(Index->getType()) && \"Vector indices not supported for integer inductions yet\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2496, __extension__ __PRETTY_FUNCTION__)) | |||
2496 | "Vector indices not supported for integer inductions yet")(static_cast <bool> (!isa<VectorType>(Index->getType ()) && "Vector indices not supported for integer inductions yet" ) ? void (0) : __assert_fail ("!isa<VectorType>(Index->getType()) && \"Vector indices not supported for integer inductions yet\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2496, __extension__ __PRETTY_FUNCTION__)); | |||
2497 | assert(Index->getType() == StartValue->getType() &&(static_cast <bool> (Index->getType() == StartValue-> getType() && "Index type does not match StartValue type" ) ? void (0) : __assert_fail ("Index->getType() == StartValue->getType() && \"Index type does not match StartValue type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2498, __extension__ __PRETTY_FUNCTION__)) | |||
2498 | "Index type does not match StartValue type")(static_cast <bool> (Index->getType() == StartValue-> getType() && "Index type does not match StartValue type" ) ? void (0) : __assert_fail ("Index->getType() == StartValue->getType() && \"Index type does not match StartValue type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2498, __extension__ __PRETTY_FUNCTION__)); | |||
2499 | if (isa<ConstantInt>(Step) && cast<ConstantInt>(Step)->isMinusOne()) | |||
2500 | return B.CreateSub(StartValue, Index); | |||
2501 | auto *Offset = CreateMul(Index, Step); | |||
2502 | return CreateAdd(StartValue, Offset); | |||
2503 | } | |||
2504 | case InductionDescriptor::IK_PtrInduction: { | |||
2505 | assert(isa<Constant>(Step) &&(static_cast <bool> (isa<Constant>(Step) && "Expected constant step for pointer induction") ? void (0) : __assert_fail ("isa<Constant>(Step) && \"Expected constant step for pointer induction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2506, __extension__ __PRETTY_FUNCTION__)) | |||
2506 | "Expected constant step for pointer induction")(static_cast <bool> (isa<Constant>(Step) && "Expected constant step for pointer induction") ? void (0) : __assert_fail ("isa<Constant>(Step) && \"Expected constant step for pointer induction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2506, __extension__ __PRETTY_FUNCTION__)); | |||
2507 | return B.CreateGEP(ID.getElementType(), StartValue, CreateMul(Index, Step)); | |||
2508 | } | |||
2509 | case InductionDescriptor::IK_FpInduction: { | |||
2510 | assert(!isa<VectorType>(Index->getType()) &&(static_cast <bool> (!isa<VectorType>(Index->getType ()) && "Vector indices not supported for FP inductions yet" ) ? void (0) : __assert_fail ("!isa<VectorType>(Index->getType()) && \"Vector indices not supported for FP inductions yet\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2511, __extension__ __PRETTY_FUNCTION__)) | |||
2511 | "Vector indices not supported for FP inductions yet")(static_cast <bool> (!isa<VectorType>(Index->getType ()) && "Vector indices not supported for FP inductions yet" ) ? void (0) : __assert_fail ("!isa<VectorType>(Index->getType()) && \"Vector indices not supported for FP inductions yet\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2511, __extension__ __PRETTY_FUNCTION__)); | |||
2512 | assert(Step->getType()->isFloatingPointTy() && "Expected FP Step value")(static_cast <bool> (Step->getType()->isFloatingPointTy () && "Expected FP Step value") ? void (0) : __assert_fail ("Step->getType()->isFloatingPointTy() && \"Expected FP Step value\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2512, __extension__ __PRETTY_FUNCTION__)); | |||
2513 | auto InductionBinOp = ID.getInductionBinOp(); | |||
2514 | assert(InductionBinOp &&(static_cast <bool> (InductionBinOp && (InductionBinOp ->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode () == Instruction::FSub) && "Original bin op should be defined for FP induction" ) ? void (0) : __assert_fail ("InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction::FSub) && \"Original bin op should be defined for FP induction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2517, __extension__ __PRETTY_FUNCTION__)) | |||
2515 | (InductionBinOp->getOpcode() == Instruction::FAdd ||(static_cast <bool> (InductionBinOp && (InductionBinOp ->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode () == Instruction::FSub) && "Original bin op should be defined for FP induction" ) ? void (0) : __assert_fail ("InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction::FSub) && \"Original bin op should be defined for FP induction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2517, __extension__ __PRETTY_FUNCTION__)) | |||
2516 | InductionBinOp->getOpcode() == Instruction::FSub) &&(static_cast <bool> (InductionBinOp && (InductionBinOp ->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode () == Instruction::FSub) && "Original bin op should be defined for FP induction" ) ? void (0) : __assert_fail ("InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction::FSub) && \"Original bin op should be defined for FP induction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2517, __extension__ __PRETTY_FUNCTION__)) | |||
2517 | "Original bin op should be defined for FP induction")(static_cast <bool> (InductionBinOp && (InductionBinOp ->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode () == Instruction::FSub) && "Original bin op should be defined for FP induction" ) ? void (0) : __assert_fail ("InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction::FSub) && \"Original bin op should be defined for FP induction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2517, __extension__ __PRETTY_FUNCTION__)); | |||
2518 | ||||
2519 | Value *MulExp = B.CreateFMul(Step, Index); | |||
2520 | return B.CreateBinOp(InductionBinOp->getOpcode(), StartValue, MulExp, | |||
2521 | "induction"); | |||
2522 | } | |||
2523 | case InductionDescriptor::IK_NoInduction: | |||
2524 | return nullptr; | |||
2525 | } | |||
2526 | llvm_unreachable("invalid enum")::llvm::llvm_unreachable_internal("invalid enum", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2526); | |||
2527 | } | |||
2528 | ||||
2529 | void InnerLoopVectorizer::packScalarIntoVectorValue(VPValue *Def, | |||
2530 | const VPIteration &Instance, | |||
2531 | VPTransformState &State) { | |||
2532 | Value *ScalarInst = State.get(Def, Instance); | |||
2533 | Value *VectorValue = State.get(Def, Instance.Part); | |||
2534 | VectorValue = Builder.CreateInsertElement( | |||
2535 | VectorValue, ScalarInst, | |||
2536 | Instance.Lane.getAsRuntimeExpr(State.Builder, VF)); | |||
2537 | State.set(Def, VectorValue, Instance.Part); | |||
2538 | } | |||
2539 | ||||
2540 | // Return whether we allow using masked interleave-groups (for dealing with | |||
2541 | // strided loads/stores that reside in predicated blocks, or for dealing | |||
2542 | // with gaps). | |||
2543 | static bool useMaskedInterleavedAccesses(const TargetTransformInfo &TTI) { | |||
2544 | // If an override option has been passed in for interleaved accesses, use it. | |||
2545 | if (EnableMaskedInterleavedMemAccesses.getNumOccurrences() > 0) | |||
2546 | return EnableMaskedInterleavedMemAccesses; | |||
2547 | ||||
2548 | return TTI.enableMaskedInterleavedAccessVectorization(); | |||
2549 | } | |||
2550 | ||||
2551 | // Try to vectorize the interleave group that \p Instr belongs to. | |||
2552 | // | |||
2553 | // E.g. Translate following interleaved load group (factor = 3): | |||
2554 | // for (i = 0; i < N; i+=3) { | |||
2555 | // R = Pic[i]; // Member of index 0 | |||
2556 | // G = Pic[i+1]; // Member of index 1 | |||
2557 | // B = Pic[i+2]; // Member of index 2 | |||
2558 | // ... // do something to R, G, B | |||
2559 | // } | |||
2560 | // To: | |||
2561 | // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B | |||
2562 | // %R.vec = shuffle %wide.vec, poison, <0, 3, 6, 9> ; R elements | |||
2563 | // %G.vec = shuffle %wide.vec, poison, <1, 4, 7, 10> ; G elements | |||
2564 | // %B.vec = shuffle %wide.vec, poison, <2, 5, 8, 11> ; B elements | |||
2565 | // | |||
2566 | // Or translate following interleaved store group (factor = 3): | |||
2567 | // for (i = 0; i < N; i+=3) { | |||
2568 | // ... do something to R, G, B | |||
2569 | // Pic[i] = R; // Member of index 0 | |||
2570 | // Pic[i+1] = G; // Member of index 1 | |||
2571 | // Pic[i+2] = B; // Member of index 2 | |||
2572 | // } | |||
2573 | // To: | |||
2574 | // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7> | |||
2575 | // %B_U.vec = shuffle %B.vec, poison, <0, 1, 2, 3, u, u, u, u> | |||
2576 | // %interleaved.vec = shuffle %R_G.vec, %B_U.vec, | |||
2577 | // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements | |||
2578 | // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B | |||
2579 | void InnerLoopVectorizer::vectorizeInterleaveGroup( | |||
2580 | const InterleaveGroup<Instruction> *Group, ArrayRef<VPValue *> VPDefs, | |||
2581 | VPTransformState &State, VPValue *Addr, ArrayRef<VPValue *> StoredValues, | |||
2582 | VPValue *BlockInMask) { | |||
2583 | Instruction *Instr = Group->getInsertPos(); | |||
2584 | const DataLayout &DL = Instr->getModule()->getDataLayout(); | |||
2585 | ||||
2586 | // Prepare for the vector type of the interleaved load/store. | |||
2587 | Type *ScalarTy = getLoadStoreType(Instr); | |||
2588 | unsigned InterleaveFactor = Group->getFactor(); | |||
2589 | assert(!VF.isScalable() && "scalable vectors not yet supported.")(static_cast <bool> (!VF.isScalable() && "scalable vectors not yet supported." ) ? void (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2589, __extension__ __PRETTY_FUNCTION__)); | |||
2590 | auto *VecTy = VectorType::get(ScalarTy, VF * InterleaveFactor); | |||
2591 | ||||
2592 | // Prepare for the new pointers. | |||
2593 | SmallVector<Value *, 2> AddrParts; | |||
2594 | unsigned Index = Group->getIndex(Instr); | |||
2595 | ||||
2596 | // TODO: extend the masked interleaved-group support to reversed access. | |||
2597 | assert((!BlockInMask || !Group->isReverse()) &&(static_cast <bool> ((!BlockInMask || !Group->isReverse ()) && "Reversed masked interleave-group not supported." ) ? void (0) : __assert_fail ("(!BlockInMask || !Group->isReverse()) && \"Reversed masked interleave-group not supported.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2598, __extension__ __PRETTY_FUNCTION__)) | |||
2598 | "Reversed masked interleave-group not supported.")(static_cast <bool> ((!BlockInMask || !Group->isReverse ()) && "Reversed masked interleave-group not supported." ) ? void (0) : __assert_fail ("(!BlockInMask || !Group->isReverse()) && \"Reversed masked interleave-group not supported.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2598, __extension__ __PRETTY_FUNCTION__)); | |||
2599 | ||||
2600 | // If the group is reverse, adjust the index to refer to the last vector lane | |||
2601 | // instead of the first. We adjust the index from the first vector lane, | |||
2602 | // rather than directly getting the pointer for lane VF - 1, because the | |||
2603 | // pointer operand of the interleaved access is supposed to be uniform. For | |||
2604 | // uniform instructions, we're only required to generate a value for the | |||
2605 | // first vector lane in each unroll iteration. | |||
2606 | if (Group->isReverse()) | |||
2607 | Index += (VF.getKnownMinValue() - 1) * Group->getFactor(); | |||
2608 | ||||
2609 | for (unsigned Part = 0; Part < UF; Part++) { | |||
2610 | Value *AddrPart = State.get(Addr, VPIteration(Part, 0)); | |||
2611 | setDebugLocFromInst(AddrPart); | |||
2612 | ||||
2613 | // Notice current instruction could be any index. Need to adjust the address | |||
2614 | // to the member of index 0. | |||
2615 | // | |||
2616 | // E.g. a = A[i+1]; // Member of index 1 (Current instruction) | |||
2617 | // b = A[i]; // Member of index 0 | |||
2618 | // Current pointer is pointed to A[i+1], adjust it to A[i]. | |||
2619 | // | |||
2620 | // E.g. A[i+1] = a; // Member of index 1 | |||
2621 | // A[i] = b; // Member of index 0 | |||
2622 | // A[i+2] = c; // Member of index 2 (Current instruction) | |||
2623 | // Current pointer is pointed to A[i+2], adjust it to A[i]. | |||
2624 | ||||
2625 | bool InBounds = false; | |||
2626 | if (auto *gep = dyn_cast<GetElementPtrInst>(AddrPart->stripPointerCasts())) | |||
2627 | InBounds = gep->isInBounds(); | |||
2628 | AddrPart = Builder.CreateGEP(ScalarTy, AddrPart, Builder.getInt32(-Index)); | |||
2629 | cast<GetElementPtrInst>(AddrPart)->setIsInBounds(InBounds); | |||
2630 | ||||
2631 | // Cast to the vector pointer type. | |||
2632 | unsigned AddressSpace = AddrPart->getType()->getPointerAddressSpace(); | |||
2633 | Type *PtrTy = VecTy->getPointerTo(AddressSpace); | |||
2634 | AddrParts.push_back(Builder.CreateBitCast(AddrPart, PtrTy)); | |||
2635 | } | |||
2636 | ||||
2637 | setDebugLocFromInst(Instr); | |||
2638 | Value *PoisonVec = PoisonValue::get(VecTy); | |||
2639 | ||||
2640 | Value *MaskForGaps = nullptr; | |||
2641 | if (Group->requiresScalarEpilogue() && !Cost->isScalarEpilogueAllowed()) { | |||
2642 | MaskForGaps = createBitMaskForGaps(Builder, VF.getKnownMinValue(), *Group); | |||
2643 | assert(MaskForGaps && "Mask for Gaps is required but it is null")(static_cast <bool> (MaskForGaps && "Mask for Gaps is required but it is null" ) ? void (0) : __assert_fail ("MaskForGaps && \"Mask for Gaps is required but it is null\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2643, __extension__ __PRETTY_FUNCTION__)); | |||
2644 | } | |||
2645 | ||||
2646 | // Vectorize the interleaved load group. | |||
2647 | if (isa<LoadInst>(Instr)) { | |||
2648 | // For each unroll part, create a wide load for the group. | |||
2649 | SmallVector<Value *, 2> NewLoads; | |||
2650 | for (unsigned Part = 0; Part < UF; Part++) { | |||
2651 | Instruction *NewLoad; | |||
2652 | if (BlockInMask || MaskForGaps) { | |||
2653 | assert(useMaskedInterleavedAccesses(*TTI) &&(static_cast <bool> (useMaskedInterleavedAccesses(*TTI) && "masked interleaved groups are not allowed.") ? void (0) : __assert_fail ("useMaskedInterleavedAccesses(*TTI) && \"masked interleaved groups are not allowed.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2654, __extension__ __PRETTY_FUNCTION__)) | |||
2654 | "masked interleaved groups are not allowed.")(static_cast <bool> (useMaskedInterleavedAccesses(*TTI) && "masked interleaved groups are not allowed.") ? void (0) : __assert_fail ("useMaskedInterleavedAccesses(*TTI) && \"masked interleaved groups are not allowed.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2654, __extension__ __PRETTY_FUNCTION__)); | |||
2655 | Value *GroupMask = MaskForGaps; | |||
2656 | if (BlockInMask) { | |||
2657 | Value *BlockInMaskPart = State.get(BlockInMask, Part); | |||
2658 | Value *ShuffledMask = Builder.CreateShuffleVector( | |||
2659 | BlockInMaskPart, | |||
2660 | createReplicatedMask(InterleaveFactor, VF.getKnownMinValue()), | |||
2661 | "interleaved.mask"); | |||
2662 | GroupMask = MaskForGaps | |||
2663 | ? Builder.CreateBinOp(Instruction::And, ShuffledMask, | |||
2664 | MaskForGaps) | |||
2665 | : ShuffledMask; | |||
2666 | } | |||
2667 | NewLoad = | |||
2668 | Builder.CreateMaskedLoad(VecTy, AddrParts[Part], Group->getAlign(), | |||
2669 | GroupMask, PoisonVec, "wide.masked.vec"); | |||
2670 | } | |||
2671 | else | |||
2672 | NewLoad = Builder.CreateAlignedLoad(VecTy, AddrParts[Part], | |||
2673 | Group->getAlign(), "wide.vec"); | |||
2674 | Group->addMetadata(NewLoad); | |||
2675 | NewLoads.push_back(NewLoad); | |||
2676 | } | |||
2677 | ||||
2678 | // For each member in the group, shuffle out the appropriate data from the | |||
2679 | // wide loads. | |||
2680 | unsigned J = 0; | |||
2681 | for (unsigned I = 0; I < InterleaveFactor; ++I) { | |||
2682 | Instruction *Member = Group->getMember(I); | |||
2683 | ||||
2684 | // Skip the gaps in the group. | |||
2685 | if (!Member) | |||
2686 | continue; | |||
2687 | ||||
2688 | auto StrideMask = | |||
2689 | createStrideMask(I, InterleaveFactor, VF.getKnownMinValue()); | |||
2690 | for (unsigned Part = 0; Part < UF; Part++) { | |||
2691 | Value *StridedVec = Builder.CreateShuffleVector( | |||
2692 | NewLoads[Part], StrideMask, "strided.vec"); | |||
2693 | ||||
2694 | // If this member has different type, cast the result type. | |||
2695 | if (Member->getType() != ScalarTy) { | |||
2696 | assert(!VF.isScalable() && "VF is assumed to be non scalable.")(static_cast <bool> (!VF.isScalable() && "VF is assumed to be non scalable." ) ? void (0) : __assert_fail ("!VF.isScalable() && \"VF is assumed to be non scalable.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2696, __extension__ __PRETTY_FUNCTION__)); | |||
2697 | VectorType *OtherVTy = VectorType::get(Member->getType(), VF); | |||
2698 | StridedVec = createBitOrPointerCast(StridedVec, OtherVTy, DL); | |||
2699 | } | |||
2700 | ||||
2701 | if (Group->isReverse()) | |||
2702 | StridedVec = Builder.CreateVectorReverse(StridedVec, "reverse"); | |||
2703 | ||||
2704 | State.set(VPDefs[J], StridedVec, Part); | |||
2705 | } | |||
2706 | ++J; | |||
2707 | } | |||
2708 | return; | |||
2709 | } | |||
2710 | ||||
2711 | // The sub vector type for current instruction. | |||
2712 | auto *SubVT = VectorType::get(ScalarTy, VF); | |||
2713 | ||||
2714 | // Vectorize the interleaved store group. | |||
2715 | MaskForGaps = createBitMaskForGaps(Builder, VF.getKnownMinValue(), *Group); | |||
2716 | assert((!MaskForGaps || useMaskedInterleavedAccesses(*TTI)) &&(static_cast <bool> ((!MaskForGaps || useMaskedInterleavedAccesses (*TTI)) && "masked interleaved groups are not allowed." ) ? void (0) : __assert_fail ("(!MaskForGaps || useMaskedInterleavedAccesses(*TTI)) && \"masked interleaved groups are not allowed.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2717, __extension__ __PRETTY_FUNCTION__)) | |||
2717 | "masked interleaved groups are not allowed.")(static_cast <bool> ((!MaskForGaps || useMaskedInterleavedAccesses (*TTI)) && "masked interleaved groups are not allowed." ) ? void (0) : __assert_fail ("(!MaskForGaps || useMaskedInterleavedAccesses(*TTI)) && \"masked interleaved groups are not allowed.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2717, __extension__ __PRETTY_FUNCTION__)); | |||
2718 | assert((!MaskForGaps || !VF.isScalable()) &&(static_cast <bool> ((!MaskForGaps || !VF.isScalable()) && "masking gaps for scalable vectors is not yet supported." ) ? void (0) : __assert_fail ("(!MaskForGaps || !VF.isScalable()) && \"masking gaps for scalable vectors is not yet supported.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2719, __extension__ __PRETTY_FUNCTION__)) | |||
2719 | "masking gaps for scalable vectors is not yet supported.")(static_cast <bool> ((!MaskForGaps || !VF.isScalable()) && "masking gaps for scalable vectors is not yet supported." ) ? void (0) : __assert_fail ("(!MaskForGaps || !VF.isScalable()) && \"masking gaps for scalable vectors is not yet supported.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2719, __extension__ __PRETTY_FUNCTION__)); | |||
2720 | for (unsigned Part = 0; Part < UF; Part++) { | |||
2721 | // Collect the stored vector from each member. | |||
2722 | SmallVector<Value *, 4> StoredVecs; | |||
2723 | for (unsigned i = 0; i < InterleaveFactor; i++) { | |||
2724 | assert((Group->getMember(i) || MaskForGaps) &&(static_cast <bool> ((Group->getMember(i) || MaskForGaps ) && "Fail to get a member from an interleaved store group" ) ? void (0) : __assert_fail ("(Group->getMember(i) || MaskForGaps) && \"Fail to get a member from an interleaved store group\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2725, __extension__ __PRETTY_FUNCTION__)) | |||
2725 | "Fail to get a member from an interleaved store group")(static_cast <bool> ((Group->getMember(i) || MaskForGaps ) && "Fail to get a member from an interleaved store group" ) ? void (0) : __assert_fail ("(Group->getMember(i) || MaskForGaps) && \"Fail to get a member from an interleaved store group\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2725, __extension__ __PRETTY_FUNCTION__)); | |||
2726 | Instruction *Member = Group->getMember(i); | |||
2727 | ||||
2728 | // Skip the gaps in the group. | |||
2729 | if (!Member) { | |||
2730 | Value *Undef = PoisonValue::get(SubVT); | |||
2731 | StoredVecs.push_back(Undef); | |||
2732 | continue; | |||
2733 | } | |||
2734 | ||||
2735 | Value *StoredVec = State.get(StoredValues[i], Part); | |||
2736 | ||||
2737 | if (Group->isReverse()) | |||
2738 | StoredVec = Builder.CreateVectorReverse(StoredVec, "reverse"); | |||
2739 | ||||
2740 | // If this member has different type, cast it to a unified type. | |||
2741 | ||||
2742 | if (StoredVec->getType() != SubVT) | |||
2743 | StoredVec = createBitOrPointerCast(StoredVec, SubVT, DL); | |||
2744 | ||||
2745 | StoredVecs.push_back(StoredVec); | |||
2746 | } | |||
2747 | ||||
2748 | // Concatenate all vectors into a wide vector. | |||
2749 | Value *WideVec = concatenateVectors(Builder, StoredVecs); | |||
2750 | ||||
2751 | // Interleave the elements in the wide vector. | |||
2752 | Value *IVec = Builder.CreateShuffleVector( | |||
2753 | WideVec, createInterleaveMask(VF.getKnownMinValue(), InterleaveFactor), | |||
2754 | "interleaved.vec"); | |||
2755 | ||||
2756 | Instruction *NewStoreInstr; | |||
2757 | if (BlockInMask || MaskForGaps) { | |||
2758 | Value *GroupMask = MaskForGaps; | |||
2759 | if (BlockInMask) { | |||
2760 | Value *BlockInMaskPart = State.get(BlockInMask, Part); | |||
2761 | Value *ShuffledMask = Builder.CreateShuffleVector( | |||
2762 | BlockInMaskPart, | |||
2763 | createReplicatedMask(InterleaveFactor, VF.getKnownMinValue()), | |||
2764 | "interleaved.mask"); | |||
2765 | GroupMask = MaskForGaps ? Builder.CreateBinOp(Instruction::And, | |||
2766 | ShuffledMask, MaskForGaps) | |||
2767 | : ShuffledMask; | |||
2768 | } | |||
2769 | NewStoreInstr = Builder.CreateMaskedStore(IVec, AddrParts[Part], | |||
2770 | Group->getAlign(), GroupMask); | |||
2771 | } else | |||
2772 | NewStoreInstr = | |||
2773 | Builder.CreateAlignedStore(IVec, AddrParts[Part], Group->getAlign()); | |||
2774 | ||||
2775 | Group->addMetadata(NewStoreInstr); | |||
2776 | } | |||
2777 | } | |||
2778 | ||||
2779 | void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, | |||
2780 | VPReplicateRecipe *RepRecipe, | |||
2781 | const VPIteration &Instance, | |||
2782 | bool IfPredicateInstr, | |||
2783 | VPTransformState &State) { | |||
2784 | assert(!Instr->getType()->isAggregateType() && "Can't handle vectors")(static_cast <bool> (!Instr->getType()->isAggregateType () && "Can't handle vectors") ? void (0) : __assert_fail ("!Instr->getType()->isAggregateType() && \"Can't handle vectors\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2784, __extension__ __PRETTY_FUNCTION__)); | |||
2785 | ||||
2786 | // llvm.experimental.noalias.scope.decl intrinsics must only be duplicated for | |||
2787 | // the first lane and part. | |||
2788 | if (isa<NoAliasScopeDeclInst>(Instr)) | |||
2789 | if (!Instance.isFirstIteration()) | |||
2790 | return; | |||
2791 | ||||
2792 | setDebugLocFromInst(Instr); | |||
2793 | ||||
2794 | // Does this instruction return a value ? | |||
2795 | bool IsVoidRetTy = Instr->getType()->isVoidTy(); | |||
2796 | ||||
2797 | Instruction *Cloned = Instr->clone(); | |||
2798 | if (!IsVoidRetTy) | |||
2799 | Cloned->setName(Instr->getName() + ".cloned"); | |||
2800 | ||||
2801 | // If the scalarized instruction contributes to the address computation of a | |||
2802 | // widen masked load/store which was in a basic block that needed predication | |||
2803 | // and is not predicated after vectorization, we can't propagate | |||
2804 | // poison-generating flags (nuw/nsw, exact, inbounds, etc.). The scalarized | |||
2805 | // instruction could feed a poison value to the base address of the widen | |||
2806 | // load/store. | |||
2807 | if (State.MayGeneratePoisonRecipes.contains(RepRecipe)) | |||
2808 | Cloned->dropPoisonGeneratingFlags(); | |||
2809 | ||||
2810 | State.Builder.SetInsertPoint(Builder.GetInsertBlock(), | |||
2811 | Builder.GetInsertPoint()); | |||
2812 | // Replace the operands of the cloned instructions with their scalar | |||
2813 | // equivalents in the new loop. | |||
2814 | for (auto &I : enumerate(RepRecipe->operands())) { | |||
2815 | auto InputInstance = Instance; | |||
2816 | VPValue *Operand = I.value(); | |||
2817 | VPReplicateRecipe *OperandR = dyn_cast<VPReplicateRecipe>(Operand); | |||
2818 | if (OperandR && OperandR->isUniform()) | |||
2819 | InputInstance.Lane = VPLane::getFirstLane(); | |||
2820 | Cloned->setOperand(I.index(), State.get(Operand, InputInstance)); | |||
2821 | } | |||
2822 | addNewMetadata(Cloned, Instr); | |||
2823 | ||||
2824 | // Place the cloned scalar in the new loop. | |||
2825 | Builder.Insert(Cloned); | |||
2826 | ||||
2827 | State.set(RepRecipe, Cloned, Instance); | |||
2828 | ||||
2829 | // If we just cloned a new assumption, add it the assumption cache. | |||
2830 | if (auto *II = dyn_cast<AssumeInst>(Cloned)) | |||
2831 | AC->registerAssumption(II); | |||
2832 | ||||
2833 | // End if-block. | |||
2834 | if (IfPredicateInstr) | |||
2835 | PredicatedInstructions.push_back(Cloned); | |||
2836 | } | |||
2837 | ||||
2838 | void InnerLoopVectorizer::createHeaderBranch(Loop *L) { | |||
2839 | BasicBlock *Header = L->getHeader(); | |||
2840 | assert(!L->getLoopLatch() && "loop should not have a latch at this point")(static_cast <bool> (!L->getLoopLatch() && "loop should not have a latch at this point" ) ? void (0) : __assert_fail ("!L->getLoopLatch() && \"loop should not have a latch at this point\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2840, __extension__ __PRETTY_FUNCTION__)); | |||
2841 | ||||
2842 | IRBuilder<> B(Header->getTerminator()); | |||
2843 | Instruction *OldInst = | |||
2844 | getDebugLocFromInstOrOperands(Legal->getPrimaryInduction()); | |||
2845 | setDebugLocFromInst(OldInst, &B); | |||
2846 | ||||
2847 | // Connect the header to the exit and header blocks and replace the old | |||
2848 | // terminator. | |||
2849 | B.CreateCondBr(B.getTrue(), L->getUniqueExitBlock(), Header); | |||
2850 | ||||
2851 | // Now we have two terminators. Remove the old one from the block. | |||
2852 | Header->getTerminator()->eraseFromParent(); | |||
2853 | } | |||
2854 | ||||
2855 | Value *InnerLoopVectorizer::getOrCreateTripCount(Loop *L) { | |||
2856 | if (TripCount) | |||
2857 | return TripCount; | |||
2858 | ||||
2859 | assert(L && "Create Trip Count for null loop.")(static_cast <bool> (L && "Create Trip Count for null loop." ) ? void (0) : __assert_fail ("L && \"Create Trip Count for null loop.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2859, __extension__ __PRETTY_FUNCTION__)); | |||
2860 | IRBuilder<> Builder(L->getLoopPreheader()->getTerminator()); | |||
2861 | // Find the loop boundaries. | |||
2862 | ScalarEvolution *SE = PSE.getSE(); | |||
2863 | const SCEV *BackedgeTakenCount = PSE.getBackedgeTakenCount(); | |||
2864 | assert(!isa<SCEVCouldNotCompute>(BackedgeTakenCount) &&(static_cast <bool> (!isa<SCEVCouldNotCompute>(BackedgeTakenCount ) && "Invalid loop count") ? void (0) : __assert_fail ("!isa<SCEVCouldNotCompute>(BackedgeTakenCount) && \"Invalid loop count\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2865, __extension__ __PRETTY_FUNCTION__)) | |||
2865 | "Invalid loop count")(static_cast <bool> (!isa<SCEVCouldNotCompute>(BackedgeTakenCount ) && "Invalid loop count") ? void (0) : __assert_fail ("!isa<SCEVCouldNotCompute>(BackedgeTakenCount) && \"Invalid loop count\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2865, __extension__ __PRETTY_FUNCTION__)); | |||
2866 | ||||
2867 | Type *IdxTy = Legal->getWidestInductionType(); | |||
2868 | assert(IdxTy && "No type for induction")(static_cast <bool> (IdxTy && "No type for induction" ) ? void (0) : __assert_fail ("IdxTy && \"No type for induction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2868, __extension__ __PRETTY_FUNCTION__)); | |||
2869 | ||||
2870 | // The exit count might have the type of i64 while the phi is i32. This can | |||
2871 | // happen if we have an induction variable that is sign extended before the | |||
2872 | // compare. The only way that we get a backedge taken count is that the | |||
2873 | // induction variable was signed and as such will not overflow. In such a case | |||
2874 | // truncation is legal. | |||
2875 | if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) > | |||
2876 | IdxTy->getPrimitiveSizeInBits()) | |||
2877 | BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, IdxTy); | |||
2878 | BackedgeTakenCount = SE->getNoopOrZeroExtend(BackedgeTakenCount, IdxTy); | |||
2879 | ||||
2880 | // Get the total trip count from the count by adding 1. | |||
2881 | const SCEV *ExitCount = SE->getAddExpr( | |||
2882 | BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType())); | |||
2883 | ||||
2884 | const DataLayout &DL = L->getHeader()->getModule()->getDataLayout(); | |||
2885 | ||||
2886 | // Expand the trip count and place the new instructions in the preheader. | |||
2887 | // Notice that the pre-header does not change, only the loop body. | |||
2888 | SCEVExpander Exp(*SE, DL, "induction"); | |||
2889 | ||||
2890 | // Count holds the overall loop count (N). | |||
2891 | TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(), | |||
2892 | L->getLoopPreheader()->getTerminator()); | |||
2893 | ||||
2894 | if (TripCount->getType()->isPointerTy()) | |||
2895 | TripCount = | |||
2896 | CastInst::CreatePointerCast(TripCount, IdxTy, "exitcount.ptrcnt.to.int", | |||
2897 | L->getLoopPreheader()->getTerminator()); | |||
2898 | ||||
2899 | return TripCount; | |||
2900 | } | |||
2901 | ||||
2902 | Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) { | |||
2903 | if (VectorTripCount) | |||
2904 | return VectorTripCount; | |||
2905 | ||||
2906 | Value *TC = getOrCreateTripCount(L); | |||
2907 | IRBuilder<> Builder(L->getLoopPreheader()->getTerminator()); | |||
| ||||
2908 | ||||
2909 | Type *Ty = TC->getType(); | |||
2910 | // This is where we can make the step a runtime constant. | |||
2911 | Value *Step = createStepForVF(Builder, Ty, VF, UF); | |||
2912 | ||||
2913 | // If the tail is to be folded by masking, round the number of iterations N | |||
2914 | // up to a multiple of Step instead of rounding down. This is done by first | |||
2915 | // adding Step-1 and then rounding down. Note that it's ok if this addition | |||
2916 | // overflows: the vector induction variable will eventually wrap to zero given | |||
2917 | // that it starts at zero and its Step is a power of two; the loop will then | |||
2918 | // exit, with the last early-exit vector comparison also producing all-true. | |||
2919 | if (Cost->foldTailByMasking()) { | |||
2920 | assert(isPowerOf2_32(VF.getKnownMinValue() * UF) &&(static_cast <bool> (isPowerOf2_32(VF.getKnownMinValue( ) * UF) && "VF*UF must be a power of 2 when folding tail by masking" ) ? void (0) : __assert_fail ("isPowerOf2_32(VF.getKnownMinValue() * UF) && \"VF*UF must be a power of 2 when folding tail by masking\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2921, __extension__ __PRETTY_FUNCTION__)) | |||
2921 | "VF*UF must be a power of 2 when folding tail by masking")(static_cast <bool> (isPowerOf2_32(VF.getKnownMinValue( ) * UF) && "VF*UF must be a power of 2 when folding tail by masking" ) ? void (0) : __assert_fail ("isPowerOf2_32(VF.getKnownMinValue() * UF) && \"VF*UF must be a power of 2 when folding tail by masking\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2921, __extension__ __PRETTY_FUNCTION__)); | |||
2922 | Value *NumLanes = getRuntimeVF(Builder, Ty, VF * UF); | |||
2923 | TC = Builder.CreateAdd( | |||
2924 | TC, Builder.CreateSub(NumLanes, ConstantInt::get(Ty, 1)), "n.rnd.up"); | |||
2925 | } | |||
2926 | ||||
2927 | // Now we need to generate the expression for the part of the loop that the | |||
2928 | // vectorized body will execute. This is equal to N - (N % Step) if scalar | |||
2929 | // iterations are not required for correctness, or N - Step, otherwise. Step | |||
2930 | // is equal to the vectorization factor (number of SIMD elements) times the | |||
2931 | // unroll factor (number of SIMD instructions). | |||
2932 | Value *R = Builder.CreateURem(TC, Step, "n.mod.vf"); | |||
2933 | ||||
2934 | // There are cases where we *must* run at least one iteration in the remainder | |||
2935 | // loop. See the cost model for when this can happen. If the step evenly | |||
2936 | // divides the trip count, we set the remainder to be equal to the step. If | |||
2937 | // the step does not evenly divide the trip count, no adjustment is necessary | |||
2938 | // since there will already be scalar iterations. Note that the minimum | |||
2939 | // iterations check ensures that N >= Step. | |||
2940 | if (Cost->requiresScalarEpilogue(VF)) { | |||
2941 | auto *IsZero = Builder.CreateICmpEQ(R, ConstantInt::get(R->getType(), 0)); | |||
2942 | R = Builder.CreateSelect(IsZero, Step, R); | |||
2943 | } | |||
2944 | ||||
2945 | VectorTripCount = Builder.CreateSub(TC, R, "n.vec"); | |||
2946 | ||||
2947 | return VectorTripCount; | |||
2948 | } | |||
2949 | ||||
2950 | Value *InnerLoopVectorizer::createBitOrPointerCast(Value *V, VectorType *DstVTy, | |||
2951 | const DataLayout &DL) { | |||
2952 | // Verify that V is a vector type with same number of elements as DstVTy. | |||
2953 | auto *DstFVTy = cast<FixedVectorType>(DstVTy); | |||
2954 | unsigned VF = DstFVTy->getNumElements(); | |||
2955 | auto *SrcVecTy = cast<FixedVectorType>(V->getType()); | |||
2956 | assert((VF == SrcVecTy->getNumElements()) && "Vector dimensions do not match")(static_cast <bool> ((VF == SrcVecTy->getNumElements ()) && "Vector dimensions do not match") ? void (0) : __assert_fail ("(VF == SrcVecTy->getNumElements()) && \"Vector dimensions do not match\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2956, __extension__ __PRETTY_FUNCTION__)); | |||
2957 | Type *SrcElemTy = SrcVecTy->getElementType(); | |||
2958 | Type *DstElemTy = DstFVTy->getElementType(); | |||
2959 | assert((DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy)) &&(static_cast <bool> ((DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy)) && "Vector elements must have same size" ) ? void (0) : __assert_fail ("(DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy)) && \"Vector elements must have same size\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2960, __extension__ __PRETTY_FUNCTION__)) | |||
2960 | "Vector elements must have same size")(static_cast <bool> ((DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy)) && "Vector elements must have same size" ) ? void (0) : __assert_fail ("(DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy)) && \"Vector elements must have same size\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2960, __extension__ __PRETTY_FUNCTION__)); | |||
2961 | ||||
2962 | // Do a direct cast if element types are castable. | |||
2963 | if (CastInst::isBitOrNoopPointerCastable(SrcElemTy, DstElemTy, DL)) { | |||
2964 | return Builder.CreateBitOrPointerCast(V, DstFVTy); | |||
2965 | } | |||
2966 | // V cannot be directly casted to desired vector type. | |||
2967 | // May happen when V is a floating point vector but DstVTy is a vector of | |||
2968 | // pointers or vice-versa. Handle this using a two-step bitcast using an | |||
2969 | // intermediate Integer type for the bitcast i.e. Ptr <-> Int <-> Float. | |||
2970 | assert((DstElemTy->isPointerTy() != SrcElemTy->isPointerTy()) &&(static_cast <bool> ((DstElemTy->isPointerTy() != SrcElemTy ->isPointerTy()) && "Only one type should be a pointer type" ) ? void (0) : __assert_fail ("(DstElemTy->isPointerTy() != SrcElemTy->isPointerTy()) && \"Only one type should be a pointer type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2971, __extension__ __PRETTY_FUNCTION__)) | |||
2971 | "Only one type should be a pointer type")(static_cast <bool> ((DstElemTy->isPointerTy() != SrcElemTy ->isPointerTy()) && "Only one type should be a pointer type" ) ? void (0) : __assert_fail ("(DstElemTy->isPointerTy() != SrcElemTy->isPointerTy()) && \"Only one type should be a pointer type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2971, __extension__ __PRETTY_FUNCTION__)); | |||
2972 | assert((DstElemTy->isFloatingPointTy() != SrcElemTy->isFloatingPointTy()) &&(static_cast <bool> ((DstElemTy->isFloatingPointTy() != SrcElemTy->isFloatingPointTy()) && "Only one type should be a floating point type" ) ? void (0) : __assert_fail ("(DstElemTy->isFloatingPointTy() != SrcElemTy->isFloatingPointTy()) && \"Only one type should be a floating point type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2973, __extension__ __PRETTY_FUNCTION__)) | |||
2973 | "Only one type should be a floating point type")(static_cast <bool> ((DstElemTy->isFloatingPointTy() != SrcElemTy->isFloatingPointTy()) && "Only one type should be a floating point type" ) ? void (0) : __assert_fail ("(DstElemTy->isFloatingPointTy() != SrcElemTy->isFloatingPointTy()) && \"Only one type should be a floating point type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2973, __extension__ __PRETTY_FUNCTION__)); | |||
2974 | Type *IntTy = | |||
2975 | IntegerType::getIntNTy(V->getContext(), DL.getTypeSizeInBits(SrcElemTy)); | |||
2976 | auto *VecIntTy = FixedVectorType::get(IntTy, VF); | |||
2977 | Value *CastVal = Builder.CreateBitOrPointerCast(V, VecIntTy); | |||
2978 | return Builder.CreateBitOrPointerCast(CastVal, DstFVTy); | |||
2979 | } | |||
2980 | ||||
2981 | void InnerLoopVectorizer::emitMinimumIterationCountCheck(Loop *L, | |||
2982 | BasicBlock *Bypass) { | |||
2983 | Value *Count = getOrCreateTripCount(L); | |||
2984 | // Reuse existing vector loop preheader for TC checks. | |||
2985 | // Note that new preheader block is generated for vector loop. | |||
2986 | BasicBlock *const TCCheckBlock = LoopVectorPreHeader; | |||
2987 | IRBuilder<> Builder(TCCheckBlock->getTerminator()); | |||
2988 | ||||
2989 | // Generate code to check if the loop's trip count is less than VF * UF, or | |||
2990 | // equal to it in case a scalar epilogue is required; this implies that the | |||
2991 | // vector trip count is zero. This check also covers the case where adding one | |||
2992 | // to the backedge-taken count overflowed leading to an incorrect trip count | |||
2993 | // of zero. In this case we will also jump to the scalar loop. | |||
2994 | auto P = Cost->requiresScalarEpilogue(VF) ? ICmpInst::ICMP_ULE | |||
2995 | : ICmpInst::ICMP_ULT; | |||
2996 | ||||
2997 | // If tail is to be folded, vector loop takes care of all iterations. | |||
2998 | Value *CheckMinIters = Builder.getFalse(); | |||
2999 | if (!Cost->foldTailByMasking()) { | |||
3000 | Value *Step = createStepForVF(Builder, Count->getType(), VF, UF); | |||
3001 | CheckMinIters = Builder.CreateICmp(P, Count, Step, "min.iters.check"); | |||
3002 | } | |||
3003 | // Create new preheader for vector loop. | |||
3004 | LoopVectorPreHeader = | |||
3005 | SplitBlock(TCCheckBlock, TCCheckBlock->getTerminator(), DT, LI, nullptr, | |||
3006 | "vector.ph"); | |||
3007 | ||||
3008 | assert(DT->properlyDominates(DT->getNode(TCCheckBlock),(static_cast <bool> (DT->properlyDominates(DT->getNode (TCCheckBlock), DT->getNode(Bypass)->getIDom()) && "TC check is expected to dominate Bypass") ? void (0) : __assert_fail ("DT->properlyDominates(DT->getNode(TCCheckBlock), DT->getNode(Bypass)->getIDom()) && \"TC check is expected to dominate Bypass\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3010, __extension__ __PRETTY_FUNCTION__)) | |||
3009 | DT->getNode(Bypass)->getIDom()) &&(static_cast <bool> (DT->properlyDominates(DT->getNode (TCCheckBlock), DT->getNode(Bypass)->getIDom()) && "TC check is expected to dominate Bypass") ? void (0) : __assert_fail ("DT->properlyDominates(DT->getNode(TCCheckBlock), DT->getNode(Bypass)->getIDom()) && \"TC check is expected to dominate Bypass\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3010, __extension__ __PRETTY_FUNCTION__)) | |||
3010 | "TC check is expected to dominate Bypass")(static_cast <bool> (DT->properlyDominates(DT->getNode (TCCheckBlock), DT->getNode(Bypass)->getIDom()) && "TC check is expected to dominate Bypass") ? void (0) : __assert_fail ("DT->properlyDominates(DT->getNode(TCCheckBlock), DT->getNode(Bypass)->getIDom()) && \"TC check is expected to dominate Bypass\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3010, __extension__ __PRETTY_FUNCTION__)); | |||
3011 | ||||
3012 | // Update dominator for Bypass & LoopExit (if needed). | |||
3013 | DT->changeImmediateDominator(Bypass, TCCheckBlock); | |||
3014 | if (!Cost->requiresScalarEpilogue(VF)) | |||
3015 | // If there is an epilogue which must run, there's no edge from the | |||
3016 | // middle block to exit blocks and thus no need to update the immediate | |||
3017 | // dominator of the exit blocks. | |||
3018 | DT->changeImmediateDominator(LoopExitBlock, TCCheckBlock); | |||
3019 | ||||
3020 | ReplaceInstWithInst( | |||
3021 | TCCheckBlock->getTerminator(), | |||
3022 | BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters)); | |||
3023 | LoopBypassBlocks.push_back(TCCheckBlock); | |||
3024 | } | |||
3025 | ||||
3026 | BasicBlock *InnerLoopVectorizer::emitSCEVChecks(BasicBlock *Bypass) { | |||
3027 | ||||
3028 | BasicBlock *const SCEVCheckBlock = | |||
3029 | RTChecks.emitSCEVChecks(Bypass, LoopVectorPreHeader, LoopExitBlock); | |||
3030 | if (!SCEVCheckBlock) | |||
3031 | return nullptr; | |||
3032 | ||||
3033 | assert(!(SCEVCheckBlock->getParent()->hasOptSize() ||(static_cast <bool> (!(SCEVCheckBlock->getParent()-> hasOptSize() || (OptForSizeBasedOnProfile && Cost-> Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && "Cannot SCEV check stride or overflow when optimizing for size" ) ? void (0) : __assert_fail ("!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && \"Cannot SCEV check stride or overflow when optimizing for size\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3036, __extension__ __PRETTY_FUNCTION__)) | |||
3034 | (OptForSizeBasedOnProfile &&(static_cast <bool> (!(SCEVCheckBlock->getParent()-> hasOptSize() || (OptForSizeBasedOnProfile && Cost-> Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && "Cannot SCEV check stride or overflow when optimizing for size" ) ? void (0) : __assert_fail ("!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && \"Cannot SCEV check stride or overflow when optimizing for size\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3036, __extension__ __PRETTY_FUNCTION__)) | |||
3035 | Cost->Hints->getForce() != LoopVectorizeHints::FK_Enabled)) &&(static_cast <bool> (!(SCEVCheckBlock->getParent()-> hasOptSize() || (OptForSizeBasedOnProfile && Cost-> Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && "Cannot SCEV check stride or overflow when optimizing for size" ) ? void (0) : __assert_fail ("!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && \"Cannot SCEV check stride or overflow when optimizing for size\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3036, __extension__ __PRETTY_FUNCTION__)) | |||
3036 | "Cannot SCEV check stride or overflow when optimizing for size")(static_cast <bool> (!(SCEVCheckBlock->getParent()-> hasOptSize() || (OptForSizeBasedOnProfile && Cost-> Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && "Cannot SCEV check stride or overflow when optimizing for size" ) ? void (0) : __assert_fail ("!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && \"Cannot SCEV check stride or overflow when optimizing for size\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3036, __extension__ __PRETTY_FUNCTION__)); | |||
3037 | ||||
3038 | ||||
3039 | // Update dominator only if this is first RT check. | |||
3040 | if (LoopBypassBlocks.empty()) { | |||
3041 | DT->changeImmediateDominator(Bypass, SCEVCheckBlock); | |||
3042 | if (!Cost->requiresScalarEpilogue(VF)) | |||
3043 | // If there is an epilogue which must run, there's no edge from the | |||
3044 | // middle block to exit blocks and thus no need to update the immediate | |||
3045 | // dominator of the exit blocks. | |||
3046 | DT->changeImmediateDominator(LoopExitBlock, SCEVCheckBlock); | |||
3047 | } | |||
3048 | ||||
3049 | LoopBypassBlocks.push_back(SCEVCheckBlock); | |||
3050 | AddedSafetyChecks = true; | |||
3051 | return SCEVCheckBlock; | |||
3052 | } | |||
3053 | ||||
3054 | BasicBlock *InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L, | |||
3055 | BasicBlock *Bypass) { | |||
3056 | // VPlan-native path does not do any analysis for runtime checks currently. | |||
3057 | if (EnableVPlanNativePath) | |||
3058 | return nullptr; | |||
3059 | ||||
3060 | BasicBlock *const MemCheckBlock = | |||
3061 | RTChecks.emitMemRuntimeChecks(L, Bypass, LoopVectorPreHeader); | |||
3062 | ||||
3063 | // Check if we generated code that checks in runtime if arrays overlap. We put | |||
3064 | // the checks into a separate block to make the more common case of few | |||
3065 | // elements faster. | |||
3066 | if (!MemCheckBlock) | |||
3067 | return nullptr; | |||
3068 | ||||
3069 | if (MemCheckBlock->getParent()->hasOptSize() || OptForSizeBasedOnProfile) { | |||
3070 | assert(Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled &&(static_cast <bool> (Cost->Hints->getForce() == LoopVectorizeHints ::FK_Enabled && "Cannot emit memory checks when optimizing for size, unless forced " "to vectorize.") ? void (0) : __assert_fail ("Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled && \"Cannot emit memory checks when optimizing for size, unless forced \" \"to vectorize.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3072, __extension__ __PRETTY_FUNCTION__)) | |||
3071 | "Cannot emit memory checks when optimizing for size, unless forced "(static_cast <bool> (Cost->Hints->getForce() == LoopVectorizeHints ::FK_Enabled && "Cannot emit memory checks when optimizing for size, unless forced " "to vectorize.") ? void (0) : __assert_fail ("Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled && \"Cannot emit memory checks when optimizing for size, unless forced \" \"to vectorize.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3072, __extension__ __PRETTY_FUNCTION__)) | |||
3072 | "to vectorize.")(static_cast <bool> (Cost->Hints->getForce() == LoopVectorizeHints ::FK_Enabled && "Cannot emit memory checks when optimizing for size, unless forced " "to vectorize.") ? void (0) : __assert_fail ("Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled && \"Cannot emit memory checks when optimizing for size, unless forced \" \"to vectorize.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3072, __extension__ __PRETTY_FUNCTION__)); | |||
3073 | ORE->emit([&]() { | |||
3074 | return OptimizationRemarkAnalysis(DEBUG_TYPE"loop-vectorize", "VectorizationCodeSize", | |||
3075 | L->getStartLoc(), L->getHeader()) | |||
3076 | << "Code-size may be reduced by not forcing " | |||
3077 | "vectorization, or by source-code modifications " | |||
3078 | "eliminating the need for runtime checks " | |||
3079 | "(e.g., adding 'restrict')."; | |||
3080 | }); | |||
3081 | } | |||
3082 | ||||
3083 | LoopBypassBlocks.push_back(MemCheckBlock); | |||
3084 | ||||
3085 | AddedSafetyChecks = true; | |||
3086 | ||||
3087 | // We currently don't use LoopVersioning for the actual loop cloning but we | |||
3088 | // still use it to add the noalias metadata. | |||
3089 | LVer = std::make_unique<LoopVersioning>( | |||
3090 | *Legal->getLAI(), | |||
3091 | Legal->getLAI()->getRuntimePointerChecking()->getChecks(), OrigLoop, LI, | |||
3092 | DT, PSE.getSE()); | |||
3093 | LVer->prepareNoAliasMetadata(); | |||
3094 | return MemCheckBlock; | |||
3095 | } | |||
3096 | ||||
3097 | Loop *InnerLoopVectorizer::createVectorLoopSkeleton(StringRef Prefix) { | |||
3098 | LoopScalarBody = OrigLoop->getHeader(); | |||
3099 | LoopVectorPreHeader = OrigLoop->getLoopPreheader(); | |||
3100 | assert(LoopVectorPreHeader && "Invalid loop structure")(static_cast <bool> (LoopVectorPreHeader && "Invalid loop structure" ) ? void (0) : __assert_fail ("LoopVectorPreHeader && \"Invalid loop structure\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3100, __extension__ __PRETTY_FUNCTION__)); | |||
3101 | LoopExitBlock = OrigLoop->getUniqueExitBlock(); // may be nullptr | |||
3102 | assert((LoopExitBlock || Cost->requiresScalarEpilogue(VF)) &&(static_cast <bool> ((LoopExitBlock || Cost->requiresScalarEpilogue (VF)) && "multiple exit loop without required epilogue?" ) ? void (0) : __assert_fail ("(LoopExitBlock || Cost->requiresScalarEpilogue(VF)) && \"multiple exit loop without required epilogue?\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3103, __extension__ __PRETTY_FUNCTION__)) | |||
3103 | "multiple exit loop without required epilogue?")(static_cast <bool> ((LoopExitBlock || Cost->requiresScalarEpilogue (VF)) && "multiple exit loop without required epilogue?" ) ? void (0) : __assert_fail ("(LoopExitBlock || Cost->requiresScalarEpilogue(VF)) && \"multiple exit loop without required epilogue?\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3103, __extension__ __PRETTY_FUNCTION__)); | |||
3104 | ||||
3105 | LoopMiddleBlock = | |||
3106 | SplitBlock(LoopVectorPreHeader, LoopVectorPreHeader->getTerminator(), DT, | |||
3107 | LI, nullptr, Twine(Prefix) + "middle.block"); | |||
3108 | LoopScalarPreHeader = | |||
3109 | SplitBlock(LoopMiddleBlock, LoopMiddleBlock->getTerminator(), DT, LI, | |||
3110 | nullptr, Twine(Prefix) + "scalar.ph"); | |||
3111 | ||||
3112 | auto *ScalarLatchTerm = OrigLoop->getLoopLatch()->getTerminator(); | |||
3113 | ||||
3114 | // Set up the middle block terminator. Two cases: | |||
3115 | // 1) If we know that we must execute the scalar epilogue, emit an | |||
3116 | // unconditional branch. | |||
3117 | // 2) Otherwise, we must have a single unique exit block (due to how we | |||
3118 | // implement the multiple exit case). In this case, set up a conditonal | |||
3119 | // branch from the middle block to the loop scalar preheader, and the | |||
3120 | // exit block. completeLoopSkeleton will update the condition to use an | |||
3121 | // iteration check, if required to decide whether to execute the remainder. | |||
3122 | BranchInst *BrInst = Cost->requiresScalarEpilogue(VF) ? | |||
3123 | BranchInst::Create(LoopScalarPreHeader) : | |||
3124 | BranchInst::Create(LoopExitBlock, LoopScalarPreHeader, | |||
3125 | Builder.getTrue()); | |||
3126 | BrInst->setDebugLoc(ScalarLatchTerm->getDebugLoc()); | |||
3127 | ReplaceInstWithInst(LoopMiddleBlock->getTerminator(), BrInst); | |||
3128 | ||||
3129 | // We intentionally don't let SplitBlock to update LoopInfo since | |||
3130 | // LoopVectorBody should belong to another loop than LoopVectorPreHeader. | |||
3131 | // LoopVectorBody is explicitly added to the correct place few lines later. | |||
3132 | BasicBlock *LoopVectorBody = | |||
3133 | SplitBlock(LoopVectorPreHeader, LoopVectorPreHeader->getTerminator(), DT, | |||
3134 | nullptr, nullptr, Twine(Prefix) + "vector.body"); | |||
3135 | ||||
3136 | // Update dominator for loop exit. | |||
3137 | if (!Cost->requiresScalarEpilogue(VF)) | |||
3138 | // If there is an epilogue which must run, there's no edge from the | |||
3139 | // middle block to exit blocks and thus no need to update the immediate | |||
3140 | // dominator of the exit blocks. | |||
3141 | DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock); | |||
3142 | ||||
3143 | // Create and register the new vector loop. | |||
3144 | Loop *Lp = LI->AllocateLoop(); | |||
3145 | Loop *ParentLoop = OrigLoop->getParentLoop(); | |||
3146 | ||||
3147 | // Insert the new loop into the loop nest and register the new basic blocks | |||
3148 | // before calling any utilities such as SCEV that require valid LoopInfo. | |||
3149 | if (ParentLoop) { | |||
3150 | ParentLoop->addChildLoop(Lp); | |||
3151 | } else { | |||
3152 | LI->addTopLevelLoop(Lp); | |||
3153 | } | |||
3154 | Lp->addBasicBlockToLoop(LoopVectorBody, *LI); | |||
3155 | return Lp; | |||
3156 | } | |||
3157 | ||||
3158 | void InnerLoopVectorizer::createInductionResumeValues( | |||
3159 | Loop *L, std::pair<BasicBlock *, Value *> AdditionalBypass) { | |||
3160 | assert(((AdditionalBypass.first && AdditionalBypass.second) ||(static_cast <bool> (((AdditionalBypass.first && AdditionalBypass.second) || (!AdditionalBypass.first && !AdditionalBypass.second)) && "Inconsistent information about additional bypass." ) ? void (0) : __assert_fail ("((AdditionalBypass.first && AdditionalBypass.second) || (!AdditionalBypass.first && !AdditionalBypass.second)) && \"Inconsistent information about additional bypass.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3162, __extension__ __PRETTY_FUNCTION__)) | |||
3161 | (!AdditionalBypass.first && !AdditionalBypass.second)) &&(static_cast <bool> (((AdditionalBypass.first && AdditionalBypass.second) || (!AdditionalBypass.first && !AdditionalBypass.second)) && "Inconsistent information about additional bypass." ) ? void (0) : __assert_fail ("((AdditionalBypass.first && AdditionalBypass.second) || (!AdditionalBypass.first && !AdditionalBypass.second)) && \"Inconsistent information about additional bypass.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3162, __extension__ __PRETTY_FUNCTION__)) | |||
3162 | "Inconsistent information about additional bypass.")(static_cast <bool> (((AdditionalBypass.first && AdditionalBypass.second) || (!AdditionalBypass.first && !AdditionalBypass.second)) && "Inconsistent information about additional bypass." ) ? void (0) : __assert_fail ("((AdditionalBypass.first && AdditionalBypass.second) || (!AdditionalBypass.first && !AdditionalBypass.second)) && \"Inconsistent information about additional bypass.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3162, __extension__ __PRETTY_FUNCTION__)); | |||
3163 | ||||
3164 | Value *VectorTripCount = getOrCreateVectorTripCount(L); | |||
3165 | assert(VectorTripCount && L && "Expected valid arguments")(static_cast <bool> (VectorTripCount && L && "Expected valid arguments") ? void (0) : __assert_fail ("VectorTripCount && L && \"Expected valid arguments\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3165, __extension__ __PRETTY_FUNCTION__)); | |||
3166 | // We are going to resume the execution of the scalar loop. | |||
3167 | // Go over all of the induction variables that we found and fix the | |||
3168 | // PHIs that are left in the scalar version of the loop. | |||
3169 | // The starting values of PHI nodes depend on the counter of the last | |||
3170 | // iteration in the vectorized loop. | |||
3171 | // If we come from a bypass edge then we need to start from the original | |||
3172 | // start value. | |||
3173 | Instruction *OldInduction = Legal->getPrimaryInduction(); | |||
3174 | for (auto &InductionEntry : Legal->getInductionVars()) { | |||
3175 | PHINode *OrigPhi = InductionEntry.first; | |||
3176 | InductionDescriptor II = InductionEntry.second; | |||
3177 | ||||
3178 | // Create phi nodes to merge from the backedge-taken check block. | |||
3179 | PHINode *BCResumeVal = | |||
3180 | PHINode::Create(OrigPhi->getType(), 3, "bc.resume.val", | |||
3181 | LoopScalarPreHeader->getTerminator()); | |||
3182 | // Copy original phi DL over to the new one. | |||
3183 | BCResumeVal->setDebugLoc(OrigPhi->getDebugLoc()); | |||
3184 | Value *&EndValue = IVEndValues[OrigPhi]; | |||
3185 | Value *EndValueFromAdditionalBypass = AdditionalBypass.second; | |||
3186 | if (OrigPhi == OldInduction) { | |||
3187 | // We know what the end value is. | |||
3188 | EndValue = VectorTripCount; | |||
3189 | } else { | |||
3190 | IRBuilder<> B(L->getLoopPreheader()->getTerminator()); | |||
3191 | ||||
3192 | // Fast-math-flags propagate from the original induction instruction. | |||
3193 | if (II.getInductionBinOp() && isa<FPMathOperator>(II.getInductionBinOp())) | |||
3194 | B.setFastMathFlags(II.getInductionBinOp()->getFastMathFlags()); | |||
3195 | ||||
3196 | Type *StepType = II.getStep()->getType(); | |||
3197 | Instruction::CastOps CastOp = | |||
3198 | CastInst::getCastOpcode(VectorTripCount, true, StepType, true); | |||
3199 | Value *CRD = B.CreateCast(CastOp, VectorTripCount, StepType, "cast.crd"); | |||
3200 | Value *Step = | |||
3201 | CreateStepValue(II.getStep(), *PSE.getSE(), &*B.GetInsertPoint()); | |||
3202 | EndValue = emitTransformedIndex(B, CRD, II.getStartValue(), Step, II); | |||
3203 | EndValue->setName("ind.end"); | |||
3204 | ||||
3205 | // Compute the end value for the additional bypass (if applicable). | |||
3206 | if (AdditionalBypass.first) { | |||
3207 | B.SetInsertPoint(&(*AdditionalBypass.first->getFirstInsertionPt())); | |||
3208 | CastOp = CastInst::getCastOpcode(AdditionalBypass.second, true, | |||
3209 | StepType, true); | |||
3210 | Value *Step = | |||
3211 | CreateStepValue(II.getStep(), *PSE.getSE(), &*B.GetInsertPoint()); | |||
3212 | CRD = | |||
3213 | B.CreateCast(CastOp, AdditionalBypass.second, StepType, "cast.crd"); | |||
3214 | EndValueFromAdditionalBypass = | |||
3215 | emitTransformedIndex(B, CRD, II.getStartValue(), Step, II); | |||
3216 | EndValueFromAdditionalBypass->setName("ind.end"); | |||
3217 | } | |||
3218 | } | |||
3219 | // The new PHI merges the original incoming value, in case of a bypass, | |||
3220 | // or the value at the end of the vectorized loop. | |||
3221 | BCResumeVal->addIncoming(EndValue, LoopMiddleBlock); | |||
3222 | ||||
3223 | // Fix the scalar body counter (PHI node). | |||
3224 | // The old induction's phi node in the scalar body needs the truncated | |||
3225 | // value. | |||
3226 | for (BasicBlock *BB : LoopBypassBlocks) | |||
3227 | BCResumeVal->addIncoming(II.getStartValue(), BB); | |||
3228 | ||||
3229 | if (AdditionalBypass.first) | |||
3230 | BCResumeVal->setIncomingValueForBlock(AdditionalBypass.first, | |||
3231 | EndValueFromAdditionalBypass); | |||
3232 | ||||
3233 | OrigPhi->setIncomingValueForBlock(LoopScalarPreHeader, BCResumeVal); | |||
3234 | } | |||
3235 | } | |||
3236 | ||||
3237 | BasicBlock *InnerLoopVectorizer::completeLoopSkeleton(Loop *L, | |||
3238 | MDNode *OrigLoopID) { | |||
3239 | assert(L && "Expected valid loop.")(static_cast <bool> (L && "Expected valid loop." ) ? void (0) : __assert_fail ("L && \"Expected valid loop.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3239, __extension__ __PRETTY_FUNCTION__)); | |||
3240 | ||||
3241 | // The trip counts should be cached by now. | |||
3242 | Value *Count = getOrCreateTripCount(L); | |||
3243 | Value *VectorTripCount = getOrCreateVectorTripCount(L); | |||
3244 | ||||
3245 | auto *ScalarLatchTerm = OrigLoop->getLoopLatch()->getTerminator(); | |||
3246 | ||||
3247 | // Add a check in the middle block to see if we have completed | |||
3248 | // all of the iterations in the first vector loop. Three cases: | |||
3249 | // 1) If we require a scalar epilogue, there is no conditional branch as | |||
3250 | // we unconditionally branch to the scalar preheader. Do nothing. | |||
3251 | // 2) If (N - N%VF) == N, then we *don't* need to run the remainder. | |||
3252 | // Thus if tail is to be folded, we know we don't need to run the | |||
3253 | // remainder and we can use the previous value for the condition (true). | |||
3254 | // 3) Otherwise, construct a runtime check. | |||
3255 | if (!Cost->requiresScalarEpilogue(VF) && !Cost->foldTailByMasking()) { | |||
3256 | Instruction *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, | |||
3257 | Count, VectorTripCount, "cmp.n", | |||
3258 | LoopMiddleBlock->getTerminator()); | |||
3259 | ||||
3260 | // Here we use the same DebugLoc as the scalar loop latch terminator instead | |||
3261 | // of the corresponding compare because they may have ended up with | |||
3262 | // different line numbers and we want to avoid awkward line stepping while | |||
3263 | // debugging. Eg. if the compare has got a line number inside the loop. | |||
3264 | CmpN->setDebugLoc(ScalarLatchTerm->getDebugLoc()); | |||
3265 | cast<BranchInst>(LoopMiddleBlock->getTerminator())->setCondition(CmpN); | |||
3266 | } | |||
3267 | ||||
3268 | // Get ready to start creating new instructions into the vectorized body. | |||
3269 | assert(LoopVectorPreHeader == L->getLoopPreheader() &&(static_cast <bool> (LoopVectorPreHeader == L->getLoopPreheader () && "Inconsistent vector loop preheader") ? void (0 ) : __assert_fail ("LoopVectorPreHeader == L->getLoopPreheader() && \"Inconsistent vector loop preheader\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3270, __extension__ __PRETTY_FUNCTION__)) | |||
3270 | "Inconsistent vector loop preheader")(static_cast <bool> (LoopVectorPreHeader == L->getLoopPreheader () && "Inconsistent vector loop preheader") ? void (0 ) : __assert_fail ("LoopVectorPreHeader == L->getLoopPreheader() && \"Inconsistent vector loop preheader\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3270, __extension__ __PRETTY_FUNCTION__)); | |||
3271 | ||||
3272 | #ifdef EXPENSIVE_CHECKS | |||
3273 | assert(DT->verify(DominatorTree::VerificationLevel::Fast))(static_cast <bool> (DT->verify(DominatorTree::VerificationLevel ::Fast)) ? void (0) : __assert_fail ("DT->verify(DominatorTree::VerificationLevel::Fast)" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3273, __extension__ __PRETTY_FUNCTION__)); | |||
3274 | LI->verify(*DT); | |||
3275 | #endif | |||
3276 | ||||
3277 | return LoopVectorPreHeader; | |||
3278 | } | |||
3279 | ||||
3280 | std::pair<BasicBlock *, Value *> | |||
3281 | InnerLoopVectorizer::createVectorizedLoopSkeleton() { | |||
3282 | /* | |||
3283 | In this function we generate a new loop. The new loop will contain | |||
3284 | the vectorized instructions while the old loop will continue to run the | |||
3285 | scalar remainder. | |||
3286 | ||||
3287 | [ ] <-- loop iteration number check. | |||
3288 | / | | |||
3289 | / v | |||
3290 | | [ ] <-- vector loop bypass (may consist of multiple blocks). | |||
3291 | | / | | |||
3292 | | / v | |||
3293 | || [ ] <-- vector pre header. | |||
3294 | |/ | | |||
3295 | | v | |||
3296 | | [ ] \ | |||
3297 | | [ ]_| <-- vector loop. | |||
3298 | | | | |||
3299 | | v | |||
3300 | \ -[ ] <--- middle-block. | |||
3301 | \/ | | |||
3302 | /\ v | |||
3303 | | ->[ ] <--- new preheader. | |||
3304 | | | | |||
3305 | (opt) v <-- edge from middle to exit iff epilogue is not required. | |||
3306 | | [ ] \ | |||
3307 | | [ ]_| <-- old scalar loop to handle remainder (scalar epilogue). | |||
3308 | \ | | |||
3309 | \ v | |||
3310 | >[ ] <-- exit block(s). | |||
3311 | ... | |||
3312 | */ | |||
3313 | ||||
3314 | // Get the metadata of the original loop before it gets modified. | |||
3315 | MDNode *OrigLoopID = OrigLoop->getLoopID(); | |||
3316 | ||||
3317 | // Workaround! Compute the trip count of the original loop and cache it | |||
3318 | // before we start modifying the CFG. This code has a systemic problem | |||
3319 | // wherein it tries to run analysis over partially constructed IR; this is | |||
3320 | // wrong, and not simply for SCEV. The trip count of the original loop | |||
3321 | // simply happens to be prone to hitting this in practice. In theory, we | |||
3322 | // can hit the same issue for any SCEV, or ValueTracking query done during | |||
3323 | // mutation. See PR49900. | |||
3324 | getOrCreateTripCount(OrigLoop); | |||
3325 | ||||
3326 | // Create an empty vector loop, and prepare basic blocks for the runtime | |||
3327 | // checks. | |||
3328 | Loop *Lp = createVectorLoopSkeleton(""); | |||
3329 | ||||
3330 | // Now, compare the new count to zero. If it is zero skip the vector loop and | |||
3331 | // jump to the scalar loop. This check also covers the case where the | |||
3332 | // backedge-taken count is uint##_max: adding one to it will overflow leading | |||
3333 | // to an incorrect trip count of zero. In this (rare) case we will also jump | |||
3334 | // to the scalar loop. | |||
3335 | emitMinimumIterationCountCheck(Lp, LoopScalarPreHeader); | |||
3336 | ||||
3337 | // Generate the code to check any assumptions that we've made for SCEV | |||
3338 | // expressions. | |||
3339 | emitSCEVChecks(LoopScalarPreHeader); | |||
3340 | ||||
3341 | // Generate the code that checks in runtime if arrays overlap. We put the | |||
3342 | // checks into a separate block to make the more common case of few elements | |||
3343 | // faster. | |||
3344 | emitMemRuntimeChecks(Lp, LoopScalarPreHeader); | |||
3345 | ||||
3346 | createHeaderBranch(Lp); | |||
3347 | ||||
3348 | // Emit phis for the new starting index of the scalar loop. | |||
3349 | createInductionResumeValues(Lp); | |||
3350 | ||||
3351 | return {completeLoopSkeleton(Lp, OrigLoopID), nullptr}; | |||
3352 | } | |||
3353 | ||||
3354 | // Fix up external users of the induction variable. At this point, we are | |||
3355 | // in LCSSA form, with all external PHIs that use the IV having one input value, | |||
3356 | // coming from the remainder loop. We need those PHIs to also have a correct | |||
3357 | // value for the IV when arriving directly from the middle block. | |||
3358 | void InnerLoopVectorizer::fixupIVUsers(PHINode *OrigPhi, | |||
3359 | const InductionDescriptor &II, | |||
3360 | Value *CountRoundDown, Value *EndValue, | |||
3361 | BasicBlock *MiddleBlock, | |||
3362 | BasicBlock *VectorHeader) { | |||
3363 | // There are two kinds of external IV usages - those that use the value | |||
3364 | // computed in the last iteration (the PHI) and those that use the penultimate | |||
3365 | // value (the value that feeds into the phi from the loop latch). | |||
3366 | // We allow both, but they, obviously, have different values. | |||
3367 | ||||
3368 | assert(OrigLoop->getUniqueExitBlock() && "Expected a single exit block")(static_cast <bool> (OrigLoop->getUniqueExitBlock() && "Expected a single exit block") ? void (0) : __assert_fail ( "OrigLoop->getUniqueExitBlock() && \"Expected a single exit block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3368, __extension__ __PRETTY_FUNCTION__)); | |||
3369 | ||||
3370 | DenseMap<Value *, Value *> MissingVals; | |||
3371 | ||||
3372 | // An external user of the last iteration's value should see the value that | |||
3373 | // the remainder loop uses to initialize its own IV. | |||
3374 | Value *PostInc = OrigPhi->getIncomingValueForBlock(OrigLoop->getLoopLatch()); | |||
3375 | for (User *U : PostInc->users()) { | |||
3376 | Instruction *UI = cast<Instruction>(U); | |||
3377 | if (!OrigLoop->contains(UI)) { | |||
3378 | assert(isa<PHINode>(UI) && "Expected LCSSA form")(static_cast <bool> (isa<PHINode>(UI) && "Expected LCSSA form" ) ? void (0) : __assert_fail ("isa<PHINode>(UI) && \"Expected LCSSA form\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3378, __extension__ __PRETTY_FUNCTION__)); | |||
3379 | MissingVals[UI] = EndValue; | |||
3380 | } | |||
3381 | } | |||
3382 | ||||
3383 | // An external user of the penultimate value need to see EndValue - Step. | |||
3384 | // The simplest way to get this is to recompute it from the constituent SCEVs, | |||
3385 | // that is Start + (Step * (CRD - 1)). | |||
3386 | for (User *U : OrigPhi->users()) { | |||
3387 | auto *UI = cast<Instruction>(U); | |||
3388 | if (!OrigLoop->contains(UI)) { | |||
3389 | assert(isa<PHINode>(UI) && "Expected LCSSA form")(static_cast <bool> (isa<PHINode>(UI) && "Expected LCSSA form" ) ? void (0) : __assert_fail ("isa<PHINode>(UI) && \"Expected LCSSA form\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3389, __extension__ __PRETTY_FUNCTION__)); | |||
3390 | ||||
3391 | IRBuilder<> B(MiddleBlock->getTerminator()); | |||
3392 | ||||
3393 | // Fast-math-flags propagate from the original induction instruction. | |||
3394 | if (II.getInductionBinOp() && isa<FPMathOperator>(II.getInductionBinOp())) | |||
3395 | B.setFastMathFlags(II.getInductionBinOp()->getFastMathFlags()); | |||
3396 | ||||
3397 | Value *CountMinusOne = B.CreateSub( | |||
3398 | CountRoundDown, ConstantInt::get(CountRoundDown->getType(), 1)); | |||
3399 | Value *CMO = | |||
3400 | !II.getStep()->getType()->isIntegerTy() | |||
3401 | ? B.CreateCast(Instruction::SIToFP, CountMinusOne, | |||
3402 | II.getStep()->getType()) | |||
3403 | : B.CreateSExtOrTrunc(CountMinusOne, II.getStep()->getType()); | |||
3404 | CMO->setName("cast.cmo"); | |||
3405 | ||||
3406 | Value *Step = CreateStepValue(II.getStep(), *PSE.getSE(), | |||
3407 | VectorHeader->getTerminator()); | |||
3408 | Value *Escape = | |||
3409 | emitTransformedIndex(B, CMO, II.getStartValue(), Step, II); | |||
3410 | Escape->setName("ind.escape"); | |||
3411 | MissingVals[UI] = Escape; | |||
3412 | } | |||
3413 | } | |||
3414 | ||||
3415 | for (auto &I : MissingVals) { | |||
3416 | PHINode *PHI = cast<PHINode>(I.first); | |||
3417 | // One corner case we have to handle is two IVs "chasing" each-other, | |||
3418 | // that is %IV2 = phi [...], [ %IV1, %latch ] | |||
3419 | // In this case, if IV1 has an external use, we need to avoid adding both | |||
3420 | // "last value of IV1" and "penultimate value of IV2". So, verify that we | |||
3421 | // don't already have an incoming value for the middle block. | |||
3422 | if (PHI->getBasicBlockIndex(MiddleBlock) == -1) | |||
3423 | PHI->addIncoming(I.second, MiddleBlock); | |||
3424 | } | |||
3425 | } | |||
3426 | ||||
3427 | namespace { | |||
3428 | ||||
3429 | struct CSEDenseMapInfo { | |||
3430 | static bool canHandle(const Instruction *I) { | |||
3431 | return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || | |||
3432 | isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I); | |||
3433 | } | |||
3434 | ||||
3435 | static inline Instruction *getEmptyKey() { | |||
3436 | return DenseMapInfo<Instruction *>::getEmptyKey(); | |||
3437 | } | |||
3438 | ||||
3439 | static inline Instruction *getTombstoneKey() { | |||
3440 | return DenseMapInfo<Instruction *>::getTombstoneKey(); | |||
3441 | } | |||
3442 | ||||
3443 | static unsigned getHashValue(const Instruction *I) { | |||
3444 | assert(canHandle(I) && "Unknown instruction!")(static_cast <bool> (canHandle(I) && "Unknown instruction!" ) ? void (0) : __assert_fail ("canHandle(I) && \"Unknown instruction!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3444, __extension__ __PRETTY_FUNCTION__)); | |||
3445 | return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(), | |||
3446 | I->value_op_end())); | |||
3447 | } | |||
3448 | ||||
3449 | static bool isEqual(const Instruction *LHS, const Instruction *RHS) { | |||
3450 | if (LHS == getEmptyKey() || RHS == getEmptyKey() || | |||
3451 | LHS == getTombstoneKey() || RHS == getTombstoneKey()) | |||
3452 | return LHS == RHS; | |||
3453 | return LHS->isIdenticalTo(RHS); | |||
3454 | } | |||
3455 | }; | |||
3456 | ||||
3457 | } // end anonymous namespace | |||
3458 | ||||
3459 | ///Perform cse of induction variable instructions. | |||
3460 | static void cse(BasicBlock *BB) { | |||
3461 | // Perform simple cse. | |||
3462 | SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap; | |||
3463 | for (Instruction &In : llvm::make_early_inc_range(*BB)) { | |||
3464 | if (!CSEDenseMapInfo::canHandle(&In)) | |||
3465 | continue; | |||
3466 | ||||
3467 | // Check if we can replace this instruction with any of the | |||
3468 | // visited instructions. | |||
3469 | if (Instruction *V = CSEMap.lookup(&In)) { | |||
3470 | In.replaceAllUsesWith(V); | |||
3471 | In.eraseFromParent(); | |||
3472 | continue; | |||
3473 | } | |||
3474 | ||||
3475 | CSEMap[&In] = &In; | |||
3476 | } | |||
3477 | } | |||
3478 | ||||
3479 | InstructionCost | |||
3480 | LoopVectorizationCostModel::getVectorCallCost(CallInst *CI, ElementCount VF, | |||
3481 | bool &NeedToScalarize) const { | |||
3482 | Function *F = CI->getCalledFunction(); | |||
3483 | Type *ScalarRetTy = CI->getType(); | |||
3484 | SmallVector<Type *, 4> Tys, ScalarTys; | |||
3485 | for (auto &ArgOp : CI->args()) | |||
3486 | ScalarTys.push_back(ArgOp->getType()); | |||
3487 | ||||
3488 | // Estimate cost of scalarized vector call. The source operands are assumed | |||
3489 | // to be vectors, so we need to extract individual elements from there, | |||
3490 | // execute VF scalar calls, and then gather the result into the vector return | |||
3491 | // value. | |||
3492 | InstructionCost ScalarCallCost = | |||
3493 | TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys, TTI::TCK_RecipThroughput); | |||
3494 | if (VF.isScalar()) | |||
3495 | return ScalarCallCost; | |||
3496 | ||||
3497 | // Compute corresponding vector type for return value and arguments. | |||
3498 | Type *RetTy = ToVectorTy(ScalarRetTy, VF); | |||
3499 | for (Type *ScalarTy : ScalarTys) | |||
3500 | Tys.push_back(ToVectorTy(ScalarTy, VF)); | |||
3501 | ||||
3502 | // Compute costs of unpacking argument values for the scalar calls and | |||
3503 | // packing the return values to a vector. | |||
3504 | InstructionCost ScalarizationCost = getScalarizationOverhead(CI, VF); | |||
3505 | ||||
3506 | InstructionCost Cost = | |||
3507 | ScalarCallCost * VF.getKnownMinValue() + ScalarizationCost; | |||
3508 | ||||
3509 | // If we can't emit a vector call for this function, then the currently found | |||
3510 | // cost is the cost we need to return. | |||
3511 | NeedToScalarize = true; | |||
3512 | VFShape Shape = VFShape::get(*CI, VF, false /*HasGlobalPred*/); | |||
3513 | Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape); | |||
3514 | ||||
3515 | if (!TLI || CI->isNoBuiltin() || !VecFunc) | |||
3516 | return Cost; | |||
3517 | ||||
3518 | // If the corresponding vector cost is cheaper, return its cost. | |||
3519 | InstructionCost VectorCallCost = | |||
3520 | TTI.getCallInstrCost(nullptr, RetTy, Tys, TTI::TCK_RecipThroughput); | |||
3521 | if (VectorCallCost < Cost) { | |||
3522 | NeedToScalarize = false; | |||
3523 | Cost = VectorCallCost; | |||
3524 | } | |||
3525 | return Cost; | |||
3526 | } | |||
3527 | ||||
3528 | static Type *MaybeVectorizeType(Type *Elt, ElementCount VF) { | |||
3529 | if (VF.isScalar() || (!Elt->isIntOrPtrTy() && !Elt->isFloatingPointTy())) | |||
3530 | return Elt; | |||
3531 | return VectorType::get(Elt, VF); | |||
3532 | } | |||
3533 | ||||
3534 | InstructionCost | |||
3535 | LoopVectorizationCostModel::getVectorIntrinsicCost(CallInst *CI, | |||
3536 | ElementCount VF) const { | |||
3537 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
3538 | assert(ID && "Expected intrinsic call!")(static_cast <bool> (ID && "Expected intrinsic call!" ) ? void (0) : __assert_fail ("ID && \"Expected intrinsic call!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3538, __extension__ __PRETTY_FUNCTION__)); | |||
3539 | Type *RetTy = MaybeVectorizeType(CI->getType(), VF); | |||
3540 | FastMathFlags FMF; | |||
3541 | if (auto *FPMO = dyn_cast<FPMathOperator>(CI)) | |||
3542 | FMF = FPMO->getFastMathFlags(); | |||
3543 | ||||
3544 | SmallVector<const Value *> Arguments(CI->args()); | |||
3545 | FunctionType *FTy = CI->getCalledFunction()->getFunctionType(); | |||
3546 | SmallVector<Type *> ParamTys; | |||
3547 | std::transform(FTy->param_begin(), FTy->param_end(), | |||
3548 | std::back_inserter(ParamTys), | |||
3549 | [&](Type *Ty) { return MaybeVectorizeType(Ty, VF); }); | |||
3550 | ||||
3551 | IntrinsicCostAttributes CostAttrs(ID, RetTy, Arguments, ParamTys, FMF, | |||
3552 | dyn_cast<IntrinsicInst>(CI)); | |||
3553 | return TTI.getIntrinsicInstrCost(CostAttrs, | |||
3554 | TargetTransformInfo::TCK_RecipThroughput); | |||
3555 | } | |||
3556 | ||||
3557 | static Type *smallestIntegerVectorType(Type *T1, Type *T2) { | |||
3558 | auto *I1 = cast<IntegerType>(cast<VectorType>(T1)->getElementType()); | |||
3559 | auto *I2 = cast<IntegerType>(cast<VectorType>(T2)->getElementType()); | |||
3560 | return I1->getBitWidth() < I2->getBitWidth() ? T1 : T2; | |||
3561 | } | |||
3562 | ||||
3563 | static Type *largestIntegerVectorType(Type *T1, Type *T2) { | |||
3564 | auto *I1 = cast<IntegerType>(cast<VectorType>(T1)->getElementType()); | |||
3565 | auto *I2 = cast<IntegerType>(cast<VectorType>(T2)->getElementType()); | |||
3566 | return I1->getBitWidth() > I2->getBitWidth() ? T1 : T2; | |||
3567 | } | |||
3568 | ||||
3569 | void InnerLoopVectorizer::truncateToMinimalBitwidths(VPTransformState &State) { | |||
3570 | // For every instruction `I` in MinBWs, truncate the operands, create a | |||
3571 | // truncated version of `I` and reextend its result. InstCombine runs | |||
3572 | // later and will remove any ext/trunc pairs. | |||
3573 | SmallPtrSet<Value *, 4> Erased; | |||
3574 | for (const auto &KV : Cost->getMinimalBitwidths()) { | |||
3575 | // If the value wasn't vectorized, we must maintain the original scalar | |||
3576 | // type. The absence of the value from State indicates that it | |||
3577 | // wasn't vectorized. | |||
3578 | // FIXME: Should not rely on getVPValue at this point. | |||
3579 | VPValue *Def = State.Plan->getVPValue(KV.first, true); | |||
3580 | if (!State.hasAnyVectorValue(Def)) | |||
3581 | continue; | |||
3582 | for (unsigned Part = 0; Part < UF; ++Part) { | |||
3583 | Value *I = State.get(Def, Part); | |||
3584 | if (Erased.count(I) || I->use_empty() || !isa<Instruction>(I)) | |||
3585 | continue; | |||
3586 | Type *OriginalTy = I->getType(); | |||
3587 | Type *ScalarTruncatedTy = | |||
3588 | IntegerType::get(OriginalTy->getContext(), KV.second); | |||
3589 | auto *TruncatedTy = VectorType::get( | |||
3590 | ScalarTruncatedTy, cast<VectorType>(OriginalTy)->getElementCount()); | |||
3591 | if (TruncatedTy == OriginalTy) | |||
3592 | continue; | |||
3593 | ||||
3594 | IRBuilder<> B(cast<Instruction>(I)); | |||
3595 | auto ShrinkOperand = [&](Value *V) -> Value * { | |||
3596 | if (auto *ZI = dyn_cast<ZExtInst>(V)) | |||
3597 | if (ZI->getSrcTy() == TruncatedTy) | |||
3598 | return ZI->getOperand(0); | |||
3599 | return B.CreateZExtOrTrunc(V, TruncatedTy); | |||
3600 | }; | |||
3601 | ||||
3602 | // The actual instruction modification depends on the instruction type, | |||
3603 | // unfortunately. | |||
3604 | Value *NewI = nullptr; | |||
3605 | if (auto *BO = dyn_cast<BinaryOperator>(I)) { | |||
3606 | NewI = B.CreateBinOp(BO->getOpcode(), ShrinkOperand(BO->getOperand(0)), | |||
3607 | ShrinkOperand(BO->getOperand(1))); | |||
3608 | ||||
3609 | // Any wrapping introduced by shrinking this operation shouldn't be | |||
3610 | // considered undefined behavior. So, we can't unconditionally copy | |||
3611 | // arithmetic wrapping flags to NewI. | |||
3612 | cast<BinaryOperator>(NewI)->copyIRFlags(I, /*IncludeWrapFlags=*/false); | |||
3613 | } else if (auto *CI = dyn_cast<ICmpInst>(I)) { | |||
3614 | NewI = | |||
3615 | B.CreateICmp(CI->getPredicate(), ShrinkOperand(CI->getOperand(0)), | |||
3616 | ShrinkOperand(CI->getOperand(1))); | |||
3617 | } else if (auto *SI = dyn_cast<SelectInst>(I)) { | |||
3618 | NewI = B.CreateSelect(SI->getCondition(), | |||
3619 | ShrinkOperand(SI->getTrueValue()), | |||
3620 | ShrinkOperand(SI->getFalseValue())); | |||
3621 | } else if (auto *CI = dyn_cast<CastInst>(I)) { | |||
3622 | switch (CI->getOpcode()) { | |||
3623 | default: | |||
3624 | llvm_unreachable("Unhandled cast!")::llvm::llvm_unreachable_internal("Unhandled cast!", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3624); | |||
3625 | case Instruction::Trunc: | |||
3626 | NewI = ShrinkOperand(CI->getOperand(0)); | |||
3627 | break; | |||
3628 | case Instruction::SExt: | |||
3629 | NewI = B.CreateSExtOrTrunc( | |||
3630 | CI->getOperand(0), | |||
3631 | smallestIntegerVectorType(OriginalTy, TruncatedTy)); | |||
3632 | break; | |||
3633 | case Instruction::ZExt: | |||
3634 | NewI = B.CreateZExtOrTrunc( | |||
3635 | CI->getOperand(0), | |||
3636 | smallestIntegerVectorType(OriginalTy, TruncatedTy)); | |||
3637 | break; | |||
3638 | } | |||
3639 | } else if (auto *SI = dyn_cast<ShuffleVectorInst>(I)) { | |||
3640 | auto Elements0 = | |||
3641 | cast<VectorType>(SI->getOperand(0)->getType())->getElementCount(); | |||
3642 | auto *O0 = B.CreateZExtOrTrunc( | |||
3643 | SI->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements0)); | |||
3644 | auto Elements1 = | |||
3645 | cast<VectorType>(SI->getOperand(1)->getType())->getElementCount(); | |||
3646 | auto *O1 = B.CreateZExtOrTrunc( | |||
3647 | SI->getOperand(1), VectorType::get(ScalarTruncatedTy, Elements1)); | |||
3648 | ||||
3649 | NewI = B.CreateShuffleVector(O0, O1, SI->getShuffleMask()); | |||
3650 | } else if (isa<LoadInst>(I) || isa<PHINode>(I)) { | |||
3651 | // Don't do anything with the operands, just extend the result. | |||
3652 | continue; | |||
3653 | } else if (auto *IE = dyn_cast<InsertElementInst>(I)) { | |||
3654 | auto Elements = | |||
3655 | cast<VectorType>(IE->getOperand(0)->getType())->getElementCount(); | |||
3656 | auto *O0 = B.CreateZExtOrTrunc( | |||
3657 | IE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements)); | |||
3658 | auto *O1 = B.CreateZExtOrTrunc(IE->getOperand(1), ScalarTruncatedTy); | |||
3659 | NewI = B.CreateInsertElement(O0, O1, IE->getOperand(2)); | |||
3660 | } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) { | |||
3661 | auto Elements = | |||
3662 | cast<VectorType>(EE->getOperand(0)->getType())->getElementCount(); | |||
3663 | auto *O0 = B.CreateZExtOrTrunc( | |||
3664 | EE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements)); | |||
3665 | NewI = B.CreateExtractElement(O0, EE->getOperand(2)); | |||
3666 | } else { | |||
3667 | // If we don't know what to do, be conservative and don't do anything. | |||
3668 | continue; | |||
3669 | } | |||
3670 | ||||
3671 | // Lastly, extend the result. | |||
3672 | NewI->takeName(cast<Instruction>(I)); | |||
3673 | Value *Res = B.CreateZExtOrTrunc(NewI, OriginalTy); | |||
3674 | I->replaceAllUsesWith(Res); | |||
3675 | cast<Instruction>(I)->eraseFromParent(); | |||
3676 | Erased.insert(I); | |||
3677 | State.reset(Def, Res, Part); | |||
3678 | } | |||
3679 | } | |||
3680 | ||||
3681 | // We'll have created a bunch of ZExts that are now parentless. Clean up. | |||
3682 | for (const auto &KV : Cost->getMinimalBitwidths()) { | |||
3683 | // If the value wasn't vectorized, we must maintain the original scalar | |||
3684 | // type. The absence of the value from State indicates that it | |||
3685 | // wasn't vectorized. | |||
3686 | // FIXME: Should not rely on getVPValue at this point. | |||
3687 | VPValue *Def = State.Plan->getVPValue(KV.first, true); | |||
3688 | if (!State.hasAnyVectorValue(Def)) | |||
3689 | continue; | |||
3690 | for (unsigned Part = 0; Part < UF; ++Part) { | |||
3691 | Value *I = State.get(Def, Part); | |||
3692 | ZExtInst *Inst = dyn_cast<ZExtInst>(I); | |||
3693 | if (Inst && Inst->use_empty()) { | |||
3694 | Value *NewI = Inst->getOperand(0); | |||
3695 | Inst->eraseFromParent(); | |||
3696 | State.reset(Def, NewI, Part); | |||
3697 | } | |||
3698 | } | |||
3699 | } | |||
3700 | } | |||
3701 | ||||
3702 | void InnerLoopVectorizer::fixVectorizedLoop(VPTransformState &State) { | |||
3703 | // Insert truncates and extends for any truncated instructions as hints to | |||
3704 | // InstCombine. | |||
3705 | if (VF.isVector()) | |||
3706 | truncateToMinimalBitwidths(State); | |||
3707 | ||||
3708 | // Fix widened non-induction PHIs by setting up the PHI operands. | |||
3709 | if (OrigPHIsToFix.size()) { | |||
3710 | assert(EnableVPlanNativePath &&(static_cast <bool> (EnableVPlanNativePath && "Unexpected non-induction PHIs for fixup in non VPlan-native path" ) ? void (0) : __assert_fail ("EnableVPlanNativePath && \"Unexpected non-induction PHIs for fixup in non VPlan-native path\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3711, __extension__ __PRETTY_FUNCTION__)) | |||
3711 | "Unexpected non-induction PHIs for fixup in non VPlan-native path")(static_cast <bool> (EnableVPlanNativePath && "Unexpected non-induction PHIs for fixup in non VPlan-native path" ) ? void (0) : __assert_fail ("EnableVPlanNativePath && \"Unexpected non-induction PHIs for fixup in non VPlan-native path\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3711, __extension__ __PRETTY_FUNCTION__)); | |||
3712 | fixNonInductionPHIs(State); | |||
3713 | } | |||
3714 | ||||
3715 | // At this point every instruction in the original loop is widened to a | |||
3716 | // vector form. Now we need to fix the recurrences in the loop. These PHI | |||
3717 | // nodes are currently empty because we did not want to introduce cycles. | |||
3718 | // This is the second stage of vectorizing recurrences. | |||
3719 | fixCrossIterationPHIs(State); | |||
3720 | ||||
3721 | // Forget the original basic block. | |||
3722 | PSE.getSE()->forgetLoop(OrigLoop); | |||
3723 | ||||
3724 | Loop *VectorLoop = LI->getLoopFor(State.CFG.PrevBB); | |||
3725 | // If we inserted an edge from the middle block to the unique exit block, | |||
3726 | // update uses outside the loop (phis) to account for the newly inserted | |||
3727 | // edge. | |||
3728 | if (!Cost->requiresScalarEpilogue(VF)) { | |||
3729 | // Fix-up external users of the induction variables. | |||
3730 | for (auto &Entry : Legal->getInductionVars()) | |||
3731 | fixupIVUsers( | |||
3732 | Entry.first, Entry.second, getOrCreateVectorTripCount(VectorLoop), | |||
3733 | IVEndValues[Entry.first], LoopMiddleBlock, VectorLoop->getHeader()); | |||
3734 | ||||
3735 | fixLCSSAPHIs(State); | |||
3736 | } | |||
3737 | ||||
3738 | for (Instruction *PI : PredicatedInstructions) | |||
3739 | sinkScalarOperands(&*PI); | |||
3740 | ||||
3741 | // Remove redundant induction instructions. | |||
3742 | cse(VectorLoop->getHeader()); | |||
3743 | ||||
3744 | // Set/update profile weights for the vector and remainder loops as original | |||
3745 | // loop iterations are now distributed among them. Note that original loop | |||
3746 | // represented by LoopScalarBody becomes remainder loop after vectorization. | |||
3747 | // | |||
3748 | // For cases like foldTailByMasking() and requiresScalarEpiloque() we may | |||
3749 | // end up getting slightly roughened result but that should be OK since | |||
3750 | // profile is not inherently precise anyway. Note also possible bypass of | |||
3751 | // vector code caused by legality checks is ignored, assigning all the weight | |||
3752 | // to the vector loop, optimistically. | |||
3753 | // | |||
3754 | // For scalable vectorization we can't know at compile time how many iterations | |||
3755 | // of the loop are handled in one vector iteration, so instead assume a pessimistic | |||
3756 | // vscale of '1'. | |||
3757 | setProfileInfoAfterUnrolling(LI->getLoopFor(LoopScalarBody), VectorLoop, | |||
3758 | LI->getLoopFor(LoopScalarBody), | |||
3759 | VF.getKnownMinValue() * UF); | |||
3760 | } | |||
3761 | ||||
3762 | void InnerLoopVectorizer::fixCrossIterationPHIs(VPTransformState &State) { | |||
3763 | // In order to support recurrences we need to be able to vectorize Phi nodes. | |||
3764 | // Phi nodes have cycles, so we need to vectorize them in two stages. This is | |||
3765 | // stage #2: We now need to fix the recurrences by adding incoming edges to | |||
3766 | // the currently empty PHI nodes. At this point every instruction in the | |||
3767 | // original loop is widened to a vector form so we can use them to construct | |||
3768 | // the incoming edges. | |||
3769 | VPBasicBlock *Header = State.Plan->getEntry()->getEntryBasicBlock(); | |||
3770 | for (VPRecipeBase &R : Header->phis()) { | |||
3771 | if (auto *ReductionPhi = dyn_cast<VPReductionPHIRecipe>(&R)) | |||
3772 | fixReduction(ReductionPhi, State); | |||
3773 | else if (auto *FOR = dyn_cast<VPFirstOrderRecurrencePHIRecipe>(&R)) | |||
3774 | fixFirstOrderRecurrence(FOR, State); | |||
3775 | } | |||
3776 | } | |||
3777 | ||||
3778 | void InnerLoopVectorizer::fixFirstOrderRecurrence( | |||
3779 | VPFirstOrderRecurrencePHIRecipe *PhiR, VPTransformState &State) { | |||
3780 | // This is the second phase of vectorizing first-order recurrences. An | |||
3781 | // overview of the transformation is described below. Suppose we have the | |||
3782 | // following loop. | |||
3783 | // | |||
3784 | // for (int i = 0; i < n; ++i) | |||
3785 | // b[i] = a[i] - a[i - 1]; | |||
3786 | // | |||
3787 | // There is a first-order recurrence on "a". For this loop, the shorthand | |||
3788 | // scalar IR looks like: | |||
3789 | // | |||
3790 | // scalar.ph: | |||
3791 | // s_init = a[-1] | |||
3792 | // br scalar.body | |||
3793 | // | |||
3794 | // scalar.body: | |||
3795 | // i = phi [0, scalar.ph], [i+1, scalar.body] | |||
3796 | // s1 = phi [s_init, scalar.ph], [s2, scalar.body] | |||
3797 | // s2 = a[i] | |||
3798 | // b[i] = s2 - s1 | |||
3799 | // br cond, scalar.body, ... | |||
3800 | // | |||
3801 | // In this example, s1 is a recurrence because it's value depends on the | |||
3802 | // previous iteration. In the first phase of vectorization, we created a | |||
3803 | // vector phi v1 for s1. We now complete the vectorization and produce the | |||
3804 | // shorthand vector IR shown below (for VF = 4, UF = 1). | |||
3805 | // | |||
3806 | // vector.ph: | |||
3807 | // v_init = vector(..., ..., ..., a[-1]) | |||
3808 | // br vector.body | |||
3809 | // | |||
3810 | // vector.body | |||
3811 | // i = phi [0, vector.ph], [i+4, vector.body] | |||
3812 | // v1 = phi [v_init, vector.ph], [v2, vector.body] | |||
3813 | // v2 = a[i, i+1, i+2, i+3]; | |||
3814 | // v3 = vector(v1(3), v2(0, 1, 2)) | |||
3815 | // b[i, i+1, i+2, i+3] = v2 - v3 | |||
3816 | // br cond, vector.body, middle.block | |||
3817 | // | |||
3818 | // middle.block: | |||
3819 | // x = v2(3) | |||
3820 | // br scalar.ph | |||
3821 | // | |||
3822 | // scalar.ph: | |||
3823 | // s_init = phi [x, middle.block], [a[-1], otherwise] | |||
3824 | // br scalar.body | |||
3825 | // | |||
3826 | // After execution completes the vector loop, we extract the next value of | |||
3827 | // the recurrence (x) to use as the initial value in the scalar loop. | |||
3828 | ||||
3829 | // Extract the last vector element in the middle block. This will be the | |||
3830 | // initial value for the recurrence when jumping to the scalar loop. | |||
3831 | VPValue *PreviousDef = PhiR->getBackedgeValue(); | |||
3832 | Value *Incoming = State.get(PreviousDef, UF - 1); | |||
3833 | auto *ExtractForScalar = Incoming; | |||
3834 | auto *IdxTy = Builder.getInt32Ty(); | |||
3835 | if (VF.isVector()) { | |||
3836 | auto *One = ConstantInt::get(IdxTy, 1); | |||
3837 | Builder.SetInsertPoint(LoopMiddleBlock->getTerminator()); | |||
3838 | auto *RuntimeVF = getRuntimeVF(Builder, IdxTy, VF); | |||
3839 | auto *LastIdx = Builder.CreateSub(RuntimeVF, One); | |||
3840 | ExtractForScalar = Builder.CreateExtractElement(ExtractForScalar, LastIdx, | |||
3841 | "vector.recur.extract"); | |||
3842 | } | |||
3843 | // Extract the second last element in the middle block if the | |||
3844 | // Phi is used outside the loop. We need to extract the phi itself | |||
3845 | // and not the last element (the phi update in the current iteration). This | |||
3846 | // will be the value when jumping to the exit block from the LoopMiddleBlock, | |||
3847 | // when the scalar loop is not run at all. | |||
3848 | Value *ExtractForPhiUsedOutsideLoop = nullptr; | |||
3849 | if (VF.isVector()) { | |||
3850 | auto *RuntimeVF = getRuntimeVF(Builder, IdxTy, VF); | |||
3851 | auto *Idx = Builder.CreateSub(RuntimeVF, ConstantInt::get(IdxTy, 2)); | |||
3852 | ExtractForPhiUsedOutsideLoop = Builder.CreateExtractElement( | |||
3853 | Incoming, Idx, "vector.recur.extract.for.phi"); | |||
3854 | } else if (UF > 1) | |||
3855 | // When loop is unrolled without vectorizing, initialize | |||
3856 | // ExtractForPhiUsedOutsideLoop with the value just prior to unrolled value | |||
3857 | // of `Incoming`. This is analogous to the vectorized case above: extracting | |||
3858 | // the second last element when VF > 1. | |||
3859 | ExtractForPhiUsedOutsideLoop = State.get(PreviousDef, UF - 2); | |||
3860 | ||||
3861 | // Fix the initial value of the original recurrence in the scalar loop. | |||
3862 | Builder.SetInsertPoint(&*LoopScalarPreHeader->begin()); | |||
3863 | PHINode *Phi = cast<PHINode>(PhiR->getUnderlyingValue()); | |||
3864 | auto *Start = Builder.CreatePHI(Phi->getType(), 2, "scalar.recur.init"); | |||
3865 | auto *ScalarInit = PhiR->getStartValue()->getLiveInIRValue(); | |||
3866 | for (auto *BB : predecessors(LoopScalarPreHeader)) { | |||
3867 | auto *Incoming = BB == LoopMiddleBlock ? ExtractForScalar : ScalarInit; | |||
3868 | Start->addIncoming(Incoming, BB); | |||
3869 | } | |||
3870 | ||||
3871 | Phi->setIncomingValueForBlock(LoopScalarPreHeader, Start); | |||
3872 | Phi->setName("scalar.recur"); | |||
3873 | ||||
3874 | // Finally, fix users of the recurrence outside the loop. The users will need | |||
3875 | // either the last value of the scalar recurrence or the last value of the | |||
3876 | // vector recurrence we extracted in the middle block. Since the loop is in | |||
3877 | // LCSSA form, we just need to find all the phi nodes for the original scalar | |||
3878 | // recurrence in the exit block, and then add an edge for the middle block. | |||
3879 | // Note that LCSSA does not imply single entry when the original scalar loop | |||
3880 | // had multiple exiting edges (as we always run the last iteration in the | |||
3881 | // scalar epilogue); in that case, there is no edge from middle to exit and | |||
3882 | // and thus no phis which needed updated. | |||
3883 | if (!Cost->requiresScalarEpilogue(VF)) | |||
3884 | for (PHINode &LCSSAPhi : LoopExitBlock->phis()) | |||
3885 | if (llvm::is_contained(LCSSAPhi.incoming_values(), Phi)) | |||
3886 | LCSSAPhi.addIncoming(ExtractForPhiUsedOutsideLoop, LoopMiddleBlock); | |||
3887 | } | |||
3888 | ||||
3889 | void InnerLoopVectorizer::fixReduction(VPReductionPHIRecipe *PhiR, | |||
3890 | VPTransformState &State) { | |||
3891 | PHINode *OrigPhi = cast<PHINode>(PhiR->getUnderlyingValue()); | |||
3892 | // Get it's reduction variable descriptor. | |||
3893 | assert(Legal->isReductionVariable(OrigPhi) &&(static_cast <bool> (Legal->isReductionVariable(OrigPhi ) && "Unable to find the reduction variable") ? void ( 0) : __assert_fail ("Legal->isReductionVariable(OrigPhi) && \"Unable to find the reduction variable\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3894, __extension__ __PRETTY_FUNCTION__)) | |||
3894 | "Unable to find the reduction variable")(static_cast <bool> (Legal->isReductionVariable(OrigPhi ) && "Unable to find the reduction variable") ? void ( 0) : __assert_fail ("Legal->isReductionVariable(OrigPhi) && \"Unable to find the reduction variable\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3894, __extension__ __PRETTY_FUNCTION__)); | |||
3895 | const RecurrenceDescriptor &RdxDesc = PhiR->getRecurrenceDescriptor(); | |||
3896 | ||||
3897 | RecurKind RK = RdxDesc.getRecurrenceKind(); | |||
3898 | TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue(); | |||
3899 | Instruction *LoopExitInst = RdxDesc.getLoopExitInstr(); | |||
3900 | setDebugLocFromInst(ReductionStartValue); | |||
3901 | ||||
3902 | VPValue *LoopExitInstDef = PhiR->getBackedgeValue(); | |||
3903 | // This is the vector-clone of the value that leaves the loop. | |||
3904 | Type *VecTy = State.get(LoopExitInstDef, 0)->getType(); | |||
3905 | ||||
3906 | // Wrap flags are in general invalid after vectorization, clear them. | |||
3907 | clearReductionWrapFlags(RdxDesc, State); | |||
3908 | ||||
3909 | // Before each round, move the insertion point right between | |||
3910 | // the PHIs and the values we are going to write. | |||
3911 | // This allows us to write both PHINodes and the extractelement | |||
3912 | // instructions. | |||
3913 | Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt()); | |||
3914 | ||||
3915 | setDebugLocFromInst(LoopExitInst); | |||
3916 | ||||
3917 | Type *PhiTy = OrigPhi->getType(); | |||
3918 | BasicBlock *VectorLoopLatch = | |||
3919 | LI->getLoopFor(State.CFG.PrevBB)->getLoopLatch(); | |||
3920 | // If tail is folded by masking, the vector value to leave the loop should be | |||
3921 | // a Select choosing between the vectorized LoopExitInst and vectorized Phi, | |||
3922 | // instead of the former. For an inloop reduction the reduction will already | |||
3923 | // be predicated, and does not need to be handled here. | |||
3924 | if (Cost->foldTailByMasking() && !PhiR->isInLoop()) { | |||
3925 | for (unsigned Part = 0; Part < UF; ++Part) { | |||
3926 | Value *VecLoopExitInst = State.get(LoopExitInstDef, Part); | |||
3927 | Value *Sel = nullptr; | |||
3928 | for (User *U : VecLoopExitInst->users()) { | |||
3929 | if (isa<SelectInst>(U)) { | |||
3930 | assert(!Sel && "Reduction exit feeding two selects")(static_cast <bool> (!Sel && "Reduction exit feeding two selects" ) ? void (0) : __assert_fail ("!Sel && \"Reduction exit feeding two selects\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3930, __extension__ __PRETTY_FUNCTION__)); | |||
3931 | Sel = U; | |||
3932 | } else | |||
3933 | assert(isa<PHINode>(U) && "Reduction exit must feed Phi's or select")(static_cast <bool> (isa<PHINode>(U) && "Reduction exit must feed Phi's or select" ) ? void (0) : __assert_fail ("isa<PHINode>(U) && \"Reduction exit must feed Phi's or select\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3933, __extension__ __PRETTY_FUNCTION__)); | |||
3934 | } | |||
3935 | assert(Sel && "Reduction exit feeds no select")(static_cast <bool> (Sel && "Reduction exit feeds no select" ) ? void (0) : __assert_fail ("Sel && \"Reduction exit feeds no select\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3935, __extension__ __PRETTY_FUNCTION__)); | |||
3936 | State.reset(LoopExitInstDef, Sel, Part); | |||
3937 | ||||
3938 | // If the target can create a predicated operator for the reduction at no | |||
3939 | // extra cost in the loop (for example a predicated vadd), it can be | |||
3940 | // cheaper for the select to remain in the loop than be sunk out of it, | |||
3941 | // and so use the select value for the phi instead of the old | |||
3942 | // LoopExitValue. | |||
3943 | if (PreferPredicatedReductionSelect || | |||
3944 | TTI->preferPredicatedReductionSelect( | |||
3945 | RdxDesc.getOpcode(), PhiTy, | |||
3946 | TargetTransformInfo::ReductionFlags())) { | |||
3947 | auto *VecRdxPhi = | |||
3948 | cast<PHINode>(State.get(PhiR, Part)); | |||
3949 | VecRdxPhi->setIncomingValueForBlock(VectorLoopLatch, Sel); | |||
3950 | } | |||
3951 | } | |||
3952 | } | |||
3953 | ||||
3954 | // If the vector reduction can be performed in a smaller type, we truncate | |||
3955 | // then extend the loop exit value to enable InstCombine to evaluate the | |||
3956 | // entire expression in the smaller type. | |||
3957 | if (VF.isVector() && PhiTy != RdxDesc.getRecurrenceType()) { | |||
3958 | assert(!PhiR->isInLoop() && "Unexpected truncated inloop reduction!")(static_cast <bool> (!PhiR->isInLoop() && "Unexpected truncated inloop reduction!" ) ? void (0) : __assert_fail ("!PhiR->isInLoop() && \"Unexpected truncated inloop reduction!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3958, __extension__ __PRETTY_FUNCTION__)); | |||
3959 | Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF); | |||
3960 | Builder.SetInsertPoint(VectorLoopLatch->getTerminator()); | |||
3961 | VectorParts RdxParts(UF); | |||
3962 | for (unsigned Part = 0; Part < UF; ++Part) { | |||
3963 | RdxParts[Part] = State.get(LoopExitInstDef, Part); | |||
3964 | Value *Trunc = Builder.CreateTrunc(RdxParts[Part], RdxVecTy); | |||
3965 | Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy) | |||
3966 | : Builder.CreateZExt(Trunc, VecTy); | |||
3967 | for (User *U : llvm::make_early_inc_range(RdxParts[Part]->users())) | |||
3968 | if (U != Trunc) { | |||
3969 | U->replaceUsesOfWith(RdxParts[Part], Extnd); | |||
3970 | RdxParts[Part] = Extnd; | |||
3971 | } | |||
3972 | } | |||
3973 | Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt()); | |||
3974 | for (unsigned Part = 0; Part < UF; ++Part) { | |||
3975 | RdxParts[Part] = Builder.CreateTrunc(RdxParts[Part], RdxVecTy); | |||
3976 | State.reset(LoopExitInstDef, RdxParts[Part], Part); | |||
3977 | } | |||
3978 | } | |||
3979 | ||||
3980 | // Reduce all of the unrolled parts into a single vector. | |||
3981 | Value *ReducedPartRdx = State.get(LoopExitInstDef, 0); | |||
3982 | unsigned Op = RecurrenceDescriptor::getOpcode(RK); | |||
3983 | ||||
3984 | // The middle block terminator has already been assigned a DebugLoc here (the | |||
3985 | // OrigLoop's single latch terminator). We want the whole middle block to | |||
3986 | // appear to execute on this line because: (a) it is all compiler generated, | |||
3987 | // (b) these instructions are always executed after evaluating the latch | |||
3988 | // conditional branch, and (c) other passes may add new predecessors which | |||
3989 | // terminate on this line. This is the easiest way to ensure we don't | |||
3990 | // accidentally cause an extra step back into the loop while debugging. | |||
3991 | setDebugLocFromInst(LoopMiddleBlock->getTerminator()); | |||
3992 | if (PhiR->isOrdered()) | |||
3993 | ReducedPartRdx = State.get(LoopExitInstDef, UF - 1); | |||
3994 | else { | |||
3995 | // Floating-point operations should have some FMF to enable the reduction. | |||
3996 | IRBuilderBase::FastMathFlagGuard FMFG(Builder); | |||
3997 | Builder.setFastMathFlags(RdxDesc.getFastMathFlags()); | |||
3998 | for (unsigned Part = 1; Part < UF; ++Part) { | |||
3999 | Value *RdxPart = State.get(LoopExitInstDef, Part); | |||
4000 | if (Op != Instruction::ICmp && Op != Instruction::FCmp) { | |||
4001 | ReducedPartRdx = Builder.CreateBinOp( | |||
4002 | (Instruction::BinaryOps)Op, RdxPart, ReducedPartRdx, "bin.rdx"); | |||
4003 | } else if (RecurrenceDescriptor::isSelectCmpRecurrenceKind(RK)) | |||
4004 | ReducedPartRdx = createSelectCmpOp(Builder, ReductionStartValue, RK, | |||
4005 | ReducedPartRdx, RdxPart); | |||
4006 | else | |||
4007 | ReducedPartRdx = createMinMaxOp(Builder, RK, ReducedPartRdx, RdxPart); | |||
4008 | } | |||
4009 | } | |||
4010 | ||||
4011 | // Create the reduction after the loop. Note that inloop reductions create the | |||
4012 | // target reduction in the loop using a Reduction recipe. | |||
4013 | if (VF.isVector() && !PhiR->isInLoop()) { | |||
4014 | ReducedPartRdx = | |||
4015 | createTargetReduction(Builder, TTI, RdxDesc, ReducedPartRdx, OrigPhi); | |||
4016 | // If the reduction can be performed in a smaller type, we need to extend | |||
4017 | // the reduction to the wider type before we branch to the original loop. | |||
4018 | if (PhiTy != RdxDesc.getRecurrenceType()) | |||
4019 | ReducedPartRdx = RdxDesc.isSigned() | |||
4020 | ? Builder.CreateSExt(ReducedPartRdx, PhiTy) | |||
4021 | : Builder.CreateZExt(ReducedPartRdx, PhiTy); | |||
4022 | } | |||
4023 | ||||
4024 | PHINode *ResumePhi = | |||
4025 | dyn_cast<PHINode>(PhiR->getStartValue()->getUnderlyingValue()); | |||
4026 | ||||
4027 | // Create a phi node that merges control-flow from the backedge-taken check | |||
4028 | // block and the middle block. | |||
4029 | PHINode *BCBlockPhi = PHINode::Create(PhiTy, 2, "bc.merge.rdx", | |||
4030 | LoopScalarPreHeader->getTerminator()); | |||
4031 | ||||
4032 | // If we are fixing reductions in the epilogue loop then we should already | |||
4033 | // have created a bc.merge.rdx Phi after the main vector body. Ensure that | |||
4034 | // we carry over the incoming values correctly. | |||
4035 | for (auto *Incoming : predecessors(LoopScalarPreHeader)) { | |||
4036 | if (Incoming == LoopMiddleBlock) | |||
4037 | BCBlockPhi->addIncoming(ReducedPartRdx, Incoming); | |||
4038 | else if (ResumePhi && llvm::is_contained(ResumePhi->blocks(), Incoming)) | |||
4039 | BCBlockPhi->addIncoming(ResumePhi->getIncomingValueForBlock(Incoming), | |||
4040 | Incoming); | |||
4041 | else | |||
4042 | BCBlockPhi->addIncoming(ReductionStartValue, Incoming); | |||
4043 | } | |||
4044 | ||||
4045 | // Set the resume value for this reduction | |||
4046 | ReductionResumeValues.insert({&RdxDesc, BCBlockPhi}); | |||
4047 | ||||
4048 | // Now, we need to fix the users of the reduction variable | |||
4049 | // inside and outside of the scalar remainder loop. | |||
4050 | ||||
4051 | // We know that the loop is in LCSSA form. We need to update the PHI nodes | |||
4052 | // in the exit blocks. See comment on analogous loop in | |||
4053 | // fixFirstOrderRecurrence for a more complete explaination of the logic. | |||
4054 | if (!Cost->requiresScalarEpilogue(VF)) | |||
4055 | for (PHINode &LCSSAPhi : LoopExitBlock->phis()) | |||
4056 | if (llvm::is_contained(LCSSAPhi.incoming_values(), LoopExitInst)) | |||
4057 | LCSSAPhi.addIncoming(ReducedPartRdx, LoopMiddleBlock); | |||
4058 | ||||
4059 | // Fix the scalar loop reduction variable with the incoming reduction sum | |||
4060 | // from the vector body and from the backedge value. | |||
4061 | int IncomingEdgeBlockIdx = | |||
4062 | OrigPhi->getBasicBlockIndex(OrigLoop->getLoopLatch()); | |||
4063 | assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index")(static_cast <bool> (IncomingEdgeBlockIdx >= 0 && "Invalid block index") ? void (0) : __assert_fail ("IncomingEdgeBlockIdx >= 0 && \"Invalid block index\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4063, __extension__ __PRETTY_FUNCTION__)); | |||
4064 | // Pick the other block. | |||
4065 | int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); | |||
4066 | OrigPhi->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi); | |||
4067 | OrigPhi->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst); | |||
4068 | } | |||
4069 | ||||
4070 | void InnerLoopVectorizer::clearReductionWrapFlags(const RecurrenceDescriptor &RdxDesc, | |||
4071 | VPTransformState &State) { | |||
4072 | RecurKind RK = RdxDesc.getRecurrenceKind(); | |||
4073 | if (RK != RecurKind::Add && RK != RecurKind::Mul) | |||
4074 | return; | |||
4075 | ||||
4076 | Instruction *LoopExitInstr = RdxDesc.getLoopExitInstr(); | |||
4077 | assert(LoopExitInstr && "null loop exit instruction")(static_cast <bool> (LoopExitInstr && "null loop exit instruction" ) ? void (0) : __assert_fail ("LoopExitInstr && \"null loop exit instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4077, __extension__ __PRETTY_FUNCTION__)); | |||
4078 | SmallVector<Instruction *, 8> Worklist; | |||
4079 | SmallPtrSet<Instruction *, 8> Visited; | |||
4080 | Worklist.push_back(LoopExitInstr); | |||
4081 | Visited.insert(LoopExitInstr); | |||
4082 | ||||
4083 | while (!Worklist.empty()) { | |||
4084 | Instruction *Cur = Worklist.pop_back_val(); | |||
4085 | if (isa<OverflowingBinaryOperator>(Cur)) | |||
4086 | for (unsigned Part = 0; Part < UF; ++Part) { | |||
4087 | // FIXME: Should not rely on getVPValue at this point. | |||
4088 | Value *V = State.get(State.Plan->getVPValue(Cur, true), Part); | |||
4089 | cast<Instruction>(V)->dropPoisonGeneratingFlags(); | |||
4090 | } | |||
4091 | ||||
4092 | for (User *U : Cur->users()) { | |||
4093 | Instruction *UI = cast<Instruction>(U); | |||
4094 | if ((Cur != LoopExitInstr || OrigLoop->contains(UI->getParent())) && | |||
4095 | Visited.insert(UI).second) | |||
4096 | Worklist.push_back(UI); | |||
4097 | } | |||
4098 | } | |||
4099 | } | |||
4100 | ||||
4101 | void InnerLoopVectorizer::fixLCSSAPHIs(VPTransformState &State) { | |||
4102 | for (PHINode &LCSSAPhi : LoopExitBlock->phis()) { | |||
4103 | if (LCSSAPhi.getBasicBlockIndex(LoopMiddleBlock) != -1) | |||
4104 | // Some phis were already hand updated by the reduction and recurrence | |||
4105 | // code above, leave them alone. | |||
4106 | continue; | |||
4107 | ||||
4108 | auto *IncomingValue = LCSSAPhi.getIncomingValue(0); | |||
4109 | // Non-instruction incoming values will have only one value. | |||
4110 | ||||
4111 | VPLane Lane = VPLane::getFirstLane(); | |||
4112 | if (isa<Instruction>(IncomingValue) && | |||
4113 | !Cost->isUniformAfterVectorization(cast<Instruction>(IncomingValue), | |||
4114 | VF)) | |||
4115 | Lane = VPLane::getLastLaneForVF(VF); | |||
4116 | ||||
4117 | // Can be a loop invariant incoming value or the last scalar value to be | |||
4118 | // extracted from the vectorized loop. | |||
4119 | // FIXME: Should not rely on getVPValue at this point. | |||
4120 | Builder.SetInsertPoint(LoopMiddleBlock->getTerminator()); | |||
4121 | Value *lastIncomingValue = | |||
4122 | OrigLoop->isLoopInvariant(IncomingValue) | |||
4123 | ? IncomingValue | |||
4124 | : State.get(State.Plan->getVPValue(IncomingValue, true), | |||
4125 | VPIteration(UF - 1, Lane)); | |||
4126 | LCSSAPhi.addIncoming(lastIncomingValue, LoopMiddleBlock); | |||
4127 | } | |||
4128 | } | |||
4129 | ||||
4130 | void InnerLoopVectorizer::sinkScalarOperands(Instruction *PredInst) { | |||
4131 | // The basic block and loop containing the predicated instruction. | |||
4132 | auto *PredBB = PredInst->getParent(); | |||
4133 | auto *VectorLoop = LI->getLoopFor(PredBB); | |||
4134 | ||||
4135 | // Initialize a worklist with the operands of the predicated instruction. | |||
4136 | SetVector<Value *> Worklist(PredInst->op_begin(), PredInst->op_end()); | |||
4137 | ||||
4138 | // Holds instructions that we need to analyze again. An instruction may be | |||
4139 | // reanalyzed if we don't yet know if we can sink it or not. | |||
4140 | SmallVector<Instruction *, 8> InstsToReanalyze; | |||
4141 | ||||
4142 | // Returns true if a given use occurs in the predicated block. Phi nodes use | |||
4143 | // their operands in their corresponding predecessor blocks. | |||
4144 | auto isBlockOfUsePredicated = [&](Use &U) -> bool { | |||
4145 | auto *I = cast<Instruction>(U.getUser()); | |||
4146 | BasicBlock *BB = I->getParent(); | |||
4147 | if (auto *Phi = dyn_cast<PHINode>(I)) | |||
4148 | BB = Phi->getIncomingBlock( | |||
4149 | PHINode::getIncomingValueNumForOperand(U.getOperandNo())); | |||
4150 | return BB == PredBB; | |||
4151 | }; | |||
4152 | ||||
4153 | // Iteratively sink the scalarized operands of the predicated instruction | |||
4154 | // into the block we created for it. When an instruction is sunk, it's | |||
4155 | // operands are then added to the worklist. The algorithm ends after one pass | |||
4156 | // through the worklist doesn't sink a single instruction. | |||
4157 | bool Changed; | |||
4158 | do { | |||
4159 | // Add the instructions that need to be reanalyzed to the worklist, and | |||
4160 | // reset the changed indicator. | |||
4161 | Worklist.insert(InstsToReanalyze.begin(), InstsToReanalyze.end()); | |||
4162 | InstsToReanalyze.clear(); | |||
4163 | Changed = false; | |||
4164 | ||||
4165 | while (!Worklist.empty()) { | |||
4166 | auto *I = dyn_cast<Instruction>(Worklist.pop_back_val()); | |||
4167 | ||||
4168 | // We can't sink an instruction if it is a phi node, is not in the loop, | |||
4169 | // or may have side effects. | |||
4170 | if (!I || isa<PHINode>(I) || !VectorLoop->contains(I) || | |||
4171 | I->mayHaveSideEffects()) | |||
4172 | continue; | |||
4173 | ||||
4174 | // If the instruction is already in PredBB, check if we can sink its | |||
4175 | // operands. In that case, VPlan's sinkScalarOperands() succeeded in | |||
4176 | // sinking the scalar instruction I, hence it appears in PredBB; but it | |||
4177 | // may have failed to sink I's operands (recursively), which we try | |||
4178 | // (again) here. | |||
4179 | if (I->getParent() == PredBB) { | |||
4180 | Worklist.insert(I->op_begin(), I->op_end()); | |||
4181 | continue; | |||
4182 | } | |||
4183 | ||||
4184 | // It's legal to sink the instruction if all its uses occur in the | |||
4185 | // predicated block. Otherwise, there's nothing to do yet, and we may | |||
4186 | // need to reanalyze the instruction. | |||
4187 | if (!llvm::all_of(I->uses(), isBlockOfUsePredicated)) { | |||
4188 | InstsToReanalyze.push_back(I); | |||
4189 | continue; | |||
4190 | } | |||
4191 | ||||
4192 | // Move the instruction to the beginning of the predicated block, and add | |||
4193 | // it's operands to the worklist. | |||
4194 | I->moveBefore(&*PredBB->getFirstInsertionPt()); | |||
4195 | Worklist.insert(I->op_begin(), I->op_end()); | |||
4196 | ||||
4197 | // The sinking may have enabled other instructions to be sunk, so we will | |||
4198 | // need to iterate. | |||
4199 | Changed = true; | |||
4200 | } | |||
4201 | } while (Changed); | |||
4202 | } | |||
4203 | ||||
4204 | void InnerLoopVectorizer::fixNonInductionPHIs(VPTransformState &State) { | |||
4205 | for (PHINode *OrigPhi : OrigPHIsToFix) { | |||
4206 | VPWidenPHIRecipe *VPPhi = | |||
4207 | cast<VPWidenPHIRecipe>(State.Plan->getVPValue(OrigPhi)); | |||
4208 | PHINode *NewPhi = cast<PHINode>(State.get(VPPhi, 0)); | |||
4209 | // Make sure the builder has a valid insert point. | |||
4210 | Builder.SetInsertPoint(NewPhi); | |||
4211 | for (unsigned i = 0; i < VPPhi->getNumOperands(); ++i) { | |||
4212 | VPValue *Inc = VPPhi->getIncomingValue(i); | |||
4213 | VPBasicBlock *VPBB = VPPhi->getIncomingBlock(i); | |||
4214 | NewPhi->addIncoming(State.get(Inc, 0), State.CFG.VPBB2IRBB[VPBB]); | |||
4215 | } | |||
4216 | } | |||
4217 | } | |||
4218 | ||||
4219 | bool InnerLoopVectorizer::useOrderedReductions( | |||
4220 | const RecurrenceDescriptor &RdxDesc) { | |||
4221 | return Cost->useOrderedReductions(RdxDesc); | |||
4222 | } | |||
4223 | ||||
4224 | void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN, | |||
4225 | VPWidenPHIRecipe *PhiR, | |||
4226 | VPTransformState &State) { | |||
4227 | PHINode *P = cast<PHINode>(PN); | |||
4228 | if (EnableVPlanNativePath) { | |||
4229 | // Currently we enter here in the VPlan-native path for non-induction | |||
4230 | // PHIs where all control flow is uniform. We simply widen these PHIs. | |||
4231 | // Create a vector phi with no operands - the vector phi operands will be | |||
4232 | // set at the end of vector code generation. | |||
4233 | Type *VecTy = (State.VF.isScalar()) | |||
4234 | ? PN->getType() | |||
4235 | : VectorType::get(PN->getType(), State.VF); | |||
4236 | Value *VecPhi = Builder.CreatePHI(VecTy, PN->getNumOperands(), "vec.phi"); | |||
4237 | State.set(PhiR, VecPhi, 0); | |||
4238 | OrigPHIsToFix.push_back(P); | |||
4239 | ||||
4240 | return; | |||
4241 | } | |||
4242 | ||||
4243 | assert(PN->getParent() == OrigLoop->getHeader() &&(static_cast <bool> (PN->getParent() == OrigLoop-> getHeader() && "Non-header phis should have been handled elsewhere" ) ? void (0) : __assert_fail ("PN->getParent() == OrigLoop->getHeader() && \"Non-header phis should have been handled elsewhere\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4244, __extension__ __PRETTY_FUNCTION__)) | |||
4244 | "Non-header phis should have been handled elsewhere")(static_cast <bool> (PN->getParent() == OrigLoop-> getHeader() && "Non-header phis should have been handled elsewhere" ) ? void (0) : __assert_fail ("PN->getParent() == OrigLoop->getHeader() && \"Non-header phis should have been handled elsewhere\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4244, __extension__ __PRETTY_FUNCTION__)); | |||
4245 | ||||
4246 | // In order to support recurrences we need to be able to vectorize Phi nodes. | |||
4247 | // Phi nodes have cycles, so we need to vectorize them in two stages. This is | |||
4248 | // stage #1: We create a new vector PHI node with no incoming edges. We'll use | |||
4249 | // this value when we vectorize all of the instructions that use the PHI. | |||
4250 | ||||
4251 | assert(!Legal->isReductionVariable(P) &&(static_cast <bool> (!Legal->isReductionVariable(P) && "reductions should be handled elsewhere") ? void (0) : __assert_fail ("!Legal->isReductionVariable(P) && \"reductions should be handled elsewhere\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4252, __extension__ __PRETTY_FUNCTION__)) | |||
4252 | "reductions should be handled elsewhere")(static_cast <bool> (!Legal->isReductionVariable(P) && "reductions should be handled elsewhere") ? void (0) : __assert_fail ("!Legal->isReductionVariable(P) && \"reductions should be handled elsewhere\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4252, __extension__ __PRETTY_FUNCTION__)); | |||
4253 | ||||
4254 | setDebugLocFromInst(P); | |||
4255 | ||||
4256 | // This PHINode must be an induction variable. | |||
4257 | // Make sure that we know about it. | |||
4258 | assert(Legal->getInductionVars().count(P) && "Not an induction variable")(static_cast <bool> (Legal->getInductionVars().count (P) && "Not an induction variable") ? void (0) : __assert_fail ("Legal->getInductionVars().count(P) && \"Not an induction variable\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4258, __extension__ __PRETTY_FUNCTION__)); | |||
4259 | ||||
4260 | InductionDescriptor II = Legal->getInductionVars().lookup(P); | |||
4261 | const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout(); | |||
4262 | ||||
4263 | auto *IVR = PhiR->getParent()->getPlan()->getCanonicalIV(); | |||
4264 | PHINode *CanonicalIV = cast<PHINode>(State.get(IVR, 0)); | |||
4265 | ||||
4266 | // FIXME: The newly created binary instructions should contain nsw/nuw flags, | |||
4267 | // which can be found from the original scalar operations. | |||
4268 | switch (II.getKind()) { | |||
4269 | case InductionDescriptor::IK_NoInduction: | |||
4270 | llvm_unreachable("Unknown induction")::llvm::llvm_unreachable_internal("Unknown induction", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4270); | |||
4271 | case InductionDescriptor::IK_IntInduction: | |||
4272 | case InductionDescriptor::IK_FpInduction: | |||
4273 | llvm_unreachable("Integer/fp induction is handled elsewhere.")::llvm::llvm_unreachable_internal("Integer/fp induction is handled elsewhere." , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4273); | |||
4274 | case InductionDescriptor::IK_PtrInduction: { | |||
4275 | // Handle the pointer induction variable case. | |||
4276 | assert(P->getType()->isPointerTy() && "Unexpected type.")(static_cast <bool> (P->getType()->isPointerTy() && "Unexpected type.") ? void (0) : __assert_fail ("P->getType()->isPointerTy() && \"Unexpected type.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4276, __extension__ __PRETTY_FUNCTION__)); | |||
4277 | ||||
4278 | if (all_of(PhiR->users(), [PhiR](const VPUser *U) { | |||
4279 | return cast<VPRecipeBase>(U)->usesScalars(PhiR); | |||
4280 | })) { | |||
4281 | // This is the normalized GEP that starts counting at zero. | |||
4282 | Value *PtrInd = | |||
4283 | Builder.CreateSExtOrTrunc(CanonicalIV, II.getStep()->getType()); | |||
4284 | // Determine the number of scalars we need to generate for each unroll | |||
4285 | // iteration. If the instruction is uniform, we only need to generate the | |||
4286 | // first lane. Otherwise, we generate all VF values. | |||
4287 | bool IsUniform = vputils::onlyFirstLaneUsed(PhiR); | |||
4288 | assert((IsUniform || !State.VF.isScalable()) &&(static_cast <bool> ((IsUniform || !State.VF.isScalable ()) && "Cannot scalarize a scalable VF") ? void (0) : __assert_fail ("(IsUniform || !State.VF.isScalable()) && \"Cannot scalarize a scalable VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4289, __extension__ __PRETTY_FUNCTION__)) | |||
4289 | "Cannot scalarize a scalable VF")(static_cast <bool> ((IsUniform || !State.VF.isScalable ()) && "Cannot scalarize a scalable VF") ? void (0) : __assert_fail ("(IsUniform || !State.VF.isScalable()) && \"Cannot scalarize a scalable VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4289, __extension__ __PRETTY_FUNCTION__)); | |||
4290 | unsigned Lanes = IsUniform ? 1 : State.VF.getFixedValue(); | |||
4291 | ||||
4292 | for (unsigned Part = 0; Part < UF; ++Part) { | |||
4293 | Value *PartStart = | |||
4294 | createStepForVF(Builder, PtrInd->getType(), VF, Part); | |||
4295 | ||||
4296 | for (unsigned Lane = 0; Lane < Lanes; ++Lane) { | |||
4297 | Value *Idx = Builder.CreateAdd( | |||
4298 | PartStart, ConstantInt::get(PtrInd->getType(), Lane)); | |||
4299 | Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx); | |||
4300 | ||||
4301 | Value *Step = CreateStepValue(II.getStep(), *PSE.getSE(), | |||
4302 | State.CFG.PrevBB->getTerminator()); | |||
4303 | Value *SclrGep = emitTransformedIndex(Builder, GlobalIdx, | |||
4304 | II.getStartValue(), Step, II); | |||
4305 | SclrGep->setName("next.gep"); | |||
4306 | State.set(PhiR, SclrGep, VPIteration(Part, Lane)); | |||
4307 | } | |||
4308 | } | |||
4309 | return; | |||
4310 | } | |||
4311 | assert(isa<SCEVConstant>(II.getStep()) &&(static_cast <bool> (isa<SCEVConstant>(II.getStep ()) && "Induction step not a SCEV constant!") ? void ( 0) : __assert_fail ("isa<SCEVConstant>(II.getStep()) && \"Induction step not a SCEV constant!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4312, __extension__ __PRETTY_FUNCTION__)) | |||
4312 | "Induction step not a SCEV constant!")(static_cast <bool> (isa<SCEVConstant>(II.getStep ()) && "Induction step not a SCEV constant!") ? void ( 0) : __assert_fail ("isa<SCEVConstant>(II.getStep()) && \"Induction step not a SCEV constant!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4312, __extension__ __PRETTY_FUNCTION__)); | |||
4313 | Type *PhiType = II.getStep()->getType(); | |||
4314 | ||||
4315 | // Build a pointer phi | |||
4316 | Value *ScalarStartValue = PhiR->getStartValue()->getLiveInIRValue(); | |||
4317 | Type *ScStValueType = ScalarStartValue->getType(); | |||
4318 | PHINode *NewPointerPhi = | |||
4319 | PHINode::Create(ScStValueType, 2, "pointer.phi", CanonicalIV); | |||
4320 | NewPointerPhi->addIncoming(ScalarStartValue, LoopVectorPreHeader); | |||
4321 | ||||
4322 | // A pointer induction, performed by using a gep | |||
4323 | BasicBlock *LoopLatch = LI->getLoopFor(State.CFG.PrevBB)->getLoopLatch(); | |||
4324 | Instruction *InductionLoc = LoopLatch->getTerminator(); | |||
4325 | const SCEV *ScalarStep = II.getStep(); | |||
4326 | SCEVExpander Exp(*PSE.getSE(), DL, "induction"); | |||
4327 | Value *ScalarStepValue = | |||
4328 | Exp.expandCodeFor(ScalarStep, PhiType, InductionLoc); | |||
4329 | Value *RuntimeVF = getRuntimeVF(Builder, PhiType, VF); | |||
4330 | Value *NumUnrolledElems = | |||
4331 | Builder.CreateMul(RuntimeVF, ConstantInt::get(PhiType, State.UF)); | |||
4332 | Value *InductionGEP = GetElementPtrInst::Create( | |||
4333 | II.getElementType(), NewPointerPhi, | |||
4334 | Builder.CreateMul(ScalarStepValue, NumUnrolledElems), "ptr.ind", | |||
4335 | InductionLoc); | |||
4336 | NewPointerPhi->addIncoming(InductionGEP, LoopLatch); | |||
4337 | ||||
4338 | // Create UF many actual address geps that use the pointer | |||
4339 | // phi as base and a vectorized version of the step value | |||
4340 | // (<step*0, ..., step*N>) as offset. | |||
4341 | for (unsigned Part = 0; Part < State.UF; ++Part) { | |||
4342 | Type *VecPhiType = VectorType::get(PhiType, State.VF); | |||
4343 | Value *StartOffsetScalar = | |||
4344 | Builder.CreateMul(RuntimeVF, ConstantInt::get(PhiType, Part)); | |||
4345 | Value *StartOffset = | |||
4346 | Builder.CreateVectorSplat(State.VF, StartOffsetScalar); | |||
4347 | // Create a vector of consecutive numbers from zero to VF. | |||
4348 | StartOffset = | |||
4349 | Builder.CreateAdd(StartOffset, Builder.CreateStepVector(VecPhiType)); | |||
4350 | ||||
4351 | Value *GEP = Builder.CreateGEP( | |||
4352 | II.getElementType(), NewPointerPhi, | |||
4353 | Builder.CreateMul( | |||
4354 | StartOffset, Builder.CreateVectorSplat(State.VF, ScalarStepValue), | |||
4355 | "vector.gep")); | |||
4356 | State.set(PhiR, GEP, Part); | |||
4357 | } | |||
4358 | } | |||
4359 | } | |||
4360 | } | |||
4361 | ||||
4362 | /// A helper function for checking whether an integer division-related | |||
4363 | /// instruction may divide by zero (in which case it must be predicated if | |||
4364 | /// executed conditionally in the scalar code). | |||
4365 | /// TODO: It may be worthwhile to generalize and check isKnownNonZero(). | |||
4366 | /// Non-zero divisors that are non compile-time constants will not be | |||
4367 | /// converted into multiplication, so we will still end up scalarizing | |||
4368 | /// the division, but can do so w/o predication. | |||
4369 | static bool mayDivideByZero(Instruction &I) { | |||
4370 | assert((I.getOpcode() == Instruction::UDiv ||(static_cast <bool> ((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction ::URem || I.getOpcode() == Instruction::SRem) && "Unexpected instruction" ) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4374, __extension__ __PRETTY_FUNCTION__)) | |||
4371 | I.getOpcode() == Instruction::SDiv ||(static_cast <bool> ((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction ::URem || I.getOpcode() == Instruction::SRem) && "Unexpected instruction" ) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4374, __extension__ __PRETTY_FUNCTION__)) | |||
4372 | I.getOpcode() == Instruction::URem ||(static_cast <bool> ((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction ::URem || I.getOpcode() == Instruction::SRem) && "Unexpected instruction" ) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4374, __extension__ __PRETTY_FUNCTION__)) | |||
4373 | I.getOpcode() == Instruction::SRem) &&(static_cast <bool> ((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction ::URem || I.getOpcode() == Instruction::SRem) && "Unexpected instruction" ) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4374, __extension__ __PRETTY_FUNCTION__)) | |||
4374 | "Unexpected instruction")(static_cast <bool> ((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction ::URem || I.getOpcode() == Instruction::SRem) && "Unexpected instruction" ) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4374, __extension__ __PRETTY_FUNCTION__)); | |||
4375 | Value *Divisor = I.getOperand(1); | |||
4376 | auto *CInt = dyn_cast<ConstantInt>(Divisor); | |||
4377 | return !CInt || CInt->isZero(); | |||
4378 | } | |||
4379 | ||||
4380 | void InnerLoopVectorizer::widenCallInstruction(CallInst &I, VPValue *Def, | |||
4381 | VPUser &ArgOperands, | |||
4382 | VPTransformState &State) { | |||
4383 | assert(!isa<DbgInfoIntrinsic>(I) &&(static_cast <bool> (!isa<DbgInfoIntrinsic>(I) && "DbgInfoIntrinsic should have been dropped during VPlan construction" ) ? void (0) : __assert_fail ("!isa<DbgInfoIntrinsic>(I) && \"DbgInfoIntrinsic should have been dropped during VPlan construction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4384, __extension__ __PRETTY_FUNCTION__)) | |||
4384 | "DbgInfoIntrinsic should have been dropped during VPlan construction")(static_cast <bool> (!isa<DbgInfoIntrinsic>(I) && "DbgInfoIntrinsic should have been dropped during VPlan construction" ) ? void (0) : __assert_fail ("!isa<DbgInfoIntrinsic>(I) && \"DbgInfoIntrinsic should have been dropped during VPlan construction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4384, __extension__ __PRETTY_FUNCTION__)); | |||
4385 | setDebugLocFromInst(&I); | |||
4386 | ||||
4387 | Module *M = I.getParent()->getParent()->getParent(); | |||
4388 | auto *CI = cast<CallInst>(&I); | |||
4389 | ||||
4390 | SmallVector<Type *, 4> Tys; | |||
4391 | for (Value *ArgOperand : CI->args()) | |||
4392 | Tys.push_back(ToVectorTy(ArgOperand->getType(), VF.getKnownMinValue())); | |||
4393 | ||||
4394 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
4395 | ||||
4396 | // The flag shows whether we use Intrinsic or a usual Call for vectorized | |||
4397 | // version of the instruction. | |||
4398 | // Is it beneficial to perform intrinsic call compared to lib call? | |||
4399 | bool NeedToScalarize = false; | |||
4400 | InstructionCost CallCost = Cost->getVectorCallCost(CI, VF, NeedToScalarize); | |||
4401 | InstructionCost IntrinsicCost = ID ? Cost->getVectorIntrinsicCost(CI, VF) : 0; | |||
4402 | bool UseVectorIntrinsic = ID && IntrinsicCost <= CallCost; | |||
4403 | assert((UseVectorIntrinsic || !NeedToScalarize) &&(static_cast <bool> ((UseVectorIntrinsic || !NeedToScalarize ) && "Instruction should be scalarized elsewhere.") ? void (0) : __assert_fail ("(UseVectorIntrinsic || !NeedToScalarize) && \"Instruction should be scalarized elsewhere.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4404, __extension__ __PRETTY_FUNCTION__)) | |||
4404 | "Instruction should be scalarized elsewhere.")(static_cast <bool> ((UseVectorIntrinsic || !NeedToScalarize ) && "Instruction should be scalarized elsewhere.") ? void (0) : __assert_fail ("(UseVectorIntrinsic || !NeedToScalarize) && \"Instruction should be scalarized elsewhere.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4404, __extension__ __PRETTY_FUNCTION__)); | |||
4405 | assert((IntrinsicCost.isValid() || CallCost.isValid()) &&(static_cast <bool> ((IntrinsicCost.isValid() || CallCost .isValid()) && "Either the intrinsic cost or vector call cost must be valid" ) ? void (0) : __assert_fail ("(IntrinsicCost.isValid() || CallCost.isValid()) && \"Either the intrinsic cost or vector call cost must be valid\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4406, __extension__ __PRETTY_FUNCTION__)) | |||
4406 | "Either the intrinsic cost or vector call cost must be valid")(static_cast <bool> ((IntrinsicCost.isValid() || CallCost .isValid()) && "Either the intrinsic cost or vector call cost must be valid" ) ? void (0) : __assert_fail ("(IntrinsicCost.isValid() || CallCost.isValid()) && \"Either the intrinsic cost or vector call cost must be valid\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4406, __extension__ __PRETTY_FUNCTION__)); | |||
4407 | ||||
4408 | for (unsigned Part = 0; Part < UF; ++Part) { | |||
4409 | SmallVector<Type *, 2> TysForDecl = {CI->getType()}; | |||
4410 | SmallVector<Value *, 4> Args; | |||
4411 | for (auto &I : enumerate(ArgOperands.operands())) { | |||
4412 | // Some intrinsics have a scalar argument - don't replace it with a | |||
4413 | // vector. | |||
4414 | Value *Arg; | |||
4415 | if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, I.index())) | |||
4416 | Arg = State.get(I.value(), Part); | |||
4417 | else { | |||
4418 | Arg = State.get(I.value(), VPIteration(0, 0)); | |||
4419 | if (hasVectorInstrinsicOverloadedScalarOpd(ID, I.index())) | |||
4420 | TysForDecl.push_back(Arg->getType()); | |||
4421 | } | |||
4422 | Args.push_back(Arg); | |||
4423 | } | |||
4424 | ||||
4425 | Function *VectorF; | |||
4426 | if (UseVectorIntrinsic) { | |||
4427 | // Use vector version of the intrinsic. | |||
4428 | if (VF.isVector()) | |||
4429 | TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF); | |||
4430 | VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl); | |||
4431 | assert(VectorF && "Can't retrieve vector intrinsic.")(static_cast <bool> (VectorF && "Can't retrieve vector intrinsic." ) ? void (0) : __assert_fail ("VectorF && \"Can't retrieve vector intrinsic.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4431, __extension__ __PRETTY_FUNCTION__)); | |||
4432 | } else { | |||
4433 | // Use vector version of the function call. | |||
4434 | const VFShape Shape = VFShape::get(*CI, VF, false /*HasGlobalPred*/); | |||
4435 | #ifndef NDEBUG | |||
4436 | assert(VFDatabase(*CI).getVectorizedFunction(Shape) != nullptr &&(static_cast <bool> (VFDatabase(*CI).getVectorizedFunction (Shape) != nullptr && "Can't create vector function." ) ? void (0) : __assert_fail ("VFDatabase(*CI).getVectorizedFunction(Shape) != nullptr && \"Can't create vector function.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4437, __extension__ __PRETTY_FUNCTION__)) | |||
4437 | "Can't create vector function.")(static_cast <bool> (VFDatabase(*CI).getVectorizedFunction (Shape) != nullptr && "Can't create vector function." ) ? void (0) : __assert_fail ("VFDatabase(*CI).getVectorizedFunction(Shape) != nullptr && \"Can't create vector function.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4437, __extension__ __PRETTY_FUNCTION__)); | |||
4438 | #endif | |||
4439 | VectorF = VFDatabase(*CI).getVectorizedFunction(Shape); | |||
4440 | } | |||
4441 | SmallVector<OperandBundleDef, 1> OpBundles; | |||
4442 | CI->getOperandBundlesAsDefs(OpBundles); | |||
4443 | CallInst *V = Builder.CreateCall(VectorF, Args, OpBundles); | |||
4444 | ||||
4445 | if (isa<FPMathOperator>(V)) | |||
4446 | V->copyFastMathFlags(CI); | |||
4447 | ||||
4448 | State.set(Def, V, Part); | |||
4449 | addMetadata(V, &I); | |||
4450 | } | |||
4451 | } | |||
4452 | ||||
4453 | void LoopVectorizationCostModel::collectLoopScalars(ElementCount VF) { | |||
4454 | // We should not collect Scalars more than once per VF. Right now, this | |||
4455 | // function is called from collectUniformsAndScalars(), which already does | |||
4456 | // this check. Collecting Scalars for VF=1 does not make any sense. | |||
4457 | assert(VF.isVector() && Scalars.find(VF) == Scalars.end() &&(static_cast <bool> (VF.isVector() && Scalars.find (VF) == Scalars.end() && "This function should not be visited twice for the same VF" ) ? void (0) : __assert_fail ("VF.isVector() && Scalars.find(VF) == Scalars.end() && \"This function should not be visited twice for the same VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4458, __extension__ __PRETTY_FUNCTION__)) | |||
4458 | "This function should not be visited twice for the same VF")(static_cast <bool> (VF.isVector() && Scalars.find (VF) == Scalars.end() && "This function should not be visited twice for the same VF" ) ? void (0) : __assert_fail ("VF.isVector() && Scalars.find(VF) == Scalars.end() && \"This function should not be visited twice for the same VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4458, __extension__ __PRETTY_FUNCTION__)); | |||
4459 | ||||
4460 | // This avoids any chances of creating a REPLICATE recipe during planning | |||
4461 | // since that would result in generation of scalarized code during execution, | |||
4462 | // which is not supported for scalable vectors. | |||
4463 | if (VF.isScalable()) { | |||
4464 | Scalars[VF].insert(Uniforms[VF].begin(), Uniforms[VF].end()); | |||
4465 | return; | |||
4466 | } | |||
4467 | ||||
4468 | SmallSetVector<Instruction *, 8> Worklist; | |||
4469 | ||||
4470 | // These sets are used to seed the analysis with pointers used by memory | |||
4471 | // accesses that will remain scalar. | |||
4472 | SmallSetVector<Instruction *, 8> ScalarPtrs; | |||
4473 | SmallPtrSet<Instruction *, 8> PossibleNonScalarPtrs; | |||
4474 | auto *Latch = TheLoop->getLoopLatch(); | |||
4475 | ||||
4476 | // A helper that returns true if the use of Ptr by MemAccess will be scalar. | |||
4477 | // The pointer operands of loads and stores will be scalar as long as the | |||
4478 | // memory access is not a gather or scatter operation. The value operand of a | |||
4479 | // store will remain scalar if the store is scalarized. | |||
4480 | auto isScalarUse = [&](Instruction *MemAccess, Value *Ptr) { | |||
4481 | InstWidening WideningDecision = getWideningDecision(MemAccess, VF); | |||
4482 | assert(WideningDecision != CM_Unknown &&(static_cast <bool> (WideningDecision != CM_Unknown && "Widening decision should be ready at this moment") ? void ( 0) : __assert_fail ("WideningDecision != CM_Unknown && \"Widening decision should be ready at this moment\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4483, __extension__ __PRETTY_FUNCTION__)) | |||
4483 | "Widening decision should be ready at this moment")(static_cast <bool> (WideningDecision != CM_Unknown && "Widening decision should be ready at this moment") ? void ( 0) : __assert_fail ("WideningDecision != CM_Unknown && \"Widening decision should be ready at this moment\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4483, __extension__ __PRETTY_FUNCTION__)); | |||
4484 | if (auto *Store = dyn_cast<StoreInst>(MemAccess)) | |||
4485 | if (Ptr == Store->getValueOperand()) | |||
4486 | return WideningDecision == CM_Scalarize; | |||
4487 | assert(Ptr == getLoadStorePointerOperand(MemAccess) &&(static_cast <bool> (Ptr == getLoadStorePointerOperand( MemAccess) && "Ptr is neither a value or pointer operand" ) ? void (0) : __assert_fail ("Ptr == getLoadStorePointerOperand(MemAccess) && \"Ptr is neither a value or pointer operand\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4488, __extension__ __PRETTY_FUNCTION__)) | |||
4488 | "Ptr is neither a value or pointer operand")(static_cast <bool> (Ptr == getLoadStorePointerOperand( MemAccess) && "Ptr is neither a value or pointer operand" ) ? void (0) : __assert_fail ("Ptr == getLoadStorePointerOperand(MemAccess) && \"Ptr is neither a value or pointer operand\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4488, __extension__ __PRETTY_FUNCTION__)); | |||
4489 | return WideningDecision != CM_GatherScatter; | |||
4490 | }; | |||
4491 | ||||
4492 | // A helper that returns true if the given value is a bitcast or | |||
4493 | // getelementptr instruction contained in the loop. | |||
4494 | auto isLoopVaryingBitCastOrGEP = [&](Value *V) { | |||
4495 | return ((isa<BitCastInst>(V) && V->getType()->isPointerTy()) || | |||
4496 | isa<GetElementPtrInst>(V)) && | |||
4497 | !TheLoop->isLoopInvariant(V); | |||
4498 | }; | |||
4499 | ||||
4500 | // A helper that evaluates a memory access's use of a pointer. If the use will | |||
4501 | // be a scalar use and the pointer is only used by memory accesses, we place | |||
4502 | // the pointer in ScalarPtrs. Otherwise, the pointer is placed in | |||
4503 | // PossibleNonScalarPtrs. | |||
4504 | auto evaluatePtrUse = [&](Instruction *MemAccess, Value *Ptr) { | |||
4505 | // We only care about bitcast and getelementptr instructions contained in | |||
4506 | // the loop. | |||
4507 | if (!isLoopVaryingBitCastOrGEP(Ptr)) | |||
4508 | return; | |||
4509 | ||||
4510 | // If the pointer has already been identified as scalar (e.g., if it was | |||
4511 | // also identified as uniform), there's nothing to do. | |||
4512 | auto *I = cast<Instruction>(Ptr); | |||
4513 | if (Worklist.count(I)) | |||
4514 | return; | |||
4515 | ||||
4516 | // If the use of the pointer will be a scalar use, and all users of the | |||
4517 | // pointer are memory accesses, place the pointer in ScalarPtrs. Otherwise, | |||
4518 | // place the pointer in PossibleNonScalarPtrs. | |||
4519 | if (isScalarUse(MemAccess, Ptr) && llvm::all_of(I->users(), [&](User *U) { | |||
4520 | return isa<LoadInst>(U) || isa<StoreInst>(U); | |||
4521 | })) | |||
4522 | ScalarPtrs.insert(I); | |||
4523 | else | |||
4524 | PossibleNonScalarPtrs.insert(I); | |||
4525 | }; | |||
4526 | ||||
4527 | // We seed the scalars analysis with three classes of instructions: (1) | |||
4528 | // instructions marked uniform-after-vectorization and (2) bitcast, | |||
4529 | // getelementptr and (pointer) phi instructions used by memory accesses | |||
4530 | // requiring a scalar use. | |||
4531 | // | |||
4532 | // (1) Add to the worklist all instructions that have been identified as | |||
4533 | // uniform-after-vectorization. | |||
4534 | Worklist.insert(Uniforms[VF].begin(), Uniforms[VF].end()); | |||
4535 | ||||
4536 | // (2) Add to the worklist all bitcast and getelementptr instructions used by | |||
4537 | // memory accesses requiring a scalar use. The pointer operands of loads and | |||
4538 | // stores will be scalar as long as the memory accesses is not a gather or | |||
4539 | // scatter operation. The value operand of a store will remain scalar if the | |||
4540 | // store is scalarized. | |||
4541 | for (auto *BB : TheLoop->blocks()) | |||
4542 | for (auto &I : *BB) { | |||
4543 | if (auto *Load = dyn_cast<LoadInst>(&I)) { | |||
4544 | evaluatePtrUse(Load, Load->getPointerOperand()); | |||
4545 | } else if (auto *Store = dyn_cast<StoreInst>(&I)) { | |||
4546 | evaluatePtrUse(Store, Store->getPointerOperand()); | |||
4547 | evaluatePtrUse(Store, Store->getValueOperand()); | |||
4548 | } | |||
4549 | } | |||
4550 | for (auto *I : ScalarPtrs) | |||
4551 | if (!PossibleNonScalarPtrs.count(I)) { | |||
4552 | 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); | |||
4553 | Worklist.insert(I); | |||
4554 | } | |||
4555 | ||||
4556 | // Insert the forced scalars. | |||
4557 | // FIXME: Currently widenPHIInstruction() often creates a dead vector | |||
4558 | // induction variable when the PHI user is scalarized. | |||
4559 | auto ForcedScalar = ForcedScalars.find(VF); | |||
4560 | if (ForcedScalar != ForcedScalars.end()) | |||
4561 | for (auto *I : ForcedScalar->second) | |||
4562 | Worklist.insert(I); | |||
4563 | ||||
4564 | // Expand the worklist by looking through any bitcasts and getelementptr | |||
4565 | // instructions we've already identified as scalar. This is similar to the | |||
4566 | // expansion step in collectLoopUniforms(); however, here we're only | |||
4567 | // expanding to include additional bitcasts and getelementptr instructions. | |||
4568 | unsigned Idx = 0; | |||
4569 | while (Idx != Worklist.size()) { | |||
4570 | Instruction *Dst = Worklist[Idx++]; | |||
4571 | if (!isLoopVaryingBitCastOrGEP(Dst->getOperand(0))) | |||
4572 | continue; | |||
4573 | auto *Src = cast<Instruction>(Dst->getOperand(0)); | |||
4574 | if (llvm::all_of(Src->users(), [&](User *U) -> bool { | |||
4575 | auto *J = cast<Instruction>(U); | |||
4576 | return !TheLoop->contains(J) || Worklist.count(J) || | |||
4577 | ((isa<LoadInst>(J) || isa<StoreInst>(J)) && | |||
4578 | isScalarUse(J, Src)); | |||
4579 | })) { | |||
4580 | Worklist.insert(Src); | |||
4581 | 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); | |||
4582 | } | |||
4583 | } | |||
4584 | ||||
4585 | // An induction variable will remain scalar if all users of the induction | |||
4586 | // variable and induction variable update remain scalar. | |||
4587 | for (auto &Induction : Legal->getInductionVars()) { | |||
4588 | auto *Ind = Induction.first; | |||
4589 | auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch)); | |||
4590 | ||||
4591 | // If tail-folding is applied, the primary induction variable will be used | |||
4592 | // to feed a vector compare. | |||
4593 | if (Ind == Legal->getPrimaryInduction() && foldTailByMasking()) | |||
4594 | continue; | |||
4595 | ||||
4596 | // Returns true if \p Indvar is a pointer induction that is used directly by | |||
4597 | // load/store instruction \p I. | |||
4598 | auto IsDirectLoadStoreFromPtrIndvar = [&](Instruction *Indvar, | |||
4599 | Instruction *I) { | |||
4600 | return Induction.second.getKind() == | |||
4601 | InductionDescriptor::IK_PtrInduction && | |||
4602 | (isa<LoadInst>(I) || isa<StoreInst>(I)) && | |||
4603 | Indvar == getLoadStorePointerOperand(I) && isScalarUse(I, Indvar); | |||
4604 | }; | |||
4605 | ||||
4606 | // Determine if all users of the induction variable are scalar after | |||
4607 | // vectorization. | |||
4608 | auto ScalarInd = llvm::all_of(Ind->users(), [&](User *U) -> bool { | |||
4609 | auto *I = cast<Instruction>(U); | |||
4610 | return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I) || | |||
4611 | IsDirectLoadStoreFromPtrIndvar(Ind, I); | |||
4612 | }); | |||
4613 | if (!ScalarInd) | |||
4614 | continue; | |||
4615 | ||||
4616 | // Determine if all users of the induction variable update instruction are | |||
4617 | // scalar after vectorization. | |||
4618 | auto ScalarIndUpdate = | |||
4619 | llvm::all_of(IndUpdate->users(), [&](User *U) -> bool { | |||
4620 | auto *I = cast<Instruction>(U); | |||
4621 | return I == Ind || !TheLoop->contains(I) || Worklist.count(I) || | |||
4622 | IsDirectLoadStoreFromPtrIndvar(IndUpdate, I); | |||
4623 | }); | |||
4624 | if (!ScalarIndUpdate) | |||
4625 | continue; | |||
4626 | ||||
4627 | // The induction variable and its update instruction will remain scalar. | |||
4628 | Worklist.insert(Ind); | |||
4629 | Worklist.insert(IndUpdate); | |||
4630 | 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); | |||
4631 | LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *IndUpdatedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: " << *IndUpdate << "\n"; } } while (false) | |||
4632 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: " << *IndUpdate << "\n"; } } while (false); | |||
4633 | } | |||
4634 | ||||
4635 | Scalars[VF].insert(Worklist.begin(), Worklist.end()); | |||
4636 | } | |||
4637 | ||||
4638 | bool LoopVectorizationCostModel::isScalarWithPredication( | |||
4639 | Instruction *I, ElementCount VF) const { | |||
4640 | if (!blockNeedsPredicationForAnyReason(I->getParent())) | |||
4641 | return false; | |||
4642 | switch(I->getOpcode()) { | |||
4643 | default: | |||
4644 | break; | |||
4645 | case Instruction::Load: | |||
4646 | case Instruction::Store: { | |||
4647 | if (!Legal->isMaskRequired(I)) | |||
4648 | return false; | |||
4649 | auto *Ptr = getLoadStorePointerOperand(I); | |||
4650 | auto *Ty = getLoadStoreType(I); | |||
4651 | Type *VTy = Ty; | |||
4652 | if (VF.isVector()) | |||
4653 | VTy = VectorType::get(Ty, VF); | |||
4654 | const Align Alignment = getLoadStoreAlignment(I); | |||
4655 | return isa<LoadInst>(I) ? !(isLegalMaskedLoad(Ty, Ptr, Alignment) || | |||
4656 | TTI.isLegalMaskedGather(VTy, Alignment)) | |||
4657 | : !(isLegalMaskedStore(Ty, Ptr, Alignment) || | |||
4658 | TTI.isLegalMaskedScatter(VTy, Alignment)); | |||
4659 | } | |||
4660 | case Instruction::UDiv: | |||
4661 | case Instruction::SDiv: | |||
4662 | case Instruction::SRem: | |||
4663 | case Instruction::URem: | |||
4664 | return mayDivideByZero(*I); | |||
4665 | } | |||
4666 | return false; | |||
4667 | } | |||
4668 | ||||
4669 | bool LoopVectorizationCostModel::interleavedAccessCanBeWidened( | |||
4670 | Instruction *I, ElementCount VF) { | |||
4671 | assert(isAccessInterleaved(I) && "Expecting interleaved access.")(static_cast <bool> (isAccessInterleaved(I) && "Expecting interleaved access." ) ? void (0) : __assert_fail ("isAccessInterleaved(I) && \"Expecting interleaved access.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4671, __extension__ __PRETTY_FUNCTION__)); | |||
4672 | assert(getWideningDecision(I, VF) == CM_Unknown &&(static_cast <bool> (getWideningDecision(I, VF) == CM_Unknown && "Decision should not be set yet.") ? void (0) : __assert_fail ("getWideningDecision(I, VF) == CM_Unknown && \"Decision should not be set yet.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4673, __extension__ __PRETTY_FUNCTION__)) | |||
4673 | "Decision should not be set yet.")(static_cast <bool> (getWideningDecision(I, VF) == CM_Unknown && "Decision should not be set yet.") ? void (0) : __assert_fail ("getWideningDecision(I, VF) == CM_Unknown && \"Decision should not be set yet.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4673, __extension__ __PRETTY_FUNCTION__)); | |||
4674 | auto *Group = getInterleavedAccessGroup(I); | |||
4675 | assert(Group && "Must have a group.")(static_cast <bool> (Group && "Must have a group." ) ? void (0) : __assert_fail ("Group && \"Must have a group.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4675, __extension__ __PRETTY_FUNCTION__)); | |||
4676 | ||||
4677 | // If the instruction's allocated size doesn't equal it's type size, it | |||
4678 | // requires padding and will be scalarized. | |||
4679 | auto &DL = I->getModule()->getDataLayout(); | |||
4680 | auto *ScalarTy = getLoadStoreType(I); | |||
4681 | if (hasIrregularType(ScalarTy, DL)) | |||
4682 | return false; | |||
4683 | ||||
4684 | // Check if masking is required. | |||
4685 | // A Group may need masking for one of two reasons: it resides in a block that | |||
4686 | // needs predication, or it was decided to use masking to deal with gaps | |||
4687 | // (either a gap at the end of a load-access that may result in a speculative | |||
4688 | // load, or any gaps in a store-access). | |||
4689 | bool PredicatedAccessRequiresMasking = | |||
4690 | blockNeedsPredicationForAnyReason(I->getParent()) && | |||
4691 | Legal->isMaskRequired(I); | |||
4692 | bool LoadAccessWithGapsRequiresEpilogMasking = | |||
4693 | isa<LoadInst>(I) && Group->requiresScalarEpilogue() && | |||
4694 | !isScalarEpilogueAllowed(); | |||
4695 | bool StoreAccessWithGapsRequiresMasking = | |||
4696 | isa<StoreInst>(I) && (Group->getNumMembers() < Group->getFactor()); | |||
4697 | if (!PredicatedAccessRequiresMasking && | |||
4698 | !LoadAccessWithGapsRequiresEpilogMasking && | |||
4699 | !StoreAccessWithGapsRequiresMasking) | |||
4700 | return true; | |||
4701 | ||||
4702 | // If masked interleaving is required, we expect that the user/target had | |||
4703 | // enabled it, because otherwise it either wouldn't have been created or | |||
4704 | // it should have been invalidated by the CostModel. | |||
4705 | assert(useMaskedInterleavedAccesses(TTI) &&(static_cast <bool> (useMaskedInterleavedAccesses(TTI) && "Masked interleave-groups for predicated accesses are not enabled." ) ? void (0) : __assert_fail ("useMaskedInterleavedAccesses(TTI) && \"Masked interleave-groups for predicated accesses are not enabled.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4706, __extension__ __PRETTY_FUNCTION__)) | |||
4706 | "Masked interleave-groups for predicated accesses are not enabled.")(static_cast <bool> (useMaskedInterleavedAccesses(TTI) && "Masked interleave-groups for predicated accesses are not enabled." ) ? void (0) : __assert_fail ("useMaskedInterleavedAccesses(TTI) && \"Masked interleave-groups for predicated accesses are not enabled.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4706, __extension__ __PRETTY_FUNCTION__)); | |||
4707 | ||||
4708 | if (Group->isReverse()) | |||
4709 | return false; | |||
4710 | ||||
4711 | auto *Ty = getLoadStoreType(I); | |||
4712 | const Align Alignment = getLoadStoreAlignment(I); | |||
4713 | return isa<LoadInst>(I) ? TTI.isLegalMaskedLoad(Ty, Alignment) | |||
4714 | : TTI.isLegalMaskedStore(Ty, Alignment); | |||
4715 | } | |||
4716 | ||||
4717 | bool LoopVectorizationCostModel::memoryInstructionCanBeWidened( | |||
4718 | Instruction *I, ElementCount VF) { | |||
4719 | // Get and ensure we have a valid memory instruction. | |||
4720 | assert((isa<LoadInst, StoreInst>(I)) && "Invalid memory instruction")(static_cast <bool> ((isa<LoadInst, StoreInst>(I) ) && "Invalid memory instruction") ? void (0) : __assert_fail ("(isa<LoadInst, StoreInst>(I)) && \"Invalid memory instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4720, __extension__ __PRETTY_FUNCTION__)); | |||
4721 | ||||
4722 | auto *Ptr = getLoadStorePointerOperand(I); | |||
4723 | auto *ScalarTy = getLoadStoreType(I); | |||
4724 | ||||
4725 | // In order to be widened, the pointer should be consecutive, first of all. | |||
4726 | if (!Legal->isConsecutivePtr(ScalarTy, Ptr)) | |||
4727 | return false; | |||
4728 | ||||
4729 | // If the instruction is a store located in a predicated block, it will be | |||
4730 | // scalarized. | |||
4731 | if (isScalarWithPredication(I, VF)) | |||
4732 | return false; | |||
4733 | ||||
4734 | // If the instruction's allocated size doesn't equal it's type size, it | |||
4735 | // requires padding and will be scalarized. | |||
4736 | auto &DL = I->getModule()->getDataLayout(); | |||
4737 | if (hasIrregularType(ScalarTy, DL)) | |||
4738 | return false; | |||
4739 | ||||
4740 | return true; | |||
4741 | } | |||
4742 | ||||
4743 | void LoopVectorizationCostModel::collectLoopUniforms(ElementCount VF) { | |||
4744 | // We should not collect Uniforms more than once per VF. Right now, | |||
4745 | // this function is called from collectUniformsAndScalars(), which | |||
4746 | // already does this check. Collecting Uniforms for VF=1 does not make any | |||
4747 | // sense. | |||
4748 | ||||
4749 | assert(VF.isVector() && Uniforms.find(VF) == Uniforms.end() &&(static_cast <bool> (VF.isVector() && Uniforms. find(VF) == Uniforms.end() && "This function should not be visited twice for the same VF" ) ? void (0) : __assert_fail ("VF.isVector() && Uniforms.find(VF) == Uniforms.end() && \"This function should not be visited twice for the same VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4750, __extension__ __PRETTY_FUNCTION__)) | |||
4750 | "This function should not be visited twice for the same VF")(static_cast <bool> (VF.isVector() && Uniforms. find(VF) == Uniforms.end() && "This function should not be visited twice for the same VF" ) ? void (0) : __assert_fail ("VF.isVector() && Uniforms.find(VF) == Uniforms.end() && \"This function should not be visited twice for the same VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4750, __extension__ __PRETTY_FUNCTION__)); | |||
4751 | ||||
4752 | // Visit the list of Uniforms. If we'll not find any uniform value, we'll | |||
4753 | // not analyze again. Uniforms.count(VF) will return 1. | |||
4754 | Uniforms[VF].clear(); | |||
4755 | ||||
4756 | // We now know that the loop is vectorizable! | |||
4757 | // Collect instructions inside the loop that will remain uniform after | |||
4758 | // vectorization. | |||
4759 | ||||
4760 | // Global values, params and instructions outside of current loop are out of | |||
4761 | // scope. | |||
4762 | auto isOutOfScope = [&](Value *V) -> bool { | |||
4763 | Instruction *I = dyn_cast<Instruction>(V); | |||
4764 | return (!I || !TheLoop->contains(I)); | |||
4765 | }; | |||
4766 | ||||
4767 | // Worklist containing uniform instructions demanding lane 0. | |||
4768 | SetVector<Instruction *> Worklist; | |||
4769 | BasicBlock *Latch = TheLoop->getLoopLatch(); | |||
4770 | ||||
4771 | // Add uniform instructions demanding lane 0 to the worklist. Instructions | |||
4772 | // that are scalar with predication must not be considered uniform after | |||
4773 | // vectorization, because that would create an erroneous replicating region | |||
4774 | // where only a single instance out of VF should be formed. | |||
4775 | // TODO: optimize such seldom cases if found important, see PR40816. | |||
4776 | auto addToWorklistIfAllowed = [&](Instruction *I) -> void { | |||
4777 | if (isOutOfScope(I)) { | |||
4778 | 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) | |||
4779 | << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found not uniform due to scope: " << *I << "\n"; } } while (false); | |||
4780 | return; | |||
4781 | } | |||
4782 | if (isScalarWithPredication(I, VF)) { | |||
4783 | 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) | |||
4784 | << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found not uniform being ScalarWithPredication: " << *I << "\n"; } } while (false); | |||
4785 | return; | |||
4786 | } | |||
4787 | 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); | |||
4788 | Worklist.insert(I); | |||
4789 | }; | |||
4790 | ||||
4791 | // Start with the conditional branch. If the branch condition is an | |||
4792 | // instruction contained in the loop that is only used by the branch, it is | |||
4793 | // uniform. | |||
4794 | auto *Cmp = dyn_cast<Instruction>(Latch->getTerminator()->getOperand(0)); | |||
4795 | if (Cmp && TheLoop->contains(Cmp) && Cmp->hasOneUse()) | |||
4796 | addToWorklistIfAllowed(Cmp); | |||
4797 | ||||
4798 | auto isUniformDecision = [&](Instruction *I, ElementCount VF) { | |||
4799 | InstWidening WideningDecision = getWideningDecision(I, VF); | |||
4800 | assert(WideningDecision != CM_Unknown &&(static_cast <bool> (WideningDecision != CM_Unknown && "Widening decision should be ready at this moment") ? void ( 0) : __assert_fail ("WideningDecision != CM_Unknown && \"Widening decision should be ready at this moment\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4801, __extension__ __PRETTY_FUNCTION__)) | |||
4801 | "Widening decision should be ready at this moment")(static_cast <bool> (WideningDecision != CM_Unknown && "Widening decision should be ready at this moment") ? void ( 0) : __assert_fail ("WideningDecision != CM_Unknown && \"Widening decision should be ready at this moment\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4801, __extension__ __PRETTY_FUNCTION__)); | |||
4802 | ||||
4803 | // A uniform memory op is itself uniform. We exclude uniform stores | |||
4804 | // here as they demand the last lane, not the first one. | |||
4805 | if (isa<LoadInst>(I) && Legal->isUniformMemOp(*I)) { | |||
4806 | assert(WideningDecision == CM_Scalarize)(static_cast <bool> (WideningDecision == CM_Scalarize) ? void (0) : __assert_fail ("WideningDecision == CM_Scalarize" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4806, __extension__ __PRETTY_FUNCTION__)); | |||
4807 | return true; | |||
4808 | } | |||
4809 | ||||
4810 | return (WideningDecision == CM_Widen || | |||
4811 | WideningDecision == CM_Widen_Reverse || | |||
4812 | WideningDecision == CM_Interleave); | |||
4813 | }; | |||
4814 | ||||
4815 | ||||
4816 | // Returns true if Ptr is the pointer operand of a memory access instruction | |||
4817 | // I, and I is known to not require scalarization. | |||
4818 | auto isVectorizedMemAccessUse = [&](Instruction *I, Value *Ptr) -> bool { | |||
4819 | return getLoadStorePointerOperand(I) == Ptr && isUniformDecision(I, VF); | |||
4820 | }; | |||
4821 | ||||
4822 | // Holds a list of values which are known to have at least one uniform use. | |||
4823 | // Note that there may be other uses which aren't uniform. A "uniform use" | |||
4824 | // here is something which only demands lane 0 of the unrolled iterations; | |||
4825 | // it does not imply that all lanes produce the same value (e.g. this is not | |||
4826 | // the usual meaning of uniform) | |||
4827 | SetVector<Value *> HasUniformUse; | |||
4828 | ||||
4829 | // Scan the loop for instructions which are either a) known to have only | |||
4830 | // lane 0 demanded or b) are uses which demand only lane 0 of their operand. | |||
4831 | for (auto *BB : TheLoop->blocks()) | |||
4832 | for (auto &I : *BB) { | |||
4833 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I)) { | |||
4834 | switch (II->getIntrinsicID()) { | |||
4835 | case Intrinsic::sideeffect: | |||
4836 | case Intrinsic::experimental_noalias_scope_decl: | |||
4837 | case Intrinsic::assume: | |||
4838 | case Intrinsic::lifetime_start: | |||
4839 | case Intrinsic::lifetime_end: | |||
4840 | if (TheLoop->hasLoopInvariantOperands(&I)) | |||
4841 | addToWorklistIfAllowed(&I); | |||
4842 | break; | |||
4843 | default: | |||
4844 | break; | |||
4845 | } | |||
4846 | } | |||
4847 | ||||
4848 | // ExtractValue instructions must be uniform, because the operands are | |||
4849 | // known to be loop-invariant. | |||
4850 | if (auto *EVI = dyn_cast<ExtractValueInst>(&I)) { | |||
4851 | assert(isOutOfScope(EVI->getAggregateOperand()) &&(static_cast <bool> (isOutOfScope(EVI->getAggregateOperand ()) && "Expected aggregate value to be loop invariant" ) ? void (0) : __assert_fail ("isOutOfScope(EVI->getAggregateOperand()) && \"Expected aggregate value to be loop invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4852, __extension__ __PRETTY_FUNCTION__)) | |||
4852 | "Expected aggregate value to be loop invariant")(static_cast <bool> (isOutOfScope(EVI->getAggregateOperand ()) && "Expected aggregate value to be loop invariant" ) ? void (0) : __assert_fail ("isOutOfScope(EVI->getAggregateOperand()) && \"Expected aggregate value to be loop invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4852, __extension__ __PRETTY_FUNCTION__)); | |||
4853 | addToWorklistIfAllowed(EVI); | |||
4854 | continue; | |||
4855 | } | |||
4856 | ||||
4857 | // If there's no pointer operand, there's nothing to do. | |||
4858 | auto *Ptr = getLoadStorePointerOperand(&I); | |||
4859 | if (!Ptr) | |||
4860 | continue; | |||
4861 | ||||
4862 | // A uniform memory op is itself uniform. We exclude uniform stores | |||
4863 | // here as they demand the last lane, not the first one. | |||
4864 | if (isa<LoadInst>(I) && Legal->isUniformMemOp(I)) | |||
4865 | addToWorklistIfAllowed(&I); | |||
4866 | ||||
4867 | if (isUniformDecision(&I, VF)) { | |||
4868 | assert(isVectorizedMemAccessUse(&I, Ptr) && "consistency check")(static_cast <bool> (isVectorizedMemAccessUse(&I, Ptr ) && "consistency check") ? void (0) : __assert_fail ( "isVectorizedMemAccessUse(&I, Ptr) && \"consistency check\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4868, __extension__ __PRETTY_FUNCTION__)); | |||
4869 | HasUniformUse.insert(Ptr); | |||
4870 | } | |||
4871 | } | |||
4872 | ||||
4873 | // Add to the worklist any operands which have *only* uniform (e.g. lane 0 | |||
4874 | // demanding) users. Since loops are assumed to be in LCSSA form, this | |||
4875 | // disallows uses outside the loop as well. | |||
4876 | for (auto *V : HasUniformUse) { | |||
4877 | if (isOutOfScope(V)) | |||
4878 | continue; | |||
4879 | auto *I = cast<Instruction>(V); | |||
4880 | auto UsersAreMemAccesses = | |||
4881 | llvm::all_of(I->users(), [&](User *U) -> bool { | |||
4882 | return isVectorizedMemAccessUse(cast<Instruction>(U), V); | |||
4883 | }); | |||
4884 | if (UsersAreMemAccesses) | |||
4885 | addToWorklistIfAllowed(I); | |||
4886 | } | |||
4887 | ||||
4888 | // Expand Worklist in topological order: whenever a new instruction | |||
4889 | // is added , its users should be already inside Worklist. It ensures | |||
4890 | // a uniform instruction will only be used by uniform instructions. | |||
4891 | unsigned idx = 0; | |||
4892 | while (idx != Worklist.size()) { | |||
4893 | Instruction *I = Worklist[idx++]; | |||
4894 | ||||
4895 | for (auto OV : I->operand_values()) { | |||
4896 | // isOutOfScope operands cannot be uniform instructions. | |||
4897 | if (isOutOfScope(OV)) | |||
4898 | continue; | |||
4899 | // First order recurrence Phi's should typically be considered | |||
4900 | // non-uniform. | |||
4901 | auto *OP = dyn_cast<PHINode>(OV); | |||
4902 | if (OP && Legal->isFirstOrderRecurrence(OP)) | |||
4903 | continue; | |||
4904 | // If all the users of the operand are uniform, then add the | |||
4905 | // operand into the uniform worklist. | |||
4906 | auto *OI = cast<Instruction>(OV); | |||
4907 | if (llvm::all_of(OI->users(), [&](User *U) -> bool { | |||
4908 | auto *J = cast<Instruction>(U); | |||
4909 | return Worklist.count(J) || isVectorizedMemAccessUse(J, OI); | |||
4910 | })) | |||
4911 | addToWorklistIfAllowed(OI); | |||
4912 | } | |||
4913 | } | |||
4914 | ||||
4915 | // For an instruction to be added into Worklist above, all its users inside | |||
4916 | // the loop should also be in Worklist. However, this condition cannot be | |||
4917 | // true for phi nodes that form a cyclic dependence. We must process phi | |||
4918 | // nodes separately. An induction variable will remain uniform if all users | |||
4919 | // of the induction variable and induction variable update remain uniform. | |||
4920 | // The code below handles both pointer and non-pointer induction variables. | |||
4921 | for (auto &Induction : Legal->getInductionVars()) { | |||
4922 | auto *Ind = Induction.first; | |||
4923 | auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch)); | |||
4924 | ||||
4925 | // Determine if all users of the induction variable are uniform after | |||
4926 | // vectorization. | |||
4927 | auto UniformInd = llvm::all_of(Ind->users(), [&](User *U) -> bool { | |||
4928 | auto *I = cast<Instruction>(U); | |||
4929 | return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I) || | |||
4930 | isVectorizedMemAccessUse(I, Ind); | |||
4931 | }); | |||
4932 | if (!UniformInd) | |||
4933 | continue; | |||
4934 | ||||
4935 | // Determine if all users of the induction variable update instruction are | |||
4936 | // uniform after vectorization. | |||
4937 | auto UniformIndUpdate = | |||
4938 | llvm::all_of(IndUpdate->users(), [&](User *U) -> bool { | |||
4939 | auto *I = cast<Instruction>(U); | |||
4940 | return I == Ind || !TheLoop->contains(I) || Worklist.count(I) || | |||
4941 | isVectorizedMemAccessUse(I, IndUpdate); | |||
4942 | }); | |||
4943 | if (!UniformIndUpdate) | |||
4944 | continue; | |||
4945 | ||||
4946 | // The induction variable and its update instruction will remain uniform. | |||
4947 | addToWorklistIfAllowed(Ind); | |||
4948 | addToWorklistIfAllowed(IndUpdate); | |||
4949 | } | |||
4950 | ||||
4951 | Uniforms[VF].insert(Worklist.begin(), Worklist.end()); | |||
4952 | } | |||
4953 | ||||
4954 | bool LoopVectorizationCostModel::runtimeChecksRequired() { | |||
4955 | LLVM_DEBUG(dbgs() << "LV: Performing code size checks.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Performing code size checks.\n" ; } } while (false); | |||
4956 | ||||
4957 | if (Legal->getRuntimePointerChecking()->Need) { | |||
4958 | reportVectorizationFailure("Runtime ptr check is required with -Os/-Oz", | |||
4959 | "runtime pointer checks needed. Enable vectorization of this " | |||
4960 | "loop with '#pragma clang loop vectorize(enable)' when " | |||
4961 | "compiling with -Os/-Oz", | |||
4962 | "CantVersionLoopWithOptForSize", ORE, TheLoop); | |||
4963 | return true; | |||
4964 | } | |||
4965 | ||||
4966 | if (!PSE.getPredicate().isAlwaysTrue()) { | |||
4967 | reportVectorizationFailure("Runtime SCEV check is required with -Os/-Oz", | |||
4968 | "runtime SCEV checks needed. Enable vectorization of this " | |||
4969 | "loop with '#pragma clang loop vectorize(enable)' when " | |||
4970 | "compiling with -Os/-Oz", | |||
4971 | "CantVersionLoopWithOptForSize", ORE, TheLoop); | |||
4972 | return true; | |||
4973 | } | |||
4974 | ||||
4975 | // FIXME: Avoid specializing for stride==1 instead of bailing out. | |||
4976 | if (!Legal->getLAI()->getSymbolicStrides().empty()) { | |||
4977 | reportVectorizationFailure("Runtime stride check for small trip count", | |||
4978 | "runtime stride == 1 checks needed. Enable vectorization of " | |||
4979 | "this loop without such check by compiling with -Os/-Oz", | |||
4980 | "CantVersionLoopWithOptForSize", ORE, TheLoop); | |||
4981 | return true; | |||
4982 | } | |||
4983 | ||||
4984 | return false; | |||
4985 | } | |||
4986 | ||||
4987 | ElementCount | |||
4988 | LoopVectorizationCostModel::getMaxLegalScalableVF(unsigned MaxSafeElements) { | |||
4989 | if (!TTI.supportsScalableVectors() && !ForceTargetSupportsScalableVectors) | |||
4990 | return ElementCount::getScalable(0); | |||
4991 | ||||
4992 | if (Hints->isScalableVectorizationDisabled()) { | |||
4993 | reportVectorizationInfo("Scalable vectorization is explicitly disabled", | |||
4994 | "ScalableVectorizationDisabled", ORE, TheLoop); | |||
4995 | return ElementCount::getScalable(0); | |||
4996 | } | |||
4997 | ||||
4998 | LLVM_DEBUG(dbgs() << "LV: Scalable vectorization is available\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Scalable vectorization is available\n" ; } } while (false); | |||
4999 | ||||
5000 | auto MaxScalableVF = ElementCount::getScalable( | |||
5001 | std::numeric_limits<ElementCount::ScalarTy>::max()); | |||
5002 | ||||
5003 | // Test that the loop-vectorizer can legalize all operations for this MaxVF. | |||
5004 | // FIXME: While for scalable vectors this is currently sufficient, this should | |||
5005 | // be replaced by a more detailed mechanism that filters out specific VFs, | |||
5006 | // instead of invalidating vectorization for a whole set of VFs based on the | |||
5007 | // MaxVF. | |||
5008 | ||||
5009 | // Disable scalable vectorization if the loop contains unsupported reductions. | |||
5010 | if (!canVectorizeReductions(MaxScalableVF)) { | |||
5011 | reportVectorizationInfo( | |||
5012 | "Scalable vectorization not supported for the reduction " | |||
5013 | "operations found in this loop.", | |||
5014 | "ScalableVFUnfeasible", ORE, TheLoop); | |||
5015 | return ElementCount::getScalable(0); | |||
5016 | } | |||
5017 | ||||
5018 | // Disable scalable vectorization if the loop contains any instructions | |||
5019 | // with element types not supported for scalable vectors. | |||
5020 | if (any_of(ElementTypesInLoop, [&](Type *Ty) { | |||
5021 | return !Ty->isVoidTy() && | |||
5022 | !this->TTI.isElementTypeLegalForScalableVector(Ty); | |||
5023 | })) { | |||
5024 | reportVectorizationInfo("Scalable vectorization is not supported " | |||
5025 | "for all element types found in this loop.", | |||
5026 | "ScalableVFUnfeasible", ORE, TheLoop); | |||
5027 | return ElementCount::getScalable(0); | |||
5028 | } | |||
5029 | ||||
5030 | if (Legal->isSafeForAnyVectorWidth()) | |||
5031 | return MaxScalableVF; | |||
5032 | ||||
5033 | // Limit MaxScalableVF by the maximum safe dependence distance. | |||
5034 | Optional<unsigned> MaxVScale = TTI.getMaxVScale(); | |||
5035 | if (!MaxVScale && TheFunction->hasFnAttribute(Attribute::VScaleRange)) | |||
5036 | MaxVScale = | |||
5037 | TheFunction->getFnAttribute(Attribute::VScaleRange).getVScaleRangeMax(); | |||
5038 | MaxScalableVF = ElementCount::getScalable( | |||
5039 | MaxVScale ? (MaxSafeElements / MaxVScale.getValue()) : 0); | |||
5040 | if (!MaxScalableVF) | |||
5041 | reportVectorizationInfo( | |||
5042 | "Max legal vector width too small, scalable vectorization " | |||
5043 | "unfeasible.", | |||
5044 | "ScalableVFUnfeasible", ORE, TheLoop); | |||
5045 | ||||
5046 | return MaxScalableVF; | |||
5047 | } | |||
5048 | ||||
5049 | FixedScalableVFPair LoopVectorizationCostModel::computeFeasibleMaxVF( | |||
5050 | unsigned ConstTripCount, ElementCount UserVF, bool FoldTailByMasking) { | |||
5051 | MinBWs = computeMinimumValueSizes(TheLoop->getBlocks(), *DB, &TTI); | |||
5052 | unsigned SmallestType, WidestType; | |||
5053 | std::tie(SmallestType, WidestType) = getSmallestAndWidestTypes(); | |||
5054 | ||||
5055 | // Get the maximum safe dependence distance in bits computed by LAA. | |||
5056 | // It is computed by MaxVF * sizeOf(type) * 8, where type is taken from | |||
5057 | // the memory accesses that is most restrictive (involved in the smallest | |||
5058 | // dependence distance). | |||
5059 | unsigned MaxSafeElements = | |||
5060 | PowerOf2Floor(Legal->getMaxSafeVectorWidthInBits() / WidestType); | |||
5061 | ||||
5062 | auto MaxSafeFixedVF = ElementCount::getFixed(MaxSafeElements); | |||
5063 | auto MaxSafeScalableVF = getMaxLegalScalableVF(MaxSafeElements); | |||
5064 | ||||
5065 | LLVM_DEBUG(dbgs() << "LV: The max safe fixed VF is: " << MaxSafeFixedVFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The max safe fixed VF is: " << MaxSafeFixedVF << ".\n"; } } while (false) | |||
5066 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The max safe fixed VF is: " << MaxSafeFixedVF << ".\n"; } } while (false); | |||
5067 | LLVM_DEBUG(dbgs() << "LV: The max safe scalable VF is: " << MaxSafeScalableVFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The max safe scalable VF is: " << MaxSafeScalableVF << ".\n"; } } while (false) | |||
5068 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The max safe scalable VF is: " << MaxSafeScalableVF << ".\n"; } } while (false); | |||
5069 | ||||
5070 | // First analyze the UserVF, fall back if the UserVF should be ignored. | |||
5071 | if (UserVF) { | |||
5072 | auto MaxSafeUserVF = | |||
5073 | UserVF.isScalable() ? MaxSafeScalableVF : MaxSafeFixedVF; | |||
5074 | ||||
5075 | if (ElementCount::isKnownLE(UserVF, MaxSafeUserVF)) { | |||
5076 | // If `VF=vscale x N` is safe, then so is `VF=N` | |||
5077 | if (UserVF.isScalable()) | |||
5078 | return FixedScalableVFPair( | |||
5079 | ElementCount::getFixed(UserVF.getKnownMinValue()), UserVF); | |||
5080 | else | |||
5081 | return UserVF; | |||
5082 | } | |||
5083 | ||||
5084 | assert(ElementCount::isKnownGT(UserVF, MaxSafeUserVF))(static_cast <bool> (ElementCount::isKnownGT(UserVF, MaxSafeUserVF )) ? void (0) : __assert_fail ("ElementCount::isKnownGT(UserVF, MaxSafeUserVF)" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5084, __extension__ __PRETTY_FUNCTION__)); | |||
5085 | ||||
5086 | // Only clamp if the UserVF is not scalable. If the UserVF is scalable, it | |||
5087 | // is better to ignore the hint and let the compiler choose a suitable VF. | |||
5088 | if (!UserVF.isScalable()) { | |||
5089 | LLVM_DEBUG(dbgs() << "LV: User VF=" << UserVFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: User VF=" << UserVF << " is unsafe, clamping to max safe VF=" << MaxSafeFixedVF << ".\n"; } } while (false) | |||
5090 | << " is unsafe, clamping to max safe VF="do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: User VF=" << UserVF << " is unsafe, clamping to max safe VF=" << MaxSafeFixedVF << ".\n"; } } while (false) | |||
5091 | << MaxSafeFixedVF << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: User VF=" << UserVF << " is unsafe, clamping to max safe VF=" << MaxSafeFixedVF << ".\n"; } } while (false); | |||
5092 | ORE->emit([&]() { | |||
5093 | return OptimizationRemarkAnalysis(DEBUG_TYPE"loop-vectorize", "VectorizationFactor", | |||
5094 | TheLoop->getStartLoc(), | |||
5095 | TheLoop->getHeader()) | |||
5096 | << "User-specified vectorization factor " | |||
5097 | << ore::NV("UserVectorizationFactor", UserVF) | |||
5098 | << " is unsafe, clamping to maximum safe vectorization factor " | |||
5099 | << ore::NV("VectorizationFactor", MaxSafeFixedVF); | |||
5100 | }); | |||
5101 | return MaxSafeFixedVF; | |||
5102 | } | |||
5103 | ||||
5104 | if (!TTI.supportsScalableVectors() && !ForceTargetSupportsScalableVectors) { | |||
5105 | LLVM_DEBUG(dbgs() << "LV: User VF=" << UserVFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: User VF=" << UserVF << " is ignored because scalable vectors are not " "available.\n"; } } while (false) | |||
5106 | << " is ignored because scalable vectors are not "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: User VF=" << UserVF << " is ignored because scalable vectors are not " "available.\n"; } } while (false) | |||
5107 | "available.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: User VF=" << UserVF << " is ignored because scalable vectors are not " "available.\n"; } } while (false); | |||
5108 | ORE->emit([&]() { | |||
5109 | return OptimizationRemarkAnalysis(DEBUG_TYPE"loop-vectorize", "VectorizationFactor", | |||
5110 | TheLoop->getStartLoc(), | |||
5111 | TheLoop->getHeader()) | |||
5112 | << "User-specified vectorization factor " | |||
5113 | << ore::NV("UserVectorizationFactor", UserVF) | |||
5114 | << " is ignored because the target does not support scalable " | |||
5115 | "vectors. The compiler will pick a more suitable value."; | |||
5116 | }); | |||
5117 | } else { | |||
5118 | LLVM_DEBUG(dbgs() << "LV: User VF=" << UserVFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: User VF=" << UserVF << " is unsafe. Ignoring scalable UserVF.\n"; } } while (false) | |||
5119 | << " is unsafe. Ignoring scalable UserVF.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: User VF=" << UserVF << " is unsafe. Ignoring scalable UserVF.\n"; } } while (false); | |||
5120 | ORE->emit([&]() { | |||
5121 | return OptimizationRemarkAnalysis(DEBUG_TYPE"loop-vectorize", "VectorizationFactor", | |||
5122 | TheLoop->getStartLoc(), | |||
5123 | TheLoop->getHeader()) | |||
5124 | << "User-specified vectorization factor " | |||
5125 | << ore::NV("UserVectorizationFactor", UserVF) | |||
5126 | << " is unsafe. Ignoring the hint to let the compiler pick a " | |||
5127 | "more suitable value."; | |||
5128 | }); | |||
5129 | } | |||
5130 | } | |||
5131 | ||||
5132 | LLVM_DEBUG(dbgs() << "LV: The Smallest and Widest types: " << SmallestTypedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The Smallest and Widest types: " << SmallestType << " / " << WidestType << " bits.\n"; } } while (false) | |||
5133 | << " / " << WidestType << " bits.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The Smallest and Widest types: " << SmallestType << " / " << WidestType << " bits.\n"; } } while (false); | |||
5134 | ||||
5135 | FixedScalableVFPair Result(ElementCount::getFixed(1), | |||
5136 | ElementCount::getScalable(0)); | |||
5137 | if (auto MaxVF = | |||
5138 | getMaximizedVFForTarget(ConstTripCount, SmallestType, WidestType, | |||
5139 | MaxSafeFixedVF, FoldTailByMasking)) | |||
5140 | Result.FixedVF = MaxVF; | |||
5141 | ||||
5142 | if (auto MaxVF = | |||
5143 | getMaximizedVFForTarget(ConstTripCount, SmallestType, WidestType, | |||
5144 | MaxSafeScalableVF, FoldTailByMasking)) | |||
5145 | if (MaxVF.isScalable()) { | |||
5146 | Result.ScalableVF = MaxVF; | |||
5147 | LLVM_DEBUG(dbgs() << "LV: Found feasible scalable VF = " << MaxVFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found feasible scalable VF = " << MaxVF << "\n"; } } while (false) | |||
5148 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found feasible scalable VF = " << MaxVF << "\n"; } } while (false); | |||
5149 | } | |||
5150 | ||||
5151 | return Result; | |||
5152 | } | |||
5153 | ||||
5154 | FixedScalableVFPair | |||
5155 | LoopVectorizationCostModel::computeMaxVF(ElementCount UserVF, unsigned UserIC) { | |||
5156 | if (Legal->getRuntimePointerChecking()->Need && TTI.hasBranchDivergence()) { | |||
5157 | // TODO: It may by useful to do since it's still likely to be dynamically | |||
5158 | // uniform if the target can skip. | |||
5159 | reportVectorizationFailure( | |||
5160 | "Not inserting runtime ptr check for divergent target", | |||
5161 | "runtime pointer checks needed. Not enabled for divergent target", | |||
5162 | "CantVersionLoopWithDivergentTarget", ORE, TheLoop); | |||
5163 | return FixedScalableVFPair::getNone(); | |||
5164 | } | |||
5165 | ||||
5166 | unsigned TC = PSE.getSE()->getSmallConstantTripCount(TheLoop); | |||
5167 | LLVM_DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found trip count: " << TC << '\n'; } } while (false); | |||
5168 | if (TC == 1) { | |||
5169 | reportVectorizationFailure("Single iteration (non) loop", | |||
5170 | "loop trip count is one, irrelevant for vectorization", | |||
5171 | "SingleIterationLoop", ORE, TheLoop); | |||
5172 | return FixedScalableVFPair::getNone(); | |||
5173 | } | |||
5174 | ||||
5175 | switch (ScalarEpilogueStatus) { | |||
5176 | case CM_ScalarEpilogueAllowed: | |||
5177 | return computeFeasibleMaxVF(TC, UserVF, false); | |||
5178 | case CM_ScalarEpilogueNotAllowedUsePredicate: | |||
5179 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
5180 | case CM_ScalarEpilogueNotNeededUsePredicate: | |||
5181 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: vector predicate hint/switch found.\n" << "LV: Not allowing scalar epilogue, creating predicated " << "vector loop.\n"; } } while (false) | |||
5182 | dbgs() << "LV: vector predicate hint/switch found.\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: vector predicate hint/switch found.\n" << "LV: Not allowing scalar epilogue, creating predicated " << "vector loop.\n"; } } while (false) | |||
5183 | << "LV: Not allowing scalar epilogue, creating predicated "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: vector predicate hint/switch found.\n" << "LV: Not allowing scalar epilogue, creating predicated " << "vector loop.\n"; } } while (false) | |||
5184 | << "vector loop.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: vector predicate hint/switch found.\n" << "LV: Not allowing scalar epilogue, creating predicated " << "vector loop.\n"; } } while (false); | |||
5185 | break; | |||
5186 | case CM_ScalarEpilogueNotAllowedLowTripLoop: | |||
5187 | // fallthrough as a special case of OptForSize | |||
5188 | case CM_ScalarEpilogueNotAllowedOptSize: | |||
5189 | if (ScalarEpilogueStatus == CM_ScalarEpilogueNotAllowedOptSize) | |||
5190 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not allowing scalar epilogue due to -Os/-Oz.\n" ; } } while (false) | |||
5191 | dbgs() << "LV: Not allowing scalar epilogue due to -Os/-Oz.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not allowing scalar epilogue due to -Os/-Oz.\n" ; } } while (false); | |||
5192 | else | |||
5193 | LLVM_DEBUG(dbgs() << "LV: Not allowing scalar epilogue due to low trip "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not allowing scalar epilogue due to low trip " << "count.\n"; } } while (false) | |||
5194 | << "count.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not allowing scalar epilogue due to low trip " << "count.\n"; } } while (false); | |||
5195 | ||||
5196 | // Bail if runtime checks are required, which are not good when optimising | |||
5197 | // for size. | |||
5198 | if (runtimeChecksRequired()) | |||
5199 | return FixedScalableVFPair::getNone(); | |||
5200 | ||||
5201 | break; | |||
5202 | } | |||
5203 | ||||
5204 | // The only loops we can vectorize without a scalar epilogue, are loops with | |||
5205 | // a bottom-test and a single exiting block. We'd have to handle the fact | |||
5206 | // that not every instruction executes on the last iteration. This will | |||
5207 | // require a lane mask which varies through the vector loop body. (TODO) | |||
5208 | if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) { | |||
5209 | // If there was a tail-folding hint/switch, but we can't fold the tail by | |||
5210 | // masking, fallback to a vectorization with a scalar epilogue. | |||
5211 | if (ScalarEpilogueStatus == CM_ScalarEpilogueNotNeededUsePredicate) { | |||
5212 | LLVM_DEBUG(dbgs() << "LV: Cannot fold tail by masking: vectorize with a "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Cannot fold tail by masking: vectorize with a " "scalar epilogue instead.\n"; } } while (false) | |||
5213 | "scalar epilogue instead.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Cannot fold tail by masking: vectorize with a " "scalar epilogue instead.\n"; } } while (false); | |||
5214 | ScalarEpilogueStatus = CM_ScalarEpilogueAllowed; | |||
5215 | return computeFeasibleMaxVF(TC, UserVF, false); | |||
5216 | } | |||
5217 | return FixedScalableVFPair::getNone(); | |||
5218 | } | |||
5219 | ||||
5220 | // Now try the tail folding | |||
5221 | ||||
5222 | // Invalidate interleave groups that require an epilogue if we can't mask | |||
5223 | // the interleave-group. | |||
5224 | if (!useMaskedInterleavedAccesses(TTI)) { | |||
5225 | assert(WideningDecisions.empty() && Uniforms.empty() && Scalars.empty() &&(static_cast <bool> (WideningDecisions.empty() && Uniforms.empty() && Scalars.empty() && "No decisions should have been taken at this point" ) ? void (0) : __assert_fail ("WideningDecisions.empty() && Uniforms.empty() && Scalars.empty() && \"No decisions should have been taken at this point\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5226, __extension__ __PRETTY_FUNCTION__)) | |||
5226 | "No decisions should have been taken at this point")(static_cast <bool> (WideningDecisions.empty() && Uniforms.empty() && Scalars.empty() && "No decisions should have been taken at this point" ) ? void (0) : __assert_fail ("WideningDecisions.empty() && Uniforms.empty() && Scalars.empty() && \"No decisions should have been taken at this point\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5226, __extension__ __PRETTY_FUNCTION__)); | |||
5227 | // Note: There is no need to invalidate any cost modeling decisions here, as | |||
5228 | // non where taken so far. | |||
5229 | InterleaveInfo.invalidateGroupsRequiringScalarEpilogue(); | |||
5230 | } | |||
5231 | ||||
5232 | FixedScalableVFPair MaxFactors = computeFeasibleMaxVF(TC, UserVF, true); | |||
5233 | // Avoid tail folding if the trip count is known to be a multiple of any VF | |||
5234 | // we chose. | |||
5235 | // FIXME: The condition below pessimises the case for fixed-width vectors, | |||
5236 | // when scalable VFs are also candidates for vectorization. | |||
5237 | if (MaxFactors.FixedVF.isVector() && !MaxFactors.ScalableVF) { | |||
5238 | ElementCount MaxFixedVF = MaxFactors.FixedVF; | |||
5239 | assert((UserVF.isNonZero() || isPowerOf2_32(MaxFixedVF.getFixedValue())) &&(static_cast <bool> ((UserVF.isNonZero() || isPowerOf2_32 (MaxFixedVF.getFixedValue())) && "MaxFixedVF must be a power of 2" ) ? void (0) : __assert_fail ("(UserVF.isNonZero() || isPowerOf2_32(MaxFixedVF.getFixedValue())) && \"MaxFixedVF must be a power of 2\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5240, __extension__ __PRETTY_FUNCTION__)) | |||
5240 | "MaxFixedVF must be a power of 2")(static_cast <bool> ((UserVF.isNonZero() || isPowerOf2_32 (MaxFixedVF.getFixedValue())) && "MaxFixedVF must be a power of 2" ) ? void (0) : __assert_fail ("(UserVF.isNonZero() || isPowerOf2_32(MaxFixedVF.getFixedValue())) && \"MaxFixedVF must be a power of 2\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5240, __extension__ __PRETTY_FUNCTION__)); | |||
5241 | unsigned MaxVFtimesIC = UserIC ? MaxFixedVF.getFixedValue() * UserIC | |||
5242 | : MaxFixedVF.getFixedValue(); | |||
5243 | ScalarEvolution *SE = PSE.getSE(); | |||
5244 | const SCEV *BackedgeTakenCount = PSE.getBackedgeTakenCount(); | |||
5245 | const SCEV *ExitCount = SE->getAddExpr( | |||
5246 | BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType())); | |||
5247 | const SCEV *Rem = SE->getURemExpr( | |||
5248 | SE->applyLoopGuards(ExitCount, TheLoop), | |||
5249 | SE->getConstant(BackedgeTakenCount->getType(), MaxVFtimesIC)); | |||
5250 | if (Rem->isZero()) { | |||
5251 | // Accept MaxFixedVF if we do not have a tail. | |||
5252 | LLVM_DEBUG(dbgs() << "LV: No tail will remain for any chosen VF.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: No tail will remain for any chosen VF.\n" ; } } while (false); | |||
5253 | return MaxFactors; | |||
5254 | } | |||
5255 | } | |||
5256 | ||||
5257 | // For scalable vectors don't use tail folding for low trip counts or | |||
5258 | // optimizing for code size. We only permit this if the user has explicitly | |||
5259 | // requested it. | |||
5260 | if (ScalarEpilogueStatus != CM_ScalarEpilogueNotNeededUsePredicate && | |||
5261 | ScalarEpilogueStatus != CM_ScalarEpilogueNotAllowedUsePredicate && | |||
5262 | MaxFactors.ScalableVF.isVector()) | |||
5263 | MaxFactors.ScalableVF = ElementCount::getScalable(0); | |||
5264 | ||||
5265 | // If we don't know the precise trip count, or if the trip count that we | |||
5266 | // found modulo the vectorization factor is not zero, try to fold the tail | |||
5267 | // by masking. | |||
5268 | // FIXME: look for a smaller MaxVF that does divide TC rather than masking. | |||
5269 | if (Legal->prepareToFoldTailByMasking()) { | |||
5270 | FoldTailByMasking = true; | |||
5271 | return MaxFactors; | |||
5272 | } | |||
5273 | ||||
5274 | // If there was a tail-folding hint/switch, but we can't fold the tail by | |||
5275 | // masking, fallback to a vectorization with a scalar epilogue. | |||
5276 | if (ScalarEpilogueStatus == CM_ScalarEpilogueNotNeededUsePredicate) { | |||
5277 | LLVM_DEBUG(dbgs() << "LV: Cannot fold tail by masking: vectorize with a "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Cannot fold tail by masking: vectorize with a " "scalar epilogue instead.\n"; } } while (false) | |||
5278 | "scalar epilogue instead.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Cannot fold tail by masking: vectorize with a " "scalar epilogue instead.\n"; } } while (false); | |||
5279 | ScalarEpilogueStatus = CM_ScalarEpilogueAllowed; | |||
5280 | return MaxFactors; | |||
5281 | } | |||
5282 | ||||
5283 | if (ScalarEpilogueStatus == CM_ScalarEpilogueNotAllowedUsePredicate) { | |||
5284 | LLVM_DEBUG(dbgs() << "LV: Can't fold tail by masking: don't vectorize\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Can't fold tail by masking: don't vectorize\n" ; } } while (false); | |||
5285 | return FixedScalableVFPair::getNone(); | |||
5286 | } | |||
5287 | ||||
5288 | if (TC == 0) { | |||
5289 | reportVectorizationFailure( | |||
5290 | "Unable to calculate the loop count due to complex control flow", | |||
5291 | "unable to calculate the loop count due to complex control flow", | |||
5292 | "UnknownLoopCountComplexCFG", ORE, TheLoop); | |||
5293 | return FixedScalableVFPair::getNone(); | |||
5294 | } | |||
5295 | ||||
5296 | reportVectorizationFailure( | |||
5297 | "Cannot optimize for size and vectorize at the same time.", | |||
5298 | "cannot optimize for size and vectorize at the same time. " | |||
5299 | "Enable vectorization of this loop with '#pragma clang loop " | |||
5300 | "vectorize(enable)' when compiling with -Os/-Oz", | |||
5301 | "NoTailLoopWithOptForSize", ORE, TheLoop); | |||
5302 | return FixedScalableVFPair::getNone(); | |||
5303 | } | |||
5304 | ||||
5305 | ElementCount LoopVectorizationCostModel::getMaximizedVFForTarget( | |||
5306 | unsigned ConstTripCount, unsigned SmallestType, unsigned WidestType, | |||
5307 | const ElementCount &MaxSafeVF, bool FoldTailByMasking) { | |||
5308 | bool ComputeScalableMaxVF = MaxSafeVF.isScalable(); | |||
5309 | TypeSize WidestRegister = TTI.getRegisterBitWidth( | |||
5310 | ComputeScalableMaxVF ? TargetTransformInfo::RGK_ScalableVector | |||
5311 | : TargetTransformInfo::RGK_FixedWidthVector); | |||
5312 | ||||
5313 | // Convenience function to return the minimum of two ElementCounts. | |||
5314 | auto MinVF = [](const ElementCount &LHS, const ElementCount &RHS) { | |||
5315 | assert((LHS.isScalable() == RHS.isScalable()) &&(static_cast <bool> ((LHS.isScalable() == RHS.isScalable ()) && "Scalable flags must match") ? void (0) : __assert_fail ("(LHS.isScalable() == RHS.isScalable()) && \"Scalable flags must match\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5316, __extension__ __PRETTY_FUNCTION__)) | |||
5316 | "Scalable flags must match")(static_cast <bool> ((LHS.isScalable() == RHS.isScalable ()) && "Scalable flags must match") ? void (0) : __assert_fail ("(LHS.isScalable() == RHS.isScalable()) && \"Scalable flags must match\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5316, __extension__ __PRETTY_FUNCTION__)); | |||
5317 | return ElementCount::isKnownLT(LHS, RHS) ? LHS : RHS; | |||
5318 | }; | |||
5319 | ||||
5320 | // Ensure MaxVF is a power of 2; the dependence distance bound may not be. | |||
5321 | // Note that both WidestRegister and WidestType may not be a powers of 2. | |||
5322 | auto MaxVectorElementCount = ElementCount::get( | |||
5323 | PowerOf2Floor(WidestRegister.getKnownMinSize() / WidestType), | |||
5324 | ComputeScalableMaxVF); | |||
5325 | MaxVectorElementCount = MinVF(MaxVectorElementCount, MaxSafeVF); | |||
5326 | LLVM_DEBUG(dbgs() << "LV: The Widest register safe to use is: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The Widest register safe to use is: " << (MaxVectorElementCount * WidestType) << " bits.\n" ; } } while (false) | |||
5327 | << (MaxVectorElementCount * WidestType) << " bits.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The Widest register safe to use is: " << (MaxVectorElementCount * WidestType) << " bits.\n" ; } } while (false); | |||
5328 | ||||
5329 | if (!MaxVectorElementCount) { | |||
5330 | LLVM_DEBUG(dbgs() << "LV: The target has no "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The target has no " << (ComputeScalableMaxVF ? "scalable" : "fixed") << " vector registers.\n"; } } while (false) | |||
5331 | << (ComputeScalableMaxVF ? "scalable" : "fixed")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The target has no " << (ComputeScalableMaxVF ? "scalable" : "fixed") << " vector registers.\n"; } } while (false) | |||
5332 | << " vector registers.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The target has no " << (ComputeScalableMaxVF ? "scalable" : "fixed") << " vector registers.\n"; } } while (false); | |||
5333 | return ElementCount::getFixed(1); | |||
5334 | } | |||
5335 | ||||
5336 | const auto TripCountEC = ElementCount::getFixed(ConstTripCount); | |||
5337 | if (ConstTripCount && | |||
5338 | ElementCount::isKnownLE(TripCountEC, MaxVectorElementCount) && | |||
5339 | (!FoldTailByMasking || isPowerOf2_32(ConstTripCount))) { | |||
5340 | // If loop trip count (TC) is known at compile time there is no point in | |||
5341 | // choosing VF greater than TC (as done in the loop below). Select maximum | |||
5342 | // power of two which doesn't exceed TC. | |||
5343 | // If MaxVectorElementCount is scalable, we only fall back on a fixed VF | |||
5344 | // when the TC is less than or equal to the known number of lanes. | |||
5345 | auto ClampedConstTripCount = PowerOf2Floor(ConstTripCount); | |||
5346 | LLVM_DEBUG(dbgs() << "LV: Clamping the MaxVF to maximum power of two not "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Clamping the MaxVF to maximum power of two not " "exceeding the constant trip count: " << ClampedConstTripCount << "\n"; } } while (false) | |||
5347 | "exceeding the constant trip count: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Clamping the MaxVF to maximum power of two not " "exceeding the constant trip count: " << ClampedConstTripCount << "\n"; } } while (false) | |||
5348 | << ClampedConstTripCount << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Clamping the MaxVF to maximum power of two not " "exceeding the constant trip count: " << ClampedConstTripCount << "\n"; } } while (false); | |||
5349 | return ElementCount::getFixed(ClampedConstTripCount); | |||
5350 | } | |||
5351 | ||||
5352 | ElementCount MaxVF = MaxVectorElementCount; | |||
5353 | if (TTI.shouldMaximizeVectorBandwidth() || | |||
5354 | (MaximizeBandwidth && isScalarEpilogueAllowed())) { | |||
5355 | auto MaxVectorElementCountMaxBW = ElementCount::get( | |||
5356 | PowerOf2Floor(WidestRegister.getKnownMinSize() / SmallestType), | |||
5357 | ComputeScalableMaxVF); | |||
5358 | MaxVectorElementCountMaxBW = MinVF(MaxVectorElementCountMaxBW, MaxSafeVF); | |||
5359 | ||||
5360 | // Collect all viable vectorization factors larger than the default MaxVF | |||
5361 | // (i.e. MaxVectorElementCount). | |||
5362 | SmallVector<ElementCount, 8> VFs; | |||
5363 | for (ElementCount VS = MaxVectorElementCount * 2; | |||
5364 | ElementCount::isKnownLE(VS, MaxVectorElementCountMaxBW); VS *= 2) | |||
5365 | VFs.push_back(VS); | |||
5366 | ||||
5367 | // For each VF calculate its register usage. | |||
5368 | auto RUs = calculateRegisterUsage(VFs); | |||
5369 | ||||
5370 | // Select the largest VF which doesn't require more registers than existing | |||
5371 | // ones. | |||
5372 | for (int i = RUs.size() - 1; i >= 0; --i) { | |||
5373 | bool Selected = true; | |||
5374 | for (auto &pair : RUs[i].MaxLocalUsers) { | |||
5375 | unsigned TargetNumRegisters = TTI.getNumberOfRegisters(pair.first); | |||
5376 | if (pair.second > TargetNumRegisters) | |||
5377 | Selected = false; | |||
5378 | } | |||
5379 | if (Selected) { | |||
5380 | MaxVF = VFs[i]; | |||
5381 | break; | |||
5382 | } | |||
5383 | } | |||
5384 | if (ElementCount MinVF = | |||
5385 | TTI.getMinimumVF(SmallestType, ComputeScalableMaxVF)) { | |||
5386 | if (ElementCount::isKnownLT(MaxVF, MinVF)) { | |||
5387 | LLVM_DEBUG(dbgs() << "LV: Overriding calculated MaxVF(" << MaxVFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Overriding calculated MaxVF(" << MaxVF << ") with target's minimum: " << MinVF << '\n'; } } while (false) | |||
5388 | << ") with target's minimum: " << MinVF << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Overriding calculated MaxVF(" << MaxVF << ") with target's minimum: " << MinVF << '\n'; } } while (false); | |||
5389 | MaxVF = MinVF; | |||
5390 | } | |||
5391 | } | |||
5392 | } | |||
5393 | return MaxVF; | |||
5394 | } | |||
5395 | ||||
5396 | Optional<unsigned> LoopVectorizationCostModel::getVScaleForTuning() const { | |||
5397 | if (TheFunction->hasFnAttribute(Attribute::VScaleRange)) { | |||
5398 | auto Attr = TheFunction->getFnAttribute(Attribute::VScaleRange); | |||
5399 | auto Min = Attr.getVScaleRangeMin(); | |||
5400 | auto Max = Attr.getVScaleRangeMax(); | |||
5401 | if (Max && Min == Max) | |||
5402 | return Max; | |||
5403 | } | |||
5404 | ||||
5405 | return TTI.getVScaleForTuning(); | |||
5406 | } | |||
5407 | ||||
5408 | bool LoopVectorizationCostModel::isMoreProfitable( | |||
5409 | const VectorizationFactor &A, const VectorizationFactor &B) const { | |||
5410 | InstructionCost CostA = A.Cost; | |||
5411 | InstructionCost CostB = B.Cost; | |||
5412 | ||||
5413 | unsigned MaxTripCount = PSE.getSE()->getSmallConstantMaxTripCount(TheLoop); | |||
5414 | ||||
5415 | if (!A.Width.isScalable() && !B.Width.isScalable() && FoldTailByMasking && | |||
5416 | MaxTripCount) { | |||
5417 | // If we are folding the tail and the trip count is a known (possibly small) | |||
5418 | // constant, the trip count will be rounded up to an integer number of | |||
5419 | // iterations. The total cost will be PerIterationCost*ceil(TripCount/VF), | |||
5420 | // which we compare directly. When not folding the tail, the total cost will | |||
5421 | // be PerIterationCost*floor(TC/VF) + Scalar remainder cost, and so is | |||
5422 | // approximated with the per-lane cost below instead of using the tripcount | |||
5423 | // as here. | |||
5424 | auto RTCostA = CostA * divideCeil(MaxTripCount, A.Width.getFixedValue()); | |||
5425 | auto RTCostB = CostB * divideCeil(MaxTripCount, B.Width.getFixedValue()); | |||
5426 | return RTCostA < RTCostB; | |||
5427 | } | |||
5428 | ||||
5429 | // Improve estimate for the vector width if it is scalable. | |||
5430 | unsigned EstimatedWidthA = A.Width.getKnownMinValue(); | |||
5431 | unsigned EstimatedWidthB = B.Width.getKnownMinValue(); | |||
5432 | if (Optional<unsigned> VScale = getVScaleForTuning()) { | |||
5433 | if (A.Width.isScalable()) | |||
5434 | EstimatedWidthA *= VScale.getValue(); | |||
5435 | if (B.Width.isScalable()) | |||
5436 | EstimatedWidthB *= VScale.getValue(); | |||
5437 | } | |||
5438 | ||||
5439 | // Assume vscale may be larger than 1 (or the value being tuned for), | |||
5440 | // so that scalable vectorization is slightly favorable over fixed-width | |||
5441 | // vectorization. | |||
5442 | if (A.Width.isScalable() && !B.Width.isScalable()) | |||
5443 | return (CostA * B.Width.getFixedValue()) <= (CostB * EstimatedWidthA); | |||
5444 | ||||
5445 | // To avoid the need for FP division: | |||
5446 | // (CostA / A.Width) < (CostB / B.Width) | |||
5447 | // <=> (CostA * B.Width) < (CostB * A.Width) | |||
5448 | return (CostA * EstimatedWidthB) < (CostB * EstimatedWidthA); | |||
5449 | } | |||
5450 | ||||
5451 | VectorizationFactor LoopVectorizationCostModel::selectVectorizationFactor( | |||
5452 | const ElementCountSet &VFCandidates) { | |||
5453 | InstructionCost ExpectedCost = expectedCost(ElementCount::getFixed(1)).first; | |||
5454 | LLVM_DEBUG(dbgs() << "LV: Scalar loop costs: " << ExpectedCost << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Scalar loop costs: " << ExpectedCost << ".\n"; } } while (false); | |||
5455 | assert(ExpectedCost.isValid() && "Unexpected invalid cost for scalar loop")(static_cast <bool> (ExpectedCost.isValid() && "Unexpected invalid cost for scalar loop" ) ? void (0) : __assert_fail ("ExpectedCost.isValid() && \"Unexpected invalid cost for scalar loop\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5455, __extension__ __PRETTY_FUNCTION__)); | |||
5456 | assert(VFCandidates.count(ElementCount::getFixed(1)) &&(static_cast <bool> (VFCandidates.count(ElementCount::getFixed (1)) && "Expected Scalar VF to be a candidate") ? void (0) : __assert_fail ("VFCandidates.count(ElementCount::getFixed(1)) && \"Expected Scalar VF to be a candidate\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5457, __extension__ __PRETTY_FUNCTION__)) | |||
5457 | "Expected Scalar VF to be a candidate")(static_cast <bool> (VFCandidates.count(ElementCount::getFixed (1)) && "Expected Scalar VF to be a candidate") ? void (0) : __assert_fail ("VFCandidates.count(ElementCount::getFixed(1)) && \"Expected Scalar VF to be a candidate\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5457, __extension__ __PRETTY_FUNCTION__)); | |||
5458 | ||||
5459 | const VectorizationFactor ScalarCost(ElementCount::getFixed(1), ExpectedCost); | |||
5460 | VectorizationFactor ChosenFactor = ScalarCost; | |||
5461 | ||||
5462 | bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled; | |||
5463 | if (ForceVectorization && VFCandidates.size() > 1) { | |||
5464 | // Ignore scalar width, because the user explicitly wants vectorization. | |||
5465 | // Initialize cost to max so that VF = 2 is, at least, chosen during cost | |||
5466 | // evaluation. | |||
5467 | ChosenFactor.Cost = InstructionCost::getMax(); | |||
5468 | } | |||
5469 | ||||
5470 | SmallVector<InstructionVFPair> InvalidCosts; | |||
5471 | for (const auto &i : VFCandidates) { | |||
5472 | // The cost for scalar VF=1 is already calculated, so ignore it. | |||
5473 | if (i.isScalar()) | |||
5474 | continue; | |||
5475 | ||||
5476 | VectorizationCostTy C = expectedCost(i, &InvalidCosts); | |||
5477 | VectorizationFactor Candidate(i, C.first); | |||
5478 | ||||
5479 | #ifndef NDEBUG | |||
5480 | unsigned AssumedMinimumVscale = 1; | |||
5481 | if (Optional<unsigned> VScale = getVScaleForTuning()) | |||
5482 | AssumedMinimumVscale = VScale.getValue(); | |||
5483 | unsigned Width = | |||
5484 | Candidate.Width.isScalable() | |||
5485 | ? Candidate.Width.getKnownMinValue() * AssumedMinimumVscale | |||
5486 | : Candidate.Width.getFixedValue(); | |||
5487 | LLVM_DEBUG(dbgs() << "LV: Vector loop of width " << ido { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Vector loop of width " << i << " costs: " << (Candidate.Cost / Width ); } } while (false) | |||
5488 | << " costs: " << (Candidate.Cost / Width))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Vector loop of width " << i << " costs: " << (Candidate.Cost / Width ); } } while (false); | |||
5489 | if (i.isScalable()) | |||
5490 | LLVM_DEBUG(dbgs() << " (assuming a minimum vscale of "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << " (assuming a minimum vscale of " << AssumedMinimumVscale << ")"; } } while (false ) | |||
5491 | << AssumedMinimumVscale << ")")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << " (assuming a minimum vscale of " << AssumedMinimumVscale << ")"; } } while (false ); | |||
5492 | LLVM_DEBUG(dbgs() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << ".\n"; } } while (false ); | |||
5493 | #endif | |||
5494 | ||||
5495 | if (!C.second && !ForceVectorization) { | |||
5496 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not considering vector loop of width " << i << " because it will not generate any vector instructions.\n" ; } } while (false) | |||
5497 | dbgs() << "LV: Not considering vector loop of width " << ido { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not considering vector loop of width " << i << " because it will not generate any vector instructions.\n" ; } } while (false) | |||
5498 | << " because it will not generate any vector instructions.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not considering vector loop of width " << i << " because it will not generate any vector instructions.\n" ; } } while (false); | |||
5499 | continue; | |||
5500 | } | |||
5501 | ||||
5502 | // If profitable add it to ProfitableVF list. | |||
5503 | if (isMoreProfitable(Candidate, ScalarCost)) | |||
5504 | ProfitableVFs.push_back(Candidate); | |||
5505 | ||||
5506 | if (isMoreProfitable(Candidate, ChosenFactor)) | |||
5507 | ChosenFactor = Candidate; | |||
5508 | } | |||
5509 | ||||
5510 | // Emit a report of VFs with invalid costs in the loop. | |||
5511 | if (!InvalidCosts.empty()) { | |||
5512 | // Group the remarks per instruction, keeping the instruction order from | |||
5513 | // InvalidCosts. | |||
5514 | std::map<Instruction *, unsigned> Numbering; | |||
5515 | unsigned I = 0; | |||
5516 | for (auto &Pair : InvalidCosts) | |||
5517 | if (!Numbering.count(Pair.first)) | |||
5518 | Numbering[Pair.first] = I++; | |||
5519 | ||||
5520 | // Sort the list, first on instruction(number) then on VF. | |||
5521 | llvm::sort(InvalidCosts, | |||
5522 | [&Numbering](InstructionVFPair &A, InstructionVFPair &B) { | |||
5523 | if (Numbering[A.first] != Numbering[B.first]) | |||
5524 | return Numbering[A.first] < Numbering[B.first]; | |||
5525 | ElementCountComparator ECC; | |||
5526 | return ECC(A.second, B.second); | |||
5527 | }); | |||
5528 | ||||
5529 | // For a list of ordered instruction-vf pairs: | |||
5530 | // [(load, vf1), (load, vf2), (store, vf1)] | |||
5531 | // Group the instructions together to emit separate remarks for: | |||
5532 | // load (vf1, vf2) | |||
5533 | // store (vf1) | |||
5534 | auto Tail = ArrayRef<InstructionVFPair>(InvalidCosts); | |||
5535 | auto Subset = ArrayRef<InstructionVFPair>(); | |||
5536 | do { | |||
5537 | if (Subset.empty()) | |||
5538 | Subset = Tail.take_front(1); | |||
5539 | ||||
5540 | Instruction *I = Subset.front().first; | |||
5541 | ||||
5542 | // If the next instruction is different, or if there are no other pairs, | |||
5543 | // emit a remark for the collated subset. e.g. | |||
5544 | // [(load, vf1), (load, vf2))] | |||
5545 | // to emit: | |||
5546 | // remark: invalid costs for 'load' at VF=(vf, vf2) | |||
5547 | if (Subset == Tail || Tail[Subset.size()].first != I) { | |||
5548 | std::string OutString; | |||
5549 | raw_string_ostream OS(OutString); | |||
5550 | assert(!Subset.empty() && "Unexpected empty range")(static_cast <bool> (!Subset.empty() && "Unexpected empty range" ) ? void (0) : __assert_fail ("!Subset.empty() && \"Unexpected empty range\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5550, __extension__ __PRETTY_FUNCTION__)); | |||
5551 | OS << "Instruction with invalid costs prevented vectorization at VF=("; | |||
5552 | for (auto &Pair : Subset) | |||
5553 | OS << (Pair.second == Subset.front().second ? "" : ", ") | |||
5554 | << Pair.second; | |||
5555 | OS << "):"; | |||
5556 | if (auto *CI = dyn_cast<CallInst>(I)) | |||
5557 | OS << " call to " << CI->getCalledFunction()->getName(); | |||
5558 | else | |||
5559 | OS << " " << I->getOpcodeName(); | |||
5560 | OS.flush(); | |||
5561 | reportVectorizationInfo(OutString, "InvalidCost", ORE, TheLoop, I); | |||
5562 | Tail = Tail.drop_front(Subset.size()); | |||
5563 | Subset = {}; | |||
5564 | } else | |||
5565 | // Grow the subset by one element | |||
5566 | Subset = Tail.take_front(Subset.size() + 1); | |||
5567 | } while (!Tail.empty()); | |||
5568 | } | |||
5569 | ||||
5570 | if (!EnableCondStoresVectorization && NumPredStores) { | |||
5571 | reportVectorizationFailure("There are conditional stores.", | |||
5572 | "store that is conditionally executed prevents vectorization", | |||
5573 | "ConditionalStore", ORE, TheLoop); | |||
5574 | ChosenFactor = ScalarCost; | |||
5575 | } | |||
5576 | ||||
5577 | LLVM_DEBUG(if (ForceVectorization && !ChosenFactor.Width.isScalar() &&do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { if (ForceVectorization && !ChosenFactor .Width.isScalar() && ChosenFactor.Cost >= ScalarCost .Cost) dbgs() << "LV: Vectorization seems to be not beneficial, " << "but was forced by a user.\n"; } } while (false) | |||
5578 | ChosenFactor.Cost >= ScalarCost.Cost) dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { if (ForceVectorization && !ChosenFactor .Width.isScalar() && ChosenFactor.Cost >= ScalarCost .Cost) dbgs() << "LV: Vectorization seems to be not beneficial, " << "but was forced by a user.\n"; } } while (false) | |||
5579 | << "LV: Vectorization seems to be not beneficial, "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { if (ForceVectorization && !ChosenFactor .Width.isScalar() && ChosenFactor.Cost >= ScalarCost .Cost) dbgs() << "LV: Vectorization seems to be not beneficial, " << "but was forced by a user.\n"; } } while (false) | |||
5580 | << "but was forced by a user.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { if (ForceVectorization && !ChosenFactor .Width.isScalar() && ChosenFactor.Cost >= ScalarCost .Cost) dbgs() << "LV: Vectorization seems to be not beneficial, " << "but was forced by a user.\n"; } } while (false); | |||
5581 | LLVM_DEBUG(dbgs() << "LV: Selecting VF: " << ChosenFactor.Width << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Selecting VF: " << ChosenFactor.Width << ".\n"; } } while (false); | |||
5582 | return ChosenFactor; | |||
5583 | } | |||
5584 | ||||
5585 | bool LoopVectorizationCostModel::isCandidateForEpilogueVectorization( | |||
5586 | const Loop &L, ElementCount VF) const { | |||
5587 | // Cross iteration phis such as reductions need special handling and are | |||
5588 | // currently unsupported. | |||
5589 | if (any_of(L.getHeader()->phis(), | |||
5590 | [&](PHINode &Phi) { return Legal->isFirstOrderRecurrence(&Phi); })) | |||
5591 | return false; | |||
5592 | ||||
5593 | // Phis with uses outside of the loop require special handling and are | |||
5594 | // currently unsupported. | |||
5595 | for (auto &Entry : Legal->getInductionVars()) { | |||
5596 | // Look for uses of the value of the induction at the last iteration. | |||
5597 | Value *PostInc = Entry.first->getIncomingValueForBlock(L.getLoopLatch()); | |||
5598 | for (User *U : PostInc->users()) | |||
5599 | if (!L.contains(cast<Instruction>(U))) | |||
5600 | return false; | |||
5601 | // Look for uses of penultimate value of the induction. | |||
5602 | for (User *U : Entry.first->users()) | |||
5603 | if (!L.contains(cast<Instruction>(U))) | |||
5604 | return false; | |||
5605 | } | |||
5606 | ||||
5607 | // Induction variables that are widened require special handling that is | |||
5608 | // currently not supported. | |||
5609 | if (any_of(Legal->getInductionVars(), [&](auto &Entry) { | |||
5610 | return !(this->isScalarAfterVectorization(Entry.first, VF) || | |||
5611 | this->isProfitableToScalarize(Entry.first, VF)); | |||
5612 | })) | |||
5613 | return false; | |||
5614 | ||||
5615 | // Epilogue vectorization code has not been auditted to ensure it handles | |||
5616 | // non-latch exits properly. It may be fine, but it needs auditted and | |||
5617 | // tested. | |||
5618 | if (L.getExitingBlock() != L.getLoopLatch()) | |||
5619 | return false; | |||
5620 | ||||
5621 | return true; | |||
5622 | } | |||
5623 | ||||
5624 | bool LoopVectorizationCostModel::isEpilogueVectorizationProfitable( | |||
5625 | const ElementCount VF) const { | |||
5626 | // FIXME: We need a much better cost-model to take different parameters such | |||
5627 | // as register pressure, code size increase and cost of extra branches into | |||
5628 | // account. For now we apply a very crude heuristic and only consider loops | |||
5629 | // with vectorization factors larger than a certain value. | |||
5630 | // We also consider epilogue vectorization unprofitable for targets that don't | |||
5631 | // consider interleaving beneficial (eg. MVE). | |||
5632 | if (TTI.getMaxInterleaveFactor(VF.getKnownMinValue()) <= 1) | |||
5633 | return false; | |||
5634 | // FIXME: We should consider changing the threshold for scalable | |||
5635 | // vectors to take VScaleForTuning into account. | |||
5636 | if (VF.getKnownMinValue() >= EpilogueVectorizationMinVF) | |||
5637 | return true; | |||
5638 | return false; | |||
5639 | } | |||
5640 | ||||
5641 | VectorizationFactor | |||
5642 | LoopVectorizationCostModel::selectEpilogueVectorizationFactor( | |||
5643 | const ElementCount MainLoopVF, const LoopVectorizationPlanner &LVP) { | |||
5644 | VectorizationFactor Result = VectorizationFactor::Disabled(); | |||
5645 | if (!EnableEpilogueVectorization) { | |||
5646 | LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization is disabled.\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization is disabled.\n" ;; } } while (false); | |||
5647 | return Result; | |||
5648 | } | |||
5649 | ||||
5650 | if (!isScalarEpilogueAllowed()) { | |||
5651 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Unable to vectorize epilogue because no epilogue is " "allowed.\n";; } } while (false) | |||
5652 | dbgs() << "LEV: Unable to vectorize epilogue because no epilogue is "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Unable to vectorize epilogue because no epilogue is " "allowed.\n";; } } while (false) | |||
5653 | "allowed.\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Unable to vectorize epilogue because no epilogue is " "allowed.\n";; } } while (false); | |||
5654 | return Result; | |||
5655 | } | |||
5656 | ||||
5657 | // Not really a cost consideration, but check for unsupported cases here to | |||
5658 | // simplify the logic. | |||
5659 | if (!isCandidateForEpilogueVectorization(*TheLoop, MainLoopVF)) { | |||
5660 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Unable to vectorize epilogue because the loop is " "not a supported candidate.\n";; } } while (false) | |||
5661 | dbgs() << "LEV: Unable to vectorize epilogue because the loop is "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Unable to vectorize epilogue because the loop is " "not a supported candidate.\n";; } } while (false) | |||
5662 | "not a supported candidate.\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Unable to vectorize epilogue because the loop is " "not a supported candidate.\n";; } } while (false); | |||
5663 | return Result; | |||
5664 | } | |||
5665 | ||||
5666 | if (EpilogueVectorizationForceVF > 1) { | |||
5667 | LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization factor is forced.\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization factor is forced.\n" ;; } } while (false); | |||
5668 | ElementCount ForcedEC = ElementCount::getFixed(EpilogueVectorizationForceVF); | |||
5669 | if (LVP.hasPlanWithVF(ForcedEC)) | |||
5670 | return {ForcedEC, 0}; | |||
5671 | else { | |||
5672 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization forced factor is not viable.\n" ;; } } while (false) | |||
5673 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization forced factor is not viable.\n" ;; } } while (false) | |||
5674 | << "LEV: Epilogue vectorization forced factor is not viable.\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization forced factor is not viable.\n" ;; } } while (false); | |||
5675 | return Result; | |||
5676 | } | |||
5677 | } | |||
5678 | ||||
5679 | if (TheLoop->getHeader()->getParent()->hasOptSize() || | |||
5680 | TheLoop->getHeader()->getParent()->hasMinSize()) { | |||
5681 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization skipped due to opt for size.\n" ;; } } while (false) | |||
5682 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization skipped due to opt for size.\n" ;; } } while (false) | |||
5683 | << "LEV: Epilogue vectorization skipped due to opt for size.\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization skipped due to opt for size.\n" ;; } } while (false); | |||
5684 | return Result; | |||
5685 | } | |||
5686 | ||||
5687 | if (!isEpilogueVectorizationProfitable(MainLoopVF)) { | |||
5688 | LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization is not profitable for "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization is not profitable for " "this loop\n"; } } while (false) | |||
5689 | "this loop\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization is not profitable for " "this loop\n"; } } while (false); | |||
5690 | return Result; | |||
5691 | } | |||
5692 | ||||
5693 | // If MainLoopVF = vscale x 2, and vscale is expected to be 4, then we know | |||
5694 | // the main loop handles 8 lanes per iteration. We could still benefit from | |||
5695 | // vectorizing the epilogue loop with VF=4. | |||
5696 | ElementCount EstimatedRuntimeVF = MainLoopVF; | |||
5697 | if (MainLoopVF.isScalable()) { | |||
5698 | EstimatedRuntimeVF = ElementCount::getFixed(MainLoopVF.getKnownMinValue()); | |||
5699 | if (Optional<unsigned> VScale = getVScaleForTuning()) | |||
5700 | EstimatedRuntimeVF *= VScale.getValue(); | |||
5701 | } | |||
5702 | ||||
5703 | for (auto &NextVF : ProfitableVFs) | |||
5704 | if (((!NextVF.Width.isScalable() && MainLoopVF.isScalable() && | |||
5705 | ElementCount::isKnownLT(NextVF.Width, EstimatedRuntimeVF)) || | |||
5706 | ElementCount::isKnownLT(NextVF.Width, MainLoopVF)) && | |||
5707 | (Result.Width.isScalar() || isMoreProfitable(NextVF, Result)) && | |||
5708 | LVP.hasPlanWithVF(NextVF.Width)) | |||
5709 | Result = NextVF; | |||
5710 | ||||
5711 | if (Result != VectorizationFactor::Disabled()) | |||
5712 | LLVM_DEBUG(dbgs() << "LEV: Vectorizing epilogue loop with VF = "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Vectorizing epilogue loop with VF = " << Result.Width << "\n";; } } while (false) | |||
5713 | << Result.Width << "\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Vectorizing epilogue loop with VF = " << Result.Width << "\n";; } } while (false); | |||
5714 | return Result; | |||
5715 | } | |||
5716 | ||||
5717 | std::pair<unsigned, unsigned> | |||
5718 | LoopVectorizationCostModel::getSmallestAndWidestTypes() { | |||
5719 | unsigned MinWidth = -1U; | |||
5720 | unsigned MaxWidth = 8; | |||
5721 | const DataLayout &DL = TheFunction->getParent()->getDataLayout(); | |||
5722 | // For in-loop reductions, no element types are added to ElementTypesInLoop | |||
5723 | // if there are no loads/stores in the loop. In this case, check through the | |||
5724 | // reduction variables to determine the maximum width. | |||
5725 | if (ElementTypesInLoop.empty() && !Legal->getReductionVars().empty()) { | |||
5726 | // Reset MaxWidth so that we can find the smallest type used by recurrences | |||
5727 | // in the loop. | |||
5728 | MaxWidth = -1U; | |||
5729 | for (auto &PhiDescriptorPair : Legal->getReductionVars()) { | |||
5730 | const RecurrenceDescriptor &RdxDesc = PhiDescriptorPair.second; | |||
5731 | // When finding the min width used by the recurrence we need to account | |||
5732 | // for casts on the input operands of the recurrence. | |||
5733 | MaxWidth = std::min<unsigned>( | |||
5734 | MaxWidth, std::min<unsigned>( | |||
5735 | RdxDesc.getMinWidthCastToRecurrenceTypeInBits(), | |||
5736 | RdxDesc.getRecurrenceType()->getScalarSizeInBits())); | |||
5737 | } | |||
5738 | } else { | |||
5739 | for (Type *T : ElementTypesInLoop) { | |||
5740 | MinWidth = std::min<unsigned>( | |||
5741 | MinWidth, DL.getTypeSizeInBits(T->getScalarType()).getFixedSize()); | |||
5742 | MaxWidth = std::max<unsigned>( | |||
5743 | MaxWidth, DL.getTypeSizeInBits(T->getScalarType()).getFixedSize()); | |||
5744 | } | |||
5745 | } | |||
5746 | return {MinWidth, MaxWidth}; | |||
5747 | } | |||
5748 | ||||
5749 | void LoopVectorizationCostModel::collectElementTypesForWidening() { | |||
5750 | ElementTypesInLoop.clear(); | |||
5751 | // For each block. | |||
5752 | for (BasicBlock *BB : TheLoop->blocks()) { | |||
5753 | // For each instruction in the loop. | |||
5754 | for (Instruction &I : BB->instructionsWithoutDebug()) { | |||
5755 | Type *T = I.getType(); | |||
5756 | ||||
5757 | // Skip ignored values. | |||
5758 | if (ValuesToIgnore.count(&I)) | |||
5759 | continue; | |||
5760 | ||||
5761 | // Only examine Loads, Stores and PHINodes. | |||
5762 | if (!isa<LoadInst>(I) && !isa<StoreInst>(I) && !isa<PHINode>(I)) | |||
5763 | continue; | |||
5764 | ||||
5765 | // Examine PHI nodes that are reduction variables. Update the type to | |||
5766 | // account for the recurrence type. | |||
5767 | if (auto *PN = dyn_cast<PHINode>(&I)) { | |||
5768 | if (!Legal->isReductionVariable(PN)) | |||
5769 | continue; | |||
5770 | const RecurrenceDescriptor &RdxDesc = | |||
5771 | Legal->getReductionVars().find(PN)->second; | |||
5772 | if (PreferInLoopReductions || useOrderedReductions(RdxDesc) || | |||
5773 | TTI.preferInLoopReduction(RdxDesc.getOpcode(), | |||
5774 | RdxDesc.getRecurrenceType(), | |||
5775 | TargetTransformInfo::ReductionFlags())) | |||
5776 | continue; | |||
5777 | T = RdxDesc.getRecurrenceType(); | |||
5778 | } | |||
5779 | ||||
5780 | // Examine the stored values. | |||
5781 | if (auto *ST = dyn_cast<StoreInst>(&I)) | |||
5782 | T = ST->getValueOperand()->getType(); | |||
5783 | ||||
5784 | assert(T->isSized() &&(static_cast <bool> (T->isSized() && "Expected the load/store/recurrence type to be sized" ) ? void (0) : __assert_fail ("T->isSized() && \"Expected the load/store/recurrence type to be sized\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5785, __extension__ __PRETTY_FUNCTION__)) | |||
5785 | "Expected the load/store/recurrence type to be sized")(static_cast <bool> (T->isSized() && "Expected the load/store/recurrence type to be sized" ) ? void (0) : __assert_fail ("T->isSized() && \"Expected the load/store/recurrence type to be sized\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5785, __extension__ __PRETTY_FUNCTION__)); | |||
5786 | ||||
5787 | ElementTypesInLoop.insert(T); | |||
5788 | } | |||
5789 | } | |||
5790 | } | |||
5791 | ||||
5792 | unsigned LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF, | |||
5793 | unsigned LoopCost) { | |||
5794 | // -- The interleave heuristics -- | |||
5795 | // We interleave the loop in order to expose ILP and reduce the loop overhead. | |||
5796 | // There are many micro-architectural considerations that we can't predict | |||
5797 | // at this level. For example, frontend pressure (on decode or fetch) due to | |||
5798 | // code size, or the number and capabilities of the execution ports. | |||
5799 | // | |||
5800 | // We use the following heuristics to select the interleave count: | |||
5801 | // 1. If the code has reductions, then we interleave to break the cross | |||
5802 | // iteration dependency. | |||
5803 | // 2. If the loop is really small, then we interleave to reduce the loop | |||
5804 | // overhead. | |||
5805 | // 3. We don't interleave if we think that we will spill registers to memory | |||
5806 | // due to the increased register pressure. | |||
5807 | ||||
5808 | if (!isScalarEpilogueAllowed()) | |||
5809 | return 1; | |||
5810 | ||||
5811 | // We used the distance for the interleave count. | |||
5812 | if (Legal->getMaxSafeDepDistBytes() != -1U) | |||
5813 | return 1; | |||
5814 | ||||
5815 | auto BestKnownTC = getSmallBestKnownTC(*PSE.getSE(), TheLoop); | |||
5816 | const bool HasReductions = !Legal->getReductionVars().empty(); | |||
5817 | // Do not interleave loops with a relatively small known or estimated trip | |||
5818 | // count. But we will interleave when InterleaveSmallLoopScalarReduction is | |||
5819 | // enabled, and the code has scalar reductions(HasReductions && VF = 1), | |||
5820 | // because with the above conditions interleaving can expose ILP and break | |||
5821 | // cross iteration dependences for reductions. | |||
5822 | if (BestKnownTC && (*BestKnownTC < TinyTripCountInterleaveThreshold) && | |||
5823 | !(InterleaveSmallLoopScalarReduction && HasReductions && VF.isScalar())) | |||
5824 | return 1; | |||
5825 | ||||
5826 | RegisterUsage R = calculateRegisterUsage({VF})[0]; | |||
5827 | // We divide by these constants so assume that we have at least one | |||
5828 | // instruction that uses at least one register. | |||
5829 | for (auto& pair : R.MaxLocalUsers) { | |||
5830 | pair.second = std::max(pair.second, 1U); | |||
5831 | } | |||
5832 | ||||
5833 | // We calculate the interleave count using the following formula. | |||
5834 | // Subtract the number of loop invariants from the number of available | |||
5835 | // registers. These registers are used by all of the interleaved instances. | |||
5836 | // Next, divide the remaining registers by the number of registers that is | |||
5837 | // required by the loop, in order to estimate how many parallel instances | |||
5838 | // fit without causing spills. All of this is rounded down if necessary to be | |||
5839 | // a power of two. We want power of two interleave count to simplify any | |||
5840 | // addressing operations or alignment considerations. | |||
5841 | // We also want power of two interleave counts to ensure that the induction | |||
5842 | // variable of the vector loop wraps to zero, when tail is folded by masking; | |||
5843 | // this currently happens when OptForSize, in which case IC is set to 1 above. | |||
5844 | unsigned IC = UINT_MAX(2147483647 *2U +1U); | |||
5845 | ||||
5846 | for (auto& pair : R.MaxLocalUsers) { | |||
5847 | unsigned TargetNumRegisters = TTI.getNumberOfRegisters(pair.first); | |||
5848 | LLVM_DEBUG(dbgs() << "LV: The target has " << TargetNumRegistersdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The target has " << TargetNumRegisters << " registers of " << TTI.getRegisterClassName (pair.first) << " register class\n"; } } while (false) | |||
5849 | << " registers of "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The target has " << TargetNumRegisters << " registers of " << TTI.getRegisterClassName (pair.first) << " register class\n"; } } while (false) | |||
5850 | << TTI.getRegisterClassName(pair.first) << " register class\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The target has " << TargetNumRegisters << " registers of " << TTI.getRegisterClassName (pair.first) << " register class\n"; } } while (false); | |||
5851 | if (VF.isScalar()) { | |||
5852 | if (ForceTargetNumScalarRegs.getNumOccurrences() > 0) | |||
5853 | TargetNumRegisters = ForceTargetNumScalarRegs; | |||
5854 | } else { | |||
5855 | if (ForceTargetNumVectorRegs.getNumOccurrences() > 0) | |||
5856 | TargetNumRegisters = ForceTargetNumVectorRegs; | |||
5857 | } | |||
5858 | unsigned MaxLocalUsers = pair.second; | |||
5859 | unsigned LoopInvariantRegs = 0; | |||
5860 | if (R.LoopInvariantRegs.find(pair.first) != R.LoopInvariantRegs.end()) | |||
5861 | LoopInvariantRegs = R.LoopInvariantRegs[pair.first]; | |||
5862 | ||||
5863 | unsigned TmpIC = PowerOf2Floor((TargetNumRegisters - LoopInvariantRegs) / MaxLocalUsers); | |||
5864 | // Don't count the induction variable as interleaved. | |||
5865 | if (EnableIndVarRegisterHeur) { | |||
5866 | TmpIC = | |||
5867 | PowerOf2Floor((TargetNumRegisters - LoopInvariantRegs - 1) / | |||
5868 | std::max(1U, (MaxLocalUsers - 1))); | |||
5869 | } | |||
5870 | ||||
5871 | IC = std::min(IC, TmpIC); | |||
5872 | } | |||
5873 | ||||
5874 | // Clamp the interleave ranges to reasonable counts. | |||
5875 | unsigned MaxInterleaveCount = | |||
5876 | TTI.getMaxInterleaveFactor(VF.getKnownMinValue()); | |||
5877 | ||||
5878 | // Check if the user has overridden the max. | |||
5879 | if (VF.isScalar()) { | |||
5880 | if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0) | |||
5881 | MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor; | |||
5882 | } else { | |||
5883 | if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0) | |||
5884 | MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor; | |||
5885 | } | |||
5886 | ||||
5887 | // If trip count is known or estimated compile time constant, limit the | |||
5888 | // interleave count to be less than the trip count divided by VF, provided it | |||
5889 | // is at least 1. | |||
5890 | // | |||
5891 | // For scalable vectors we can't know if interleaving is beneficial. It may | |||
5892 | // not be beneficial for small loops if none of the lanes in the second vector | |||
5893 | // iterations is enabled. However, for larger loops, there is likely to be a | |||
5894 | // similar benefit as for fixed-width vectors. For now, we choose to leave | |||
5895 | // the InterleaveCount as if vscale is '1', although if some information about | |||
5896 | // the vector is known (e.g. min vector size), we can make a better decision. | |||
5897 | if (BestKnownTC) { | |||
5898 | MaxInterleaveCount = | |||
5899 | std::min(*BestKnownTC / VF.getKnownMinValue(), MaxInterleaveCount); | |||
5900 | // Make sure MaxInterleaveCount is greater than 0. | |||
5901 | MaxInterleaveCount = std::max(1u, MaxInterleaveCount); | |||
5902 | } | |||
5903 | ||||
5904 | assert(MaxInterleaveCount > 0 &&(static_cast <bool> (MaxInterleaveCount > 0 && "Maximum interleave count must be greater than 0") ? void (0 ) : __assert_fail ("MaxInterleaveCount > 0 && \"Maximum interleave count must be greater than 0\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5905, __extension__ __PRETTY_FUNCTION__)) | |||
5905 | "Maximum interleave count must be greater than 0")(static_cast <bool> (MaxInterleaveCount > 0 && "Maximum interleave count must be greater than 0") ? void (0 ) : __assert_fail ("MaxInterleaveCount > 0 && \"Maximum interleave count must be greater than 0\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5905, __extension__ __PRETTY_FUNCTION__)); | |||
5906 | ||||
5907 | // Clamp the calculated IC to be between the 1 and the max interleave count | |||
5908 | // that the target and trip count allows. | |||
5909 | if (IC > MaxInterleaveCount) | |||
5910 | IC = MaxInterleaveCount; | |||
5911 | else | |||
5912 | // Make sure IC is greater than 0. | |||
5913 | IC = std::max(1u, IC); | |||
5914 | ||||
5915 | assert(IC > 0 && "Interleave count must be greater than 0.")(static_cast <bool> (IC > 0 && "Interleave count must be greater than 0." ) ? void (0) : __assert_fail ("IC > 0 && \"Interleave count must be greater than 0.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5915, __extension__ __PRETTY_FUNCTION__)); | |||
5916 | ||||
5917 | // If we did not calculate the cost for VF (because the user selected the VF) | |||
5918 | // then we calculate the cost of VF here. | |||
5919 | if (LoopCost == 0) { | |||
5920 | InstructionCost C = expectedCost(VF).first; | |||
5921 | assert(C.isValid() && "Expected to have chosen a VF with valid cost")(static_cast <bool> (C.isValid() && "Expected to have chosen a VF with valid cost" ) ? void (0) : __assert_fail ("C.isValid() && \"Expected to have chosen a VF with valid cost\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5921, __extension__ __PRETTY_FUNCTION__)); | |||
5922 | LoopCost = *C.getValue(); | |||
5923 | } | |||
5924 | ||||
5925 | assert(LoopCost && "Non-zero loop cost expected")(static_cast <bool> (LoopCost && "Non-zero loop cost expected" ) ? void (0) : __assert_fail ("LoopCost && \"Non-zero loop cost expected\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5925, __extension__ __PRETTY_FUNCTION__)); | |||
5926 | ||||
5927 | // Interleave if we vectorized this loop and there is a reduction that could | |||
5928 | // benefit from interleaving. | |||
5929 | if (VF.isVector() && HasReductions) { | |||
5930 | LLVM_DEBUG(dbgs() << "LV: Interleaving because of reductions.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving because of reductions.\n" ; } } while (false); | |||
5931 | return IC; | |||
5932 | } | |||
5933 | ||||
5934 | // For any scalar loop that either requires runtime checks or predication we | |||
5935 | // are better off leaving this to the unroller. Note that if we've already | |||
5936 | // vectorized the loop we will have done the runtime check and so interleaving | |||
5937 | // won't require further checks. | |||
5938 | bool ScalarInterleavingRequiresPredication = | |||
5939 | (VF.isScalar() && any_of(TheLoop->blocks(), [this](BasicBlock *BB) { | |||
5940 | return Legal->blockNeedsPredication(BB); | |||
5941 | })); | |||
5942 | bool ScalarInterleavingRequiresRuntimePointerCheck = | |||
5943 | (VF.isScalar() && Legal->getRuntimePointerChecking()->Need); | |||
5944 | ||||
5945 | // We want to interleave small loops in order to reduce the loop overhead and | |||
5946 | // potentially expose ILP opportunities. | |||
5947 | LLVM_DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop cost is " << LoopCost << '\n' << "LV: IC is " << IC << '\n' << "LV: VF is " << VF << '\n'; } } while (false) | |||
5948 | << "LV: IC is " << IC << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop cost is " << LoopCost << '\n' << "LV: IC is " << IC << '\n' << "LV: VF is " << VF << '\n'; } } while (false) | |||
5949 | << "LV: VF is " << VF << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop cost is " << LoopCost << '\n' << "LV: IC is " << IC << '\n' << "LV: VF is " << VF << '\n'; } } while (false); | |||
5950 | const bool AggressivelyInterleaveReductions = | |||
5951 | TTI.enableAggressiveInterleaving(HasReductions); | |||
5952 | if (!ScalarInterleavingRequiresRuntimePointerCheck && | |||
5953 | !ScalarInterleavingRequiresPredication && LoopCost < SmallLoopCost) { | |||
5954 | // We assume that the cost overhead is 1 and we use the cost model | |||
5955 | // to estimate the cost of the loop and interleave until the cost of the | |||
5956 | // loop overhead is about 5% of the cost of the loop. | |||
5957 | unsigned SmallIC = | |||
5958 | std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost)); | |||
5959 | ||||
5960 | // Interleave until store/load ports (estimated by max interleave count) are | |||
5961 | // saturated. | |||
5962 | unsigned NumStores = Legal->getNumStores(); | |||
5963 | unsigned NumLoads = Legal->getNumLoads(); | |||
5964 | unsigned StoresIC = IC / (NumStores ? NumStores : 1); | |||
5965 | unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1); | |||
5966 | ||||
5967 | // There is little point in interleaving for reductions containing selects | |||
5968 | // and compares when VF=1 since it may just create more overhead than it's | |||
5969 | // worth for loops with small trip counts. This is because we still have to | |||
5970 | // do the final reduction after the loop. | |||
5971 | bool HasSelectCmpReductions = | |||
5972 | HasReductions && | |||
5973 | any_of(Legal->getReductionVars(), [&](auto &Reduction) -> bool { | |||
5974 | const RecurrenceDescriptor &RdxDesc = Reduction.second; | |||
5975 | return RecurrenceDescriptor::isSelectCmpRecurrenceKind( | |||
5976 | RdxDesc.getRecurrenceKind()); | |||
5977 | }); | |||
5978 | if (HasSelectCmpReductions) { | |||
5979 | LLVM_DEBUG(dbgs() << "LV: Not interleaving select-cmp reductions.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not interleaving select-cmp reductions.\n" ; } } while (false); | |||
5980 | return 1; | |||
5981 | } | |||
5982 | ||||
5983 | // If we have a scalar reduction (vector reductions are already dealt with | |||
5984 | // by this point), we can increase the critical path length if the loop | |||
5985 | // we're interleaving is inside another loop. For tree-wise reductions | |||
5986 | // set the limit to 2, and for ordered reductions it's best to disable | |||
5987 | // interleaving entirely. | |||
5988 | if (HasReductions && TheLoop->getLoopDepth() > 1) { | |||
5989 | bool HasOrderedReductions = | |||
5990 | any_of(Legal->getReductionVars(), [&](auto &Reduction) -> bool { | |||
5991 | const RecurrenceDescriptor &RdxDesc = Reduction.second; | |||
5992 | return RdxDesc.isOrdered(); | |||
5993 | }); | |||
5994 | if (HasOrderedReductions) { | |||
5995 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not interleaving scalar ordered reductions.\n" ; } } while (false) | |||
5996 | dbgs() << "LV: Not interleaving scalar ordered reductions.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not interleaving scalar ordered reductions.\n" ; } } while (false); | |||
5997 | return 1; | |||
5998 | } | |||
5999 | ||||
6000 | unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC); | |||
6001 | SmallIC = std::min(SmallIC, F); | |||
6002 | StoresIC = std::min(StoresIC, F); | |||
6003 | LoadsIC = std::min(LoadsIC, F); | |||
6004 | } | |||
6005 | ||||
6006 | if (EnableLoadStoreRuntimeInterleave && | |||
6007 | std::max(StoresIC, LoadsIC) > SmallIC) { | |||
6008 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving to saturate store or load ports.\n" ; } } while (false) | |||
6009 | dbgs() << "LV: Interleaving to saturate store or load ports.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving to saturate store or load ports.\n" ; } } while (false); | |||
6010 | return std::max(StoresIC, LoadsIC); | |||
6011 | } | |||
6012 | ||||
6013 | // If there are scalar reductions and TTI has enabled aggressive | |||
6014 | // interleaving for reductions, we will interleave to expose ILP. | |||
6015 | if (InterleaveSmallLoopScalarReduction && VF.isScalar() && | |||
6016 | AggressivelyInterleaveReductions) { | |||
6017 | LLVM_DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving to expose ILP.\n" ; } } while (false); | |||
6018 | // Interleave no less than SmallIC but not as aggressive as the normal IC | |||
6019 | // to satisfy the rare situation when resources are too limited. | |||
6020 | return std::max(IC / 2, SmallIC); | |||
6021 | } else { | |||
6022 | LLVM_DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving to reduce branch cost.\n" ; } } while (false); | |||
6023 | return SmallIC; | |||
6024 | } | |||
6025 | } | |||
6026 | ||||
6027 | // Interleave if this is a large loop (small loops are already dealt with by | |||
6028 | // this point) that could benefit from interleaving. | |||
6029 | if (AggressivelyInterleaveReductions) { | |||
6030 | LLVM_DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving to expose ILP.\n" ; } } while (false); | |||
6031 | return IC; | |||
6032 | } | |||
6033 | ||||
6034 | LLVM_DEBUG(dbgs() << "LV: Not Interleaving.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not Interleaving.\n" ; } } while (false); | |||
6035 | return 1; | |||
6036 | } | |||
6037 | ||||
6038 | SmallVector<LoopVectorizationCostModel::RegisterUsage, 8> | |||
6039 | LoopVectorizationCostModel::calculateRegisterUsage(ArrayRef<ElementCount> VFs) { | |||
6040 | // This function calculates the register usage by measuring the highest number | |||
6041 | // of values that are alive at a single location. Obviously, this is a very | |||
6042 | // rough estimation. We scan the loop in a topological order in order and | |||
6043 | // assign a number to each instruction. We use RPO to ensure that defs are | |||
6044 | // met before their users. We assume that each instruction that has in-loop | |||
6045 | // users starts an interval. We record every time that an in-loop value is | |||
6046 | // used, so we have a list of the first and last occurrences of each | |||
6047 | // instruction. Next, we transpose this data structure into a multi map that | |||
6048 | // holds the list of intervals that *end* at a specific location. This multi | |||
6049 | // map allows us to perform a linear search. We scan the instructions linearly | |||
6050 | // and record each time that a new interval starts, by placing it in a set. | |||
6051 | // If we find this value in the multi-map then we remove it from the set. | |||
6052 | // The max register usage is the maximum size of the set. | |||
6053 | // We also search for instructions that are defined outside the loop, but are | |||
6054 | // used inside the loop. We need this number separately from the max-interval | |||
6055 | // usage number because when we unroll, loop-invariant values do not take | |||
6056 | // more register. | |||
6057 | LoopBlocksDFS DFS(TheLoop); | |||
6058 | DFS.perform(LI); | |||
6059 | ||||
6060 | RegisterUsage RU; | |||
6061 | ||||
6062 | // Each 'key' in the map opens a new interval. The values | |||
6063 | // of the map are the index of the 'last seen' usage of the | |||
6064 | // instruction that is the key. | |||
6065 | using IntervalMap = DenseMap<Instruction *, unsigned>; | |||
6066 | ||||
6067 | // Maps instruction to its index. | |||
6068 | SmallVector<Instruction *, 64> IdxToInstr; | |||
6069 | // Marks the end of each interval. | |||
6070 | IntervalMap EndPoint; | |||
6071 | // Saves the list of instruction indices that are used in the loop. | |||
6072 | SmallPtrSet<Instruction *, 8> Ends; | |||
6073 | // Saves the list of values that are used in the loop but are | |||
6074 | // defined outside the loop, such as arguments and constants. | |||
6075 | SmallPtrSet<Value *, 8> LoopInvariants; | |||
6076 | ||||
6077 | for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO())) { | |||
6078 | for (Instruction &I : BB->instructionsWithoutDebug()) { | |||
6079 | IdxToInstr.push_back(&I); | |||
6080 | ||||
6081 | // Save the end location of each USE. | |||
6082 | for (Value *U : I.operands()) { | |||
6083 | auto *Instr = dyn_cast<Instruction>(U); | |||
6084 | ||||
6085 | // Ignore non-instruction values such as arguments, constants, etc. | |||
6086 | if (!Instr) | |||
6087 | continue; | |||
6088 | ||||
6089 | // If this instruction is outside the loop then record it and continue. | |||
6090 | if (!TheLoop->contains(Instr)) { | |||
6091 | LoopInvariants.insert(Instr); | |||
6092 | continue; | |||
6093 | } | |||
6094 | ||||
6095 | // Overwrite previous end points. | |||
6096 | EndPoint[Instr] = IdxToInstr.size(); | |||
6097 | Ends.insert(Instr); | |||
6098 | } | |||
6099 | } | |||
6100 | } | |||
6101 | ||||
6102 | // Saves the list of intervals that end with the index in 'key'. | |||
6103 | using InstrList = SmallVector<Instruction *, 2>; | |||
6104 | DenseMap<unsigned, InstrList> TransposeEnds; | |||
6105 | ||||
6106 | // Transpose the EndPoints to a list of values that end at each index. | |||
6107 | for (auto &Interval : EndPoint) | |||
6108 | TransposeEnds[Interval.second].push_back(Interval.first); | |||
6109 | ||||
6110 | SmallPtrSet<Instruction *, 8> OpenIntervals; | |||
6111 | SmallVector<RegisterUsage, 8> RUs(VFs.size()); | |||
6112 | SmallVector<SmallMapVector<unsigned, unsigned, 4>, 8> MaxUsages(VFs.size()); | |||
6113 | ||||
6114 | LLVM_DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV(REG): Calculating max register usage:\n" ; } } while (false); | |||
6115 | ||||
6116 | // A lambda that gets the register usage for the given type and VF. | |||
6117 | const auto &TTICapture = TTI; | |||
6118 | auto GetRegUsage = [&TTICapture](Type *Ty, ElementCount VF) -> unsigned { | |||
6119 | if (Ty->isTokenTy() || !VectorType::isValidElementType(Ty)) | |||
6120 | return 0; | |||
6121 | InstructionCost::CostType RegUsage = | |||
6122 | *TTICapture.getRegUsageForType(VectorType::get(Ty, VF)).getValue(); | |||
6123 | assert(RegUsage >= 0 && RegUsage <= std::numeric_limits<unsigned>::max() &&(static_cast <bool> (RegUsage >= 0 && RegUsage <= std::numeric_limits<unsigned>::max() && "Nonsensical values for register usage." ) ? void (0) : __assert_fail ("RegUsage >= 0 && RegUsage <= std::numeric_limits<unsigned>::max() && \"Nonsensical values for register usage.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6124, __extension__ __PRETTY_FUNCTION__)) | |||
6124 | "Nonsensical values for register usage.")(static_cast <bool> (RegUsage >= 0 && RegUsage <= std::numeric_limits<unsigned>::max() && "Nonsensical values for register usage." ) ? void (0) : __assert_fail ("RegUsage >= 0 && RegUsage <= std::numeric_limits<unsigned>::max() && \"Nonsensical values for register usage.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6124, __extension__ __PRETTY_FUNCTION__)); | |||
6125 | return RegUsage; | |||
6126 | }; | |||
6127 | ||||
6128 | for (unsigned int i = 0, s = IdxToInstr.size(); i < s; ++i) { | |||
6129 | Instruction *I = IdxToInstr[i]; | |||
6130 | ||||
6131 | // Remove all of the instructions that end at this location. | |||
6132 | InstrList &List = TransposeEnds[i]; | |||
6133 | for (Instruction *ToRemove : List) | |||
6134 | OpenIntervals.erase(ToRemove); | |||
6135 | ||||
6136 | // Ignore instructions that are never used within the loop. | |||
6137 | if (!Ends.count(I)) | |||
6138 | continue; | |||
6139 | ||||
6140 | // Skip ignored values. | |||
6141 | if (ValuesToIgnore.count(I)) | |||
6142 | continue; | |||
6143 | ||||
6144 | // For each VF find the maximum usage of registers. | |||
6145 | for (unsigned j = 0, e = VFs.size(); j < e; ++j) { | |||
6146 | // Count the number of live intervals. | |||
6147 | SmallMapVector<unsigned, unsigned, 4> RegUsage; | |||
6148 | ||||
6149 | if (VFs[j].isScalar()) { | |||
6150 | for (auto Inst : OpenIntervals) { | |||
6151 | unsigned ClassID = TTI.getRegisterClassForType(false, Inst->getType()); | |||
6152 | if (RegUsage.find(ClassID) == RegUsage.end()) | |||
6153 | RegUsage[ClassID] = 1; | |||
6154 | else | |||
6155 | RegUsage[ClassID] += 1; | |||
6156 | } | |||
6157 | } else { | |||
6158 | collectUniformsAndScalars(VFs[j]); | |||
6159 | for (auto Inst : OpenIntervals) { | |||
6160 | // Skip ignored values for VF > 1. | |||
6161 | if (VecValuesToIgnore.count(Inst)) | |||
6162 | continue; | |||
6163 | if (isScalarAfterVectorization(Inst, VFs[j])) { | |||
6164 | unsigned ClassID = TTI.getRegisterClassForType(false, Inst->getType()); | |||
6165 | if (RegUsage.find(ClassID) == RegUsage.end()) | |||
6166 | RegUsage[ClassID] = 1; | |||
6167 | else | |||
6168 | RegUsage[ClassID] += 1; | |||
6169 | } else { | |||
6170 | unsigned ClassID = TTI.getRegisterClassForType(true, Inst->getType()); | |||
6171 | if (RegUsage.find(ClassID) == RegUsage.end()) | |||
6172 | RegUsage[ClassID] = GetRegUsage(Inst->getType(), VFs[j]); | |||
6173 | else | |||
6174 | RegUsage[ClassID] += GetRegUsage(Inst->getType(), VFs[j]); | |||
6175 | } | |||
6176 | } | |||
6177 | } | |||
6178 | ||||
6179 | for (auto& pair : RegUsage) { | |||
6180 | if (MaxUsages[j].find(pair.first) != MaxUsages[j].end()) | |||
6181 | MaxUsages[j][pair.first] = std::max(MaxUsages[j][pair.first], pair.second); | |||
6182 | else | |||
6183 | MaxUsages[j][pair.first] = pair.second; | |||
6184 | } | |||
6185 | } | |||
6186 | ||||
6187 | LLVM_DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV(REG): At #" << i << " Interval # " << OpenIntervals.size() << '\n'; } } while (false) | |||
6188 | << OpenIntervals.size() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV(REG): At #" << i << " Interval # " << OpenIntervals.size() << '\n'; } } while (false); | |||
6189 | ||||
6190 | // Add the current instruction to the list of open intervals. | |||
6191 | OpenIntervals.insert(I); | |||
6192 | } | |||
6193 | ||||
6194 | for (unsigned i = 0, e = VFs.size(); i < e; ++i) { | |||
6195 | SmallMapVector<unsigned, unsigned, 4> Invariant; | |||
6196 | ||||
6197 | for (auto Inst : LoopInvariants) { | |||
6198 | unsigned Usage = | |||
6199 | VFs[i].isScalar() ? 1 : GetRegUsage(Inst->getType(), VFs[i]); | |||
6200 | unsigned ClassID = | |||
6201 | TTI.getRegisterClassForType(VFs[i].isVector(), Inst->getType()); | |||
6202 | if (Invariant.find(ClassID) == Invariant.end()) | |||
6203 | Invariant[ClassID] = Usage; | |||
6204 | else | |||
6205 | Invariant[ClassID] += Usage; | |||
6206 | } | |||
6207 | ||||
6208 | LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | |||
6209 | dbgs() << "LV(REG): VF = " << VFs[i] << '\n';do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | |||
6210 | dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | |||
6211 | << " item\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | |||
6212 | for (const auto &pair : MaxUsages[i]) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | |||
6213 | dbgs() << "LV(REG): RegisterClass: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | |||
6214 | << TTI.getRegisterClassName(pair.first) << ", " << pair.seconddo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | |||
6215 | << " registers\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | |||
6216 | }do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | |||
6217 | dbgs() << "LV(REG): Found invariant usage: " << Invariant.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | |||
6218 | << " item\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | |||
6219 | for (const auto &pair : Invariant) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | |||
6220 | dbgs() << "LV(REG): RegisterClass: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | |||
6221 | << TTI.getRegisterClassName(pair.first) << ", " << pair.seconddo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | |||
6222 | << " registers\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | |||
6223 | }do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | |||
6224 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false); | |||
6225 | ||||
6226 | RU.LoopInvariantRegs = Invariant; | |||
6227 | RU.MaxLocalUsers = MaxUsages[i]; | |||
6228 | RUs[i] = RU; | |||
6229 | } | |||
6230 | ||||
6231 | return RUs; | |||
6232 | } | |||
6233 | ||||
6234 | bool LoopVectorizationCostModel::useEmulatedMaskMemRefHack(Instruction *I, | |||
6235 | ElementCount VF) { | |||
6236 | // TODO: Cost model for emulated masked load/store is completely | |||
6237 | // broken. This hack guides the cost model to use an artificially | |||
6238 | // high enough value to practically disable vectorization with such | |||
6239 | // operations, except where previously deployed legality hack allowed | |||
6240 | // using very low cost values. This is to avoid regressions coming simply | |||
6241 | // from moving "masked load/store" check from legality to cost model. | |||
6242 | // Masked Load/Gather emulation was previously never allowed. | |||
6243 | // Limited number of Masked Store/Scatter emulation was allowed. | |||
6244 | assert(isPredicatedInst(I, VF) && "Expecting a scalar emulated instruction")(static_cast <bool> (isPredicatedInst(I, VF) && "Expecting a scalar emulated instruction") ? void (0) : __assert_fail ("isPredicatedInst(I, VF) && \"Expecting a scalar emulated instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6244, __extension__ __PRETTY_FUNCTION__)); | |||
6245 | return isa<LoadInst>(I) || | |||
6246 | (isa<StoreInst>(I) && | |||
6247 | NumPredStores > NumberOfStoresToPredicate); | |||
6248 | } | |||
6249 | ||||
6250 | void LoopVectorizationCostModel::collectInstsToScalarize(ElementCount VF) { | |||
6251 | // If we aren't vectorizing the loop, or if we've already collected the | |||
6252 | // instructions to scalarize, there's nothing to do. Collection may already | |||
6253 | // have occurred if we have a user-selected VF and are now computing the | |||
6254 | // expected cost for interleaving. | |||
6255 | if (VF.isScalar() || VF.isZero() || | |||
6256 | InstsToScalarize.find(VF) != InstsToScalarize.end()) | |||
6257 | return; | |||
6258 | ||||
6259 | // Initialize a mapping for VF in InstsToScalalarize. If we find that it's | |||
6260 | // not profitable to scalarize any instructions, the presence of VF in the | |||
6261 | // map will indicate that we've analyzed it already. | |||
6262 | ScalarCostsTy &ScalarCostsVF = InstsToScalarize[VF]; | |||
6263 | ||||
6264 | // Find all the instructions that are scalar with predication in the loop and | |||
6265 | // determine if it would be better to not if-convert the blocks they are in. | |||
6266 | // If so, we also record the instructions to scalarize. | |||
6267 | for (BasicBlock *BB : TheLoop->blocks()) { | |||
6268 | if (!blockNeedsPredicationForAnyReason(BB)) | |||
6269 | continue; | |||
6270 | for (Instruction &I : *BB) | |||
6271 | if (isScalarWithPredication(&I, VF)) { | |||
6272 | ScalarCostsTy ScalarCosts; | |||
6273 | // Do not apply discount if scalable, because that would lead to | |||
6274 | // invalid scalarization costs. | |||
6275 | // Do not apply discount logic if hacked cost is needed | |||
6276 | // for emulated masked memrefs. | |||
6277 | if (!VF.isScalable() && !useEmulatedMaskMemRefHack(&I, VF) && | |||
6278 | computePredInstDiscount(&I, ScalarCosts, VF) >= 0) | |||
6279 | ScalarCostsVF.insert(ScalarCosts.begin(), ScalarCosts.end()); | |||
6280 | // Remember that BB will remain after vectorization. | |||
6281 | PredicatedBBsAfterVectorization.insert(BB); | |||
6282 | } | |||
6283 | } | |||
6284 | } | |||
6285 | ||||
6286 | int LoopVectorizationCostModel::computePredInstDiscount( | |||
6287 | Instruction *PredInst, ScalarCostsTy &ScalarCosts, ElementCount VF) { | |||
6288 | assert(!isUniformAfterVectorization(PredInst, VF) &&(static_cast <bool> (!isUniformAfterVectorization(PredInst , VF) && "Instruction marked uniform-after-vectorization will be predicated" ) ? void (0) : __assert_fail ("!isUniformAfterVectorization(PredInst, VF) && \"Instruction marked uniform-after-vectorization will be predicated\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6289, __extension__ __PRETTY_FUNCTION__)) | |||
6289 | "Instruction marked uniform-after-vectorization will be predicated")(static_cast <bool> (!isUniformAfterVectorization(PredInst , VF) && "Instruction marked uniform-after-vectorization will be predicated" ) ? void (0) : __assert_fail ("!isUniformAfterVectorization(PredInst, VF) && \"Instruction marked uniform-after-vectorization will be predicated\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6289, __extension__ __PRETTY_FUNCTION__)); | |||
6290 | ||||
6291 | // Initialize the discount to zero, meaning that the scalar version and the | |||
6292 | // vector version cost the same. | |||
6293 | InstructionCost Discount = 0; | |||
6294 | ||||
6295 | // Holds instructions to analyze. The instructions we visit are mapped in | |||
6296 | // ScalarCosts. Those instructions are the ones that would be scalarized if | |||
6297 | // we find that the scalar version costs less. | |||
6298 | SmallVector<Instruction *, 8> Worklist; | |||
6299 | ||||
6300 | // Returns true if the given instruction can be scalarized. | |||
6301 | auto canBeScalarized = [&](Instruction *I) -> bool { | |||
6302 | // We only attempt to scalarize instructions forming a single-use chain | |||
6303 | // from the original predicated block that would otherwise be vectorized. | |||
6304 | // Although not strictly necessary, we give up on instructions we know will | |||
6305 | // already be scalar to avoid traversing chains that are unlikely to be | |||
6306 | // beneficial. | |||
6307 | if (!I->hasOneUse() || PredInst->getParent() != I->getParent() || | |||
6308 | isScalarAfterVectorization(I, VF)) | |||
6309 | return false; | |||
6310 | ||||
6311 | // If the instruction is scalar with predication, it will be analyzed | |||
6312 | // separately. We ignore it within the context of PredInst. | |||
6313 | if (isScalarWithPredication(I, VF)) | |||
6314 | return false; | |||
6315 | ||||
6316 | // If any of the instruction's operands are uniform after vectorization, | |||
6317 | // the instruction cannot be scalarized. This prevents, for example, a | |||
6318 | // masked load from being scalarized. | |||
6319 | // | |||
6320 | // We assume we will only emit a value for lane zero of an instruction | |||
6321 | // marked uniform after vectorization, rather than VF identical values. | |||
6322 | // Thus, if we scalarize an instruction that uses a uniform, we would | |||
6323 | // create uses of values corresponding to the lanes we aren't emitting code | |||
6324 | // for. This behavior can be changed by allowing getScalarValue to clone | |||
6325 | // the lane zero values for uniforms rather than asserting. | |||
6326 | for (Use &U : I->operands()) | |||
6327 | if (auto *J = dyn_cast<Instruction>(U.get())) | |||
6328 | if (isUniformAfterVectorization(J, VF)) | |||
6329 | return false; | |||
6330 | ||||
6331 | // Otherwise, we can scalarize the instruction. | |||
6332 | return true; | |||
6333 | }; | |||
6334 | ||||
6335 | // Compute the expected cost discount from scalarizing the entire expression | |||
6336 | // feeding the predicated instruction. We currently only consider expressions | |||
6337 | // that are single-use instruction chains. | |||
6338 | Worklist.push_back(PredInst); | |||
6339 | while (!Worklist.empty()) { | |||
6340 | Instruction *I = Worklist.pop_back_val(); | |||
6341 | ||||
6342 | // If we've already analyzed the instruction, there's nothing to do. | |||
6343 | if (ScalarCosts.find(I) != ScalarCosts.end()) | |||
6344 | continue; | |||
6345 | ||||
6346 | // Compute the cost of the vector instruction. Note that this cost already | |||
6347 | // includes the scalarization overhead of the predicated instruction. | |||
6348 | InstructionCost VectorCost = getInstructionCost(I, VF).first; | |||
6349 | ||||
6350 | // Compute the cost of the scalarized instruction. This cost is the cost of | |||
6351 | // the instruction as if it wasn't if-converted and instead remained in the | |||
6352 | // predicated block. We will scale this cost by block probability after | |||
6353 | // computing the scalarization overhead. | |||
6354 | InstructionCost ScalarCost = | |||
6355 | VF.getFixedValue() * | |||
6356 | getInstructionCost(I, ElementCount::getFixed(1)).first; | |||
6357 | ||||
6358 | // Compute the scalarization overhead of needed insertelement instructions | |||
6359 | // and phi nodes. | |||
6360 | if (isScalarWithPredication(I, VF) && !I->getType()->isVoidTy()) { | |||
6361 | ScalarCost += TTI.getScalarizationOverhead( | |||
6362 | cast<VectorType>(ToVectorTy(I->getType(), VF)), | |||
6363 | APInt::getAllOnes(VF.getFixedValue()), true, false); | |||
6364 | ScalarCost += | |||
6365 | VF.getFixedValue() * | |||
6366 | TTI.getCFInstrCost(Instruction::PHI, TTI::TCK_RecipThroughput); | |||
6367 | } | |||
6368 | ||||
6369 | // Compute the scalarization overhead of needed extractelement | |||
6370 | // instructions. For each of the instruction's operands, if the operand can | |||
6371 | // be scalarized, add it to the worklist; otherwise, account for the | |||
6372 | // overhead. | |||
6373 | for (Use &U : I->operands()) | |||
6374 | if (auto *J = dyn_cast<Instruction>(U.get())) { | |||
6375 | assert(VectorType::isValidElementType(J->getType()) &&(static_cast <bool> (VectorType::isValidElementType(J-> getType()) && "Instruction has non-scalar type") ? void (0) : __assert_fail ("VectorType::isValidElementType(J->getType()) && \"Instruction has non-scalar type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6376, __extension__ __PRETTY_FUNCTION__)) | |||
6376 | "Instruction has non-scalar type")(static_cast <bool> (VectorType::isValidElementType(J-> getType()) && "Instruction has non-scalar type") ? void (0) : __assert_fail ("VectorType::isValidElementType(J->getType()) && \"Instruction has non-scalar type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6376, __extension__ __PRETTY_FUNCTION__)); | |||
6377 | if (canBeScalarized(J)) | |||
6378 | Worklist.push_back(J); | |||
6379 | else if (needsExtract(J, VF)) { | |||
6380 | ScalarCost += TTI.getScalarizationOverhead( | |||
6381 | cast<VectorType>(ToVectorTy(J->getType(), VF)), | |||
6382 | APInt::getAllOnes(VF.getFixedValue()), false, true); | |||
6383 | } | |||
6384 | } | |||
6385 | ||||
6386 | // Scale the total scalar cost by block probability. | |||
6387 | ScalarCost /= getReciprocalPredBlockProb(); | |||
6388 | ||||
6389 | // Compute the discount. A non-negative discount means the vector version | |||
6390 | // of the instruction costs more, and scalarizing would be beneficial. | |||
6391 | Discount += VectorCost - ScalarCost; | |||
6392 | ScalarCosts[I] = ScalarCost; | |||
6393 | } | |||
6394 | ||||
6395 | return *Discount.getValue(); | |||
6396 | } | |||
6397 | ||||
6398 | LoopVectorizationCostModel::VectorizationCostTy | |||
6399 | LoopVectorizationCostModel::expectedCost( | |||
6400 | ElementCount VF, SmallVectorImpl<InstructionVFPair> *Invalid) { | |||
6401 | VectorizationCostTy Cost; | |||
6402 | ||||
6403 | // For each block. | |||
6404 | for (BasicBlock *BB : TheLoop->blocks()) { | |||
6405 | VectorizationCostTy BlockCost; | |||
6406 | ||||
6407 | // For each instruction in the old loop. | |||
6408 | for (Instruction &I : BB->instructionsWithoutDebug()) { | |||
6409 | // Skip ignored values. | |||
6410 | if (ValuesToIgnore.count(&I) || | |||
6411 | (VF.isVector() && VecValuesToIgnore.count(&I))) | |||
6412 | continue; | |||
6413 | ||||
6414 | VectorizationCostTy C = getInstructionCost(&I, VF); | |||
6415 | ||||
6416 | // Check if we should override the cost. | |||
6417 | if (C.first.isValid() && | |||
6418 | ForceTargetInstructionCost.getNumOccurrences() > 0) | |||
6419 | C.first = InstructionCost(ForceTargetInstructionCost); | |||
6420 | ||||
6421 | // Keep a list of instructions with invalid costs. | |||
6422 | if (Invalid && !C.first.isValid()) | |||
6423 | Invalid->emplace_back(&I, VF); | |||
6424 | ||||
6425 | BlockCost.first += C.first; | |||
6426 | BlockCost.second |= C.second; | |||
6427 | LLVM_DEBUG(dbgs() << "LV: Found an estimated cost of " << C.firstdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found an estimated cost of " << C.first << " for VF " << VF << " For instruction: " << I << '\n'; } } while (false) | |||
6428 | << " for VF " << VF << " For instruction: " << Ido { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found an estimated cost of " << C.first << " for VF " << VF << " For instruction: " << I << '\n'; } } while (false) | |||
6429 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found an estimated cost of " << C.first << " for VF " << VF << " For instruction: " << I << '\n'; } } while (false); | |||
6430 | } | |||
6431 | ||||
6432 | // If we are vectorizing a predicated block, it will have been | |||
6433 | // if-converted. This means that the block's instructions (aside from | |||
6434 | // stores and instructions that may divide by zero) will now be | |||
6435 | // unconditionally executed. For the scalar case, we may not always execute | |||
6436 | // the predicated block, if it is an if-else block. Thus, scale the block's | |||
6437 | // cost by the probability of executing it. blockNeedsPredication from | |||
6438 | // Legal is used so as to not include all blocks in tail folded loops. | |||
6439 | if (VF.isScalar() && Legal->blockNeedsPredication(BB)) | |||
6440 | BlockCost.first /= getReciprocalPredBlockProb(); | |||
6441 | ||||
6442 | Cost.first += BlockCost.first; | |||
6443 | Cost.second |= BlockCost.second; | |||
6444 | } | |||
6445 | ||||
6446 | return Cost; | |||
6447 | } | |||
6448 | ||||
6449 | /// Gets Address Access SCEV after verifying that the access pattern | |||
6450 | /// is loop invariant except the induction variable dependence. | |||
6451 | /// | |||
6452 | /// This SCEV can be sent to the Target in order to estimate the address | |||
6453 | /// calculation cost. | |||
6454 | static const SCEV *getAddressAccessSCEV( | |||
6455 | Value *Ptr, | |||
6456 | LoopVectorizationLegality *Legal, | |||
6457 | PredicatedScalarEvolution &PSE, | |||
6458 | const Loop *TheLoop) { | |||
6459 | ||||
6460 | auto *Gep = dyn_cast<GetElementPtrInst>(Ptr); | |||
6461 | if (!Gep) | |||
6462 | return nullptr; | |||
6463 | ||||
6464 | // We are looking for a gep with all loop invariant indices except for one | |||
6465 | // which should be an induction variable. | |||
6466 | auto SE = PSE.getSE(); | |||
6467 | unsigned NumOperands = Gep->getNumOperands(); | |||
6468 | for (unsigned i = 1; i < NumOperands; ++i) { | |||
6469 | Value *Opd = Gep->getOperand(i); | |||
6470 | if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) && | |||
6471 | !Legal->isInductionVariable(Opd)) | |||
6472 | return nullptr; | |||
6473 | } | |||
6474 | ||||
6475 | // Now we know we have a GEP ptr, %inv, %ind, %inv. return the Ptr SCEV. | |||
6476 | return PSE.getSCEV(Ptr); | |||
6477 | } | |||
6478 | ||||
6479 | static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) { | |||
6480 | return Legal->hasStride(I->getOperand(0)) || | |||
6481 | Legal->hasStride(I->getOperand(1)); | |||
6482 | } | |||
6483 | ||||
6484 | InstructionCost | |||
6485 | LoopVectorizationCostModel::getMemInstScalarizationCost(Instruction *I, | |||
6486 | ElementCount VF) { | |||
6487 | assert(VF.isVector() &&(static_cast <bool> (VF.isVector() && "Scalarization cost of instruction implies vectorization." ) ? void (0) : __assert_fail ("VF.isVector() && \"Scalarization cost of instruction implies vectorization.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6488, __extension__ __PRETTY_FUNCTION__)) | |||
6488 | "Scalarization cost of instruction implies vectorization.")(static_cast <bool> (VF.isVector() && "Scalarization cost of instruction implies vectorization." ) ? void (0) : __assert_fail ("VF.isVector() && \"Scalarization cost of instruction implies vectorization.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6488, __extension__ __PRETTY_FUNCTION__)); | |||
6489 | if (VF.isScalable()) | |||
6490 | return InstructionCost::getInvalid(); | |||
6491 | ||||
6492 | Type *ValTy = getLoadStoreType(I); | |||
6493 | auto SE = PSE.getSE(); | |||
6494 | ||||
6495 | unsigned AS = getLoadStoreAddressSpace(I); | |||
6496 | Value *Ptr = getLoadStorePointerOperand(I); | |||
6497 | Type *PtrTy = ToVectorTy(Ptr->getType(), VF); | |||
6498 | // NOTE: PtrTy is a vector to signal `TTI::getAddressComputationCost` | |||
6499 | // that it is being called from this specific place. | |||
6500 | ||||
6501 | // Figure out whether the access is strided and get the stride value | |||
6502 | // if it's known in compile time | |||
6503 | const SCEV *PtrSCEV = getAddressAccessSCEV(Ptr, Legal, PSE, TheLoop); | |||
6504 | ||||
6505 | // Get the cost of the scalar memory instruction and address computation. | |||
6506 | InstructionCost Cost = | |||
6507 | VF.getKnownMinValue() * TTI.getAddressComputationCost(PtrTy, SE, PtrSCEV); | |||
6508 | ||||
6509 | // Don't pass *I here, since it is scalar but will actually be part of a | |||
6510 | // vectorized loop where the user of it is a vectorized instruction. | |||
6511 | const Align Alignment = getLoadStoreAlignment(I); | |||
6512 | Cost += VF.getKnownMinValue() * | |||
6513 | TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(), Alignment, | |||
6514 | AS, TTI::TCK_RecipThroughput); | |||
6515 | ||||
6516 | // Get the overhead of the extractelement and insertelement instructions | |||
6517 | // we might create due to scalarization. | |||
6518 | Cost += getScalarizationOverhead(I, VF); | |||
6519 | ||||
6520 | // If we have a predicated load/store, it will need extra i1 extracts and | |||
6521 | // conditional branches, but may not be executed for each vector lane. Scale | |||
6522 | // the cost by the probability of executing the predicated block. | |||
6523 | if (isPredicatedInst(I, VF)) { | |||
6524 | Cost /= getReciprocalPredBlockProb(); | |||
6525 | ||||
6526 | // Add the cost of an i1 extract and a branch | |||
6527 | auto *Vec_i1Ty = | |||
6528 | VectorType::get(IntegerType::getInt1Ty(ValTy->getContext()), VF); | |||
6529 | Cost += TTI.getScalarizationOverhead( | |||
6530 | Vec_i1Ty, APInt::getAllOnes(VF.getKnownMinValue()), | |||
6531 | /*Insert=*/false, /*Extract=*/true); | |||
6532 | Cost += TTI.getCFInstrCost(Instruction::Br, TTI::TCK_RecipThroughput); | |||
6533 | ||||
6534 | if (useEmulatedMaskMemRefHack(I, VF)) | |||
6535 | // Artificially setting to a high enough value to practically disable | |||
6536 | // vectorization with such operations. | |||
6537 | Cost = 3000000; | |||
6538 | } | |||
6539 | ||||
6540 | return Cost; | |||
6541 | } | |||
6542 | ||||
6543 | InstructionCost | |||
6544 | LoopVectorizationCostModel::getConsecutiveMemOpCost(Instruction *I, | |||
6545 | ElementCount VF) { | |||
6546 | Type *ValTy = getLoadStoreType(I); | |||
6547 | auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF)); | |||
6548 | Value *Ptr = getLoadStorePointerOperand(I); | |||
6549 | unsigned AS = getLoadStoreAddressSpace(I); | |||
6550 | int ConsecutiveStride = Legal->isConsecutivePtr(ValTy, Ptr); | |||
6551 | enum TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | |||
6552 | ||||
6553 | assert((ConsecutiveStride == 1 || ConsecutiveStride == -1) &&(static_cast <bool> ((ConsecutiveStride == 1 || ConsecutiveStride == -1) && "Stride should be 1 or -1 for consecutive memory access" ) ? void (0) : __assert_fail ("(ConsecutiveStride == 1 || ConsecutiveStride == -1) && \"Stride should be 1 or -1 for consecutive memory access\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6554, __extension__ __PRETTY_FUNCTION__)) | |||
6554 | "Stride should be 1 or -1 for consecutive memory access")(static_cast <bool> ((ConsecutiveStride == 1 || ConsecutiveStride == -1) && "Stride should be 1 or -1 for consecutive memory access" ) ? void (0) : __assert_fail ("(ConsecutiveStride == 1 || ConsecutiveStride == -1) && \"Stride should be 1 or -1 for consecutive memory access\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6554, __extension__ __PRETTY_FUNCTION__)); | |||
6555 | const Align Alignment = getLoadStoreAlignment(I); | |||
6556 | InstructionCost Cost = 0; | |||
6557 | if (Legal->isMaskRequired(I)) | |||
6558 | Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS, | |||
6559 | CostKind); | |||
6560 | else | |||
6561 | Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS, | |||
6562 | CostKind, I); | |||
6563 | ||||
6564 | bool Reverse = ConsecutiveStride < 0; | |||
6565 | if (Reverse) | |||
6566 | Cost += | |||
6567 | TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, None, 0); | |||
6568 | return Cost; | |||
6569 | } | |||
6570 | ||||
6571 | InstructionCost | |||
6572 | LoopVectorizationCostModel::getUniformMemOpCost(Instruction *I, | |||
6573 | ElementCount VF) { | |||
6574 | assert(Legal->isUniformMemOp(*I))(static_cast <bool> (Legal->isUniformMemOp(*I)) ? void (0) : __assert_fail ("Legal->isUniformMemOp(*I)", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6574, __extension__ __PRETTY_FUNCTION__)); | |||
6575 | ||||
6576 | Type *ValTy = getLoadStoreType(I); | |||
6577 | auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF)); | |||
6578 | const Align Alignment = getLoadStoreAlignment(I); | |||
6579 | unsigned AS = getLoadStoreAddressSpace(I); | |||
6580 | enum TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | |||
6581 | if (isa<LoadInst>(I)) { | |||
6582 | return TTI.getAddressComputationCost(ValTy) + | |||
6583 | TTI.getMemoryOpCost(Instruction::Load, ValTy, Alignment, AS, | |||
6584 | CostKind) + | |||
6585 | TTI.getShuffleCost(TargetTransformInfo::SK_Broadcast, VectorTy); | |||
6586 | } | |||
6587 | StoreInst *SI = cast<StoreInst>(I); | |||
6588 | ||||
6589 | bool isLoopInvariantStoreValue = Legal->isUniform(SI->getValueOperand()); | |||
6590 | return TTI.getAddressComputationCost(ValTy) + | |||
6591 | TTI.getMemoryOpCost(Instruction::Store, ValTy, Alignment, AS, | |||
6592 | CostKind) + | |||
6593 | (isLoopInvariantStoreValue | |||
6594 | ? 0 | |||
6595 | : TTI.getVectorInstrCost(Instruction::ExtractElement, VectorTy, | |||
6596 | VF.getKnownMinValue() - 1)); | |||
6597 | } | |||
6598 | ||||
6599 | InstructionCost | |||
6600 | LoopVectorizationCostModel::getGatherScatterCost(Instruction *I, | |||
6601 | ElementCount VF) { | |||
6602 | Type *ValTy = getLoadStoreType(I); | |||
6603 | auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF)); | |||
6604 | const Align Alignment = getLoadStoreAlignment(I); | |||
6605 | const Value *Ptr = getLoadStorePointerOperand(I); | |||
6606 | ||||
6607 | return TTI.getAddressComputationCost(VectorTy) + | |||
6608 | TTI.getGatherScatterOpCost( | |||
6609 | I->getOpcode(), VectorTy, Ptr, Legal->isMaskRequired(I), Alignment, | |||
6610 | TargetTransformInfo::TCK_RecipThroughput, I); | |||
6611 | } | |||
6612 | ||||
6613 | InstructionCost | |||
6614 | LoopVectorizationCostModel::getInterleaveGroupCost(Instruction *I, | |||
6615 | ElementCount VF) { | |||
6616 | // TODO: Once we have support for interleaving with scalable vectors | |||
6617 | // we can calculate the cost properly here. | |||
6618 | if (VF.isScalable()) | |||
6619 | return InstructionCost::getInvalid(); | |||
6620 | ||||
6621 | Type *ValTy = getLoadStoreType(I); | |||
6622 | auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF)); | |||
6623 | unsigned AS = getLoadStoreAddressSpace(I); | |||
6624 | ||||
6625 | auto Group = getInterleavedAccessGroup(I); | |||
6626 | assert(Group && "Fail to get an interleaved access group.")(static_cast <bool> (Group && "Fail to get an interleaved access group." ) ? void (0) : __assert_fail ("Group && \"Fail to get an interleaved access group.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6626, __extension__ __PRETTY_FUNCTION__)); | |||
6627 | ||||
6628 | unsigned InterleaveFactor = Group->getFactor(); | |||
6629 | auto *WideVecTy = VectorType::get(ValTy, VF * InterleaveFactor); | |||
6630 | ||||
6631 | // Holds the indices of existing members in the interleaved group. | |||
6632 | SmallVector<unsigned, 4> Indices; | |||
6633 | for (unsigned IF = 0; IF < InterleaveFactor; IF++) | |||
6634 | if (Group->getMember(IF)) | |||
6635 | Indices.push_back(IF); | |||
6636 | ||||
6637 | // Calculate the cost of the whole interleaved group. | |||
6638 | bool UseMaskForGaps = | |||
6639 | (Group->requiresScalarEpilogue() && !isScalarEpilogueAllowed()) || | |||
6640 | (isa<StoreInst>(I) && (Group->getNumMembers() < Group->getFactor())); | |||
6641 | InstructionCost Cost = TTI.getInterleavedMemoryOpCost( | |||
6642 | I->getOpcode(), WideVecTy, Group->getFactor(), Indices, Group->getAlign(), | |||
6643 | AS, TTI::TCK_RecipThroughput, Legal->isMaskRequired(I), UseMaskForGaps); | |||
6644 | ||||
6645 | if (Group->isReverse()) { | |||
6646 | // TODO: Add support for reversed masked interleaved access. | |||
6647 | assert(!Legal->isMaskRequired(I) &&(static_cast <bool> (!Legal->isMaskRequired(I) && "Reverse masked interleaved access not supported.") ? void ( 0) : __assert_fail ("!Legal->isMaskRequired(I) && \"Reverse masked interleaved access not supported.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6648, __extension__ __PRETTY_FUNCTION__)) | |||
6648 | "Reverse masked interleaved access not supported.")(static_cast <bool> (!Legal->isMaskRequired(I) && "Reverse masked interleaved access not supported.") ? void ( 0) : __assert_fail ("!Legal->isMaskRequired(I) && \"Reverse masked interleaved access not supported.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6648, __extension__ __PRETTY_FUNCTION__)); | |||
6649 | Cost += | |||
6650 | Group->getNumMembers() * | |||
6651 | TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, None, 0); | |||
6652 | } | |||
6653 | return Cost; | |||
6654 | } | |||
6655 | ||||
6656 | Optional<InstructionCost> LoopVectorizationCostModel::getReductionPatternCost( | |||
6657 | Instruction *I, ElementCount VF, Type *Ty, TTI::TargetCostKind CostKind) { | |||
6658 | using namespace llvm::PatternMatch; | |||
6659 | // Early exit for no inloop reductions | |||
6660 | if (InLoopReductionChains.empty() || VF.isScalar() || !isa<VectorType>(Ty)) | |||
6661 | return None; | |||
6662 | auto *VectorTy = cast<VectorType>(Ty); | |||
6663 | ||||
6664 | // We are looking for a pattern of, and finding the minimal acceptable cost: | |||
6665 | // reduce(mul(ext(A), ext(B))) or | |||
6666 | // reduce(mul(A, B)) or | |||
6667 | // reduce(ext(A)) or | |||
6668 | // reduce(A). | |||
6669 | // The basic idea is that we walk down the tree to do that, finding the root | |||
6670 | // reduction instruction in InLoopReductionImmediateChains. From there we find | |||
6671 | // the pattern of mul/ext and test the cost of the entire pattern vs the cost | |||
6672 | // of the components. If the reduction cost is lower then we return it for the | |||
6673 | // reduction instruction and 0 for the other instructions in the pattern. If | |||
6674 | // it is not we return an invalid cost specifying the orignal cost method | |||
6675 | // should be used. | |||
6676 | Instruction *RetI = I; | |||
6677 | if (match(RetI, m_ZExtOrSExt(m_Value()))) { | |||
6678 | if (!RetI->hasOneUser()) | |||
6679 | return None; | |||
6680 | RetI = RetI->user_back(); | |||
6681 | } | |||
6682 | if (match(RetI, m_Mul(m_Value(), m_Value())) && | |||
6683 | RetI->user_back()->getOpcode() == Instruction::Add) { | |||
6684 | if (!RetI->hasOneUser()) | |||
6685 | return None; | |||
6686 | RetI = RetI->user_back(); | |||
6687 | } | |||
6688 | ||||
6689 | // Test if the found instruction is a reduction, and if not return an invalid | |||
6690 | // cost specifying the parent to use the original cost modelling. | |||
6691 | if (!InLoopReductionImmediateChains.count(RetI)) | |||
6692 | return None; | |||
6693 | ||||
6694 | // Find the reduction this chain is a part of and calculate the basic cost of | |||
6695 | // the reduction on its own. | |||
6696 | Instruction *LastChain = InLoopReductionImmediateChains[RetI]; | |||
6697 | Instruction *ReductionPhi = LastChain; | |||
6698 | while (!isa<PHINode>(ReductionPhi)) | |||
6699 | ReductionPhi = InLoopReductionImmediateChains[ReductionPhi]; | |||
6700 | ||||
6701 | const RecurrenceDescriptor &RdxDesc = | |||
6702 | Legal->getReductionVars().find(cast<PHINode>(ReductionPhi))->second; | |||
6703 | ||||
6704 | InstructionCost BaseCost = TTI.getArithmeticReductionCost( | |||
6705 | RdxDesc.getOpcode(), VectorTy, RdxDesc.getFastMathFlags(), CostKind); | |||
6706 | ||||
6707 | // For a call to the llvm.fmuladd intrinsic we need to add the cost of a | |||
6708 | // normal fmul instruction to the cost of the fadd reduction. | |||
6709 | if (RdxDesc.getRecurrenceKind() == RecurKind::FMulAdd) | |||
6710 | BaseCost += | |||
6711 | TTI.getArithmeticInstrCost(Instruction::FMul, VectorTy, CostKind); | |||
6712 | ||||
6713 | // If we're using ordered reductions then we can just return the base cost | |||
6714 | // here, since getArithmeticReductionCost calculates the full ordered | |||
6715 | // reduction cost when FP reassociation is not allowed. | |||
6716 | if (useOrderedReductions(RdxDesc)) | |||
6717 | return BaseCost; | |||
6718 | ||||
6719 | // Get the operand that was not the reduction chain and match it to one of the | |||
6720 | // patterns, returning the better cost if it is found. | |||
6721 | Instruction *RedOp = RetI->getOperand(1) == LastChain | |||
6722 | ? dyn_cast<Instruction>(RetI->getOperand(0)) | |||
6723 | : dyn_cast<Instruction>(RetI->getOperand(1)); | |||
6724 | ||||
6725 | VectorTy = VectorType::get(I->getOperand(0)->getType(), VectorTy); | |||
6726 | ||||
6727 | Instruction *Op0, *Op1; | |||
6728 | if (RedOp && | |||
6729 | match(RedOp, | |||
6730 | m_ZExtOrSExt(m_Mul(m_Instruction(Op0), m_Instruction(Op1)))) && | |||
6731 | match(Op0, m_ZExtOrSExt(m_Value())) && | |||
6732 | Op0->getOpcode() == Op1->getOpcode() && | |||
6733 | Op0->getOperand(0)->getType() == Op1->getOperand(0)->getType() && | |||
6734 | !TheLoop->isLoopInvariant(Op0) && !TheLoop->isLoopInvariant(Op1) && | |||
6735 | (Op0->getOpcode() == RedOp->getOpcode() || Op0 == Op1)) { | |||
6736 | ||||
6737 | // Matched reduce(ext(mul(ext(A), ext(B))) | |||
6738 | // Note that the extend opcodes need to all match, or if A==B they will have | |||
6739 | // been converted to zext(mul(sext(A), sext(A))) as it is known positive, | |||
6740 | // which is equally fine. | |||
6741 | bool IsUnsigned = isa<ZExtInst>(Op0); | |||
6742 | auto *ExtType = VectorType::get(Op0->getOperand(0)->getType(), VectorTy); | |||
6743 | auto *MulType = VectorType::get(Op0->getType(), VectorTy); | |||
6744 | ||||
6745 | InstructionCost ExtCost = | |||
6746 | TTI.getCastInstrCost(Op0->getOpcode(), MulType, ExtType, | |||
6747 | TTI::CastContextHint::None, CostKind, Op0); | |||
6748 | InstructionCost MulCost = | |||
6749 | TTI.getArithmeticInstrCost(Instruction::Mul, MulType, CostKind); | |||
6750 | InstructionCost Ext2Cost = | |||
6751 | TTI.getCastInstrCost(RedOp->getOpcode(), VectorTy, MulType, | |||
6752 | TTI::CastContextHint::None, CostKind, RedOp); | |||
6753 | ||||
6754 | InstructionCost RedCost = TTI.getExtendedAddReductionCost( | |||
6755 | /*IsMLA=*/true, IsUnsigned, RdxDesc.getRecurrenceType(), ExtType, | |||
6756 | CostKind); | |||
6757 | ||||
6758 | if (RedCost.isValid() && | |||
6759 | RedCost < ExtCost * 2 + MulCost + Ext2Cost + BaseCost) | |||
6760 | return I == RetI ? RedCost : 0; | |||
6761 | } else if (RedOp && match(RedOp, m_ZExtOrSExt(m_Value())) && | |||
6762 | !TheLoop->isLoopInvariant(RedOp)) { | |||
6763 | // Matched reduce(ext(A)) | |||
6764 | bool IsUnsigned = isa<ZExtInst>(RedOp); | |||
6765 | auto *ExtType = VectorType::get(RedOp->getOperand(0)->getType(), VectorTy); | |||
6766 | InstructionCost RedCost = TTI.getExtendedAddReductionCost( | |||
6767 | /*IsMLA=*/false, IsUnsigned, RdxDesc.getRecurrenceType(), ExtType, | |||
6768 | CostKind); | |||
6769 | ||||
6770 | InstructionCost ExtCost = | |||
6771 | TTI.getCastInstrCost(RedOp->getOpcode(), VectorTy, ExtType, | |||
6772 | TTI::CastContextHint::None, CostKind, RedOp); | |||
6773 | if (RedCost.isValid() && RedCost < BaseCost + ExtCost) | |||
6774 | return I == RetI ? RedCost : 0; | |||
6775 | } else if (RedOp && | |||
6776 | match(RedOp, m_Mul(m_Instruction(Op0), m_Instruction(Op1)))) { | |||
6777 | if (match(Op0, m_ZExtOrSExt(m_Value())) && | |||
6778 | Op0->getOpcode() == Op1->getOpcode() && | |||
6779 | !TheLoop->isLoopInvariant(Op0) && !TheLoop->isLoopInvariant(Op1)) { | |||
6780 | bool IsUnsigned = isa<ZExtInst>(Op0); | |||
6781 | Type *Op0Ty = Op0->getOperand(0)->getType(); | |||
6782 | Type *Op1Ty = Op1->getOperand(0)->getType(); | |||
6783 | Type *LargestOpTy = | |||
6784 | Op0Ty->getIntegerBitWidth() < Op1Ty->getIntegerBitWidth() ? Op1Ty | |||
6785 | : Op0Ty; | |||
6786 | auto *ExtType = VectorType::get(LargestOpTy, VectorTy); | |||
6787 | ||||
6788 | // Matched reduce(mul(ext(A), ext(B))), where the two ext may be of | |||
6789 | // different sizes. We take the largest type as the ext to reduce, and add | |||
6790 | // the remaining cost as, for example reduce(mul(ext(ext(A)), ext(B))). | |||
6791 | InstructionCost ExtCost0 = TTI.getCastInstrCost( | |||
6792 | Op0->getOpcode(), VectorTy, VectorType::get(Op0Ty, VectorTy), | |||
6793 | TTI::CastContextHint::None, CostKind, Op0); | |||
6794 | InstructionCost ExtCost1 = TTI.getCastInstrCost( | |||
6795 | Op1->getOpcode(), VectorTy, VectorType::get(Op1Ty, VectorTy), | |||
6796 | TTI::CastContextHint::None, CostKind, Op1); | |||
6797 | InstructionCost MulCost = | |||
6798 | TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind); | |||
6799 | ||||
6800 | InstructionCost RedCost = TTI.getExtendedAddReductionCost( | |||
6801 | /*IsMLA=*/true, IsUnsigned, RdxDesc.getRecurrenceType(), ExtType, | |||
6802 | CostKind); | |||
6803 | InstructionCost ExtraExtCost = 0; | |||
6804 | if (Op0Ty != LargestOpTy || Op1Ty != LargestOpTy) { | |||
6805 | Instruction *ExtraExtOp = (Op0Ty != LargestOpTy) ? Op0 : Op1; | |||
6806 | ExtraExtCost = TTI.getCastInstrCost( | |||
6807 | ExtraExtOp->getOpcode(), ExtType, | |||
6808 | VectorType::get(ExtraExtOp->getOperand(0)->getType(), VectorTy), | |||
6809 | TTI::CastContextHint::None, CostKind, ExtraExtOp); | |||
6810 | } | |||
6811 | ||||
6812 | if (RedCost.isValid() && | |||
6813 | (RedCost + ExtraExtCost) < (ExtCost0 + ExtCost1 + MulCost + BaseCost)) | |||
6814 | return I == RetI ? RedCost : 0; | |||
6815 | } else if (!match(I, m_ZExtOrSExt(m_Value()))) { | |||
6816 | // Matched reduce(mul()) | |||
6817 | InstructionCost MulCost = | |||
6818 | TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind); | |||
6819 | ||||
6820 | InstructionCost RedCost = TTI.getExtendedAddReductionCost( | |||
6821 | /*IsMLA=*/true, true, RdxDesc.getRecurrenceType(), VectorTy, | |||
6822 | CostKind); | |||
6823 | ||||
6824 | if (RedCost.isValid() && RedCost < MulCost + BaseCost) | |||
6825 | return I == RetI ? RedCost : 0; | |||
6826 | } | |||
6827 | } | |||
6828 | ||||
6829 | return I == RetI ? Optional<InstructionCost>(BaseCost) : None; | |||
6830 | } | |||
6831 | ||||
6832 | InstructionCost | |||
6833 | LoopVectorizationCostModel::getMemoryInstructionCost(Instruction *I, | |||
6834 | ElementCount VF) { | |||
6835 | // Calculate scalar cost only. Vectorization cost should be ready at this | |||
6836 | // moment. | |||
6837 | if (VF.isScalar()) { | |||
6838 | Type *ValTy = getLoadStoreType(I); | |||
6839 | const Align Alignment = getLoadStoreAlignment(I); | |||
6840 | unsigned AS = getLoadStoreAddressSpace(I); | |||
6841 | ||||
6842 | return TTI.getAddressComputationCost(ValTy) + | |||
6843 | TTI.getMemoryOpCost(I->getOpcode(), ValTy, Alignment, AS, | |||
6844 | TTI::TCK_RecipThroughput, I); | |||
6845 | } | |||
6846 | return getWideningCost(I, VF); | |||
6847 | } | |||
6848 | ||||
6849 | LoopVectorizationCostModel::VectorizationCostTy | |||
6850 | LoopVectorizationCostModel::getInstructionCost(Instruction *I, | |||
6851 | ElementCount VF) { | |||
6852 | // If we know that this instruction will remain uniform, check the cost of | |||
6853 | // the scalar version. | |||
6854 | if (isUniformAfterVectorization(I, VF)) | |||
6855 | VF = ElementCount::getFixed(1); | |||
6856 | ||||
6857 | if (VF.isVector() && isProfitableToScalarize(I, VF)) | |||
6858 | return VectorizationCostTy(InstsToScalarize[VF][I], false); | |||
6859 | ||||
6860 | // Forced scalars do not have any scalarization overhead. | |||
6861 | auto ForcedScalar = ForcedScalars.find(VF); | |||
6862 | if (VF.isVector() && ForcedScalar != ForcedScalars.end()) { | |||
6863 | auto InstSet = ForcedScalar->second; | |||
6864 | if (InstSet.count(I)) | |||
6865 | return VectorizationCostTy( | |||
6866 | (getInstructionCost(I, ElementCount::getFixed(1)).first * | |||
6867 | VF.getKnownMinValue()), | |||
6868 | false); | |||
6869 | } | |||
6870 | ||||
6871 | Type *VectorTy; | |||
6872 | InstructionCost C = getInstructionCost(I, VF, VectorTy); | |||
6873 | ||||
6874 | bool TypeNotScalarized = false; | |||
6875 | if (VF.isVector() && VectorTy->isVectorTy()) { | |||
6876 | unsigned NumParts = TTI.getNumberOfParts(VectorTy); | |||
6877 | if (NumParts) | |||
6878 | TypeNotScalarized = NumParts < VF.getKnownMinValue(); | |||
6879 | else | |||
6880 | C = InstructionCost::getInvalid(); | |||
6881 | } | |||
6882 | return VectorizationCostTy(C, TypeNotScalarized); | |||
6883 | } | |||
6884 | ||||
6885 | InstructionCost | |||
6886 | LoopVectorizationCostModel::getScalarizationOverhead(Instruction *I, | |||
6887 | ElementCount VF) const { | |||
6888 | ||||
6889 | // There is no mechanism yet to create a scalable scalarization loop, | |||
6890 | // so this is currently Invalid. | |||
6891 | if (VF.isScalable()) | |||
6892 | return InstructionCost::getInvalid(); | |||
6893 | ||||
6894 | if (VF.isScalar()) | |||
6895 | return 0; | |||
6896 | ||||
6897 | InstructionCost Cost = 0; | |||
6898 | Type *RetTy = ToVectorTy(I->getType(), VF); | |||
6899 | if (!RetTy->isVoidTy() && | |||
6900 | (!isa<LoadInst>(I) || !TTI.supportsEfficientVectorElementLoadStore())) | |||
6901 | Cost += TTI.getScalarizationOverhead( | |||
6902 | cast<VectorType>(RetTy), APInt::getAllOnes(VF.getKnownMinValue()), true, | |||
6903 | false); | |||
6904 | ||||
6905 | // Some targets keep addresses scalar. | |||
6906 | if (isa<LoadInst>(I) && !TTI.prefersVectorizedAddressing()) | |||
6907 | return Cost; | |||
6908 | ||||
6909 | // Some targets support efficient element stores. | |||
6910 | if (isa<StoreInst>(I) && TTI.supportsEfficientVectorElementLoadStore()) | |||
6911 | return Cost; | |||
6912 | ||||
6913 | // Collect operands to consider. | |||
6914 | CallInst *CI = dyn_cast<CallInst>(I); | |||
6915 | Instruction::op_range Ops = CI ? CI->args() : I->operands(); | |||
6916 | ||||
6917 | // Skip operands that do not require extraction/scalarization and do not incur | |||
6918 | // any overhead. | |||
6919 | SmallVector<Type *> Tys; | |||
6920 | for (auto *V : filterExtractingOperands(Ops, VF)) | |||
6921 | Tys.push_back(MaybeVectorizeType(V->getType(), VF)); | |||
6922 | return Cost + TTI.getOperandsScalarizationOverhead( | |||
6923 | filterExtractingOperands(Ops, VF), Tys); | |||
6924 | } | |||
6925 | ||||
6926 | void LoopVectorizationCostModel::setCostBasedWideningDecision(ElementCount VF) { | |||
6927 | if (VF.isScalar()) | |||
6928 | return; | |||
6929 | NumPredStores = 0; | |||
6930 | for (BasicBlock *BB : TheLoop->blocks()) { | |||
6931 | // For each instruction in the old loop. | |||
6932 | for (Instruction &I : *BB) { | |||
6933 | Value *Ptr = getLoadStorePointerOperand(&I); | |||
6934 | if (!Ptr) | |||
6935 | continue; | |||
6936 | ||||
6937 | // TODO: We should generate better code and update the cost model for | |||
6938 | // predicated uniform stores. Today they are treated as any other | |||
6939 | // predicated store (see added test cases in | |||
6940 | // invariant-store-vectorization.ll). | |||
6941 | if (isa<StoreInst>(&I) && isScalarWithPredication(&I, VF)) | |||
6942 | NumPredStores++; | |||
6943 | ||||
6944 | if (Legal->isUniformMemOp(I)) { | |||
6945 | // TODO: Avoid replicating loads and stores instead of | |||
6946 | // relying on instcombine to remove them. | |||
6947 | // Load: Scalar load + broadcast | |||
6948 | // Store: Scalar store + isLoopInvariantStoreValue ? 0 : extract | |||
6949 | InstructionCost Cost; | |||
6950 | if (isa<StoreInst>(&I) && VF.isScalable() && | |||
6951 | isLegalGatherOrScatter(&I, VF)) { | |||
6952 | Cost = getGatherScatterCost(&I, VF); | |||
6953 | setWideningDecision(&I, VF, CM_GatherScatter, Cost); | |||
6954 | } else { | |||
6955 | assert((isa<LoadInst>(&I) || !VF.isScalable()) &&(static_cast <bool> ((isa<LoadInst>(&I) || !VF .isScalable()) && "Cannot yet scalarize uniform stores" ) ? void (0) : __assert_fail ("(isa<LoadInst>(&I) || !VF.isScalable()) && \"Cannot yet scalarize uniform stores\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6956, __extension__ __PRETTY_FUNCTION__)) | |||
6956 | "Cannot yet scalarize uniform stores")(static_cast <bool> ((isa<LoadInst>(&I) || !VF .isScalable()) && "Cannot yet scalarize uniform stores" ) ? void (0) : __assert_fail ("(isa<LoadInst>(&I) || !VF.isScalable()) && \"Cannot yet scalarize uniform stores\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6956, __extension__ __PRETTY_FUNCTION__)); | |||
6957 | Cost = getUniformMemOpCost(&I, VF); | |||
6958 | setWideningDecision(&I, VF, CM_Scalarize, Cost); | |||
6959 | } | |||
6960 | continue; | |||
6961 | } | |||
6962 | ||||
6963 | // We assume that widening is the best solution when possible. | |||
6964 | if (memoryInstructionCanBeWidened(&I, VF)) { | |||
6965 | InstructionCost Cost = getConsecutiveMemOpCost(&I, VF); | |||
6966 | int ConsecutiveStride = Legal->isConsecutivePtr( | |||
6967 | getLoadStoreType(&I), getLoadStorePointerOperand(&I)); | |||
6968 | assert((ConsecutiveStride == 1 || ConsecutiveStride == -1) &&(static_cast <bool> ((ConsecutiveStride == 1 || ConsecutiveStride == -1) && "Expected consecutive stride.") ? void (0) : __assert_fail ("(ConsecutiveStride == 1 || ConsecutiveStride == -1) && \"Expected consecutive stride.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6969, __extension__ __PRETTY_FUNCTION__)) | |||
6969 | "Expected consecutive stride.")(static_cast <bool> ((ConsecutiveStride == 1 || ConsecutiveStride == -1) && "Expected consecutive stride.") ? void (0) : __assert_fail ("(ConsecutiveStride == 1 || ConsecutiveStride == -1) && \"Expected consecutive stride.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6969, __extension__ __PRETTY_FUNCTION__)); | |||
6970 | InstWidening Decision = | |||
6971 | ConsecutiveStride == 1 ? CM_Widen : CM_Widen_Reverse; | |||
6972 | setWideningDecision(&I, VF, Decision, Cost); | |||
6973 | continue; | |||
6974 | } | |||
6975 | ||||
6976 | // Choose between Interleaving, Gather/Scatter or Scalarization. | |||
6977 | InstructionCost InterleaveCost = InstructionCost::getInvalid(); | |||
6978 | unsigned NumAccesses = 1; | |||
6979 | if (isAccessInterleaved(&I)) { | |||
6980 | auto Group = getInterleavedAccessGroup(&I); | |||
6981 | assert(Group && "Fail to get an interleaved access group.")(static_cast <bool> (Group && "Fail to get an interleaved access group." ) ? void (0) : __assert_fail ("Group && \"Fail to get an interleaved access group.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6981, __extension__ __PRETTY_FUNCTION__)); | |||
6982 | ||||
6983 | // Make one decision for the whole group. | |||
6984 | if (getWideningDecision(&I, VF) != CM_Unknown) | |||
6985 | continue; | |||
6986 | ||||
6987 | NumAccesses = Group->getNumMembers(); | |||
6988 | if (interleavedAccessCanBeWidened(&I, VF)) | |||
6989 | InterleaveCost = getInterleaveGroupCost(&I, VF); | |||
6990 | } | |||
6991 | ||||
6992 | InstructionCost GatherScatterCost = | |||
6993 | isLegalGatherOrScatter(&I, VF) | |||
6994 | ? getGatherScatterCost(&I, VF) * NumAccesses | |||
6995 | : InstructionCost::getInvalid(); | |||
6996 | ||||
6997 | InstructionCost ScalarizationCost = | |||
6998 | getMemInstScalarizationCost(&I, VF) * NumAccesses; | |||
6999 | ||||
7000 | // Choose better solution for the current VF, | |||
7001 | // write down this decision and use it during vectorization. | |||
7002 | InstructionCost Cost; | |||
7003 | InstWidening Decision; | |||
7004 | if (InterleaveCost <= GatherScatterCost && | |||
7005 | InterleaveCost < ScalarizationCost) { | |||
7006 | Decision = CM_Interleave; | |||
7007 | Cost = InterleaveCost; | |||
7008 | } else if (GatherScatterCost < ScalarizationCost) { | |||
7009 | Decision = CM_GatherScatter; | |||
7010 | Cost = GatherScatterCost; | |||
7011 | } else { | |||
7012 | Decision = CM_Scalarize; | |||
7013 | Cost = ScalarizationCost; | |||
7014 | } | |||
7015 | // If the instructions belongs to an interleave group, the whole group | |||
7016 | // receives the same decision. The whole group receives the cost, but | |||
7017 | // the cost will actually be assigned to one instruction. | |||
7018 | if (auto Group = getInterleavedAccessGroup(&I)) | |||
7019 | setWideningDecision(Group, VF, Decision, Cost); | |||
7020 | else | |||
7021 | setWideningDecision(&I, VF, Decision, Cost); | |||
7022 | } | |||
7023 | } | |||
7024 | ||||
7025 | // Make sure that any load of address and any other address computation | |||
7026 | // remains scalar unless there is gather/scatter support. This avoids | |||
7027 | // inevitable extracts into address registers, and also has the benefit of | |||
7028 | // activating LSR more, since that pass can't optimize vectorized | |||
7029 | // addresses. | |||
7030 | if (TTI.prefersVectorizedAddressing()) | |||
7031 | return; | |||
7032 | ||||
7033 | // Start with all scalar pointer uses. | |||
7034 | SmallPtrSet<Instruction *, 8> AddrDefs; | |||
7035 | for (BasicBlock *BB : TheLoop->blocks()) | |||
7036 | for (Instruction &I : *BB) { | |||
7037 | Instruction *PtrDef = | |||
7038 | dyn_cast_or_null<Instruction>(getLoadStorePointerOperand(&I)); | |||
7039 | if (PtrDef && TheLoop->contains(PtrDef) && | |||
7040 | getWideningDecision(&I, VF) != CM_GatherScatter) | |||
7041 | AddrDefs.insert(PtrDef); | |||
7042 | } | |||
7043 | ||||
7044 | // Add all instructions used to generate the addresses. | |||
7045 | SmallVector<Instruction *, 4> Worklist; | |||
7046 | append_range(Worklist, AddrDefs); | |||
7047 | while (!Worklist.empty()) { | |||
7048 | Instruction *I = Worklist.pop_back_val(); | |||
7049 | for (auto &Op : I->operands()) | |||
7050 | if (auto *InstOp = dyn_cast<Instruction>(Op)) | |||
7051 | if ((InstOp->getParent() == I->getParent()) && !isa<PHINode>(InstOp) && | |||
7052 | AddrDefs.insert(InstOp).second) | |||
7053 | Worklist.push_back(InstOp); | |||
7054 | } | |||
7055 | ||||
7056 | for (auto *I : AddrDefs) { | |||
7057 | if (isa<LoadInst>(I)) { | |||
7058 | // Setting the desired widening decision should ideally be handled in | |||
7059 | // by cost functions, but since this involves the task of finding out | |||
7060 | // if the loaded register is involved in an address computation, it is | |||
7061 | // instead changed here when we know this is the case. | |||
7062 | InstWidening Decision = getWideningDecision(I, VF); | |||
7063 | if (Decision == CM_Widen || Decision == CM_Widen_Reverse) | |||
7064 | // Scalarize a widened load of address. | |||
7065 | setWideningDecision( | |||
7066 | I, VF, CM_Scalarize, | |||
7067 | (VF.getKnownMinValue() * | |||
7068 | getMemoryInstructionCost(I, ElementCount::getFixed(1)))); | |||
7069 | else if (auto Group = getInterleavedAccessGroup(I)) { | |||
7070 | // Scalarize an interleave group of address loads. | |||
7071 | for (unsigned I = 0; I < Group->getFactor(); ++I) { | |||
7072 | if (Instruction *Member = Group->getMember(I)) | |||
7073 | setWideningDecision( | |||
7074 | Member, VF, CM_Scalarize, | |||
7075 | (VF.getKnownMinValue() * | |||
7076 | getMemoryInstructionCost(Member, ElementCount::getFixed(1)))); | |||
7077 | } | |||
7078 | } | |||
7079 | } else | |||
7080 | // Make sure I gets scalarized and a cost estimate without | |||
7081 | // scalarization overhead. | |||
7082 | ForcedScalars[VF].insert(I); | |||
7083 | } | |||
7084 | } | |||
7085 | ||||
7086 | InstructionCost | |||
7087 | LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF, | |||
7088 | Type *&VectorTy) { | |||
7089 | Type *RetTy = I->getType(); | |||
7090 | if (canTruncateToMinimalBitwidth(I, VF)) | |||
7091 | RetTy = IntegerType::get(RetTy->getContext(), MinBWs[I]); | |||
7092 | auto SE = PSE.getSE(); | |||
7093 | TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | |||
7094 | ||||
7095 | auto hasSingleCopyAfterVectorization = [this](Instruction *I, | |||
7096 | ElementCount VF) -> bool { | |||
7097 | if (VF.isScalar()) | |||
7098 | return true; | |||
7099 | ||||
7100 | auto Scalarized = InstsToScalarize.find(VF); | |||
7101 | assert(Scalarized != InstsToScalarize.end() &&(static_cast <bool> (Scalarized != InstsToScalarize.end () && "VF not yet analyzed for scalarization profitability" ) ? void (0) : __assert_fail ("Scalarized != InstsToScalarize.end() && \"VF not yet analyzed for scalarization profitability\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7102, __extension__ __PRETTY_FUNCTION__)) | |||
7102 | "VF not yet analyzed for scalarization profitability")(static_cast <bool> (Scalarized != InstsToScalarize.end () && "VF not yet analyzed for scalarization profitability" ) ? void (0) : __assert_fail ("Scalarized != InstsToScalarize.end() && \"VF not yet analyzed for scalarization profitability\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7102, __extension__ __PRETTY_FUNCTION__)); | |||
7103 | return !Scalarized->second.count(I) && | |||
7104 | llvm::all_of(I->users(), [&](User *U) { | |||
7105 | auto *UI = cast<Instruction>(U); | |||
7106 | return !Scalarized->second.count(UI); | |||
7107 | }); | |||
7108 | }; | |||
7109 | (void) hasSingleCopyAfterVectorization; | |||
7110 | ||||
7111 | if (isScalarAfterVectorization(I, VF)) { | |||
7112 | // With the exception of GEPs and PHIs, after scalarization there should | |||
7113 | // only be one copy of the instruction generated in the loop. This is | |||
7114 | // because the VF is either 1, or any instructions that need scalarizing | |||
7115 | // have already been dealt with by the the time we get here. As a result, | |||
7116 | // it means we don't have to multiply the instruction cost by VF. | |||
7117 | assert(I->getOpcode() == Instruction::GetElementPtr ||(static_cast <bool> (I->getOpcode() == Instruction:: GetElementPtr || I->getOpcode() == Instruction::PHI || (I-> getOpcode() == Instruction::BitCast && I->getType( )->isPointerTy()) || hasSingleCopyAfterVectorization(I, VF )) ? void (0) : __assert_fail ("I->getOpcode() == Instruction::GetElementPtr || I->getOpcode() == Instruction::PHI || (I->getOpcode() == Instruction::BitCast && I->getType()->isPointerTy()) || hasSingleCopyAfterVectorization(I, VF)" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7121, __extension__ __PRETTY_FUNCTION__)) | |||
7118 | I->getOpcode() == Instruction::PHI ||(static_cast <bool> (I->getOpcode() == Instruction:: GetElementPtr || I->getOpcode() == Instruction::PHI || (I-> getOpcode() == Instruction::BitCast && I->getType( )->isPointerTy()) || hasSingleCopyAfterVectorization(I, VF )) ? void (0) : __assert_fail ("I->getOpcode() == Instruction::GetElementPtr || I->getOpcode() == Instruction::PHI || (I->getOpcode() == Instruction::BitCast && I->getType()->isPointerTy()) || hasSingleCopyAfterVectorization(I, VF)" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7121, __extension__ __PRETTY_FUNCTION__)) | |||
7119 | (I->getOpcode() == Instruction::BitCast &&(static_cast <bool> (I->getOpcode() == Instruction:: GetElementPtr || I->getOpcode() == Instruction::PHI || (I-> getOpcode() == Instruction::BitCast && I->getType( )->isPointerTy()) || hasSingleCopyAfterVectorization(I, VF )) ? void (0) : __assert_fail ("I->getOpcode() == Instruction::GetElementPtr || I->getOpcode() == Instruction::PHI || (I->getOpcode() == Instruction::BitCast && I->getType()->isPointerTy()) || hasSingleCopyAfterVectorization(I, VF)" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7121, __extension__ __PRETTY_FUNCTION__)) | |||
7120 | I->getType()->isPointerTy()) ||(static_cast <bool> (I->getOpcode() == Instruction:: GetElementPtr || I->getOpcode() == Instruction::PHI || (I-> getOpcode() == Instruction::BitCast && I->getType( )->isPointerTy()) || hasSingleCopyAfterVectorization(I, VF )) ? void (0) : __assert_fail ("I->getOpcode() == Instruction::GetElementPtr || I->getOpcode() == Instruction::PHI || (I->getOpcode() == Instruction::BitCast && I->getType()->isPointerTy()) || hasSingleCopyAfterVectorization(I, VF)" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7121, __extension__ __PRETTY_FUNCTION__)) | |||
7121 | hasSingleCopyAfterVectorization(I, VF))(static_cast <bool> (I->getOpcode() == Instruction:: GetElementPtr || I->getOpcode() == Instruction::PHI || (I-> getOpcode() == Instruction::BitCast && I->getType( )->isPointerTy()) || hasSingleCopyAfterVectorization(I, VF )) ? void (0) : __assert_fail ("I->getOpcode() == Instruction::GetElementPtr || I->getOpcode() == Instruction::PHI || (I->getOpcode() == Instruction::BitCast && I->getType()->isPointerTy()) || hasSingleCopyAfterVectorization(I, VF)" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7121, __extension__ __PRETTY_FUNCTION__)); | |||
7122 | VectorTy = RetTy; | |||
7123 | } else | |||
7124 | VectorTy = ToVectorTy(RetTy, VF); | |||
7125 | ||||
7126 | // TODO: We need to estimate the cost of intrinsic calls. | |||
7127 | switch (I->getOpcode()) { | |||
7128 | case Instruction::GetElementPtr: | |||
7129 | // We mark this instruction as zero-cost because the cost of GEPs in | |||
7130 | // vectorized code depends on whether the corresponding memory instruction | |||
7131 | // is scalarized or not. Therefore, we handle GEPs with the memory | |||
7132 | // instruction cost. | |||
7133 | return 0; | |||
7134 | case Instruction::Br: { | |||
7135 | // In cases of scalarized and predicated instructions, there will be VF | |||
7136 | // predicated blocks in the vectorized loop. Each branch around these | |||
7137 | // blocks requires also an extract of its vector compare i1 element. | |||
7138 | bool ScalarPredicatedBB = false; | |||
7139 | BranchInst *BI = cast<BranchInst>(I); | |||
7140 | if (VF.isVector() && BI->isConditional() && | |||
7141 | (PredicatedBBsAfterVectorization.count(BI->getSuccessor(0)) || | |||
7142 | PredicatedBBsAfterVectorization.count(BI->getSuccessor(1)))) | |||
7143 | ScalarPredicatedBB = true; | |||
7144 | ||||
7145 | if (ScalarPredicatedBB) { | |||
7146 | // Not possible to scalarize scalable vector with predicated instructions. | |||
7147 | if (VF.isScalable()) | |||
7148 | return InstructionCost::getInvalid(); | |||
7149 | // Return cost for branches around scalarized and predicated blocks. | |||
7150 | auto *Vec_i1Ty = | |||
7151 | VectorType::get(IntegerType::getInt1Ty(RetTy->getContext()), VF); | |||
7152 | return ( | |||
7153 | TTI.getScalarizationOverhead( | |||
7154 | Vec_i1Ty, APInt::getAllOnes(VF.getFixedValue()), false, true) + | |||
7155 | (TTI.getCFInstrCost(Instruction::Br, CostKind) * VF.getFixedValue())); | |||
7156 | } else if (I->getParent() == TheLoop->getLoopLatch() || VF.isScalar()) | |||
7157 | // The back-edge branch will remain, as will all scalar branches. | |||
7158 | return TTI.getCFInstrCost(Instruction::Br, CostKind); | |||
7159 | else | |||
7160 | // This branch will be eliminated by if-conversion. | |||
7161 | return 0; | |||
7162 | // Note: We currently assume zero cost for an unconditional branch inside | |||
7163 | // a predicated block since it will become a fall-through, although we | |||
7164 | // may decide in the future to call TTI for all branches. | |||
7165 | } | |||
7166 | case Instruction::PHI: { | |||
7167 | auto *Phi = cast<PHINode>(I); | |||
7168 | ||||
7169 | // First-order recurrences are replaced by vector shuffles inside the loop. | |||
7170 | // NOTE: Don't use ToVectorTy as SK_ExtractSubvector expects a vector type. | |||
7171 | if (VF.isVector() && Legal->isFirstOrderRecurrence(Phi)) | |||
7172 | return TTI.getShuffleCost( | |||
7173 | TargetTransformInfo::SK_ExtractSubvector, cast<VectorType>(VectorTy), | |||
7174 | None, VF.getKnownMinValue() - 1, FixedVectorType::get(RetTy, 1)); | |||
7175 | ||||
7176 | // Phi nodes in non-header blocks (not inductions, reductions, etc.) are | |||
7177 | // converted into select instructions. We require N - 1 selects per phi | |||
7178 | // node, where N is the number of incoming values. | |||
7179 | if (VF.isVector() && Phi->getParent() != TheLoop->getHeader()) | |||
7180 | return (Phi->getNumIncomingValues() - 1) * | |||
7181 | TTI.getCmpSelInstrCost( | |||
7182 | Instruction::Select, ToVectorTy(Phi->getType(), VF), | |||
7183 | ToVectorTy(Type::getInt1Ty(Phi->getContext()), VF), | |||
7184 | CmpInst::BAD_ICMP_PREDICATE, CostKind); | |||
7185 | ||||
7186 | return TTI.getCFInstrCost(Instruction::PHI, CostKind); | |||
7187 | } | |||
7188 | case Instruction::UDiv: | |||
7189 | case Instruction::SDiv: | |||
7190 | case Instruction::URem: | |||
7191 | case Instruction::SRem: | |||
7192 | // If we have a predicated instruction, it may not be executed for each | |||
7193 | // vector lane. Get the scalarization cost and scale this amount by the | |||
7194 | // probability of executing the predicated block. If the instruction is not | |||
7195 | // predicated, we fall through to the next case. | |||
7196 | if (VF.isVector() && isScalarWithPredication(I, VF)) { | |||
7197 | InstructionCost Cost = 0; | |||
7198 | ||||
7199 | // These instructions have a non-void type, so account for the phi nodes | |||
7200 | // that we will create. This cost is likely to be zero. The phi node | |||
7201 | // cost, if any, should be scaled by the block probability because it | |||
7202 | // models a copy at the end of each predicated block. | |||
7203 | Cost += VF.getKnownMinValue() * | |||
7204 | TTI.getCFInstrCost(Instruction::PHI, CostKind); | |||
7205 | ||||
7206 | // The cost of the non-predicated instruction. | |||
7207 | Cost += VF.getKnownMinValue() * | |||
7208 | TTI.getArithmeticInstrCost(I->getOpcode(), RetTy, CostKind); | |||
7209 | ||||
7210 | // The cost of insertelement and extractelement instructions needed for | |||
7211 | // scalarization. | |||
7212 | Cost += getScalarizationOverhead(I, VF); | |||
7213 | ||||
7214 | // Scale the cost by the probability of executing the predicated blocks. | |||
7215 | // This assumes the predicated block for each vector lane is equally | |||
7216 | // likely. | |||
7217 | return Cost / getReciprocalPredBlockProb(); | |||
7218 | } | |||
7219 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
7220 | case Instruction::Add: | |||
7221 | case Instruction::FAdd: | |||
7222 | case Instruction::Sub: | |||
7223 | case Instruction::FSub: | |||
7224 | case Instruction::Mul: | |||
7225 | case Instruction::FMul: | |||
7226 | case Instruction::FDiv: | |||
7227 | case Instruction::FRem: | |||
7228 | case Instruction::Shl: | |||
7229 | case Instruction::LShr: | |||
7230 | case Instruction::AShr: | |||
7231 | case Instruction::And: | |||
7232 | case Instruction::Or: | |||
7233 | case Instruction::Xor: { | |||
7234 | // Since we will replace the stride by 1 the multiplication should go away. | |||
7235 | if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal)) | |||
7236 | return 0; | |||
7237 | ||||
7238 | // Detect reduction patterns | |||
7239 | if (auto RedCost = getReductionPatternCost(I, VF, VectorTy, CostKind)) | |||
7240 | return *RedCost; | |||
7241 | ||||
7242 | // Certain instructions can be cheaper to vectorize if they have a constant | |||
7243 | // second vector operand. One example of this are shifts on x86. | |||
7244 | Value *Op2 = I->getOperand(1); | |||
7245 | TargetTransformInfo::OperandValueProperties Op2VP; | |||
7246 | TargetTransformInfo::OperandValueKind Op2VK = | |||
7247 | TTI.getOperandInfo(Op2, Op2VP); | |||
7248 | if (Op2VK == TargetTransformInfo::OK_AnyValue && Legal->isUniform(Op2)) | |||
7249 | Op2VK = TargetTransformInfo::OK_UniformValue; | |||
7250 | ||||
7251 | SmallVector<const Value *, 4> Operands(I->operand_values()); | |||
7252 | return TTI.getArithmeticInstrCost( | |||
7253 | I->getOpcode(), VectorTy, CostKind, TargetTransformInfo::OK_AnyValue, | |||
7254 | Op2VK, TargetTransformInfo::OP_None, Op2VP, Operands, I); | |||
7255 | } | |||
7256 | case Instruction::FNeg: { | |||
7257 | return TTI.getArithmeticInstrCost( | |||
7258 | I->getOpcode(), VectorTy, CostKind, TargetTransformInfo::OK_AnyValue, | |||
7259 | TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None, | |||
7260 | TargetTransformInfo::OP_None, I->getOperand(0), I); | |||
7261 | } | |||
7262 | case Instruction::Select: { | |||
7263 | SelectInst *SI = cast<SelectInst>(I); | |||
7264 | const SCEV *CondSCEV = SE->getSCEV(SI->getCondition()); | |||
7265 | bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop)); | |||
7266 | ||||
7267 | const Value *Op0, *Op1; | |||
7268 | using namespace llvm::PatternMatch; | |||
7269 | if (!ScalarCond && (match(I, m_LogicalAnd(m_Value(Op0), m_Value(Op1))) || | |||
7270 | match(I, m_LogicalOr(m_Value(Op0), m_Value(Op1))))) { | |||
7271 | // select x, y, false --> x & y | |||
7272 | // select x, true, y --> x | y | |||
7273 | TTI::OperandValueProperties Op1VP = TTI::OP_None; | |||
7274 | TTI::OperandValueProperties Op2VP = TTI::OP_None; | |||
7275 | TTI::OperandValueKind Op1VK = TTI::getOperandInfo(Op0, Op1VP); | |||
7276 | TTI::OperandValueKind Op2VK = TTI::getOperandInfo(Op1, Op2VP); | |||
7277 | assert(Op0->getType()->getScalarSizeInBits() == 1 &&(static_cast <bool> (Op0->getType()->getScalarSizeInBits () == 1 && Op1->getType()->getScalarSizeInBits( ) == 1) ? void (0) : __assert_fail ("Op0->getType()->getScalarSizeInBits() == 1 && Op1->getType()->getScalarSizeInBits() == 1" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7278, __extension__ __PRETTY_FUNCTION__)) | |||
7278 | Op1->getType()->getScalarSizeInBits() == 1)(static_cast <bool> (Op0->getType()->getScalarSizeInBits () == 1 && Op1->getType()->getScalarSizeInBits( ) == 1) ? void (0) : __assert_fail ("Op0->getType()->getScalarSizeInBits() == 1 && Op1->getType()->getScalarSizeInBits() == 1" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7278, __extension__ __PRETTY_FUNCTION__)); | |||
7279 | ||||
7280 | SmallVector<const Value *, 2> Operands{Op0, Op1}; | |||
7281 | return TTI.getArithmeticInstrCost( | |||
7282 | match(I, m_LogicalOr()) ? Instruction::Or : Instruction::And, VectorTy, | |||
7283 | CostKind, Op1VK, Op2VK, Op1VP, Op2VP, Operands, I); | |||
7284 | } | |||
7285 | ||||
7286 | Type *CondTy = SI->getCondition()->getType(); | |||
7287 | if (!ScalarCond) | |||
7288 | CondTy = VectorType::get(CondTy, VF); | |||
7289 | ||||
7290 | CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE; | |||
7291 | if (auto *Cmp = dyn_cast<CmpInst>(SI->getCondition())) | |||
7292 | Pred = Cmp->getPredicate(); | |||
7293 | return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy, Pred, | |||
7294 | CostKind, I); | |||
7295 | } | |||
7296 | case Instruction::ICmp: | |||
7297 | case Instruction::FCmp: { | |||
7298 | Type *ValTy = I->getOperand(0)->getType(); | |||
7299 | Instruction *Op0AsInstruction = dyn_cast<Instruction>(I->getOperand(0)); | |||
7300 | if (canTruncateToMinimalBitwidth(Op0AsInstruction, VF)) | |||
7301 | ValTy = IntegerType::get(ValTy->getContext(), MinBWs[Op0AsInstruction]); | |||
7302 | VectorTy = ToVectorTy(ValTy, VF); | |||
7303 | return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, nullptr, | |||
7304 | cast<CmpInst>(I)->getPredicate(), CostKind, | |||
7305 | I); | |||
7306 | } | |||
7307 | case Instruction::Store: | |||
7308 | case Instruction::Load: { | |||
7309 | ElementCount Width = VF; | |||
7310 | if (Width.isVector()) { | |||
7311 | InstWidening Decision = getWideningDecision(I, Width); | |||
7312 | assert(Decision != CM_Unknown &&(static_cast <bool> (Decision != CM_Unknown && "CM decision should be taken at this point" ) ? void (0) : __assert_fail ("Decision != CM_Unknown && \"CM decision should be taken at this point\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7313, __extension__ __PRETTY_FUNCTION__)) | |||
7313 | "CM decision should be taken at this point")(static_cast <bool> (Decision != CM_Unknown && "CM decision should be taken at this point" ) ? void (0) : __assert_fail ("Decision != CM_Unknown && \"CM decision should be taken at this point\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7313, __extension__ __PRETTY_FUNCTION__)); | |||
7314 | if (Decision == CM_Scalarize) | |||
7315 | Width = ElementCount::getFixed(1); | |||
7316 | } | |||
7317 | VectorTy = ToVectorTy(getLoadStoreType(I), Width); | |||
7318 | return getMemoryInstructionCost(I, VF); | |||
7319 | } | |||
7320 | case Instruction::BitCast: | |||
7321 | if (I->getType()->isPointerTy()) | |||
7322 | return 0; | |||
7323 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
7324 | case Instruction::ZExt: | |||
7325 | case Instruction::SExt: | |||
7326 | case Instruction::FPToUI: | |||
7327 | case Instruction::FPToSI: | |||
7328 | case Instruction::FPExt: | |||
7329 | case Instruction::PtrToInt: | |||
7330 | case Instruction::IntToPtr: | |||
7331 | case Instruction::SIToFP: | |||
7332 | case Instruction::UIToFP: | |||
7333 | case Instruction::Trunc: | |||
7334 | case Instruction::FPTrunc: { | |||
7335 | // Computes the CastContextHint from a Load/Store instruction. | |||
7336 | auto ComputeCCH = [&](Instruction *I) -> TTI::CastContextHint { | |||
7337 | assert((isa<LoadInst>(I) || isa<StoreInst>(I)) &&(static_cast <bool> ((isa<LoadInst>(I) || isa< StoreInst>(I)) && "Expected a load or a store!") ? void (0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Expected a load or a store!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7338, __extension__ __PRETTY_FUNCTION__)) | |||
7338 | "Expected a load or a store!")(static_cast <bool> ((isa<LoadInst>(I) || isa< StoreInst>(I)) && "Expected a load or a store!") ? void (0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Expected a load or a store!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7338, __extension__ __PRETTY_FUNCTION__)); | |||
7339 | ||||
7340 | if (VF.isScalar() || !TheLoop->contains(I)) | |||
7341 | return TTI::CastContextHint::Normal; | |||
7342 | ||||
7343 | switch (getWideningDecision(I, VF)) { | |||
7344 | case LoopVectorizationCostModel::CM_GatherScatter: | |||
7345 | return TTI::CastContextHint::GatherScatter; | |||
7346 | case LoopVectorizationCostModel::CM_Interleave: | |||
7347 | return TTI::CastContextHint::Interleave; | |||
7348 | case LoopVectorizationCostModel::CM_Scalarize: | |||
7349 | case LoopVectorizationCostModel::CM_Widen: | |||
7350 | return Legal->isMaskRequired(I) ? TTI::CastContextHint::Masked | |||
7351 | : TTI::CastContextHint::Normal; | |||
7352 | case LoopVectorizationCostModel::CM_Widen_Reverse: | |||
7353 | return TTI::CastContextHint::Reversed; | |||
7354 | case LoopVectorizationCostModel::CM_Unknown: | |||
7355 | llvm_unreachable("Instr did not go through cost modelling?")::llvm::llvm_unreachable_internal("Instr did not go through cost modelling?" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7355); | |||
7356 | } | |||
7357 | ||||
7358 | llvm_unreachable("Unhandled case!")::llvm::llvm_unreachable_internal("Unhandled case!", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7358); | |||
7359 | }; | |||
7360 | ||||
7361 | unsigned Opcode = I->getOpcode(); | |||
7362 | TTI::CastContextHint CCH = TTI::CastContextHint::None; | |||
7363 | // For Trunc, the context is the only user, which must be a StoreInst. | |||
7364 | if (Opcode == Instruction::Trunc || Opcode == Instruction::FPTrunc) { | |||
7365 | if (I->hasOneUse()) | |||
7366 | if (StoreInst *Store = dyn_cast<StoreInst>(*I->user_begin())) | |||
7367 | CCH = ComputeCCH(Store); | |||
7368 | } | |||
7369 | // For Z/Sext, the context is the operand, which must be a LoadInst. | |||
7370 | else if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt || | |||
7371 | Opcode == Instruction::FPExt) { | |||
7372 | if (LoadInst *Load = dyn_cast<LoadInst>(I->getOperand(0))) | |||
7373 | CCH = ComputeCCH(Load); | |||
7374 | } | |||
7375 | ||||
7376 | // We optimize the truncation of induction variables having constant | |||
7377 | // integer steps. The cost of these truncations is the same as the scalar | |||
7378 | // operation. | |||
7379 | if (isOptimizableIVTruncate(I, VF)) { | |||
7380 | auto *Trunc = cast<TruncInst>(I); | |||
7381 | return TTI.getCastInstrCost(Instruction::Trunc, Trunc->getDestTy(), | |||
7382 | Trunc->getSrcTy(), CCH, CostKind, Trunc); | |||
7383 | } | |||
7384 | ||||
7385 | // Detect reduction patterns | |||
7386 | if (auto RedCost = getReductionPatternCost(I, VF, VectorTy, CostKind)) | |||
7387 | return *RedCost; | |||
7388 | ||||
7389 | Type *SrcScalarTy = I->getOperand(0)->getType(); | |||
7390 | Type *SrcVecTy = | |||
7391 | VectorTy->isVectorTy() ? ToVectorTy(SrcScalarTy, VF) : SrcScalarTy; | |||
7392 | if (canTruncateToMinimalBitwidth(I, VF)) { | |||
7393 | // This cast is going to be shrunk. This may remove the cast or it might | |||
7394 | // turn it into slightly different cast. For example, if MinBW == 16, | |||
7395 | // "zext i8 %1 to i32" becomes "zext i8 %1 to i16". | |||
7396 | // | |||
7397 | // Calculate the modified src and dest types. | |||
7398 | Type *MinVecTy = VectorTy; | |||
7399 | if (Opcode == Instruction::Trunc) { | |||
7400 | SrcVecTy = smallestIntegerVectorType(SrcVecTy, MinVecTy); | |||
7401 | VectorTy = | |||
7402 | largestIntegerVectorType(ToVectorTy(I->getType(), VF), MinVecTy); | |||
7403 | } else if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) { | |||
7404 | SrcVecTy = largestIntegerVectorType(SrcVecTy, MinVecTy); | |||
7405 | VectorTy = | |||
7406 | smallestIntegerVectorType(ToVectorTy(I->getType(), VF), MinVecTy); | |||
7407 | } | |||
7408 | } | |||
7409 | ||||
7410 | return TTI.getCastInstrCost(Opcode, VectorTy, SrcVecTy, CCH, CostKind, I); | |||
7411 | } | |||
7412 | case Instruction::Call: { | |||
7413 | if (RecurrenceDescriptor::isFMulAddIntrinsic(I)) | |||
7414 | if (auto RedCost = getReductionPatternCost(I, VF, VectorTy, CostKind)) | |||
7415 | return *RedCost; | |||
7416 | bool NeedToScalarize; | |||
7417 | CallInst *CI = cast<CallInst>(I); | |||
7418 | InstructionCost CallCost = getVectorCallCost(CI, VF, NeedToScalarize); | |||
7419 | if (getVectorIntrinsicIDForCall(CI, TLI)) { | |||
7420 | InstructionCost IntrinsicCost = getVectorIntrinsicCost(CI, VF); | |||
7421 | return std::min(CallCost, IntrinsicCost); | |||
7422 | } | |||
7423 | return CallCost; | |||
7424 | } | |||
7425 | case Instruction::ExtractValue: | |||
7426 | return TTI.getInstructionCost(I, TTI::TCK_RecipThroughput); | |||
7427 | case Instruction::Alloca: | |||
7428 | // We cannot easily widen alloca to a scalable alloca, as | |||
7429 | // the result would need to be a vector of pointers. | |||
7430 | if (VF.isScalable()) | |||
7431 | return InstructionCost::getInvalid(); | |||
7432 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
7433 | default: | |||
7434 | // This opcode is unknown. Assume that it is the same as 'mul'. | |||
7435 | return TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind); | |||
7436 | } // end of switch. | |||
7437 | } | |||
7438 | ||||
7439 | char LoopVectorize::ID = 0; | |||
7440 | ||||
7441 | static const char lv_name[] = "Loop Vectorization"; | |||
7442 | ||||
7443 | INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)static void *initializeLoopVectorizePassOnce(PassRegistry & Registry) { | |||
7444 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); | |||
7445 | INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)initializeBasicAAWrapperPassPass(Registry); | |||
7446 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry); | |||
7447 | INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)initializeGlobalsAAWrapperPassPass(Registry); | |||
7448 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry); | |||
7449 | INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)initializeBlockFrequencyInfoWrapperPassPass(Registry); | |||
7450 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry); | |||
7451 | INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry); | |||
7452 | INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry); | |||
7453 | INITIALIZE_PASS_DEPENDENCY(LoopAccessLegacyAnalysis)initializeLoopAccessLegacyAnalysisPass(Registry); | |||
7454 | INITIALIZE_PASS_DEPENDENCY(DemandedBitsWrapperPass)initializeDemandedBitsWrapperPassPass(Registry); | |||
7455 | INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)initializeOptimizationRemarkEmitterWrapperPassPass(Registry); | |||
7456 | INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)initializeProfileSummaryInfoWrapperPassPass(Registry); | |||
7457 | INITIALIZE_PASS_DEPENDENCY(InjectTLIMappingsLegacy)initializeInjectTLIMappingsLegacyPass(Registry); | |||
7458 | INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)PassInfo *PI = new PassInfo( lv_name, "loop-vectorize", & LoopVectorize::ID, PassInfo::NormalCtor_t(callDefaultCtor< LoopVectorize>), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeLoopVectorizePassFlag ; void llvm::initializeLoopVectorizePass(PassRegistry &Registry ) { llvm::call_once(InitializeLoopVectorizePassFlag, initializeLoopVectorizePassOnce , std::ref(Registry)); } | |||
7459 | ||||
7460 | namespace llvm { | |||
7461 | ||||
7462 | Pass *createLoopVectorizePass() { return new LoopVectorize(); } | |||
7463 | ||||
7464 | Pass *createLoopVectorizePass(bool InterleaveOnlyWhenForced, | |||
7465 | bool VectorizeOnlyWhenForced) { | |||
7466 | return new LoopVectorize(InterleaveOnlyWhenForced, VectorizeOnlyWhenForced); | |||
7467 | } | |||
7468 | ||||
7469 | } // end namespace llvm | |||
7470 | ||||
7471 | bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) { | |||
7472 | // Check if the pointer operand of a load or store instruction is | |||
7473 | // consecutive. | |||
7474 | if (auto *Ptr = getLoadStorePointerOperand(Inst)) | |||
7475 | return Legal->isConsecutivePtr(getLoadStoreType(Inst), Ptr); | |||
7476 | return false; | |||
7477 | } | |||
7478 | ||||
7479 | void LoopVectorizationCostModel::collectValuesToIgnore() { | |||
7480 | // Ignore ephemeral values. | |||
7481 | CodeMetrics::collectEphemeralValues(TheLoop, AC, ValuesToIgnore); | |||
7482 | ||||
7483 | // Ignore type-promoting instructions we identified during reduction | |||
7484 | // detection. | |||
7485 | for (auto &Reduction : Legal->getReductionVars()) { | |||
7486 | const RecurrenceDescriptor &RedDes = Reduction.second; | |||
7487 | const SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts(); | |||
7488 | VecValuesToIgnore.insert(Casts.begin(), Casts.end()); | |||
7489 | } | |||
7490 | // Ignore type-casting instructions we identified during induction | |||
7491 | // detection. | |||
7492 | for (auto &Induction : Legal->getInductionVars()) { | |||
7493 | const InductionDescriptor &IndDes = Induction.second; | |||
7494 | const SmallVectorImpl<Instruction *> &Casts = IndDes.getCastInsts(); | |||
7495 | VecValuesToIgnore.insert(Casts.begin(), Casts.end()); | |||
7496 | } | |||
7497 | } | |||
7498 | ||||
7499 | void LoopVectorizationCostModel::collectInLoopReductions() { | |||
7500 | for (auto &Reduction : Legal->getReductionVars()) { | |||
7501 | PHINode *Phi = Reduction.first; | |||
7502 | const RecurrenceDescriptor &RdxDesc = Reduction.second; | |||
7503 | ||||
7504 | // We don't collect reductions that are type promoted (yet). | |||
7505 | if (RdxDesc.getRecurrenceType() != Phi->getType()) | |||
7506 | continue; | |||
7507 | ||||
7508 | // If the target would prefer this reduction to happen "in-loop", then we | |||
7509 | // want to record it as such. | |||
7510 | unsigned Opcode = RdxDesc.getOpcode(); | |||
7511 | if (!PreferInLoopReductions && !useOrderedReductions(RdxDesc) && | |||
7512 | !TTI.preferInLoopReduction(Opcode, Phi->getType(), | |||
7513 | TargetTransformInfo::ReductionFlags())) | |||
7514 | continue; | |||
7515 | ||||
7516 | // Check that we can correctly put the reductions into the loop, by | |||
7517 | // finding the chain of operations that leads from the phi to the loop | |||
7518 | // exit value. | |||
7519 | SmallVector<Instruction *, 4> ReductionOperations = | |||
7520 | RdxDesc.getReductionOpChain(Phi, TheLoop); | |||
7521 | bool InLoop = !ReductionOperations.empty(); | |||
7522 | if (InLoop) { | |||
7523 | InLoopReductionChains[Phi] = ReductionOperations; | |||
7524 | // Add the elements to InLoopReductionImmediateChains for cost modelling. | |||
7525 | Instruction *LastChain = Phi; | |||
7526 | for (auto *I : ReductionOperations) { | |||
7527 | InLoopReductionImmediateChains[I] = LastChain; | |||
7528 | LastChain = I; | |||
7529 | } | |||
7530 | } | |||
7531 | LLVM_DEBUG(dbgs() << "LV: Using " << (InLoop ? "inloop" : "out of loop")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Using " << ( InLoop ? "inloop" : "out of loop") << " reduction for phi: " << *Phi << "\n"; } } while (false) | |||
7532 | << " reduction for phi: " << *Phi << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Using " << ( InLoop ? "inloop" : "out of loop") << " reduction for phi: " << *Phi << "\n"; } } while (false); | |||
7533 | } | |||
7534 | } | |||
7535 | ||||
7536 | // TODO: we could return a pair of values that specify the max VF and | |||
7537 | // min VF, to be used in `buildVPlans(MinVF, MaxVF)` instead of | |||
7538 | // `buildVPlans(VF, VF)`. We cannot do it because VPLAN at the moment | |||
7539 | // doesn't have a cost model that can choose which plan to execute if | |||
7540 | // more than one is generated. | |||
7541 | static unsigned determineVPlanVF(const unsigned WidestVectorRegBits, | |||
7542 | LoopVectorizationCostModel &CM) { | |||
7543 | unsigned WidestType; | |||
7544 | std::tie(std::ignore, WidestType) = CM.getSmallestAndWidestTypes(); | |||
7545 | return WidestVectorRegBits / WidestType; | |||
7546 | } | |||
7547 | ||||
7548 | VectorizationFactor | |||
7549 | LoopVectorizationPlanner::planInVPlanNativePath(ElementCount UserVF) { | |||
7550 | assert(!UserVF.isScalable() && "scalable vectors not yet supported")(static_cast <bool> (!UserVF.isScalable() && "scalable vectors not yet supported" ) ? void (0) : __assert_fail ("!UserVF.isScalable() && \"scalable vectors not yet supported\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7550, __extension__ __PRETTY_FUNCTION__)); | |||
7551 | ElementCount VF = UserVF; | |||
7552 | // Outer loop handling: They may require CFG and instruction level | |||
7553 | // transformations before even evaluating whether vectorization is profitable. | |||
7554 | // Since we cannot modify the incoming IR, we need to build VPlan upfront in | |||
7555 | // the vectorization pipeline. | |||
7556 | if (!OrigLoop->isInnermost()) { | |||
7557 | // If the user doesn't provide a vectorization factor, determine a | |||
7558 | // reasonable one. | |||
7559 | if (UserVF.isZero()) { | |||
7560 | VF = ElementCount::getFixed(determineVPlanVF( | |||
7561 | TTI->getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector) | |||
7562 | .getFixedSize(), | |||
7563 | CM)); | |||
7564 | LLVM_DEBUG(dbgs() << "LV: VPlan computed VF " << VF << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: VPlan computed VF " << VF << ".\n"; } } while (false); | |||
7565 | ||||
7566 | // Make sure we have a VF > 1 for stress testing. | |||
7567 | if (VPlanBuildStressTest && (VF.isScalar() || VF.isZero())) { | |||
7568 | LLVM_DEBUG(dbgs() << "LV: VPlan stress testing: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: VPlan stress testing: " << "overriding computed VF.\n"; } } while (false) | |||
7569 | << "overriding computed VF.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: VPlan stress testing: " << "overriding computed VF.\n"; } } while (false); | |||
7570 | VF = ElementCount::getFixed(4); | |||
7571 | } | |||
7572 | } | |||
7573 | assert(EnableVPlanNativePath && "VPlan-native path is not enabled.")(static_cast <bool> (EnableVPlanNativePath && "VPlan-native path is not enabled." ) ? void (0) : __assert_fail ("EnableVPlanNativePath && \"VPlan-native path is not enabled.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7573, __extension__ __PRETTY_FUNCTION__)); | |||
7574 | assert(isPowerOf2_32(VF.getKnownMinValue()) &&(static_cast <bool> (isPowerOf2_32(VF.getKnownMinValue( )) && "VF needs to be a power of two") ? void (0) : __assert_fail ("isPowerOf2_32(VF.getKnownMinValue()) && \"VF needs to be a power of two\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7575, __extension__ __PRETTY_FUNCTION__)) | |||
7575 | "VF needs to be a power of two")(static_cast <bool> (isPowerOf2_32(VF.getKnownMinValue( )) && "VF needs to be a power of two") ? void (0) : __assert_fail ("isPowerOf2_32(VF.getKnownMinValue()) && \"VF needs to be a power of two\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7575, __extension__ __PRETTY_FUNCTION__)); | |||
7576 | LLVM_DEBUG(dbgs() << "LV: Using " << (!UserVF.isZero() ? "user " : "")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Using " << ( !UserVF.isZero() ? "user " : "") << "VF " << VF << " to build VPlans.\n"; } } while (false) | |||
7577 | << "VF " << VF << " to build VPlans.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Using " << ( !UserVF.isZero() ? "user " : "") << "VF " << VF << " to build VPlans.\n"; } } while (false); | |||
7578 | buildVPlans(VF, VF); | |||
7579 | ||||
7580 | // For VPlan build stress testing, we bail out after VPlan construction. | |||
7581 | if (VPlanBuildStressTest) | |||
7582 | return VectorizationFactor::Disabled(); | |||
7583 | ||||
7584 | return {VF, 0 /*Cost*/}; | |||
7585 | } | |||
7586 | ||||
7587 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not vectorizing. Inner loops aren't supported in the " "VPlan-native path.\n"; } } while (false) | |||
7588 | dbgs() << "LV: Not vectorizing. Inner loops aren't supported in the "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not vectorizing. Inner loops aren't supported in the " "VPlan-native path.\n"; } } while (false) | |||
7589 | "VPlan-native path.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not vectorizing. Inner loops aren't supported in the " "VPlan-native path.\n"; } } while (false); | |||
7590 | return VectorizationFactor::Disabled(); | |||
7591 | } | |||
7592 | ||||
7593 | Optional<VectorizationFactor> | |||
7594 | LoopVectorizationPlanner::plan(ElementCount UserVF, unsigned UserIC) { | |||
7595 | assert(OrigLoop->isInnermost() && "Inner loop expected.")(static_cast <bool> (OrigLoop->isInnermost() && "Inner loop expected.") ? void (0) : __assert_fail ("OrigLoop->isInnermost() && \"Inner loop expected.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7595, __extension__ __PRETTY_FUNCTION__)); | |||
7596 | FixedScalableVFPair MaxFactors = CM.computeMaxVF(UserVF, UserIC); | |||
7597 | if (!MaxFactors) // Cases that should not to be vectorized nor interleaved. | |||
7598 | return None; | |||
7599 | ||||
7600 | // Invalidate interleave groups if all blocks of loop will be predicated. | |||
7601 | if (CM.blockNeedsPredicationForAnyReason(OrigLoop->getHeader()) && | |||
7602 | !useMaskedInterleavedAccesses(*TTI)) { | |||
7603 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Invalidate all interleaved groups due to fold-tail by masking " "which requires masked-interleaved support.\n"; } } while (false ) | |||
7604 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Invalidate all interleaved groups due to fold-tail by masking " "which requires masked-interleaved support.\n"; } } while (false ) | |||
7605 | << "LV: Invalidate all interleaved groups due to fold-tail by masking "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Invalidate all interleaved groups due to fold-tail by masking " "which requires masked-interleaved support.\n"; } } while (false ) | |||
7606 | "which requires masked-interleaved support.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Invalidate all interleaved groups due to fold-tail by masking " "which requires masked-interleaved support.\n"; } } while (false ); | |||
7607 | if (CM.InterleaveInfo.invalidateGroups()) | |||
7608 | // Invalidating interleave groups also requires invalidating all decisions | |||
7609 | // based on them, which includes widening decisions and uniform and scalar | |||
7610 | // values. | |||
7611 | CM.invalidateCostModelingDecisions(); | |||
7612 | } | |||
7613 | ||||
7614 | ElementCount MaxUserVF = | |||
7615 | UserVF.isScalable() ? MaxFactors.ScalableVF : MaxFactors.FixedVF; | |||
7616 | bool UserVFIsLegal = ElementCount::isKnownLE(UserVF, MaxUserVF); | |||
7617 | if (!UserVF.isZero() && UserVFIsLegal) { | |||
7618 | assert(isPowerOf2_32(UserVF.getKnownMinValue()) &&(static_cast <bool> (isPowerOf2_32(UserVF.getKnownMinValue ()) && "VF needs to be a power of two") ? void (0) : __assert_fail ("isPowerOf2_32(UserVF.getKnownMinValue()) && \"VF needs to be a power of two\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7619, __extension__ __PRETTY_FUNCTION__)) | |||
7619 | "VF needs to be a power of two")(static_cast <bool> (isPowerOf2_32(UserVF.getKnownMinValue ()) && "VF needs to be a power of two") ? void (0) : __assert_fail ("isPowerOf2_32(UserVF.getKnownMinValue()) && \"VF needs to be a power of two\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7619, __extension__ __PRETTY_FUNCTION__)); | |||
7620 | // Collect the instructions (and their associated costs) that will be more | |||
7621 | // profitable to scalarize. | |||
7622 | if (CM.selectUserVectorizationFactor(UserVF)) { | |||
7623 | LLVM_DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Using user VF " << UserVF << ".\n"; } } while (false); | |||
7624 | CM.collectInLoopReductions(); | |||
7625 | buildVPlansWithVPRecipes(UserVF, UserVF); | |||
7626 | LLVM_DEBUG(printPlans(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { printPlans(dbgs()); } } while (false); | |||
7627 | return {{UserVF, 0}}; | |||
7628 | } else | |||
7629 | reportVectorizationInfo("UserVF ignored because of invalid costs.", | |||
7630 | "InvalidCost", ORE, OrigLoop); | |||
7631 | } | |||
7632 | ||||
7633 | // Populate the set of Vectorization Factor Candidates. | |||
7634 | ElementCountSet VFCandidates; | |||
7635 | for (auto VF = ElementCount::getFixed(1); | |||
7636 | ElementCount::isKnownLE(VF, MaxFactors.FixedVF); VF *= 2) | |||
7637 | VFCandidates.insert(VF); | |||
7638 | for (auto VF = ElementCount::getScalable(1); | |||
7639 | ElementCount::isKnownLE(VF, MaxFactors.ScalableVF); VF *= 2) | |||
7640 | VFCandidates.insert(VF); | |||
7641 | ||||
7642 | for (const auto &VF : VFCandidates) { | |||
7643 | // Collect Uniform and Scalar instructions after vectorization with VF. | |||
7644 | CM.collectUniformsAndScalars(VF); | |||
7645 | ||||
7646 | // Collect the instructions (and their associated costs) that will be more | |||
7647 | // profitable to scalarize. | |||
7648 | if (VF.isVector()) | |||
7649 | CM.collectInstsToScalarize(VF); | |||
7650 | } | |||
7651 | ||||
7652 | CM.collectInLoopReductions(); | |||
7653 | buildVPlansWithVPRecipes(ElementCount::getFixed(1), MaxFactors.FixedVF); | |||
7654 | buildVPlansWithVPRecipes(ElementCount::getScalable(1), MaxFactors.ScalableVF); | |||
7655 | ||||
7656 | LLVM_DEBUG(printPlans(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { printPlans(dbgs()); } } while (false); | |||
7657 | if (!MaxFactors.hasVector()) | |||
7658 | return VectorizationFactor::Disabled(); | |||
7659 | ||||
7660 | // Select the optimal vectorization factor. | |||
7661 | auto SelectedVF = CM.selectVectorizationFactor(VFCandidates); | |||
7662 | ||||
7663 | // Check if it is profitable to vectorize with runtime checks. | |||
7664 | unsigned NumRuntimePointerChecks = Requirements.getNumRuntimePointerChecks(); | |||
7665 | if (SelectedVF.Width.getKnownMinValue() > 1 && NumRuntimePointerChecks) { | |||
7666 | bool PragmaThresholdReached = | |||
7667 | NumRuntimePointerChecks > PragmaVectorizeMemoryCheckThreshold; | |||
7668 | bool ThresholdReached = | |||
7669 | NumRuntimePointerChecks > VectorizerParams::RuntimeMemoryCheckThreshold; | |||
7670 | if ((ThresholdReached && !Hints.allowReordering()) || | |||
7671 | PragmaThresholdReached) { | |||
7672 | ORE->emit([&]() { | |||
7673 | return OptimizationRemarkAnalysisAliasing( | |||
7674 | DEBUG_TYPE"loop-vectorize", "CantReorderMemOps", OrigLoop->getStartLoc(), | |||
7675 | OrigLoop->getHeader()) | |||
7676 | << "loop not vectorized: cannot prove it is safe to reorder " | |||
7677 | "memory operations"; | |||
7678 | }); | |||
7679 | LLVM_DEBUG(dbgs() << "LV: Too many memory checks needed.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Too many memory checks needed.\n" ; } } while (false); | |||
7680 | Hints.emitRemarkWithHints(); | |||
7681 | return VectorizationFactor::Disabled(); | |||
7682 | } | |||
7683 | } | |||
7684 | return SelectedVF; | |||
7685 | } | |||
7686 | ||||
7687 | VPlan &LoopVectorizationPlanner::getBestPlanFor(ElementCount VF) const { | |||
7688 | assert(count_if(VPlans,(static_cast <bool> (count_if(VPlans, [VF](const VPlanPtr &Plan) { return Plan->hasVF(VF); }) == 1 && "Best VF has not a single VPlan." ) ? void (0) : __assert_fail ("count_if(VPlans, [VF](const VPlanPtr &Plan) { return Plan->hasVF(VF); }) == 1 && \"Best VF has not a single VPlan.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7691, __extension__ __PRETTY_FUNCTION__)) | |||
7689 | [VF](const VPlanPtr &Plan) { return Plan->hasVF(VF); }) ==(static_cast <bool> (count_if(VPlans, [VF](const VPlanPtr &Plan) { return Plan->hasVF(VF); }) == 1 && "Best VF has not a single VPlan." ) ? void (0) : __assert_fail ("count_if(VPlans, [VF](const VPlanPtr &Plan) { return Plan->hasVF(VF); }) == 1 && \"Best VF has not a single VPlan.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7691, __extension__ __PRETTY_FUNCTION__)) | |||
7690 | 1 &&(static_cast <bool> (count_if(VPlans, [VF](const VPlanPtr &Plan) { return Plan->hasVF(VF); }) == 1 && "Best VF has not a single VPlan." ) ? void (0) : __assert_fail ("count_if(VPlans, [VF](const VPlanPtr &Plan) { return Plan->hasVF(VF); }) == 1 && \"Best VF has not a single VPlan.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7691, __extension__ __PRETTY_FUNCTION__)) | |||
7691 | "Best VF has not a single VPlan.")(static_cast <bool> (count_if(VPlans, [VF](const VPlanPtr &Plan) { return Plan->hasVF(VF); }) == 1 && "Best VF has not a single VPlan." ) ? void (0) : __assert_fail ("count_if(VPlans, [VF](const VPlanPtr &Plan) { return Plan->hasVF(VF); }) == 1 && \"Best VF has not a single VPlan.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7691, __extension__ __PRETTY_FUNCTION__)); | |||
7692 | ||||
7693 | for (const VPlanPtr &Plan : VPlans) { | |||
7694 | if (Plan->hasVF(VF)) | |||
7695 | return *Plan.get(); | |||
7696 | } | |||
7697 | llvm_unreachable("No plan found!")::llvm::llvm_unreachable_internal("No plan found!", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7697); | |||
7698 | } | |||
7699 | ||||
7700 | static void AddRuntimeUnrollDisableMetaData(Loop *L) { | |||
7701 | SmallVector<Metadata *, 4> MDs; | |||
7702 | // Reserve first location for self reference to the LoopID metadata node. | |||
7703 | MDs.push_back(nullptr); | |||
7704 | bool IsUnrollMetadata = false; | |||
7705 | MDNode *LoopID = L->getLoopID(); | |||
7706 | if (LoopID) { | |||
7707 | // First find existing loop unrolling disable metadata. | |||
7708 | for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) { | |||
7709 | auto *MD = dyn_cast<MDNode>(LoopID->getOperand(i)); | |||
7710 | if (MD) { | |||
7711 | const auto *S = dyn_cast<MDString>(MD->getOperand(0)); | |||
7712 | IsUnrollMetadata = | |||
7713 | S && S->getString().startswith("llvm.loop.unroll.disable"); | |||
7714 | } | |||
7715 | MDs.push_back(LoopID->getOperand(i)); | |||
7716 | } | |||
7717 | } | |||
7718 | ||||
7719 | if (!IsUnrollMetadata) { | |||
7720 | // Add runtime unroll disable metadata. | |||
7721 | LLVMContext &Context = L->getHeader()->getContext(); | |||
7722 | SmallVector<Metadata *, 1> DisableOperands; | |||
7723 | DisableOperands.push_back( | |||
7724 | MDString::get(Context, "llvm.loop.unroll.runtime.disable")); | |||
7725 | MDNode *DisableNode = MDNode::get(Context, DisableOperands); | |||
7726 | MDs.push_back(DisableNode); | |||
7727 | MDNode *NewLoopID = MDNode::get(Context, MDs); | |||
7728 | // Set operand 0 to refer to the loop id itself. | |||
7729 | NewLoopID->replaceOperandWith(0, NewLoopID); | |||
7730 | L->setLoopID(NewLoopID); | |||
7731 | } | |||
7732 | } | |||
7733 | ||||
7734 | void LoopVectorizationPlanner::executePlan(ElementCount BestVF, unsigned BestUF, | |||
7735 | VPlan &BestVPlan, | |||
7736 | InnerLoopVectorizer &ILV, | |||
7737 | DominatorTree *DT) { | |||
7738 | LLVM_DEBUG(dbgs() << "Executing best plan with VF=" << BestVF << ", UF=" << BestUFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Executing best plan with VF=" << BestVF << ", UF=" << BestUF << '\n' ; } } while (false) | |||
| ||||
7739 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Executing best plan with VF=" << BestVF << ", UF=" << BestUF << '\n' ; } } while (false); | |||
7740 | ||||
7741 | // Perform the actual loop transformation. | |||
7742 | ||||
7743 | // 1. Create a new empty loop. Unlink the old loop and connect the new one. | |||
7744 | VPTransformState State{BestVF, BestUF, LI, DT, ILV.Builder, &ILV, &BestVPlan}; | |||
7745 | Value *CanonicalIVStartValue; | |||
7746 | std::tie(State.CFG.PrevBB, CanonicalIVStartValue) = | |||
7747 | ILV.createVectorizedLoopSkeleton(); | |||
7748 | ILV.collectPoisonGeneratingRecipes(State); | |||
7749 | ||||
7750 | ILV.printDebugTracesAtStart(); | |||
7751 | ||||
7752 | //===------------------------------------------------===// | |||
7753 | // | |||
7754 | // Notice: any optimization or new instruction that go | |||
7755 | // into the code below should also be implemented in | |||
7756 | // the cost-model. | |||
7757 | // | |||
7758 | //===------------------------------------------------===// | |||
7759 | ||||
7760 | // 2. Copy and widen instructions from the old loop into the new loop. | |||
7761 | BestVPlan.prepareToExecute(ILV.getOrCreateTripCount(nullptr), | |||
7762 | ILV.getOrCreateVectorTripCount(nullptr), | |||
7763 | CanonicalIVStartValue, State); | |||
7764 | BestVPlan.execute(&State); | |||
7765 | ||||
7766 | // Keep all loop hints from the original loop on the vector loop (we'll | |||
7767 | // replace the vectorizer-specific hints below). | |||
7768 | MDNode *OrigLoopID = OrigLoop->getLoopID(); | |||
7769 | ||||
7770 | Optional<MDNode *> VectorizedLoopID = | |||
7771 | makeFollowupLoopID(OrigLoopID, {LLVMLoopVectorizeFollowupAll, | |||
7772 | LLVMLoopVectorizeFollowupVectorized}); | |||
7773 | ||||
7774 | Loop *L = LI->getLoopFor(State.CFG.PrevBB); | |||
7775 | if (VectorizedLoopID.hasValue()) | |||
7776 | L->setLoopID(VectorizedLoopID.getValue()); | |||
7777 | else { | |||
7778 | // Keep all loop hints from the original loop on the vector loop (we'll | |||
7779 | // replace the vectorizer-specific hints below). | |||
7780 | if (MDNode *LID = OrigLoop->getLoopID()) | |||
7781 | L->setLoopID(LID); | |||
7782 | ||||
7783 | LoopVectorizeHints Hints(L, true, *ORE); | |||
7784 | Hints.setAlreadyVectorized(); | |||
7785 | } | |||
7786 | // Disable runtime unrolling when vectorizing the epilogue loop. | |||
7787 | if (CanonicalIVStartValue) | |||
7788 | AddRuntimeUnrollDisableMetaData(L); | |||
7789 | ||||
7790 | // 3. Fix the vectorized code: take care of header phi's, live-outs, | |||
7791 | // predication, updating analyses. | |||
7792 | ILV.fixVectorizedLoop(State); | |||
7793 | ||||
7794 | ILV.printDebugTracesAtEnd(); | |||
7795 | } | |||
7796 | ||||
7797 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | |||
7798 | void LoopVectorizationPlanner::printPlans(raw_ostream &O) { | |||
7799 | for (const auto &Plan : VPlans) | |||
7800 | if (PrintVPlansInDotFormat) | |||
7801 | Plan->printDOT(O); | |||
7802 | else | |||
7803 | Plan->print(O); | |||
7804 | } | |||
7805 | #endif | |||
7806 | ||||
7807 | void LoopVectorizationPlanner::collectTriviallyDeadInstructions( | |||
7808 | SmallPtrSetImpl<Instruction *> &DeadInstructions) { | |||
7809 | ||||
7810 | // We create new control-flow for the vectorized loop, so the original exit | |||
7811 | // conditions will be dead after vectorization if it's only used by the | |||
7812 | // terminator | |||
7813 | SmallVector<BasicBlock*> ExitingBlocks; | |||
7814 | OrigLoop->getExitingBlocks(ExitingBlocks); | |||
7815 | for (auto *BB : ExitingBlocks) { | |||
7816 | auto *Cmp = dyn_cast<Instruction>(BB->getTerminator()->getOperand(0)); | |||
7817 | if (!Cmp || !Cmp->hasOneUse()) | |||
7818 | continue; | |||
7819 | ||||
7820 | // TODO: we should introduce a getUniqueExitingBlocks on Loop | |||
7821 | if (!DeadInstructions.insert(Cmp).second) | |||
7822 | continue; | |||
7823 | ||||
7824 | // The operands of the icmp is often a dead trunc, used by IndUpdate. | |||
7825 | // TODO: can recurse through operands in general | |||
7826 | for (Value *Op : Cmp->operands()) { | |||
7827 | if (isa<TruncInst>(Op) && Op->hasOneUse()) | |||
7828 | DeadInstructions.insert(cast<Instruction>(Op)); | |||
7829 | } | |||
7830 | } | |||
7831 | ||||
7832 | // We create new "steps" for induction variable updates to which the original | |||
7833 | // induction variables map. An original update instruction will be dead if | |||
7834 | // all its users except the induction variable are dead. | |||
7835 | auto *Latch = OrigLoop->getLoopLatch(); | |||
7836 | for (auto &Induction : Legal->getInductionVars()) { | |||
7837 | PHINode *Ind = Induction.first; | |||
7838 | auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch)); | |||
7839 | ||||
7840 | // If the tail is to be folded by masking, the primary induction variable, | |||
7841 | // if exists, isn't dead: it will be used for masking. Don't kill it. | |||
7842 | if (CM.foldTailByMasking() && IndUpdate == Legal->getPrimaryInduction()) | |||
7843 | continue; | |||
7844 | ||||
7845 | if (llvm::all_of(IndUpdate->users(), [&](User *U) -> bool { | |||
7846 | return U == Ind || DeadInstructions.count(cast<Instruction>(U)); | |||
7847 | })) | |||
7848 | DeadInstructions.insert(IndUpdate); | |||
7849 | } | |||
7850 | } | |||
7851 | ||||
7852 | Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) { return V; } | |||
7853 | ||||
7854 | //===--------------------------------------------------------------------===// | |||
7855 | // EpilogueVectorizerMainLoop | |||
7856 | //===--------------------------------------------------------------------===// | |||
7857 | ||||
7858 | /// This function is partially responsible for generating the control flow | |||
7859 | /// depicted in https://llvm.org/docs/Vectorizers.html#epilogue-vectorization. | |||
7860 | std::pair<BasicBlock *, Value *> | |||
7861 | EpilogueVectorizerMainLoop::createEpilogueVectorizedLoopSkeleton() { | |||
7862 | MDNode *OrigLoopID = OrigLoop->getLoopID(); | |||
7863 | Loop *Lp = createVectorLoopSkeleton(""); | |||
7864 | ||||
7865 | // Generate the code to check the minimum iteration count of the vector | |||
7866 | // epilogue (see below). | |||
7867 | EPI.EpilogueIterationCountCheck = | |||
7868 | emitMinimumIterationCountCheck(Lp, LoopScalarPreHeader, true); | |||
7869 | EPI.EpilogueIterationCountCheck->setName("iter.check"); | |||
7870 | ||||
7871 | // Generate the code to check any assumptions that we've made for SCEV | |||
7872 | // expressions. | |||
7873 | EPI.SCEVSafetyCheck = emitSCEVChecks(LoopScalarPreHeader); | |||
7874 | ||||
7875 | // Generate the code that checks at runtime if arrays overlap. We put the | |||
7876 | // checks into a separate block to make the more common case of few elements | |||
7877 | // faster. | |||
7878 | EPI.MemSafetyCheck = emitMemRuntimeChecks(Lp, LoopScalarPreHeader); | |||
7879 | ||||
7880 | // Generate the iteration count check for the main loop, *after* the check | |||
7881 | // for the epilogue loop, so that the path-length is shorter for the case | |||
7882 | // that goes directly through the vector epilogue. The longer-path length for | |||
7883 | // the main loop is compensated for, by the gain from vectorizing the larger | |||
7884 | // trip count. Note: the branch will get updated later on when we vectorize | |||
7885 | // the epilogue. | |||
7886 | EPI.MainLoopIterationCountCheck = | |||
7887 | emitMinimumIterationCountCheck(Lp, LoopScalarPreHeader, false); | |||
7888 | ||||
7889 | // Generate the induction variable. | |||
7890 | Value *CountRoundDown = getOrCreateVectorTripCount(Lp); | |||
7891 | EPI.VectorTripCount = CountRoundDown; | |||
7892 | createHeaderBranch(Lp); | |||
7893 | ||||
7894 | // Skip induction resume value creation here because they will be created in | |||
7895 | // the second pass. If we created them here, they wouldn't be used anyway, | |||
7896 | // because the vplan in the second pass still contains the inductions from the | |||
7897 | // original loop. | |||
7898 | ||||
7899 | return {completeLoopSkeleton(Lp, OrigLoopID), nullptr}; | |||
7900 | } | |||
7901 | ||||
7902 | void EpilogueVectorizerMainLoop::printDebugTracesAtStart() { | |||
7903 | LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (first pass)\n" << "Main Loop VF:" << EPI.MainLoopVF << ", Main Loop UF:" << EPI.MainLoopUF << ", Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false) | |||
7904 | dbgs() << "Create Skeleton for epilogue vectorized loop (first pass)\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (first pass)\n" << "Main Loop VF:" << EPI.MainLoopVF << ", Main Loop UF:" << EPI.MainLoopUF << ", Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false) | |||
7905 | << "Main Loop VF:" << EPI.MainLoopVFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (first pass)\n" << "Main Loop VF:" << EPI.MainLoopVF << ", Main Loop UF:" << EPI.MainLoopUF << ", Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false) | |||
7906 | << ", Main Loop UF:" << EPI.MainLoopUFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (first pass)\n" << "Main Loop VF:" << EPI.MainLoopVF << ", Main Loop UF:" << EPI.MainLoopUF << ", Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false) | |||
7907 | << ", Epilogue Loop VF:" << EPI.EpilogueVFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (first pass)\n" << "Main Loop VF:" << EPI.MainLoopVF << ", Main Loop UF:" << EPI.MainLoopUF << ", Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false) | |||
7908 | << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (first pass)\n" << "Main Loop VF:" << EPI.MainLoopVF << ", Main Loop UF:" << EPI.MainLoopUF << ", Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false) | |||
7909 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (first pass)\n" << "Main Loop VF:" << EPI.MainLoopVF << ", Main Loop UF:" << EPI.MainLoopUF << ", Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false); | |||
7910 | } | |||
7911 | ||||
7912 | void EpilogueVectorizerMainLoop::printDebugTracesAtEnd() { | |||
7913 | DEBUG_WITH_TYPE(VerboseDebug, {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "intermediate fn:\n" << *OrigLoop->getHeader()->getParent() << "\n"; }; } } while (false) | |||
7914 | dbgs() << "intermediate fn:\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "intermediate fn:\n" << *OrigLoop->getHeader()->getParent() << "\n"; }; } } while (false) | |||
7915 | << *OrigLoop->getHeader()->getParent() << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "intermediate fn:\n" << *OrigLoop->getHeader()->getParent() << "\n"; }; } } while (false) | |||
7916 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "intermediate fn:\n" << *OrigLoop->getHeader()->getParent() << "\n"; }; } } while (false); | |||
7917 | } | |||
7918 | ||||
7919 | BasicBlock *EpilogueVectorizerMainLoop::emitMinimumIterationCountCheck( | |||
7920 | Loop *L, BasicBlock *Bypass, bool ForEpilogue) { | |||
7921 | assert(L && "Expected valid Loop.")(static_cast <bool> (L && "Expected valid Loop." ) ? void (0) : __assert_fail ("L && \"Expected valid Loop.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7921, __extension__ __PRETTY_FUNCTION__)); | |||
7922 | assert(Bypass && "Expected valid bypass basic block.")(static_cast <bool> (Bypass && "Expected valid bypass basic block." ) ? void (0) : __assert_fail ("Bypass && \"Expected valid bypass basic block.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7922, __extension__ __PRETTY_FUNCTION__)); | |||
7923 | ElementCount VFactor = ForEpilogue ? EPI.EpilogueVF : VF; | |||
7924 | unsigned UFactor = ForEpilogue ? EPI.EpilogueUF : UF; | |||
7925 | Value *Count = getOrCreateTripCount(L); | |||
7926 | // Reuse existing vector loop preheader for TC checks. | |||
7927 | // Note that new preheader block is generated for vector loop. | |||
7928 | BasicBlock *const TCCheckBlock = LoopVectorPreHeader; | |||
7929 | IRBuilder<> Builder(TCCheckBlock->getTerminator()); | |||
7930 | ||||
7931 | // Generate code to check if the loop's trip count is less than VF * UF of the | |||
7932 | // main vector loop. | |||
7933 | auto P = Cost->requiresScalarEpilogue(ForEpilogue ? EPI.EpilogueVF : VF) ? | |||
7934 | ICmpInst::ICMP_ULE : ICmpInst::ICMP_ULT; | |||
7935 | ||||
7936 | Value *CheckMinIters = Builder.CreateICmp( | |||
7937 | P, Count, createStepForVF(Builder, Count->getType(), VFactor, UFactor), | |||
7938 | "min.iters.check"); | |||
7939 | ||||
7940 | if (!ForEpilogue) | |||
7941 | TCCheckBlock->setName("vector.main.loop.iter.check"); | |||
7942 | ||||
7943 | // Create new preheader for vector loop. | |||
7944 | LoopVectorPreHeader = SplitBlock(TCCheckBlock, TCCheckBlock->getTerminator(), | |||
7945 | DT, LI, nullptr, "vector.ph"); | |||
7946 | ||||
7947 | if (ForEpilogue) { | |||
7948 | assert(DT->properlyDominates(DT->getNode(TCCheckBlock),(static_cast <bool> (DT->properlyDominates(DT->getNode (TCCheckBlock), DT->getNode(Bypass)->getIDom()) && "TC check is expected to dominate Bypass") ? void (0) : __assert_fail ("DT->properlyDominates(DT->getNode(TCCheckBlock), DT->getNode(Bypass)->getIDom()) && \"TC check is expected to dominate Bypass\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7950, __extension__ __PRETTY_FUNCTION__)) | |||
7949 | DT->getNode(Bypass)->getIDom()) &&(static_cast <bool> (DT->properlyDominates(DT->getNode (TCCheckBlock), DT->getNode(Bypass)->getIDom()) && "TC check is expected to dominate Bypass") ? void (0) : __assert_fail ("DT->properlyDominates(DT->getNode(TCCheckBlock), DT->getNode(Bypass)->getIDom()) && \"TC check is expected to dominate Bypass\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7950, __extension__ __PRETTY_FUNCTION__)) | |||
7950 | "TC check is expected to dominate Bypass")(static_cast <bool> (DT->properlyDominates(DT->getNode (TCCheckBlock), DT->getNode(Bypass)->getIDom()) && "TC check is expected to dominate Bypass") ? void (0) : __assert_fail ("DT->properlyDominates(DT->getNode(TCCheckBlock), DT->getNode(Bypass)->getIDom()) && \"TC check is expected to dominate Bypass\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7950, __extension__ __PRETTY_FUNCTION__)); | |||
7951 | ||||
7952 | // Update dominator for Bypass & LoopExit. | |||
7953 | DT->changeImmediateDominator(Bypass, TCCheckBlock); | |||
7954 | if (!Cost->requiresScalarEpilogue(EPI.EpilogueVF)) | |||
7955 | // For loops with multiple exits, there's no edge from the middle block | |||
7956 | // to exit blocks (as the epilogue must run) and thus no need to update | |||
7957 | // the immediate dominator of the exit blocks. | |||
7958 | DT->changeImmediateDominator(LoopExitBlock, TCCheckBlock); | |||
7959 | ||||
7960 | LoopBypassBlocks.push_back(TCCheckBlock); | |||
7961 | ||||
7962 | // Save the trip count so we don't have to regenerate it in the | |||
7963 | // vec.epilog.iter.check. This is safe to do because the trip count | |||
7964 | // generated here dominates the vector epilog iter check. | |||
7965 | EPI.TripCount = Count; | |||
7966 | } | |||
7967 | ||||
7968 | ReplaceInstWithInst( | |||
7969 | TCCheckBlock->getTerminator(), | |||
7970 | BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters)); | |||
7971 | ||||
7972 | return TCCheckBlock; | |||
7973 | } | |||
7974 | ||||
7975 | //===--------------------------------------------------------------------===// | |||
7976 | // EpilogueVectorizerEpilogueLoop | |||
7977 | //===--------------------------------------------------------------------===// | |||
7978 | ||||
7979 | /// This function is partially responsible for generating the control flow | |||
7980 | /// depicted in https://llvm.org/docs/Vectorizers.html#epilogue-vectorization. | |||
7981 | std::pair<BasicBlock *, Value *> | |||
7982 | EpilogueVectorizerEpilogueLoop::createEpilogueVectorizedLoopSkeleton() { | |||
7983 | MDNode *OrigLoopID = OrigLoop->getLoopID(); | |||
7984 | Loop *Lp = createVectorLoopSkeleton("vec.epilog."); | |||
7985 | ||||
7986 | // Now, compare the remaining count and if there aren't enough iterations to | |||
7987 | // execute the vectorized epilogue skip to the scalar part. | |||
7988 | BasicBlock *VecEpilogueIterationCountCheck = LoopVectorPreHeader; | |||
7989 | VecEpilogueIterationCountCheck->setName("vec.epilog.iter.check"); | |||
7990 | LoopVectorPreHeader = | |||
7991 | SplitBlock(LoopVectorPreHeader, LoopVectorPreHeader->getTerminator(), DT, | |||
7992 | LI, nullptr, "vec.epilog.ph"); | |||
7993 | emitMinimumVectorEpilogueIterCountCheck(LoopScalarPreHeader, | |||
7994 | VecEpilogueIterationCountCheck); | |||
7995 | ||||
7996 | // Adjust the control flow taking the state info from the main loop | |||
7997 | // vectorization into account. | |||
7998 | assert(EPI.MainLoopIterationCountCheck && EPI.EpilogueIterationCountCheck &&(static_cast <bool> (EPI.MainLoopIterationCountCheck && EPI.EpilogueIterationCountCheck && "expected this to be saved from the previous pass." ) ? void (0) : __assert_fail ("EPI.MainLoopIterationCountCheck && EPI.EpilogueIterationCountCheck && \"expected this to be saved from the previous pass.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7999, __extension__ __PRETTY_FUNCTION__)) | |||
7999 | "expected this to be saved from the previous pass.")(static_cast <bool> (EPI.MainLoopIterationCountCheck && EPI.EpilogueIterationCountCheck && "expected this to be saved from the previous pass." ) ? void (0) : __assert_fail ("EPI.MainLoopIterationCountCheck && EPI.EpilogueIterationCountCheck && \"expected this to be saved from the previous pass.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7999, __extension__ __PRETTY_FUNCTION__)); | |||
8000 | EPI.MainLoopIterationCountCheck->getTerminator()->replaceUsesOfWith( | |||
8001 | VecEpilogueIterationCountCheck, LoopVectorPreHeader); | |||
8002 | ||||
8003 | DT->changeImmediateDominator(LoopVectorPreHeader, | |||
8004 | EPI.MainLoopIterationCountCheck); | |||
8005 | ||||
8006 | EPI.EpilogueIterationCountCheck->getTerminator()->replaceUsesOfWith( | |||
8007 | VecEpilogueIterationCountCheck, LoopScalarPreHeader); | |||
8008 | ||||
8009 | if (EPI.SCEVSafetyCheck) | |||
8010 | EPI.SCEVSafetyCheck->getTerminator()->replaceUsesOfWith( | |||
8011 | VecEpilogueIterationCountCheck, LoopScalarPreHeader); | |||
8012 | if (EPI.MemSafetyCheck) | |||
8013 | EPI.MemSafetyCheck->getTerminator()->replaceUsesOfWith( | |||
8014 | VecEpilogueIterationCountCheck, LoopScalarPreHeader); | |||
8015 | ||||
8016 | DT->changeImmediateDominator( | |||
8017 | VecEpilogueIterationCountCheck, | |||
8018 | VecEpilogueIterationCountCheck->getSinglePredecessor()); | |||
8019 | ||||
8020 | DT->changeImmediateDominator(LoopScalarPreHeader, | |||
8021 | EPI.EpilogueIterationCountCheck); | |||
8022 | if (!Cost->requiresScalarEpilogue(EPI.EpilogueVF)) | |||
8023 | // If there is an epilogue which must run, there's no edge from the | |||
8024 | // middle block to exit blocks and thus no need to update the immediate | |||
8025 | // dominator of the exit blocks. | |||
8026 | DT->changeImmediateDominator(LoopExitBlock, | |||
8027 | EPI.EpilogueIterationCountCheck); | |||
8028 | ||||
8029 | // Keep track of bypass blocks, as they feed start values to the induction | |||
8030 | // phis in the scalar loop preheader. | |||
8031 | if (EPI.SCEVSafetyCheck) | |||
8032 | LoopBypassBlocks.push_back(EPI.SCEVSafetyCheck); | |||
8033 | if (EPI.MemSafetyCheck) | |||
8034 | LoopBypassBlocks.push_back(EPI.MemSafetyCheck); | |||
8035 | LoopBypassBlocks.push_back(EPI.EpilogueIterationCountCheck); | |||
8036 | ||||
8037 | // The vec.epilog.iter.check block may contain Phi nodes from reductions which | |||
8038 | // merge control-flow from the latch block and the middle block. Update the | |||
8039 | // incoming values here and move the Phi into the preheader. | |||
8040 | SmallVector<PHINode *, 4> PhisInBlock; | |||
8041 | for (PHINode &Phi : VecEpilogueIterationCountCheck->phis()) | |||
8042 | PhisInBlock.push_back(&Phi); | |||
8043 | ||||
8044 | for (PHINode *Phi : PhisInBlock) { | |||
8045 | Phi->replaceIncomingBlockWith( | |||
8046 | VecEpilogueIterationCountCheck->getSinglePredecessor(), | |||
8047 | VecEpilogueIterationCountCheck); | |||
8048 | Phi->removeIncomingValue(EPI.EpilogueIterationCountCheck); | |||
8049 | if (EPI.SCEVSafetyCheck) | |||
8050 | Phi->removeIncomingValue(EPI.SCEVSafetyCheck); | |||
8051 | if (EPI.MemSafetyCheck) | |||
8052 | Phi->removeIncomingValue(EPI.MemSafetyCheck); | |||
8053 | Phi->moveBefore(LoopVectorPreHeader->getFirstNonPHI()); | |||
8054 | } | |||
8055 | ||||
8056 | // Generate a resume induction for the vector epilogue and put it in the | |||
8057 | // vector epilogue preheader | |||
8058 | Type *IdxTy = Legal->getWidestInductionType(); | |||
8059 | PHINode *EPResumeVal = PHINode::Create(IdxTy, 2, "vec.epilog.resume.val", | |||
8060 | LoopVectorPreHeader->getFirstNonPHI()); | |||
8061 | EPResumeVal->addIncoming(EPI.VectorTripCount, VecEpilogueIterationCountCheck); | |||
8062 | EPResumeVal->addIncoming(ConstantInt::get(IdxTy, 0), | |||
8063 | EPI.MainLoopIterationCountCheck); | |||
8064 | ||||
8065 | // Generate the induction variable. | |||
8066 | createHeaderBranch(Lp); | |||
8067 | ||||
8068 | // Generate induction resume values. These variables save the new starting | |||
8069 | // indexes for the scalar loop. They are used to test if there are any tail | |||
8070 | // iterations left once the vector loop has completed. | |||
8071 | // Note that when the vectorized epilogue is skipped due to iteration count | |||
8072 | // check, then the resume value for the induction variable comes from | |||
8073 | // the trip count of the main vector loop, hence passing the AdditionalBypass | |||
8074 | // argument. | |||
8075 | createInductionResumeValues(Lp, {VecEpilogueIterationCountCheck, | |||
8076 | EPI.VectorTripCount} /* AdditionalBypass */); | |||
8077 | ||||
8078 | return {completeLoopSkeleton(Lp, OrigLoopID), EPResumeVal}; | |||
8079 | } | |||
8080 | ||||
8081 | BasicBlock * | |||
8082 | EpilogueVectorizerEpilogueLoop::emitMinimumVectorEpilogueIterCountCheck( | |||
8083 | BasicBlock *Bypass, BasicBlock *Insert) { | |||
8084 | ||||
8085 | assert(EPI.TripCount &&(static_cast <bool> (EPI.TripCount && "Expected trip count to have been safed in the first pass." ) ? void (0) : __assert_fail ("EPI.TripCount && \"Expected trip count to have been safed in the first pass.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8086, __extension__ __PRETTY_FUNCTION__)) | |||
8086 | "Expected trip count to have been safed in the first pass.")(static_cast <bool> (EPI.TripCount && "Expected trip count to have been safed in the first pass." ) ? void (0) : __assert_fail ("EPI.TripCount && \"Expected trip count to have been safed in the first pass.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8086, __extension__ __PRETTY_FUNCTION__)); | |||
8087 | assert((static_cast <bool> ((!isa<Instruction>(EPI.TripCount ) || DT->dominates(cast<Instruction>(EPI.TripCount)-> getParent(), Insert)) && "saved trip count does not dominate insertion point." ) ? void (0) : __assert_fail ("(!isa<Instruction>(EPI.TripCount) || DT->dominates(cast<Instruction>(EPI.TripCount)->getParent(), Insert)) && \"saved trip count does not dominate insertion point.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8090, __extension__ __PRETTY_FUNCTION__)) | |||
8088 | (!isa<Instruction>(EPI.TripCount) ||(static_cast <bool> ((!isa<Instruction>(EPI.TripCount ) || DT->dominates(cast<Instruction>(EPI.TripCount)-> getParent(), Insert)) && "saved trip count does not dominate insertion point." ) ? void (0) : __assert_fail ("(!isa<Instruction>(EPI.TripCount) || DT->dominates(cast<Instruction>(EPI.TripCount)->getParent(), Insert)) && \"saved trip count does not dominate insertion point.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8090, __extension__ __PRETTY_FUNCTION__)) | |||
8089 | DT->dominates(cast<Instruction>(EPI.TripCount)->getParent(), Insert)) &&(static_cast <bool> ((!isa<Instruction>(EPI.TripCount ) || DT->dominates(cast<Instruction>(EPI.TripCount)-> getParent(), Insert)) && "saved trip count does not dominate insertion point." ) ? void (0) : __assert_fail ("(!isa<Instruction>(EPI.TripCount) || DT->dominates(cast<Instruction>(EPI.TripCount)->getParent(), Insert)) && \"saved trip count does not dominate insertion point.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8090, __extension__ __PRETTY_FUNCTION__)) | |||
8090 | "saved trip count does not dominate insertion point.")(static_cast <bool> ((!isa<Instruction>(EPI.TripCount ) || DT->dominates(cast<Instruction>(EPI.TripCount)-> getParent(), Insert)) && "saved trip count does not dominate insertion point." ) ? void (0) : __assert_fail ("(!isa<Instruction>(EPI.TripCount) || DT->dominates(cast<Instruction>(EPI.TripCount)->getParent(), Insert)) && \"saved trip count does not dominate insertion point.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8090, __extension__ __PRETTY_FUNCTION__)); | |||
8091 | Value *TC = EPI.TripCount; | |||
8092 | IRBuilder<> Builder(Insert->getTerminator()); | |||
8093 | Value *Count = Builder.CreateSub(TC, EPI.VectorTripCount, "n.vec.remaining"); | |||
8094 | ||||
8095 | // Generate code to check if the loop's trip count is less than VF * UF of the | |||
8096 | // vector epilogue loop. | |||
8097 | auto P = Cost->requiresScalarEpilogue(EPI.EpilogueVF) ? | |||
8098 | ICmpInst::ICMP_ULE : ICmpInst::ICMP_ULT; | |||
8099 | ||||
8100 | Value *CheckMinIters = | |||
8101 | Builder.CreateICmp(P, Count, | |||
8102 | createStepForVF(Builder, Count->getType(), | |||
8103 | EPI.EpilogueVF, EPI.EpilogueUF), | |||
8104 | "min.epilog.iters.check"); | |||
8105 | ||||
8106 | ReplaceInstWithInst( | |||
8107 | Insert->getTerminator(), | |||
8108 | BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters)); | |||
8109 | ||||
8110 | LoopBypassBlocks.push_back(Insert); | |||
8111 | return Insert; | |||
8112 | } | |||
8113 | ||||
8114 | void EpilogueVectorizerEpilogueLoop::printDebugTracesAtStart() { | |||
8115 | LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (second pass)\n" << "Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false) | |||
8116 | dbgs() << "Create Skeleton for epilogue vectorized loop (second pass)\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (second pass)\n" << "Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false) | |||
8117 | << "Epilogue Loop VF:" << EPI.EpilogueVFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (second pass)\n" << "Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false) | |||
8118 | << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (second pass)\n" << "Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false) | |||
8119 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (second pass)\n" << "Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false); | |||
8120 | } | |||
8121 | ||||
8122 | void EpilogueVectorizerEpilogueLoop::printDebugTracesAtEnd() { | |||
8123 | DEBUG_WITH_TYPE(VerboseDebug, {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "final fn:\n" << *OrigLoop ->getHeader()->getParent() << "\n"; }; } } while ( false) | |||
8124 | dbgs() << "final fn:\n" << *OrigLoop->getHeader()->getParent() << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "final fn:\n" << *OrigLoop ->getHeader()->getParent() << "\n"; }; } } while ( false) | |||
8125 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "final fn:\n" << *OrigLoop ->getHeader()->getParent() << "\n"; }; } } while ( false); | |||
8126 | } | |||
8127 | ||||
8128 | bool LoopVectorizationPlanner::getDecisionAndClampRange( | |||
8129 | const std::function<bool(ElementCount)> &Predicate, VFRange &Range) { | |||
8130 | assert(!Range.isEmpty() && "Trying to test an empty VF range.")(static_cast <bool> (!Range.isEmpty() && "Trying to test an empty VF range." ) ? void (0) : __assert_fail ("!Range.isEmpty() && \"Trying to test an empty VF range.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8130, __extension__ __PRETTY_FUNCTION__)); | |||
8131 | bool PredicateAtRangeStart = Predicate(Range.Start); | |||
8132 | ||||
8133 | for (ElementCount TmpVF = Range.Start * 2; | |||
8134 | ElementCount::isKnownLT(TmpVF, Range.End); TmpVF *= 2) | |||
8135 | if (Predicate(TmpVF) != PredicateAtRangeStart) { | |||
8136 | Range.End = TmpVF; | |||
8137 | break; | |||
8138 | } | |||
8139 | ||||
8140 | return PredicateAtRangeStart; | |||
8141 | } | |||
8142 | ||||
8143 | /// Build VPlans for the full range of feasible VF's = {\p MinVF, 2 * \p MinVF, | |||
8144 | /// 4 * \p MinVF, ..., \p MaxVF} by repeatedly building a VPlan for a sub-range | |||
8145 | /// of VF's starting at a given VF and extending it as much as possible. Each | |||
8146 | /// vectorization decision can potentially shorten this sub-range during | |||
8147 | /// buildVPlan(). | |||
8148 | void LoopVectorizationPlanner::buildVPlans(ElementCount MinVF, | |||
8149 | ElementCount MaxVF) { | |||
8150 | auto MaxVFPlusOne = MaxVF.getWithIncrement(1); | |||
8151 | for (ElementCount VF = MinVF; ElementCount::isKnownLT(VF, MaxVFPlusOne);) { | |||
8152 | VFRange SubRange = {VF, MaxVFPlusOne}; | |||
8153 | VPlans.push_back(buildVPlan(SubRange)); | |||
8154 | VF = SubRange.End; | |||
8155 | } | |||
8156 | } | |||
8157 | ||||
8158 | VPValue *VPRecipeBuilder::createEdgeMask(BasicBlock *Src, BasicBlock *Dst, | |||
8159 | VPlanPtr &Plan) { | |||
8160 | assert(is_contained(predecessors(Dst), Src) && "Invalid edge")(static_cast <bool> (is_contained(predecessors(Dst), Src ) && "Invalid edge") ? void (0) : __assert_fail ("is_contained(predecessors(Dst), Src) && \"Invalid edge\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8160, __extension__ __PRETTY_FUNCTION__)); | |||
8161 | ||||
8162 | // Look for cached value. | |||
8163 | std::pair<BasicBlock *, BasicBlock *> Edge(Src, Dst); | |||
8164 | EdgeMaskCacheTy::iterator ECEntryIt = EdgeMaskCache.find(Edge); | |||
8165 | if (ECEntryIt != EdgeMaskCache.end()) | |||
8166 | return ECEntryIt->second; | |||
8167 | ||||
8168 | VPValue *SrcMask = createBlockInMask(Src, Plan); | |||
8169 | ||||
8170 | // The terminator has to be a branch inst! | |||
8171 | BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator()); | |||
8172 | assert(BI && "Unexpected terminator found")(static_cast <bool> (BI && "Unexpected terminator found" ) ? void (0) : __assert_fail ("BI && \"Unexpected terminator found\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8172, __extension__ __PRETTY_FUNCTION__)); | |||
8173 | ||||
8174 | if (!BI->isConditional() || BI->getSuccessor(0) == BI->getSuccessor(1)) | |||
8175 | return EdgeMaskCache[Edge] = SrcMask; | |||
8176 | ||||
8177 | // If source is an exiting block, we know the exit edge is dynamically dead | |||
8178 | // in the vector loop, and thus we don't need to restrict the mask. Avoid | |||
8179 | // adding uses of an otherwise potentially dead instruction. | |||
8180 | if (OrigLoop->isLoopExiting(Src)) | |||
8181 | return EdgeMaskCache[Edge] = SrcMask; | |||
8182 | ||||
8183 | VPValue *EdgeMask = Plan->getOrAddVPValue(BI->getCondition()); | |||
8184 | assert(EdgeMask && "No Edge Mask found for condition")(static_cast <bool> (EdgeMask && "No Edge Mask found for condition" ) ? void (0) : __assert_fail ("EdgeMask && \"No Edge Mask found for condition\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8184, __extension__ __PRETTY_FUNCTION__)); | |||
8185 | ||||
8186 | if (BI->getSuccessor(0) != Dst) | |||
8187 | EdgeMask = Builder.createNot(EdgeMask, BI->getDebugLoc()); | |||
8188 | ||||
8189 | if (SrcMask) { // Otherwise block in-mask is all-one, no need to AND. | |||
8190 | // The condition is 'SrcMask && EdgeMask', which is equivalent to | |||
8191 | // 'select i1 SrcMask, i1 EdgeMask, i1 false'. | |||
8192 | // The select version does not introduce new UB if SrcMask is false and | |||
8193 | // EdgeMask is poison. Using 'and' here introduces undefined behavior. | |||
8194 | VPValue *False = Plan->getOrAddVPValue( | |||
8195 | ConstantInt::getFalse(BI->getCondition()->getType())); | |||
8196 | EdgeMask = | |||
8197 | Builder.createSelect(SrcMask, EdgeMask, False, BI->getDebugLoc()); | |||
8198 | } | |||
8199 | ||||
8200 | return EdgeMaskCache[Edge] = EdgeMask; | |||
8201 | } | |||
8202 | ||||
8203 | VPValue *VPRecipeBuilder::createBlockInMask(BasicBlock *BB, VPlanPtr &Plan) { | |||
8204 | assert(OrigLoop->contains(BB) && "Block is not a part of a loop")(static_cast <bool> (OrigLoop->contains(BB) && "Block is not a part of a loop") ? void (0) : __assert_fail ( "OrigLoop->contains(BB) && \"Block is not a part of a loop\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8204, __extension__ __PRETTY_FUNCTION__)); | |||
8205 | ||||
8206 | // Look for cached value. | |||
8207 | BlockMaskCacheTy::iterator BCEntryIt = BlockMaskCache.find(BB); | |||
8208 | if (BCEntryIt != BlockMaskCache.end()) | |||
8209 | return BCEntryIt->second; | |||
8210 | ||||
8211 | // All-one mask is modelled as no-mask following the convention for masked | |||
8212 | // load/store/gather/scatter. Initialize BlockMask to no-mask. | |||
8213 | VPValue *BlockMask = nullptr; | |||
8214 | ||||
8215 | if (OrigLoop->getHeader() == BB) { | |||
8216 | if (!CM.blockNeedsPredicationForAnyReason(BB)) | |||
8217 | return BlockMaskCache[BB] = BlockMask; // Loop incoming mask is all-one. | |||
8218 | ||||
8219 | // Introduce the early-exit compare IV <= BTC to form header block mask. | |||
8220 | // This is used instead of IV < TC because TC may wrap, unlike BTC. Start by | |||
8221 | // constructing the desired canonical IV in the header block as its first | |||
8222 | // non-phi instructions. | |||
8223 | assert(CM.foldTailByMasking() && "must fold the tail")(static_cast <bool> (CM.foldTailByMasking() && "must fold the tail" ) ? void (0) : __assert_fail ("CM.foldTailByMasking() && \"must fold the tail\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8223, __extension__ __PRETTY_FUNCTION__)); | |||
8224 | VPBasicBlock *HeaderVPBB = Plan->getEntry()->getEntryBasicBlock(); | |||
8225 | auto NewInsertionPoint = HeaderVPBB->getFirstNonPhi(); | |||
8226 | auto *IV = new VPWidenCanonicalIVRecipe(Plan->getCanonicalIV()); | |||
8227 | HeaderVPBB->insert(IV, HeaderVPBB->getFirstNonPhi()); | |||
8228 | ||||
8229 | VPBuilder::InsertPointGuard Guard(Builder); | |||
8230 | Builder.setInsertPoint(HeaderVPBB, NewInsertionPoint); | |||
8231 | if (CM.TTI.emitGetActiveLaneMask()) { | |||
8232 | VPValue *TC = Plan->getOrCreateTripCount(); | |||
8233 | BlockMask = Builder.createNaryOp(VPInstruction::ActiveLaneMask, {IV, TC}); | |||
8234 | } else { | |||
8235 | VPValue *BTC = Plan->getOrCreateBackedgeTakenCount(); | |||
8236 | BlockMask = Builder.createNaryOp(VPInstruction::ICmpULE, {IV, BTC}); | |||
8237 | } | |||
8238 | return BlockMaskCache[BB] = BlockMask; | |||
8239 | } | |||
8240 | ||||
8241 | // This is the block mask. We OR all incoming edges. | |||
8242 | for (auto *Predecessor : predecessors(BB)) { | |||
8243 | VPValue *EdgeMask = createEdgeMask(Predecessor, BB, Plan); | |||
8244 | if (!EdgeMask) // Mask of predecessor is all-one so mask of block is too. | |||
8245 | return BlockMaskCache[BB] = EdgeMask; | |||
8246 | ||||
8247 | if (!BlockMask) { // BlockMask has its initialized nullptr value. | |||
8248 | BlockMask = EdgeMask; | |||
8249 | continue; | |||
8250 | } | |||
8251 | ||||
8252 | BlockMask = Builder.createOr(BlockMask, EdgeMask, {}); | |||
8253 | } | |||
8254 | ||||
8255 | return BlockMaskCache[BB] = BlockMask; | |||
8256 | } | |||
8257 | ||||
8258 | VPRecipeBase *VPRecipeBuilder::tryToWidenMemory(Instruction *I, | |||
8259 | ArrayRef<VPValue *> Operands, | |||
8260 | VFRange &Range, | |||
8261 | VPlanPtr &Plan) { | |||
8262 | assert((isa<LoadInst>(I) || isa<StoreInst>(I)) &&(static_cast <bool> ((isa<LoadInst>(I) || isa< StoreInst>(I)) && "Must be called with either a load or store" ) ? void (0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Must be called with either a load or store\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8263, __extension__ __PRETTY_FUNCTION__)) | |||
8263 | "Must be called with either a load or store")(static_cast <bool> ((isa<LoadInst>(I) || isa< StoreInst>(I)) && "Must be called with either a load or store" ) ? void (0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Must be called with either a load or store\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8263, __extension__ __PRETTY_FUNCTION__)); | |||
8264 | ||||
8265 | auto willWiden = [&](ElementCount VF) -> bool { | |||
8266 | if (VF.isScalar()) | |||
8267 | return false; | |||
8268 | LoopVectorizationCostModel::InstWidening Decision = | |||
8269 | CM.getWideningDecision(I, VF); | |||
8270 | assert(Decision != LoopVectorizationCostModel::CM_Unknown &&(static_cast <bool> (Decision != LoopVectorizationCostModel ::CM_Unknown && "CM decision should be taken at this point." ) ? void (0) : __assert_fail ("Decision != LoopVectorizationCostModel::CM_Unknown && \"CM decision should be taken at this point.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8271, __extension__ __PRETTY_FUNCTION__)) | |||
8271 | "CM decision should be taken at this point.")(static_cast <bool> (Decision != LoopVectorizationCostModel ::CM_Unknown && "CM decision should be taken at this point." ) ? void (0) : __assert_fail ("Decision != LoopVectorizationCostModel::CM_Unknown && \"CM decision should be taken at this point.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8271, __extension__ __PRETTY_FUNCTION__)); | |||
8272 | if (Decision == LoopVectorizationCostModel::CM_Interleave) | |||
8273 | return true; | |||
8274 | if (CM.isScalarAfterVectorization(I, VF) || | |||
8275 | CM.isProfitableToScalarize(I, VF)) | |||
8276 | return false; | |||
8277 | return Decision != LoopVectorizationCostModel::CM_Scalarize; | |||
8278 | }; | |||
8279 | ||||
8280 | if (!LoopVectorizationPlanner::getDecisionAndClampRange(willWiden, Range)) | |||
8281 | return nullptr; | |||
8282 | ||||
8283 | VPValue *Mask = nullptr; | |||
8284 | if (Legal->isMaskRequired(I)) | |||
8285 | Mask = createBlockInMask(I->getParent(), Plan); | |||
8286 | ||||
8287 | // Determine if the pointer operand of the access is either consecutive or | |||
8288 | // reverse consecutive. | |||
8289 | LoopVectorizationCostModel::InstWidening Decision = | |||
8290 | CM.getWideningDecision(I, Range.Start); | |||
8291 | bool Reverse = Decision == LoopVectorizationCostModel::CM_Widen_Reverse; | |||
8292 | bool Consecutive = | |||
8293 | Reverse || Decision == LoopVectorizationCostModel::CM_Widen; | |||
8294 | ||||
8295 | if (LoadInst *Load = dyn_cast<LoadInst>(I)) | |||
8296 | return new VPWidenMemoryInstructionRecipe(*Load, Operands[0], Mask, | |||
8297 | Consecutive, Reverse); | |||
8298 | ||||
8299 | StoreInst *Store = cast<StoreInst>(I); | |||
8300 | return new VPWidenMemoryInstructionRecipe(*Store, Operands[1], Operands[0], | |||
8301 | Mask, Consecutive, Reverse); | |||
8302 | } | |||
8303 | ||||
8304 | static VPWidenIntOrFpInductionRecipe * | |||
8305 | createWidenInductionRecipe(PHINode *Phi, Instruction *PhiOrTrunc, | |||
8306 | VPValue *Start, const InductionDescriptor &IndDesc, | |||
8307 | LoopVectorizationCostModel &CM, ScalarEvolution &SE, | |||
8308 | Loop &OrigLoop, VFRange &Range) { | |||
8309 | // Returns true if an instruction \p I should be scalarized instead of | |||
8310 | // vectorized for the chosen vectorization factor. | |||
8311 | auto ShouldScalarizeInstruction = [&CM](Instruction *I, ElementCount VF) { | |||
8312 | return CM.isScalarAfterVectorization(I, VF) || | |||
8313 | CM.isProfitableToScalarize(I, VF); | |||
8314 | }; | |||
8315 | ||||
8316 | bool NeedsScalarIV = LoopVectorizationPlanner::getDecisionAndClampRange( | |||
8317 | [&](ElementCount VF) { | |||
8318 | // Returns true if we should generate a scalar version of \p IV. | |||
8319 | if (ShouldScalarizeInstruction(PhiOrTrunc, VF)) | |||
8320 | return true; | |||
8321 | auto isScalarInst = [&](User *U) -> bool { | |||
8322 | auto *I = cast<Instruction>(U); | |||
8323 | return OrigLoop.contains(I) && ShouldScalarizeInstruction(I, VF); | |||
8324 | }; | |||
8325 | return any_of(PhiOrTrunc->users(), isScalarInst); | |||
8326 | }, | |||
8327 | Range); | |||
8328 | bool NeedsScalarIVOnly = LoopVectorizationPlanner::getDecisionAndClampRange( | |||
8329 | [&](ElementCount VF) { | |||
8330 | return ShouldScalarizeInstruction(PhiOrTrunc, VF); | |||
8331 | }, | |||
8332 | Range); | |||
8333 | assert(IndDesc.getStartValue() ==(static_cast <bool> (IndDesc.getStartValue() == Phi-> getIncomingValueForBlock(OrigLoop.getLoopPreheader())) ? void (0) : __assert_fail ("IndDesc.getStartValue() == Phi->getIncomingValueForBlock(OrigLoop.getLoopPreheader())" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8334, __extension__ __PRETTY_FUNCTION__)) | |||
8334 | Phi->getIncomingValueForBlock(OrigLoop.getLoopPreheader()))(static_cast <bool> (IndDesc.getStartValue() == Phi-> getIncomingValueForBlock(OrigLoop.getLoopPreheader())) ? void (0) : __assert_fail ("IndDesc.getStartValue() == Phi->getIncomingValueForBlock(OrigLoop.getLoopPreheader())" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8334, __extension__ __PRETTY_FUNCTION__)); | |||
8335 | assert(SE.isLoopInvariant(IndDesc.getStep(), &OrigLoop) &&(static_cast <bool> (SE.isLoopInvariant(IndDesc.getStep (), &OrigLoop) && "step must be loop invariant") ? void (0) : __assert_fail ("SE.isLoopInvariant(IndDesc.getStep(), &OrigLoop) && \"step must be loop invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8336, __extension__ __PRETTY_FUNCTION__)) | |||
8336 | "step must be loop invariant")(static_cast <bool> (SE.isLoopInvariant(IndDesc.getStep (), &OrigLoop) && "step must be loop invariant") ? void (0) : __assert_fail ("SE.isLoopInvariant(IndDesc.getStep(), &OrigLoop) && \"step must be loop invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8336, __extension__ __PRETTY_FUNCTION__)); | |||
8337 | if (auto *TruncI = dyn_cast<TruncInst>(PhiOrTrunc)) { | |||
8338 | return new VPWidenIntOrFpInductionRecipe( | |||
8339 | Phi, Start, IndDesc, TruncI, NeedsScalarIV, !NeedsScalarIVOnly, SE); | |||
8340 | } | |||
8341 | assert(isa<PHINode>(PhiOrTrunc) && "must be a phi node here")(static_cast <bool> (isa<PHINode>(PhiOrTrunc) && "must be a phi node here") ? void (0) : __assert_fail ("isa<PHINode>(PhiOrTrunc) && \"must be a phi node here\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8341, __extension__ __PRETTY_FUNCTION__)); | |||
8342 | return new VPWidenIntOrFpInductionRecipe(Phi, Start, IndDesc, NeedsScalarIV, | |||
8343 | !NeedsScalarIVOnly, SE); | |||
8344 | } | |||
8345 | ||||
8346 | VPWidenIntOrFpInductionRecipe *VPRecipeBuilder::tryToOptimizeInductionPHI( | |||
8347 | PHINode *Phi, ArrayRef<VPValue *> Operands, VFRange &Range) const { | |||
8348 | ||||
8349 | // Check if this is an integer or fp induction. If so, build the recipe that | |||
8350 | // produces its scalar and vector values. | |||
8351 | if (auto *II = Legal->getIntOrFpInductionDescriptor(Phi)) | |||
8352 | return createWidenInductionRecipe(Phi, Phi, Operands[0], *II, CM, | |||
8353 | *PSE.getSE(), *OrigLoop, Range); | |||
8354 | ||||
8355 | return nullptr; | |||
8356 | } | |||
8357 | ||||
8358 | VPWidenIntOrFpInductionRecipe *VPRecipeBuilder::tryToOptimizeInductionTruncate( | |||
8359 | TruncInst *I, ArrayRef<VPValue *> Operands, VFRange &Range, | |||
8360 | VPlan &Plan) const { | |||
8361 | // Optimize the special case where the source is a constant integer | |||
8362 | // induction variable. Notice that we can only optimize the 'trunc' case | |||
8363 | // because (a) FP conversions lose precision, (b) sext/zext may wrap, and | |||
8364 | // (c) other casts depend on pointer size. | |||
8365 | ||||
8366 | // Determine whether \p K is a truncation based on an induction variable that | |||
8367 | // can be optimized. | |||
8368 | auto isOptimizableIVTruncate = | |||
8369 | [&](Instruction *K) -> std::function<bool(ElementCount)> { | |||
8370 | return [=](ElementCount VF) -> bool { | |||
8371 | return CM.isOptimizableIVTruncate(K, VF); | |||
8372 | }; | |||
8373 | }; | |||
8374 | ||||
8375 | if (LoopVectorizationPlanner::getDecisionAndClampRange( | |||
8376 | isOptimizableIVTruncate(I), Range)) { | |||
8377 | ||||
8378 | auto *Phi = cast<PHINode>(I->getOperand(0)); | |||
8379 | const InductionDescriptor &II = *Legal->getIntOrFpInductionDescriptor(Phi); | |||
8380 | VPValue *Start = Plan.getOrAddVPValue(II.getStartValue()); | |||
8381 | return createWidenInductionRecipe(Phi, I, Start, II, CM, *PSE.getSE(), | |||
8382 | *OrigLoop, Range); | |||
8383 | } | |||
8384 | return nullptr; | |||
8385 | } | |||
8386 | ||||
8387 | VPRecipeOrVPValueTy VPRecipeBuilder::tryToBlend(PHINode *Phi, | |||
8388 | ArrayRef<VPValue *> Operands, | |||
8389 | VPlanPtr &Plan) { | |||
8390 | // If all incoming values are equal, the incoming VPValue can be used directly | |||
8391 | // instead of creating a new VPBlendRecipe. | |||
8392 | VPValue *FirstIncoming = Operands[0]; | |||
8393 | if (all_of(Operands, [FirstIncoming](const VPValue *Inc) { | |||
8394 | return FirstIncoming == Inc; | |||
8395 | })) { | |||
8396 | return Operands[0]; | |||
8397 | } | |||
8398 | ||||
8399 | unsigned NumIncoming = Phi->getNumIncomingValues(); | |||
8400 | // For in-loop reductions, we do not need to create an additional select. | |||
8401 | VPValue *InLoopVal = nullptr; | |||
8402 | for (unsigned In = 0; In < NumIncoming; In++) { | |||
8403 | PHINode *PhiOp = | |||
8404 | dyn_cast_or_null<PHINode>(Operands[In]->getUnderlyingValue()); | |||
8405 | if (PhiOp && CM.isInLoopReduction(PhiOp)) { | |||
8406 | assert(!InLoopVal && "Found more than one in-loop reduction!")(static_cast <bool> (!InLoopVal && "Found more than one in-loop reduction!" ) ? void (0) : __assert_fail ("!InLoopVal && \"Found more than one in-loop reduction!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8406, __extension__ __PRETTY_FUNCTION__)); | |||
8407 | InLoopVal = Operands[In]; | |||
8408 | } | |||
8409 | } | |||
8410 | ||||
8411 | assert((!InLoopVal || NumIncoming == 2) &&(static_cast <bool> ((!InLoopVal || NumIncoming == 2) && "Found an in-loop reduction for PHI with unexpected number of " "incoming values") ? void (0) : __assert_fail ("(!InLoopVal || NumIncoming == 2) && \"Found an in-loop reduction for PHI with unexpected number of \" \"incoming values\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8413, __extension__ __PRETTY_FUNCTION__)) | |||
8412 | "Found an in-loop reduction for PHI with unexpected number of "(static_cast <bool> ((!InLoopVal || NumIncoming == 2) && "Found an in-loop reduction for PHI with unexpected number of " "incoming values") ? void (0) : __assert_fail ("(!InLoopVal || NumIncoming == 2) && \"Found an in-loop reduction for PHI with unexpected number of \" \"incoming values\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8413, __extension__ __PRETTY_FUNCTION__)) | |||
8413 | "incoming values")(static_cast <bool> ((!InLoopVal || NumIncoming == 2) && "Found an in-loop reduction for PHI with unexpected number of " "incoming values") ? void (0) : __assert_fail ("(!InLoopVal || NumIncoming == 2) && \"Found an in-loop reduction for PHI with unexpected number of \" \"incoming values\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8413, __extension__ __PRETTY_FUNCTION__)); | |||
8414 | if (InLoopVal) | |||
8415 | return Operands[Operands[0] == InLoopVal ? 1 : 0]; | |||
8416 | ||||
8417 | // We know that all PHIs in non-header blocks are converted into selects, so | |||
8418 | // we don't have to worry about the insertion order and we can just use the | |||
8419 | // builder. At this point we generate the predication tree. There may be | |||
8420 | // duplications since this is a simple recursive scan, but future | |||
8421 | // optimizations will clean it up. | |||
8422 | SmallVector<VPValue *, 2> OperandsWithMask; | |||
8423 | ||||
8424 | for (unsigned In = 0; In < NumIncoming; In++) { | |||
8425 | VPValue *EdgeMask = | |||
8426 | createEdgeMask(Phi->getIncomingBlock(In), Phi->getParent(), Plan); | |||
8427 | assert((EdgeMask || NumIncoming == 1) &&(static_cast <bool> ((EdgeMask || NumIncoming == 1) && "Multiple predecessors with one having a full mask") ? void ( 0) : __assert_fail ("(EdgeMask || NumIncoming == 1) && \"Multiple predecessors with one having a full mask\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8428, __extension__ __PRETTY_FUNCTION__)) | |||
8428 | "Multiple predecessors with one having a full mask")(static_cast <bool> ((EdgeMask || NumIncoming == 1) && "Multiple predecessors with one having a full mask") ? void ( 0) : __assert_fail ("(EdgeMask || NumIncoming == 1) && \"Multiple predecessors with one having a full mask\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8428, __extension__ __PRETTY_FUNCTION__)); | |||
8429 | OperandsWithMask.push_back(Operands[In]); | |||
8430 | if (EdgeMask) | |||
8431 | OperandsWithMask.push_back(EdgeMask); | |||
8432 | } | |||
8433 | return toVPRecipeResult(new VPBlendRecipe(Phi, OperandsWithMask)); | |||
8434 | } | |||
8435 | ||||
8436 | VPWidenCallRecipe *VPRecipeBuilder::tryToWidenCall(CallInst *CI, | |||
8437 | ArrayRef<VPValue *> Operands, | |||
8438 | VFRange &Range) const { | |||
8439 | ||||
8440 | bool IsPredicated = LoopVectorizationPlanner::getDecisionAndClampRange( | |||
8441 | [this, CI](ElementCount VF) { | |||
8442 | return CM.isScalarWithPredication(CI, VF); | |||
8443 | }, | |||
8444 | Range); | |||
8445 | ||||
8446 | if (IsPredicated) | |||
8447 | return nullptr; | |||
8448 | ||||
8449 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
8450 | if (ID && (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end || | |||
8451 | ID == Intrinsic::lifetime_start || ID == Intrinsic::sideeffect || | |||
8452 | ID == Intrinsic::pseudoprobe || | |||
8453 | ID == Intrinsic::experimental_noalias_scope_decl)) | |||
8454 | return nullptr; | |||
8455 | ||||
8456 | auto willWiden = [&](ElementCount VF) -> bool { | |||
8457 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
8458 | // The following case may be scalarized depending on the VF. | |||
8459 | // The flag shows whether we use Intrinsic or a usual Call for vectorized | |||
8460 | // version of the instruction. | |||
8461 | // Is it beneficial to perform intrinsic call compared to lib call? | |||
8462 | bool NeedToScalarize = false; | |||
8463 | InstructionCost CallCost = CM.getVectorCallCost(CI, VF, NeedToScalarize); | |||
8464 | InstructionCost IntrinsicCost = ID ? CM.getVectorIntrinsicCost(CI, VF) : 0; | |||
8465 | bool UseVectorIntrinsic = ID && IntrinsicCost <= CallCost; | |||
8466 | return UseVectorIntrinsic || !NeedToScalarize; | |||
8467 | }; | |||
8468 | ||||
8469 | if (!LoopVectorizationPlanner::getDecisionAndClampRange(willWiden, Range)) | |||
8470 | return nullptr; | |||
8471 | ||||
8472 | ArrayRef<VPValue *> Ops = Operands.take_front(CI->arg_size()); | |||
8473 | return new VPWidenCallRecipe(*CI, make_range(Ops.begin(), Ops.end())); | |||
8474 | } | |||
8475 | ||||
8476 | bool VPRecipeBuilder::shouldWiden(Instruction *I, VFRange &Range) const { | |||
8477 | assert(!isa<BranchInst>(I) && !isa<PHINode>(I) && !isa<LoadInst>(I) &&(static_cast <bool> (!isa<BranchInst>(I) && !isa<PHINode>(I) && !isa<LoadInst>(I) && !isa<StoreInst>(I) && "Instruction should have been handled earlier" ) ? void (0) : __assert_fail ("!isa<BranchInst>(I) && !isa<PHINode>(I) && !isa<LoadInst>(I) && !isa<StoreInst>(I) && \"Instruction should have been handled earlier\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8478, __extension__ __PRETTY_FUNCTION__)) | |||
8478 | !isa<StoreInst>(I) && "Instruction should have been handled earlier")(static_cast <bool> (!isa<BranchInst>(I) && !isa<PHINode>(I) && !isa<LoadInst>(I) && !isa<StoreInst>(I) && "Instruction should have been handled earlier" ) ? void (0) : __assert_fail ("!isa<BranchInst>(I) && !isa<PHINode>(I) && !isa<LoadInst>(I) && !isa<StoreInst>(I) && \"Instruction should have been handled earlier\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8478, __extension__ __PRETTY_FUNCTION__)); | |||
8479 | // Instruction should be widened, unless it is scalar after vectorization, | |||
8480 | // scalarization is profitable or it is predicated. | |||
8481 | auto WillScalarize = [this, I](ElementCount VF) -> bool { | |||
8482 | return CM.isScalarAfterVectorization(I, VF) || | |||
8483 | CM.isProfitableToScalarize(I, VF) || | |||
8484 | CM.isScalarWithPredication(I, VF); | |||
8485 | }; | |||
8486 | return !LoopVectorizationPlanner::getDecisionAndClampRange(WillScalarize, | |||
8487 | Range); | |||
8488 | } | |||
8489 | ||||
8490 | VPWidenRecipe *VPRecipeBuilder::tryToWiden(Instruction *I, | |||
8491 | ArrayRef<VPValue *> Operands) const { | |||
8492 | auto IsVectorizableOpcode = [](unsigned Opcode) { | |||
8493 | switch (Opcode) { | |||
8494 | case Instruction::Add: | |||
8495 | case Instruction::And: | |||
8496 | case Instruction::AShr: | |||
8497 | case Instruction::BitCast: | |||
8498 | case Instruction::FAdd: | |||
8499 | case Instruction::FCmp: | |||
8500 | case Instruction::FDiv: | |||
8501 | case Instruction::FMul: | |||
8502 | case Instruction::FNeg: | |||
8503 | case Instruction::FPExt: | |||
8504 | case Instruction::FPToSI: | |||
8505 | case Instruction::FPToUI: | |||
8506 | case Instruction::FPTrunc: | |||
8507 | case Instruction::FRem: | |||
8508 | case Instruction::FSub: | |||
8509 | case Instruction::ICmp: | |||
8510 | case Instruction::IntToPtr: | |||
8511 | case Instruction::LShr: | |||
8512 | case Instruction::Mul: | |||
8513 | case Instruction::Or: | |||
8514 | case Instruction::PtrToInt: | |||
8515 | case Instruction::SDiv: | |||
8516 | case Instruction::Select: | |||
8517 | case Instruction::SExt: | |||
8518 | case Instruction::Shl: | |||
8519 | case Instruction::SIToFP: | |||
8520 | case Instruction::SRem: | |||
8521 | case Instruction::Sub: | |||
8522 | case Instruction::Trunc: | |||
8523 | case Instruction::UDiv: | |||
8524 | case Instruction::UIToFP: | |||
8525 | case Instruction::URem: | |||
8526 | case Instruction::Xor: | |||
8527 | case Instruction::ZExt: | |||
8528 | return true; | |||
8529 | } | |||
8530 | return false; | |||
8531 | }; | |||
8532 | ||||
8533 | if (!IsVectorizableOpcode(I->getOpcode())) | |||
8534 | return nullptr; | |||
8535 | ||||
8536 | // Success: widen this instruction. | |||
8537 | return new VPWidenRecipe(*I, make_range(Operands.begin(), Operands.end())); | |||
8538 | } | |||
8539 | ||||
8540 | void VPRecipeBuilder::fixHeaderPhis() { | |||
8541 | BasicBlock *OrigLatch = OrigLoop->getLoopLatch(); | |||
8542 | for (VPHeaderPHIRecipe *R : PhisToFix) { | |||
8543 | auto *PN = cast<PHINode>(R->getUnderlyingValue()); | |||
8544 | VPRecipeBase *IncR = | |||
8545 | getRecipe(cast<Instruction>(PN->getIncomingValueForBlock(OrigLatch))); | |||
8546 | R->addOperand(IncR->getVPSingleValue()); | |||
8547 | } | |||
8548 | } | |||
8549 | ||||
8550 | VPBasicBlock *VPRecipeBuilder::handleReplication( | |||
8551 | Instruction *I, VFRange &Range, VPBasicBlock *VPBB, | |||
8552 | VPlanPtr &Plan) { | |||
8553 | bool IsUniform = LoopVectorizationPlanner::getDecisionAndClampRange( | |||
8554 | [&](ElementCount VF) { return CM.isUniformAfterVectorization(I, VF); }, | |||
8555 | Range); | |||
8556 | ||||
8557 | bool IsPredicated = LoopVectorizationPlanner::getDecisionAndClampRange( | |||
8558 | [&](ElementCount VF) { return CM.isPredicatedInst(I, VF, IsUniform); }, | |||
8559 | Range); | |||
8560 | ||||
8561 | // Even if the instruction is not marked as uniform, there are certain | |||
8562 | // intrinsic calls that can be effectively treated as such, so we check for | |||
8563 | // them here. Conservatively, we only do this for scalable vectors, since | |||
8564 | // for fixed-width VFs we can always fall back on full scalarization. | |||
8565 | if (!IsUniform && Range.Start.isScalable() && isa<IntrinsicInst>(I)) { | |||
8566 | switch (cast<IntrinsicInst>(I)->getIntrinsicID()) { | |||
8567 | case Intrinsic::assume: | |||
8568 | case Intrinsic::lifetime_start: | |||
8569 | case Intrinsic::lifetime_end: | |||
8570 | // For scalable vectors if one of the operands is variant then we still | |||
8571 | // want to mark as uniform, which will generate one instruction for just | |||
8572 | // the first lane of the vector. We can't scalarize the call in the same | |||
8573 | // way as for fixed-width vectors because we don't know how many lanes | |||
8574 | // there are. | |||
8575 | // | |||
8576 | // The reasons for doing it this way for scalable vectors are: | |||
8577 | // 1. For the assume intrinsic generating the instruction for the first | |||
8578 | // lane is still be better than not generating any at all. For | |||
8579 | // example, the input may be a splat across all lanes. | |||
8580 | // 2. For the lifetime start/end intrinsics the pointer operand only | |||
8581 | // does anything useful when the input comes from a stack object, | |||
8582 | // which suggests it should always be uniform. For non-stack objects | |||
8583 | // the effect is to poison the object, which still allows us to | |||
8584 | // remove the call. | |||
8585 | IsUniform = true; | |||
8586 | break; | |||
8587 | default: | |||
8588 | break; | |||
8589 | } | |||
8590 | } | |||
8591 | ||||
8592 | auto *Recipe = new VPReplicateRecipe(I, Plan->mapToVPValues(I->operands()), | |||
8593 | IsUniform, IsPredicated); | |||
8594 | setRecipe(I, Recipe); | |||
8595 | Plan->addVPValue(I, Recipe); | |||
8596 | ||||
8597 | // Find if I uses a predicated instruction. If so, it will use its scalar | |||
8598 | // value. Avoid hoisting the insert-element which packs the scalar value into | |||
8599 | // a vector value, as that happens iff all users use the vector value. | |||
8600 | for (VPValue *Op : Recipe->operands()) { | |||
8601 | auto *PredR = dyn_cast_or_null<VPPredInstPHIRecipe>(Op->getDef()); | |||
8602 | if (!PredR) | |||
8603 | continue; | |||
8604 | auto *RepR = | |||
8605 | cast_or_null<VPReplicateRecipe>(PredR->getOperand(0)->getDef()); | |||
8606 | assert(RepR->isPredicated() &&(static_cast <bool> (RepR->isPredicated() && "expected Replicate recipe to be predicated") ? void (0) : __assert_fail ("RepR->isPredicated() && \"expected Replicate recipe to be predicated\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8607, __extension__ __PRETTY_FUNCTION__)) | |||
8607 | "expected Replicate recipe to be predicated")(static_cast <bool> (RepR->isPredicated() && "expected Replicate recipe to be predicated") ? void (0) : __assert_fail ("RepR->isPredicated() && \"expected Replicate recipe to be predicated\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8607, __extension__ __PRETTY_FUNCTION__)); | |||
8608 | RepR->setAlsoPack(false); | |||
8609 | } | |||
8610 | ||||
8611 | // Finalize the recipe for Instr, first if it is not predicated. | |||
8612 | if (!IsPredicated) { | |||
8613 | LLVM_DEBUG(dbgs() << "LV: Scalarizing:" << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Scalarizing:" << *I << "\n"; } } while (false); | |||
8614 | VPBB->appendRecipe(Recipe); | |||
8615 | return VPBB; | |||
8616 | } | |||
8617 | LLVM_DEBUG(dbgs() << "LV: Scalarizing and predicating:" << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Scalarizing and predicating:" << *I << "\n"; } } while (false); | |||
8618 | ||||
8619 | VPBlockBase *SingleSucc = VPBB->getSingleSuccessor(); | |||
8620 | assert(SingleSucc && "VPBB must have a single successor when handling "(static_cast <bool> (SingleSucc && "VPBB must have a single successor when handling " "predicated replication.") ? void (0) : __assert_fail ("SingleSucc && \"VPBB must have a single successor when handling \" \"predicated replication.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8621, __extension__ __PRETTY_FUNCTION__)) | |||
8621 | "predicated replication.")(static_cast <bool> (SingleSucc && "VPBB must have a single successor when handling " "predicated replication.") ? void (0) : __assert_fail ("SingleSucc && \"VPBB must have a single successor when handling \" \"predicated replication.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8621, __extension__ __PRETTY_FUNCTION__)); | |||
8622 | VPBlockUtils::disconnectBlocks(VPBB, SingleSucc); | |||
8623 | // Record predicated instructions for above packing optimizations. | |||
8624 | VPBlockBase *Region = createReplicateRegion(I, Recipe, Plan); | |||
8625 | VPBlockUtils::insertBlockAfter(Region, VPBB); | |||
8626 | auto *RegSucc = new VPBasicBlock(); | |||
8627 | VPBlockUtils::insertBlockAfter(RegSucc, Region); | |||
8628 | VPBlockUtils::connectBlocks(RegSucc, SingleSucc); | |||
8629 | return RegSucc; | |||
8630 | } | |||
8631 | ||||
8632 | VPRegionBlock *VPRecipeBuilder::createReplicateRegion(Instruction *Instr, | |||
8633 | VPRecipeBase *PredRecipe, | |||
8634 | VPlanPtr &Plan) { | |||
8635 | // Instructions marked for predication are replicated and placed under an | |||
8636 | // if-then construct to prevent side-effects. | |||
8637 | ||||
8638 | // Generate recipes to compute the block mask for this region. | |||
8639 | VPValue *BlockInMask = createBlockInMask(Instr->getParent(), Plan); | |||
8640 | ||||
8641 | // Build the triangular if-then region. | |||
8642 | std::string RegionName = (Twine("pred.") + Instr->getOpcodeName()).str(); | |||
8643 | assert(Instr->getParent() && "Predicated instruction not in any basic block")(static_cast <bool> (Instr->getParent() && "Predicated instruction not in any basic block" ) ? void (0) : __assert_fail ("Instr->getParent() && \"Predicated instruction not in any basic block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8643, __extension__ __PRETTY_FUNCTION__)); | |||
8644 | auto *BOMRecipe = new VPBranchOnMaskRecipe(BlockInMask); | |||
8645 | auto *Entry = new VPBasicBlock(Twine(RegionName) + ".entry", BOMRecipe); | |||
8646 | auto *PHIRecipe = Instr->getType()->isVoidTy() | |||
8647 | ? nullptr | |||
8648 | : new VPPredInstPHIRecipe(Plan->getOrAddVPValue(Instr)); | |||
8649 | if (PHIRecipe) { | |||
8650 | Plan->removeVPValueFor(Instr); | |||
8651 | Plan->addVPValue(Instr, PHIRecipe); | |||
8652 | } | |||
8653 | auto *Exit = new VPBasicBlock(Twine(RegionName) + ".continue", PHIRecipe); | |||
8654 | auto *Pred = new VPBasicBlock(Twine(RegionName) + ".if", PredRecipe); | |||
8655 | VPRegionBlock *Region = new VPRegionBlock(Entry, Exit, RegionName, true); | |||
8656 | ||||
8657 | // Note: first set Entry as region entry and then connect successors starting | |||
8658 | // from it in order, to propagate the "parent" of each VPBasicBlock. | |||
8659 | VPBlockUtils::insertTwoBlocksAfter(Pred, Exit, BlockInMask, Entry); | |||
8660 | VPBlockUtils::connectBlocks(Pred, Exit); | |||
8661 | ||||
8662 | return Region; | |||
8663 | } | |||
8664 | ||||
8665 | VPRecipeOrVPValueTy | |||
8666 | VPRecipeBuilder::tryToCreateWidenRecipe(Instruction *Instr, | |||
8667 | ArrayRef<VPValue *> Operands, | |||
8668 | VFRange &Range, VPlanPtr &Plan) { | |||
8669 | // First, check for specific widening recipes that deal with calls, memory | |||
8670 | // operations, inductions and Phi nodes. | |||
8671 | if (auto *CI = dyn_cast<CallInst>(Instr)) | |||
8672 | return toVPRecipeResult(tryToWidenCall(CI, Operands, Range)); | |||
8673 | ||||
8674 | if (isa<LoadInst>(Instr) || isa<StoreInst>(Instr)) | |||
8675 | return toVPRecipeResult(tryToWidenMemory(Instr, Operands, Range, Plan)); | |||
8676 | ||||
8677 | VPRecipeBase *Recipe; | |||
8678 | if (auto Phi = dyn_cast<PHINode>(Instr)) { | |||
8679 | if (Phi->getParent() != OrigLoop->getHeader()) | |||
8680 | return tryToBlend(Phi, Operands, Plan); | |||
8681 | if ((Recipe = tryToOptimizeInductionPHI(Phi, Operands, Range))) | |||
8682 | return toVPRecipeResult(Recipe); | |||
8683 | ||||
8684 | VPHeaderPHIRecipe *PhiRecipe = nullptr; | |||
8685 | if (Legal->isReductionVariable(Phi) || Legal->isFirstOrderRecurrence(Phi)) { | |||
8686 | VPValue *StartV = Operands[0]; | |||
8687 | if (Legal->isReductionVariable(Phi)) { | |||
8688 | const RecurrenceDescriptor &RdxDesc = | |||
8689 | Legal->getReductionVars().find(Phi)->second; | |||
8690 | assert(RdxDesc.getRecurrenceStartValue() ==(static_cast <bool> (RdxDesc.getRecurrenceStartValue() == Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader ())) ? void (0) : __assert_fail ("RdxDesc.getRecurrenceStartValue() == Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader())" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8691, __extension__ __PRETTY_FUNCTION__)) | |||
8691 | Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader()))(static_cast <bool> (RdxDesc.getRecurrenceStartValue() == Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader ())) ? void (0) : __assert_fail ("RdxDesc.getRecurrenceStartValue() == Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader())" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8691, __extension__ __PRETTY_FUNCTION__)); | |||
8692 | PhiRecipe = new VPReductionPHIRecipe(Phi, RdxDesc, *StartV, | |||
8693 | CM.isInLoopReduction(Phi), | |||
8694 | CM.useOrderedReductions(RdxDesc)); | |||
8695 | } else { | |||
8696 | PhiRecipe = new VPFirstOrderRecurrencePHIRecipe(Phi, *StartV); | |||
8697 | } | |||
8698 | ||||
8699 | // Record the incoming value from the backedge, so we can add the incoming | |||
8700 | // value from the backedge after all recipes have been created. | |||
8701 | recordRecipeOf(cast<Instruction>( | |||
8702 | Phi->getIncomingValueForBlock(OrigLoop->getLoopLatch()))); | |||
8703 | PhisToFix.push_back(PhiRecipe); | |||
8704 | } else { | |||
8705 | // TODO: record backedge value for remaining pointer induction phis. | |||
8706 | assert(Phi->getType()->isPointerTy() &&(static_cast <bool> (Phi->getType()->isPointerTy( ) && "only pointer phis should be handled here") ? void (0) : __assert_fail ("Phi->getType()->isPointerTy() && \"only pointer phis should be handled here\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8707, __extension__ __PRETTY_FUNCTION__)) | |||
8707 | "only pointer phis should be handled here")(static_cast <bool> (Phi->getType()->isPointerTy( ) && "only pointer phis should be handled here") ? void (0) : __assert_fail ("Phi->getType()->isPointerTy() && \"only pointer phis should be handled here\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8707, __extension__ __PRETTY_FUNCTION__)); | |||
8708 | assert(Legal->getInductionVars().count(Phi) &&(static_cast <bool> (Legal->getInductionVars().count (Phi) && "Not an induction variable") ? void (0) : __assert_fail ("Legal->getInductionVars().count(Phi) && \"Not an induction variable\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8709, __extension__ __PRETTY_FUNCTION__)) | |||
8709 | "Not an induction variable")(static_cast <bool> (Legal->getInductionVars().count (Phi) && "Not an induction variable") ? void (0) : __assert_fail ("Legal->getInductionVars().count(Phi) && \"Not an induction variable\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8709, __extension__ __PRETTY_FUNCTION__)); | |||
8710 | InductionDescriptor II = Legal->getInductionVars().lookup(Phi); | |||
8711 | VPValue *Start = Plan->getOrAddVPValue(II.getStartValue()); | |||
8712 | PhiRecipe = new VPWidenPHIRecipe(Phi, Start); | |||
8713 | } | |||
8714 | ||||
8715 | return toVPRecipeResult(PhiRecipe); | |||
8716 | } | |||
8717 | ||||
8718 | if (isa<TruncInst>(Instr) && | |||
8719 | (Recipe = tryToOptimizeInductionTruncate(cast<TruncInst>(Instr), Operands, | |||
8720 | Range, *Plan))) | |||
8721 | return toVPRecipeResult(Recipe); | |||
8722 | ||||
8723 | if (!shouldWiden(Instr, Range)) | |||
8724 | return nullptr; | |||
8725 | ||||
8726 | if (auto GEP = dyn_cast<GetElementPtrInst>(Instr)) | |||
8727 | return toVPRecipeResult(new VPWidenGEPRecipe( | |||
8728 | GEP, make_range(Operands.begin(), Operands.end()), OrigLoop)); | |||
8729 | ||||
8730 | if (auto *SI = dyn_cast<SelectInst>(Instr)) { | |||
8731 | bool InvariantCond = | |||
8732 | PSE.getSE()->isLoopInvariant(PSE.getSCEV(SI->getOperand(0)), OrigLoop); | |||
8733 | return toVPRecipeResult(new VPWidenSelectRecipe( | |||
8734 | *SI, make_range(Operands.begin(), Operands.end()), InvariantCond)); | |||
8735 | } | |||
8736 | ||||
8737 | return toVPRecipeResult(tryToWiden(Instr, Operands)); | |||
8738 | } | |||
8739 | ||||
8740 | void LoopVectorizationPlanner::buildVPlansWithVPRecipes(ElementCount MinVF, | |||
8741 | ElementCount MaxVF) { | |||
8742 | assert(OrigLoop->isInnermost() && "Inner loop expected.")(static_cast <bool> (OrigLoop->isInnermost() && "Inner loop expected.") ? void (0) : __assert_fail ("OrigLoop->isInnermost() && \"Inner loop expected.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8742, __extension__ __PRETTY_FUNCTION__)); | |||
8743 | ||||
8744 | // Collect instructions from the original loop that will become trivially dead | |||
8745 | // in the vectorized loop. We don't need to vectorize these instructions. For | |||
8746 | // example, original induction update instructions can become dead because we | |||
8747 | // separately emit induction "steps" when generating code for the new loop. | |||
8748 | // Similarly, we create a new latch condition when setting up the structure | |||
8749 | // of the new loop, so the old one can become dead. | |||
8750 | SmallPtrSet<Instruction *, 4> DeadInstructions; | |||
8751 | collectTriviallyDeadInstructions(DeadInstructions); | |||
8752 | ||||
8753 | // Add assume instructions we need to drop to DeadInstructions, to prevent | |||
8754 | // them from being added to the VPlan. | |||
8755 | // TODO: We only need to drop assumes in blocks that get flattend. If the | |||
8756 | // control flow is preserved, we should keep them. | |||
8757 | auto &ConditionalAssumes = Legal->getConditionalAssumes(); | |||
8758 | DeadInstructions.insert(ConditionalAssumes.begin(), ConditionalAssumes.end()); | |||
8759 | ||||
8760 | MapVector<Instruction *, Instruction *> &SinkAfter = Legal->getSinkAfter(); | |||
8761 | // Dead instructions do not need sinking. Remove them from SinkAfter. | |||
8762 | for (Instruction *I : DeadInstructions) | |||
8763 | SinkAfter.erase(I); | |||
8764 | ||||
8765 | // Cannot sink instructions after dead instructions (there won't be any | |||
8766 | // recipes for them). Instead, find the first non-dead previous instruction. | |||
8767 | for (auto &P : Legal->getSinkAfter()) { | |||
8768 | Instruction *SinkTarget = P.second; | |||
8769 | Instruction *FirstInst = &*SinkTarget->getParent()->begin(); | |||
8770 | (void)FirstInst; | |||
8771 | while (DeadInstructions.contains(SinkTarget)) { | |||
8772 | assert((static_cast <bool> (SinkTarget != FirstInst && "Must find a live instruction (at least the one feeding the " "first-order recurrence PHI) before reaching beginning of the block" ) ? void (0) : __assert_fail ("SinkTarget != FirstInst && \"Must find a live instruction (at least the one feeding the \" \"first-order recurrence PHI) before reaching beginning of the block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8775, __extension__ __PRETTY_FUNCTION__)) | |||
8773 | SinkTarget != FirstInst &&(static_cast <bool> (SinkTarget != FirstInst && "Must find a live instruction (at least the one feeding the " "first-order recurrence PHI) before reaching beginning of the block" ) ? void (0) : __assert_fail ("SinkTarget != FirstInst && \"Must find a live instruction (at least the one feeding the \" \"first-order recurrence PHI) before reaching beginning of the block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8775, __extension__ __PRETTY_FUNCTION__)) | |||
8774 | "Must find a live instruction (at least the one feeding the "(static_cast <bool> (SinkTarget != FirstInst && "Must find a live instruction (at least the one feeding the " "first-order recurrence PHI) before reaching beginning of the block" ) ? void (0) : __assert_fail ("SinkTarget != FirstInst && \"Must find a live instruction (at least the one feeding the \" \"first-order recurrence PHI) before reaching beginning of the block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8775, __extension__ __PRETTY_FUNCTION__)) | |||
8775 | "first-order recurrence PHI) before reaching beginning of the block")(static_cast <bool> (SinkTarget != FirstInst && "Must find a live instruction (at least the one feeding the " "first-order recurrence PHI) before reaching beginning of the block" ) ? void (0) : __assert_fail ("SinkTarget != FirstInst && \"Must find a live instruction (at least the one feeding the \" \"first-order recurrence PHI) before reaching beginning of the block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8775, __extension__ __PRETTY_FUNCTION__)); | |||
8776 | SinkTarget = SinkTarget->getPrevNode(); | |||
8777 | assert(SinkTarget != P.first &&(static_cast <bool> (SinkTarget != P.first && "sink source equals target, no sinking required" ) ? void (0) : __assert_fail ("SinkTarget != P.first && \"sink source equals target, no sinking required\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8778, __extension__ __PRETTY_FUNCTION__)) | |||
8778 | "sink source equals target, no sinking required")(static_cast <bool> (SinkTarget != P.first && "sink source equals target, no sinking required" ) ? void (0) : __assert_fail ("SinkTarget != P.first && \"sink source equals target, no sinking required\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8778, __extension__ __PRETTY_FUNCTION__)); | |||
8779 | } | |||
8780 | P.second = SinkTarget; | |||
8781 | } | |||
8782 | ||||
8783 | auto MaxVFPlusOne = MaxVF.getWithIncrement(1); | |||
8784 | for (ElementCount VF = MinVF; ElementCount::isKnownLT(VF, MaxVFPlusOne);) { | |||
8785 | VFRange SubRange = {VF, MaxVFPlusOne}; | |||
8786 | VPlans.push_back( | |||
8787 | buildVPlanWithVPRecipes(SubRange, DeadInstructions, SinkAfter)); | |||
8788 | VF = SubRange.End; | |||
8789 | } | |||
8790 | } | |||
8791 | ||||
8792 | // Add a VPCanonicalIVPHIRecipe starting at 0 to the header, a | |||
8793 | // CanonicalIVIncrement{NUW} VPInstruction to increment it by VF * UF and a | |||
8794 | // BranchOnCount VPInstruction to the latch. | |||
8795 | static void addCanonicalIVRecipes(VPlan &Plan, Type *IdxTy, DebugLoc DL, | |||
8796 | bool HasNUW, bool IsVPlanNative) { | |||
8797 | Value *StartIdx = ConstantInt::get(IdxTy, 0); | |||
8798 | auto *StartV = Plan.getOrAddVPValue(StartIdx); | |||
8799 | ||||
8800 | auto *CanonicalIVPHI = new VPCanonicalIVPHIRecipe(StartV, DL); | |||
8801 | VPRegionBlock *TopRegion = Plan.getVectorLoopRegion(); | |||
8802 | VPBasicBlock *Header = TopRegion->getEntryBasicBlock(); | |||
8803 | if (IsVPlanNative) | |||
8804 | Header = cast<VPBasicBlock>(Header->getSingleSuccessor()); | |||
8805 | Header->insert(CanonicalIVPHI, Header->begin()); | |||
8806 | ||||
8807 | auto *CanonicalIVIncrement = | |||
8808 | new VPInstruction(HasNUW ? VPInstruction::CanonicalIVIncrementNUW | |||
8809 | : VPInstruction::CanonicalIVIncrement, | |||
8810 | {CanonicalIVPHI}, DL); | |||
8811 | CanonicalIVPHI->addOperand(CanonicalIVIncrement); | |||
8812 | ||||
8813 | VPBasicBlock *EB = TopRegion->getExitBasicBlock(); | |||
8814 | if (IsVPlanNative) { | |||
8815 | EB = cast<VPBasicBlock>(EB->getSinglePredecessor()); | |||
8816 | EB->setCondBit(nullptr); | |||
8817 | } | |||
8818 | EB->appendRecipe(CanonicalIVIncrement); | |||
8819 | ||||
8820 | auto *BranchOnCount = | |||
8821 | new VPInstruction(VPInstruction::BranchOnCount, | |||
8822 | {CanonicalIVIncrement, &Plan.getVectorTripCount()}, DL); | |||
8823 | EB->appendRecipe(BranchOnCount); | |||
8824 | } | |||
8825 | ||||
8826 | VPlanPtr LoopVectorizationPlanner::buildVPlanWithVPRecipes( | |||
8827 | VFRange &Range, SmallPtrSetImpl<Instruction *> &DeadInstructions, | |||
8828 | const MapVector<Instruction *, Instruction *> &SinkAfter) { | |||
8829 | ||||
8830 | SmallPtrSet<const InterleaveGroup<Instruction> *, 1> InterleaveGroups; | |||
8831 | ||||
8832 | VPRecipeBuilder RecipeBuilder(OrigLoop, TLI, Legal, CM, PSE, Builder); | |||
8833 | ||||
8834 | // --------------------------------------------------------------------------- | |||
8835 | // Pre-construction: record ingredients whose recipes we'll need to further | |||
8836 | // process after constructing the initial VPlan. | |||
8837 | // --------------------------------------------------------------------------- | |||
8838 | ||||
8839 | // Mark instructions we'll need to sink later and their targets as | |||
8840 | // ingredients whose recipe we'll need to record. | |||
8841 | for (auto &Entry : SinkAfter) { | |||
8842 | RecipeBuilder.recordRecipeOf(Entry.first); | |||
8843 | RecipeBuilder.recordRecipeOf(Entry.second); | |||
8844 | } | |||
8845 | for (auto &Reduction : CM.getInLoopReductionChains()) { | |||
8846 | PHINode *Phi = Reduction.first; | |||
8847 | RecurKind Kind = | |||
8848 | Legal->getReductionVars().find(Phi)->second.getRecurrenceKind(); | |||
8849 | const SmallVector<Instruction *, 4> &ReductionOperations = Reduction.second; | |||
8850 | ||||
8851 | RecipeBuilder.recordRecipeOf(Phi); | |||
8852 | for (auto &R : ReductionOperations) { | |||
8853 | RecipeBuilder.recordRecipeOf(R); | |||
8854 | // For min/max reducitons, where we have a pair of icmp/select, we also | |||
8855 | // need to record the ICmp recipe, so it can be removed later. | |||
8856 | assert(!RecurrenceDescriptor::isSelectCmpRecurrenceKind(Kind) &&(static_cast <bool> (!RecurrenceDescriptor::isSelectCmpRecurrenceKind (Kind) && "Only min/max recurrences allowed for inloop reductions" ) ? void (0) : __assert_fail ("!RecurrenceDescriptor::isSelectCmpRecurrenceKind(Kind) && \"Only min/max recurrences allowed for inloop reductions\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8857, __extension__ __PRETTY_FUNCTION__)) | |||
8857 | "Only min/max recurrences allowed for inloop reductions")(static_cast <bool> (!RecurrenceDescriptor::isSelectCmpRecurrenceKind (Kind) && "Only min/max recurrences allowed for inloop reductions" ) ? void (0) : __assert_fail ("!RecurrenceDescriptor::isSelectCmpRecurrenceKind(Kind) && \"Only min/max recurrences allowed for inloop reductions\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8857, __extension__ __PRETTY_FUNCTION__)); | |||
8858 | if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) | |||
8859 | RecipeBuilder.recordRecipeOf(cast<Instruction>(R->getOperand(0))); | |||
8860 | } | |||
8861 | } | |||
8862 | ||||
8863 | // For each interleave group which is relevant for this (possibly trimmed) | |||
8864 | // Range, add it to the set of groups to be later applied to the VPlan and add | |||
8865 | // placeholders for its members' Recipes which we'll be replacing with a | |||
8866 | // single VPInterleaveRecipe. | |||
8867 | for (InterleaveGroup<Instruction> *IG : IAI.getInterleaveGroups()) { | |||
8868 | auto applyIG = [IG, this](ElementCount VF) -> bool { | |||
8869 | return (VF.isVector() && // Query is illegal for VF == 1 | |||
8870 | CM.getWideningDecision(IG->getInsertPos(), VF) == | |||
8871 | LoopVectorizationCostModel::CM_Interleave); | |||
8872 | }; | |||
8873 | if (!getDecisionAndClampRange(applyIG, Range)) | |||
8874 | continue; | |||
8875 | InterleaveGroups.insert(IG); | |||
8876 | for (unsigned i = 0; i < IG->getFactor(); i++) | |||
8877 | if (Instruction *Member = IG->getMember(i)) | |||
8878 | RecipeBuilder.recordRecipeOf(Member); | |||
8879 | }; | |||
8880 | ||||
8881 | // --------------------------------------------------------------------------- | |||
8882 | // Build initial VPlan: Scan the body of the loop in a topological order to | |||
8883 | // visit each basic block after having visited its predecessor basic blocks. | |||
8884 | // --------------------------------------------------------------------------- | |||
8885 | ||||
8886 | // Create initial VPlan skeleton, with separate header and latch blocks. | |||
8887 | VPBasicBlock *HeaderVPBB = new VPBasicBlock(); | |||
8888 | VPBasicBlock *LatchVPBB = new VPBasicBlock("vector.latch"); | |||
8889 | VPBlockUtils::insertBlockAfter(LatchVPBB, HeaderVPBB); | |||
8890 | auto *TopRegion = new VPRegionBlock(HeaderVPBB, LatchVPBB, "vector loop"); | |||
8891 | auto Plan = std::make_unique<VPlan>(TopRegion); | |||
8892 | ||||
8893 | Instruction *DLInst = | |||
8894 | getDebugLocFromInstOrOperands(Legal->getPrimaryInduction()); | |||
8895 | addCanonicalIVRecipes(*Plan, Legal->getWidestInductionType(), | |||
8896 | DLInst ? DLInst->getDebugLoc() : DebugLoc(), | |||
8897 | !CM.foldTailByMasking(), false); | |||
8898 | ||||
8899 | // Scan the body of the loop in a topological order to visit each basic block | |||
8900 | // after having visited its predecessor basic blocks. | |||
8901 | LoopBlocksDFS DFS(OrigLoop); | |||
8902 | DFS.perform(LI); | |||
8903 | ||||
8904 | VPBasicBlock *VPBB = HeaderVPBB; | |||
8905 | SmallVector<VPWidenIntOrFpInductionRecipe *> InductionsToMove; | |||
8906 | for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO())) { | |||
8907 | // Relevant instructions from basic block BB will be grouped into VPRecipe | |||
8908 | // ingredients and fill a new VPBasicBlock. | |||
8909 | unsigned VPBBsForBB = 0; | |||
8910 | VPBB->setName(BB->getName()); | |||
8911 | Builder.setInsertPoint(VPBB); | |||
8912 | ||||
8913 | // Introduce each ingredient into VPlan. | |||
8914 | // TODO: Model and preserve debug instrinsics in VPlan. | |||
8915 | for (Instruction &I : BB->instructionsWithoutDebug()) { | |||
8916 | Instruction *Instr = &I; | |||
8917 | ||||
8918 | // First filter out irrelevant instructions, to ensure no recipes are | |||
8919 | // built for them. | |||
8920 | if (isa<BranchInst>(Instr) || DeadInstructions.count(Instr)) | |||
8921 | continue; | |||
8922 | ||||
8923 | SmallVector<VPValue *, 4> Operands; | |||
8924 | auto *Phi = dyn_cast<PHINode>(Instr); | |||
8925 | if (Phi && Phi->getParent() == OrigLoop->getHeader()) { | |||
8926 | Operands.push_back(Plan->getOrAddVPValue( | |||
8927 | Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader()))); | |||
8928 | } else { | |||
8929 | auto OpRange = Plan->mapToVPValues(Instr->operands()); | |||
8930 | Operands = {OpRange.begin(), OpRange.end()}; | |||
8931 | } | |||
8932 | if (auto RecipeOrValue = RecipeBuilder.tryToCreateWidenRecipe( | |||
8933 | Instr, Operands, Range, Plan)) { | |||
8934 | // If Instr can be simplified to an existing VPValue, use it. | |||
8935 | if (RecipeOrValue.is<VPValue *>()) { | |||
8936 | auto *VPV = RecipeOrValue.get<VPValue *>(); | |||
8937 | Plan->addVPValue(Instr, VPV); | |||
8938 | // If the re-used value is a recipe, register the recipe for the | |||
8939 | // instruction, in case the recipe for Instr needs to be recorded. | |||
8940 | if (auto *R = dyn_cast_or_null<VPRecipeBase>(VPV->getDef())) | |||
8941 | RecipeBuilder.setRecipe(Instr, R); | |||
8942 | continue; | |||
8943 | } | |||
8944 | // Otherwise, add the new recipe. | |||
8945 | VPRecipeBase *Recipe = RecipeOrValue.get<VPRecipeBase *>(); | |||
8946 | for (auto *Def : Recipe->definedValues()) { | |||
8947 | auto *UV = Def->getUnderlyingValue(); | |||
8948 | Plan->addVPValue(UV, Def); | |||
8949 | } | |||
8950 | ||||
8951 | if (isa<VPWidenIntOrFpInductionRecipe>(Recipe) && | |||
8952 | HeaderVPBB->getFirstNonPhi() != VPBB->end()) { | |||
8953 | // Keep track of VPWidenIntOrFpInductionRecipes not in the phi section | |||
8954 | // of the header block. That can happen for truncates of induction | |||
8955 | // variables. Those recipes are moved to the phi section of the header | |||
8956 | // block after applying SinkAfter, which relies on the original | |||
8957 | // position of the trunc. | |||
8958 | assert(isa<TruncInst>(Instr))(static_cast <bool> (isa<TruncInst>(Instr)) ? void (0) : __assert_fail ("isa<TruncInst>(Instr)", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 8958, __extension__ __PRETTY_FUNCTION__)); | |||
8959 | InductionsToMove.push_back( | |||
8960 | cast<VPWidenIntOrFpInductionRecipe>(Recipe)); | |||
8961 | } | |||
8962 | RecipeBuilder.setRecipe(Instr, Recipe); | |||
8963 | VPBB->appendRecipe(Recipe); | |||
8964 | continue; | |||
8965 | } | |||
8966 | ||||
8967 | // Otherwise, if all widening options failed, Instruction is to be | |||
8968 | // replicated. This may create a successor for VPBB. | |||
8969 | VPBasicBlock *NextVPBB = | |||
8970 | RecipeBuilder.handleReplication(Instr, Range, VPBB, Plan); | |||
8971 | if (NextVPBB != VPBB) { | |||
8972 | VPBB = NextVPBB; | |||
8973 | VPBB->setName(BB->hasName() ? BB->getName() + "." + Twine(VPBBsForBB++) | |||
8974 | : ""); | |||
8975 | } | |||
8976 | } | |||
8977 | ||||
8978 | VPBlockUtils::insertBlockAfter(new VPBasicBlock(), VPBB); | |||
8979 | VPBB = cast<VPBasicBlock>(VPBB->getSingleSuccessor()); | |||
8980 | } | |||
8981 | ||||
8982 | // Fold the last, empty block into its predecessor. | |||
8983 | VPBB = VPBlockUtils::tryToMergeBlockIntoPredecessor(VPBB); | |||
8984 | assert(VPBB && "expected to fold last (empty) block")(static_cast <bool> (VPBB && "expected to fold last (empty) block" ) ? void (0) : __assert_fail ("VPBB && \"expected to fold last (empty) block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8984, __extension__ __PRETTY_FUNCTION__)); | |||
8985 | // After here, VPBB should not be used. | |||
8986 | VPBB = nullptr; | |||
8987 | ||||
8988 | assert(isa<VPRegionBlock>(Plan->getEntry()) &&(static_cast <bool> (isa<VPRegionBlock>(Plan-> getEntry()) && !Plan->getEntry()->getEntryBasicBlock ()->empty() && "entry block must be set to a VPRegionBlock having a non-empty entry " "VPBasicBlock") ? void (0) : __assert_fail ("isa<VPRegionBlock>(Plan->getEntry()) && !Plan->getEntry()->getEntryBasicBlock()->empty() && \"entry block must be set to a VPRegionBlock having a non-empty entry \" \"VPBasicBlock\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8991, __extension__ __PRETTY_FUNCTION__)) | |||
8989 | !Plan->getEntry()->getEntryBasicBlock()->empty() &&(static_cast <bool> (isa<VPRegionBlock>(Plan-> getEntry()) && !Plan->getEntry()->getEntryBasicBlock ()->empty() && "entry block must be set to a VPRegionBlock having a non-empty entry " "VPBasicBlock") ? void (0) : __assert_fail ("isa<VPRegionBlock>(Plan->getEntry()) && !Plan->getEntry()->getEntryBasicBlock()->empty() && \"entry block must be set to a VPRegionBlock having a non-empty entry \" \"VPBasicBlock\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8991, __extension__ __PRETTY_FUNCTION__)) | |||
8990 | "entry block must be set to a VPRegionBlock having a non-empty entry "(static_cast <bool> (isa<VPRegionBlock>(Plan-> getEntry()) && !Plan->getEntry()->getEntryBasicBlock ()->empty() && "entry block must be set to a VPRegionBlock having a non-empty entry " "VPBasicBlock") ? void (0) : __assert_fail ("isa<VPRegionBlock>(Plan->getEntry()) && !Plan->getEntry()->getEntryBasicBlock()->empty() && \"entry block must be set to a VPRegionBlock having a non-empty entry \" \"VPBasicBlock\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8991, __extension__ __PRETTY_FUNCTION__)) | |||
8991 | "VPBasicBlock")(static_cast <bool> (isa<VPRegionBlock>(Plan-> getEntry()) && !Plan->getEntry()->getEntryBasicBlock ()->empty() && "entry block must be set to a VPRegionBlock having a non-empty entry " "VPBasicBlock") ? void (0) : __assert_fail ("isa<VPRegionBlock>(Plan->getEntry()) && !Plan->getEntry()->getEntryBasicBlock()->empty() && \"entry block must be set to a VPRegionBlock having a non-empty entry \" \"VPBasicBlock\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8991, __extension__ __PRETTY_FUNCTION__)); | |||
8992 | RecipeBuilder.fixHeaderPhis(); | |||
8993 | ||||
8994 | // --------------------------------------------------------------------------- | |||
8995 | // Transform initial VPlan: Apply previously taken decisions, in order, to | |||
8996 | // bring the VPlan to its final state. | |||
8997 | // --------------------------------------------------------------------------- | |||
8998 | ||||
8999 | // Apply Sink-After legal constraints. | |||
9000 | auto GetReplicateRegion = [](VPRecipeBase *R) -> VPRegionBlock * { | |||
9001 | auto *Region = dyn_cast_or_null<VPRegionBlock>(R->getParent()->getParent()); | |||
9002 | if (Region && Region->isReplicator()) { | |||
9003 | assert(Region->getNumSuccessors() == 1 &&(static_cast <bool> (Region->getNumSuccessors() == 1 && Region->getNumPredecessors() == 1 && "Expected SESE region!" ) ? void (0) : __assert_fail ("Region->getNumSuccessors() == 1 && Region->getNumPredecessors() == 1 && \"Expected SESE region!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9004, __extension__ __PRETTY_FUNCTION__)) | |||
9004 | Region->getNumPredecessors() == 1 && "Expected SESE region!")(static_cast <bool> (Region->getNumSuccessors() == 1 && Region->getNumPredecessors() == 1 && "Expected SESE region!" ) ? void (0) : __assert_fail ("Region->getNumSuccessors() == 1 && Region->getNumPredecessors() == 1 && \"Expected SESE region!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9004, __extension__ __PRETTY_FUNCTION__)); | |||
9005 | assert(R->getParent()->size() == 1 &&(static_cast <bool> (R->getParent()->size() == 1 && "A recipe in an original replicator region must be the only " "recipe in its block") ? void (0) : __assert_fail ("R->getParent()->size() == 1 && \"A recipe in an original replicator region must be the only \" \"recipe in its block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9007, __extension__ __PRETTY_FUNCTION__)) | |||
9006 | "A recipe in an original replicator region must be the only "(static_cast <bool> (R->getParent()->size() == 1 && "A recipe in an original replicator region must be the only " "recipe in its block") ? void (0) : __assert_fail ("R->getParent()->size() == 1 && \"A recipe in an original replicator region must be the only \" \"recipe in its block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9007, __extension__ __PRETTY_FUNCTION__)) | |||
9007 | "recipe in its block")(static_cast <bool> (R->getParent()->size() == 1 && "A recipe in an original replicator region must be the only " "recipe in its block") ? void (0) : __assert_fail ("R->getParent()->size() == 1 && \"A recipe in an original replicator region must be the only \" \"recipe in its block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9007, __extension__ __PRETTY_FUNCTION__)); | |||
9008 | return Region; | |||
9009 | } | |||
9010 | return nullptr; | |||
9011 | }; | |||
9012 | for (auto &Entry : SinkAfter) { | |||
9013 | VPRecipeBase *Sink = RecipeBuilder.getRecipe(Entry.first); | |||
9014 | VPRecipeBase *Target = RecipeBuilder.getRecipe(Entry.second); | |||
9015 | ||||
9016 | auto *TargetRegion = GetReplicateRegion(Target); | |||
9017 | auto *SinkRegion = GetReplicateRegion(Sink); | |||
9018 | if (!SinkRegion) { | |||
9019 | // If the sink source is not a replicate region, sink the recipe directly. | |||
9020 | if (TargetRegion) { | |||
9021 | // The target is in a replication region, make sure to move Sink to | |||
9022 | // the block after it, not into the replication region itself. | |||
9023 | VPBasicBlock *NextBlock = | |||
9024 | cast<VPBasicBlock>(TargetRegion->getSuccessors().front()); | |||
9025 | Sink->moveBefore(*NextBlock, NextBlock->getFirstNonPhi()); | |||
9026 | } else | |||
9027 | Sink->moveAfter(Target); | |||
9028 | continue; | |||
9029 | } | |||
9030 | ||||
9031 | // The sink source is in a replicate region. Unhook the region from the CFG. | |||
9032 | auto *SinkPred = SinkRegion->getSinglePredecessor(); | |||
9033 | auto *SinkSucc = SinkRegion->getSingleSuccessor(); | |||
9034 | VPBlockUtils::disconnectBlocks(SinkPred, SinkRegion); | |||
9035 | VPBlockUtils::disconnectBlocks(SinkRegion, SinkSucc); | |||
9036 | VPBlockUtils::connectBlocks(SinkPred, SinkSucc); | |||
9037 | ||||
9038 | if (TargetRegion) { | |||
9039 | // The target recipe is also in a replicate region, move the sink region | |||
9040 | // after the target region. | |||
9041 | auto *TargetSucc = TargetRegion->getSingleSuccessor(); | |||
9042 | VPBlockUtils::disconnectBlocks(TargetRegion, TargetSucc); | |||
9043 | VPBlockUtils::connectBlocks(TargetRegion, SinkRegion); | |||
9044 | VPBlockUtils::connectBlocks(SinkRegion, TargetSucc); | |||
9045 | } else { | |||
9046 | // The sink source is in a replicate region, we need to move the whole | |||
9047 | // replicate region, which should only contain a single recipe in the | |||
9048 | // main block. | |||
9049 | auto *SplitBlock = | |||
9050 | Target->getParent()->splitAt(std::next(Target->getIterator())); | |||
9051 | ||||
9052 | auto *SplitPred = SplitBlock->getSinglePredecessor(); | |||
9053 | ||||
9054 | VPBlockUtils::disconnectBlocks(SplitPred, SplitBlock); | |||
9055 | VPBlockUtils::connectBlocks(SplitPred, SinkRegion); | |||
9056 | VPBlockUtils::connectBlocks(SinkRegion, SplitBlock); | |||
9057 | } | |||
9058 | } | |||
9059 | ||||
9060 | VPlanTransforms::removeRedundantCanonicalIVs(*Plan); | |||
9061 | VPlanTransforms::removeRedundantInductionCasts(*Plan); | |||
9062 | ||||
9063 | // Now that sink-after is done, move induction recipes for optimized truncates | |||
9064 | // to the phi section of the header block. | |||
9065 | for (VPWidenIntOrFpInductionRecipe *Ind : InductionsToMove) | |||
9066 | Ind->moveBefore(*HeaderVPBB, HeaderVPBB->getFirstNonPhi()); | |||
9067 | ||||
9068 | // Adjust the recipes for any inloop reductions. | |||
9069 | adjustRecipesForReductions(cast<VPBasicBlock>(TopRegion->getExit()), Plan, | |||
9070 | RecipeBuilder, Range.Start); | |||
9071 | ||||
9072 | // Introduce a recipe to combine the incoming and previous values of a | |||
9073 | // first-order recurrence. | |||
9074 | for (VPRecipeBase &R : Plan->getEntry()->getEntryBasicBlock()->phis()) { | |||
9075 | auto *RecurPhi = dyn_cast<VPFirstOrderRecurrencePHIRecipe>(&R); | |||
9076 | if (!RecurPhi) | |||
9077 | continue; | |||
9078 | ||||
9079 | VPRecipeBase *PrevRecipe = RecurPhi->getBackedgeRecipe(); | |||
9080 | VPBasicBlock *InsertBlock = PrevRecipe->getParent(); | |||
9081 | auto *Region = GetReplicateRegion(PrevRecipe); | |||
9082 | if (Region) | |||
9083 | InsertBlock = cast<VPBasicBlock>(Region->getSingleSuccessor()); | |||
9084 | if (Region || PrevRecipe->isPhi()) | |||
9085 | Builder.setInsertPoint(InsertBlock, InsertBlock->getFirstNonPhi()); | |||
9086 | else | |||
9087 | Builder.setInsertPoint(InsertBlock, std::next(PrevRecipe->getIterator())); | |||
9088 | ||||
9089 | auto *RecurSplice = cast<VPInstruction>( | |||
9090 | Builder.createNaryOp(VPInstruction::FirstOrderRecurrenceSplice, | |||
9091 | {RecurPhi, RecurPhi->getBackedgeValue()})); | |||
9092 | ||||
9093 | RecurPhi->replaceAllUsesWith(RecurSplice); | |||
9094 | // Set the first operand of RecurSplice to RecurPhi again, after replacing | |||
9095 | // all users. | |||
9096 | RecurSplice->setOperand(0, RecurPhi); | |||
9097 | } | |||
9098 | ||||
9099 | // Interleave memory: for each Interleave Group we marked earlier as relevant | |||
9100 | // for this VPlan, replace the Recipes widening its memory instructions with a | |||
9101 | // single VPInterleaveRecipe at its insertion point. | |||
9102 | for (auto IG : InterleaveGroups) { | |||
9103 | auto *Recipe = cast<VPWidenMemoryInstructionRecipe>( | |||
9104 | RecipeBuilder.getRecipe(IG->getInsertPos())); | |||
9105 | SmallVector<VPValue *, 4> StoredValues; | |||
9106 | for (unsigned i = 0; i < IG->getFactor(); ++i) | |||
9107 | if (auto *SI = dyn_cast_or_null<StoreInst>(IG->getMember(i))) { | |||
9108 | auto *StoreR = | |||
9109 | cast<VPWidenMemoryInstructionRecipe>(RecipeBuilder.getRecipe(SI)); | |||
9110 | StoredValues.push_back(StoreR->getStoredValue()); | |||
9111 | } | |||
9112 | ||||
9113 | auto *VPIG = new VPInterleaveRecipe(IG, Recipe->getAddr(), StoredValues, | |||
9114 | Recipe->getMask()); | |||
9115 | VPIG->insertBefore(Recipe); | |||
9116 | unsigned J = 0; | |||
9117 | for (unsigned i = 0; i < IG->getFactor(); ++i) | |||
9118 | if (Instruction *Member = IG->getMember(i)) { | |||
9119 | if (!Member->getType()->isVoidTy()) { | |||
9120 | VPValue *OriginalV = Plan->getVPValue(Member); | |||
9121 | Plan->removeVPValueFor(Member); | |||
9122 | Plan->addVPValue(Member, VPIG->getVPValue(J)); | |||
9123 | OriginalV->replaceAllUsesWith(VPIG->getVPValue(J)); | |||
9124 | J++; | |||
9125 | } | |||
9126 | RecipeBuilder.getRecipe(Member)->eraseFromParent(); | |||
9127 | } | |||
9128 | } | |||
9129 | ||||
9130 | // From this point onwards, VPlan-to-VPlan transformations may change the plan | |||
9131 | // in ways that accessing values using original IR values is incorrect. | |||
9132 | Plan->disableValue2VPValue(); | |||
9133 | ||||
9134 | VPlanTransforms::optimizeInductions(*Plan, *PSE.getSE()); | |||
9135 | VPlanTransforms::sinkScalarOperands(*Plan); | |||
9136 | VPlanTransforms::mergeReplicateRegions(*Plan); | |||
9137 | VPlanTransforms::removeDeadRecipes(*Plan, *OrigLoop); | |||
9138 | ||||
9139 | std::string PlanName; | |||
9140 | raw_string_ostream RSO(PlanName); | |||
9141 | ElementCount VF = Range.Start; | |||
9142 | Plan->addVF(VF); | |||
9143 | RSO << "Initial VPlan for VF={" << VF; | |||
9144 | for (VF *= 2; ElementCount::isKnownLT(VF, Range.End); VF *= 2) { | |||
9145 | Plan->addVF(VF); | |||
9146 | RSO << "," << VF; | |||
9147 | } | |||
9148 | RSO << "},UF>=1"; | |||
9149 | RSO.flush(); | |||
9150 | Plan->setName(PlanName); | |||
9151 | ||||
9152 | // Fold Exit block into its predecessor if possible. | |||
9153 | // TODO: Fold block earlier once all VPlan transforms properly maintain a | |||
9154 | // VPBasicBlock as exit. | |||
9155 | VPBlockUtils::tryToMergeBlockIntoPredecessor(TopRegion->getExit()); | |||
9156 | ||||
9157 | assert(VPlanVerifier::verifyPlanIsValid(*Plan) && "VPlan is invalid")(static_cast <bool> (VPlanVerifier::verifyPlanIsValid(* Plan) && "VPlan is invalid") ? void (0) : __assert_fail ("VPlanVerifier::verifyPlanIsValid(*Plan) && \"VPlan is invalid\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9157, __extension__ __PRETTY_FUNCTION__)); | |||
9158 | return Plan; | |||
9159 | } | |||
9160 | ||||
9161 | VPlanPtr LoopVectorizationPlanner::buildVPlan(VFRange &Range) { | |||
9162 | // Outer loop handling: They may require CFG and instruction level | |||
9163 | // transformations before even evaluating whether vectorization is profitable. | |||
9164 | // Since we cannot modify the incoming IR, we need to build VPlan upfront in | |||
9165 | // the vectorization pipeline. | |||
9166 | assert(!OrigLoop->isInnermost())(static_cast <bool> (!OrigLoop->isInnermost()) ? void (0) : __assert_fail ("!OrigLoop->isInnermost()", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 9166, __extension__ __PRETTY_FUNCTION__)); | |||
9167 | assert(EnableVPlanNativePath && "VPlan-native path is not enabled.")(static_cast <bool> (EnableVPlanNativePath && "VPlan-native path is not enabled." ) ? void (0) : __assert_fail ("EnableVPlanNativePath && \"VPlan-native path is not enabled.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9167, __extension__ __PRETTY_FUNCTION__)); | |||
9168 | ||||
9169 | // Create new empty VPlan | |||
9170 | auto Plan = std::make_unique<VPlan>(); | |||
9171 | ||||
9172 | // Build hierarchical CFG | |||
9173 | VPlanHCFGBuilder HCFGBuilder(OrigLoop, LI, *Plan); | |||
9174 | HCFGBuilder.buildHierarchicalCFG(); | |||
9175 | ||||
9176 | for (ElementCount VF = Range.Start; ElementCount::isKnownLT(VF, Range.End); | |||
9177 | VF *= 2) | |||
9178 | Plan->addVF(VF); | |||
9179 | ||||
9180 | if (EnableVPlanPredication) { | |||
9181 | VPlanPredicator VPP(*Plan); | |||
9182 | VPP.predicate(); | |||
9183 | ||||
9184 | // Avoid running transformation to recipes until masked code generation in | |||
9185 | // VPlan-native path is in place. | |||
9186 | return Plan; | |||
9187 | } | |||
9188 | ||||
9189 | SmallPtrSet<Instruction *, 1> DeadInstructions; | |||
9190 | VPlanTransforms::VPInstructionsToVPRecipes( | |||
9191 | OrigLoop, Plan, | |||
9192 | [this](PHINode *P) { return Legal->getIntOrFpInductionDescriptor(P); }, | |||
9193 | DeadInstructions, *PSE.getSE()); | |||
9194 | ||||
9195 | addCanonicalIVRecipes(*Plan, Legal->getWidestInductionType(), DebugLoc(), | |||
9196 | true, true); | |||
9197 | return Plan; | |||
9198 | } | |||
9199 | ||||
9200 | // Adjust the recipes for reductions. For in-loop reductions the chain of | |||
9201 | // instructions leading from the loop exit instr to the phi need to be converted | |||
9202 | // to reductions, with one operand being vector and the other being the scalar | |||
9203 | // reduction chain. For other reductions, a select is introduced between the phi | |||
9204 | // and live-out recipes when folding the tail. | |||
9205 | void LoopVectorizationPlanner::adjustRecipesForReductions( | |||
9206 | VPBasicBlock *LatchVPBB, VPlanPtr &Plan, VPRecipeBuilder &RecipeBuilder, | |||
9207 | ElementCount MinVF) { | |||
9208 | for (auto &Reduction : CM.getInLoopReductionChains()) { | |||
9209 | PHINode *Phi = Reduction.first; | |||
9210 | const RecurrenceDescriptor &RdxDesc = | |||
9211 | Legal->getReductionVars().find(Phi)->second; | |||
9212 | const SmallVector<Instruction *, 4> &ReductionOperations = Reduction.second; | |||
9213 | ||||
9214 | if (MinVF.isScalar() && !CM.useOrderedReductions(RdxDesc)) | |||
9215 | continue; | |||
9216 | ||||
9217 | // ReductionOperations are orders top-down from the phi's use to the | |||
9218 | // LoopExitValue. We keep a track of the previous item (the Chain) to tell | |||
9219 | // which of the two operands will remain scalar and which will be reduced. | |||
9220 | // For minmax the chain will be the select instructions. | |||
9221 | Instruction *Chain = Phi; | |||
9222 | for (Instruction *R : ReductionOperations) { | |||
9223 | VPRecipeBase *WidenRecipe = RecipeBuilder.getRecipe(R); | |||
9224 | RecurKind Kind = RdxDesc.getRecurrenceKind(); | |||
9225 | ||||
9226 | VPValue *ChainOp = Plan->getVPValue(Chain); | |||
9227 | unsigned FirstOpId; | |||
9228 | assert(!RecurrenceDescriptor::isSelectCmpRecurrenceKind(Kind) &&(static_cast <bool> (!RecurrenceDescriptor::isSelectCmpRecurrenceKind (Kind) && "Only min/max recurrences allowed for inloop reductions" ) ? void (0) : __assert_fail ("!RecurrenceDescriptor::isSelectCmpRecurrenceKind(Kind) && \"Only min/max recurrences allowed for inloop reductions\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9229, __extension__ __PRETTY_FUNCTION__)) | |||
9229 | "Only min/max recurrences allowed for inloop reductions")(static_cast <bool> (!RecurrenceDescriptor::isSelectCmpRecurrenceKind (Kind) && "Only min/max recurrences allowed for inloop reductions" ) ? void (0) : __assert_fail ("!RecurrenceDescriptor::isSelectCmpRecurrenceKind(Kind) && \"Only min/max recurrences allowed for inloop reductions\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9229, __extension__ __PRETTY_FUNCTION__)); | |||
9230 | // Recognize a call to the llvm.fmuladd intrinsic. | |||
9231 | bool IsFMulAdd = (Kind == RecurKind::FMulAdd); | |||
9232 | assert((!IsFMulAdd || RecurrenceDescriptor::isFMulAddIntrinsic(R)) &&(static_cast <bool> ((!IsFMulAdd || RecurrenceDescriptor ::isFMulAddIntrinsic(R)) && "Expected instruction to be a call to the llvm.fmuladd intrinsic" ) ? void (0) : __assert_fail ("(!IsFMulAdd || RecurrenceDescriptor::isFMulAddIntrinsic(R)) && \"Expected instruction to be a call to the llvm.fmuladd intrinsic\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9233, __extension__ __PRETTY_FUNCTION__)) | |||
9233 | "Expected instruction to be a call to the llvm.fmuladd intrinsic")(static_cast <bool> ((!IsFMulAdd || RecurrenceDescriptor ::isFMulAddIntrinsic(R)) && "Expected instruction to be a call to the llvm.fmuladd intrinsic" ) ? void (0) : __assert_fail ("(!IsFMulAdd || RecurrenceDescriptor::isFMulAddIntrinsic(R)) && \"Expected instruction to be a call to the llvm.fmuladd intrinsic\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9233, __extension__ __PRETTY_FUNCTION__)); | |||
9234 | if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) { | |||
9235 | assert(isa<VPWidenSelectRecipe>(WidenRecipe) &&(static_cast <bool> (isa<VPWidenSelectRecipe>(WidenRecipe ) && "Expected to replace a VPWidenSelectSC") ? void ( 0) : __assert_fail ("isa<VPWidenSelectRecipe>(WidenRecipe) && \"Expected to replace a VPWidenSelectSC\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9236, __extension__ __PRETTY_FUNCTION__)) | |||
9236 | "Expected to replace a VPWidenSelectSC")(static_cast <bool> (isa<VPWidenSelectRecipe>(WidenRecipe ) && "Expected to replace a VPWidenSelectSC") ? void ( 0) : __assert_fail ("isa<VPWidenSelectRecipe>(WidenRecipe) && \"Expected to replace a VPWidenSelectSC\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9236, __extension__ __PRETTY_FUNCTION__)); | |||
9237 | FirstOpId = 1; | |||
9238 | } else { | |||
9239 | assert((MinVF.isScalar() || isa<VPWidenRecipe>(WidenRecipe) ||(static_cast <bool> ((MinVF.isScalar() || isa<VPWidenRecipe >(WidenRecipe) || (IsFMulAdd && isa<VPWidenCallRecipe >(WidenRecipe))) && "Expected to replace a VPWidenSC" ) ? void (0) : __assert_fail ("(MinVF.isScalar() || isa<VPWidenRecipe>(WidenRecipe) || (IsFMulAdd && isa<VPWidenCallRecipe>(WidenRecipe))) && \"Expected to replace a VPWidenSC\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9241, __extension__ __PRETTY_FUNCTION__)) | |||
9240 | (IsFMulAdd && isa<VPWidenCallRecipe>(WidenRecipe))) &&(static_cast <bool> ((MinVF.isScalar() || isa<VPWidenRecipe >(WidenRecipe) || (IsFMulAdd && isa<VPWidenCallRecipe >(WidenRecipe))) && "Expected to replace a VPWidenSC" ) ? void (0) : __assert_fail ("(MinVF.isScalar() || isa<VPWidenRecipe>(WidenRecipe) || (IsFMulAdd && isa<VPWidenCallRecipe>(WidenRecipe))) && \"Expected to replace a VPWidenSC\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9241, __extension__ __PRETTY_FUNCTION__)) | |||
9241 | "Expected to replace a VPWidenSC")(static_cast <bool> ((MinVF.isScalar() || isa<VPWidenRecipe >(WidenRecipe) || (IsFMulAdd && isa<VPWidenCallRecipe >(WidenRecipe))) && "Expected to replace a VPWidenSC" ) ? void (0) : __assert_fail ("(MinVF.isScalar() || isa<VPWidenRecipe>(WidenRecipe) || (IsFMulAdd && isa<VPWidenCallRecipe>(WidenRecipe))) && \"Expected to replace a VPWidenSC\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9241, __extension__ __PRETTY_FUNCTION__)); | |||
9242 | FirstOpId = 0; | |||
9243 | } | |||
9244 | unsigned VecOpId = | |||
9245 | R->getOperand(FirstOpId) == Chain ? FirstOpId + 1 : FirstOpId; | |||
9246 | VPValue *VecOp = Plan->getVPValue(R->getOperand(VecOpId)); | |||
9247 | ||||
9248 | auto *CondOp = CM.blockNeedsPredicationForAnyReason(R->getParent()) | |||
9249 | ? RecipeBuilder.createBlockInMask(R->getParent(), Plan) | |||
9250 | : nullptr; | |||
9251 | ||||
9252 | if (IsFMulAdd) { | |||
9253 | // If the instruction is a call to the llvm.fmuladd intrinsic then we | |||
9254 | // need to create an fmul recipe to use as the vector operand for the | |||
9255 | // fadd reduction. | |||
9256 | VPInstruction *FMulRecipe = new VPInstruction( | |||
9257 | Instruction::FMul, {VecOp, Plan->getVPValue(R->getOperand(1))}); | |||
9258 | FMulRecipe->setFastMathFlags(R->getFastMathFlags()); | |||
9259 | WidenRecipe->getParent()->insert(FMulRecipe, | |||
9260 | WidenRecipe->getIterator()); | |||
9261 | VecOp = FMulRecipe; | |||
9262 | } | |||
9263 | VPReductionRecipe *RedRecipe = | |||
9264 | new VPReductionRecipe(&RdxDesc, R, ChainOp, VecOp, CondOp, TTI); | |||
9265 | WidenRecipe->getVPSingleValue()->replaceAllUsesWith(RedRecipe); | |||
9266 | Plan->removeVPValueFor(R); | |||
9267 | Plan->addVPValue(R, RedRecipe); | |||
9268 | WidenRecipe->getParent()->insert(RedRecipe, WidenRecipe->getIterator()); | |||
9269 | WidenRecipe->getVPSingleValue()->replaceAllUsesWith(RedRecipe); | |||
9270 | WidenRecipe->eraseFromParent(); | |||
9271 | ||||
9272 | if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) { | |||
9273 | VPRecipeBase *CompareRecipe = | |||
9274 | RecipeBuilder.getRecipe(cast<Instruction>(R->getOperand(0))); | |||
9275 | assert(isa<VPWidenRecipe>(CompareRecipe) &&(static_cast <bool> (isa<VPWidenRecipe>(CompareRecipe ) && "Expected to replace a VPWidenSC") ? void (0) : __assert_fail ("isa<VPWidenRecipe>(CompareRecipe) && \"Expected to replace a VPWidenSC\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9276, __extension__ __PRETTY_FUNCTION__)) | |||
9276 | "Expected to replace a VPWidenSC")(static_cast <bool> (isa<VPWidenRecipe>(CompareRecipe ) && "Expected to replace a VPWidenSC") ? void (0) : __assert_fail ("isa<VPWidenRecipe>(CompareRecipe) && \"Expected to replace a VPWidenSC\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9276, __extension__ __PRETTY_FUNCTION__)); | |||
9277 | assert(cast<VPWidenRecipe>(CompareRecipe)->getNumUsers() == 0 &&(static_cast <bool> (cast<VPWidenRecipe>(CompareRecipe )->getNumUsers() == 0 && "Expected no remaining users" ) ? void (0) : __assert_fail ("cast<VPWidenRecipe>(CompareRecipe)->getNumUsers() == 0 && \"Expected no remaining users\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9278, __extension__ __PRETTY_FUNCTION__)) | |||
9278 | "Expected no remaining users")(static_cast <bool> (cast<VPWidenRecipe>(CompareRecipe )->getNumUsers() == 0 && "Expected no remaining users" ) ? void (0) : __assert_fail ("cast<VPWidenRecipe>(CompareRecipe)->getNumUsers() == 0 && \"Expected no remaining users\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9278, __extension__ __PRETTY_FUNCTION__)); | |||
9279 | CompareRecipe->eraseFromParent(); | |||
9280 | } | |||
9281 | Chain = R; | |||
9282 | } | |||
9283 | } | |||
9284 | ||||
9285 | // If tail is folded by masking, introduce selects between the phi | |||
9286 | // and the live-out instruction of each reduction, at the beginning of the | |||
9287 | // dedicated latch block. | |||
9288 | if (CM.foldTailByMasking()) { | |||
9289 | Builder.setInsertPoint(LatchVPBB, LatchVPBB->begin()); | |||
9290 | for (VPRecipeBase &R : Plan->getEntry()->getEntryBasicBlock()->phis()) { | |||
9291 | VPReductionPHIRecipe *PhiR = dyn_cast<VPReductionPHIRecipe>(&R); | |||
9292 | if (!PhiR || PhiR->isInLoop()) | |||
9293 | continue; | |||
9294 | VPValue *Cond = | |||
9295 | RecipeBuilder.createBlockInMask(OrigLoop->getHeader(), Plan); | |||
9296 | VPValue *Red = PhiR->getBackedgeValue(); | |||
9297 | assert(cast<VPRecipeBase>(Red->getDef())->getParent() != LatchVPBB &&(static_cast <bool> (cast<VPRecipeBase>(Red->getDef ())->getParent() != LatchVPBB && "reduction recipe must be defined before latch" ) ? void (0) : __assert_fail ("cast<VPRecipeBase>(Red->getDef())->getParent() != LatchVPBB && \"reduction recipe must be defined before latch\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9298, __extension__ __PRETTY_FUNCTION__)) | |||
9298 | "reduction recipe must be defined before latch")(static_cast <bool> (cast<VPRecipeBase>(Red->getDef ())->getParent() != LatchVPBB && "reduction recipe must be defined before latch" ) ? void (0) : __assert_fail ("cast<VPRecipeBase>(Red->getDef())->getParent() != LatchVPBB && \"reduction recipe must be defined before latch\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9298, __extension__ __PRETTY_FUNCTION__)); | |||
9299 | Builder.createNaryOp(Instruction::Select, {Cond, Red, PhiR}); | |||
9300 | } | |||
9301 | } | |||
9302 | } | |||
9303 | ||||
9304 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | |||
9305 | void VPInterleaveRecipe::print(raw_ostream &O, const Twine &Indent, | |||
9306 | VPSlotTracker &SlotTracker) const { | |||
9307 | O << Indent << "INTERLEAVE-GROUP with factor " << IG->getFactor() << " at "; | |||
9308 | IG->getInsertPos()->printAsOperand(O, false); | |||
9309 | O << ", "; | |||
9310 | getAddr()->printAsOperand(O, SlotTracker); | |||
9311 | VPValue *Mask = getMask(); | |||
9312 | if (Mask) { | |||
9313 | O << ", "; | |||
9314 | Mask->printAsOperand(O, SlotTracker); | |||
9315 | } | |||
9316 | ||||
9317 | unsigned OpIdx = 0; | |||
9318 | for (unsigned i = 0; i < IG->getFactor(); ++i) { | |||
9319 | if (!IG->getMember(i)) | |||
9320 | continue; | |||
9321 | if (getNumStoreOperands() > 0) { | |||
9322 | O << "\n" << Indent << " store "; | |||
9323 | getOperand(1 + OpIdx)->printAsOperand(O, SlotTracker); | |||
9324 | O << " to index " << i; | |||
9325 | } else { | |||
9326 | O << "\n" << Indent << " "; | |||
9327 | getVPValue(OpIdx)->printAsOperand(O, SlotTracker); | |||
9328 | O << " = load from index " << i; | |||
9329 | } | |||
9330 | ++OpIdx; | |||
9331 | } | |||
9332 | } | |||
9333 | #endif | |||
9334 | ||||
9335 | void VPWidenCallRecipe::execute(VPTransformState &State) { | |||
9336 | State.ILV->widenCallInstruction(*cast<CallInst>(getUnderlyingInstr()), this, | |||
9337 | *this, State); | |||
9338 | } | |||
9339 | ||||
9340 | void VPWidenSelectRecipe::execute(VPTransformState &State) { | |||
9341 | auto &I = *cast<SelectInst>(getUnderlyingInstr()); | |||
9342 | State.ILV->setDebugLocFromInst(&I); | |||
9343 | ||||
9344 | // The condition can be loop invariant but still defined inside the | |||
9345 | // loop. This means that we can't just use the original 'cond' value. | |||
9346 | // We have to take the 'vectorized' value and pick the first lane. | |||
9347 | // Instcombine will make this a no-op. | |||
9348 | auto *InvarCond = | |||
9349 | InvariantCond ? State.get(getOperand(0), VPIteration(0, 0)) : nullptr; | |||
9350 | ||||
9351 | for (unsigned Part = 0; Part < State.UF; ++Part) { | |||
9352 | Value *Cond = InvarCond ? InvarCond : State.get(getOperand(0), Part); | |||
9353 | Value *Op0 = State.get(getOperand(1), Part); | |||
9354 | Value *Op1 = State.get(getOperand(2), Part); | |||
9355 | Value *Sel = State.Builder.CreateSelect(Cond, Op0, Op1); | |||
9356 | State.set(this, Sel, Part); | |||
9357 | State.ILV->addMetadata(Sel, &I); | |||
9358 | } | |||
9359 | } | |||
9360 | ||||
9361 | void VPWidenRecipe::execute(VPTransformState &State) { | |||
9362 | auto &I = *cast<Instruction>(getUnderlyingValue()); | |||
9363 | auto &Builder = State.Builder; | |||
9364 | switch (I.getOpcode()) { | |||
9365 | case Instruction::Call: | |||
9366 | case Instruction::Br: | |||
9367 | case Instruction::PHI: | |||
9368 | case Instruction::GetElementPtr: | |||
9369 | case Instruction::Select: | |||
9370 | llvm_unreachable("This instruction is handled by a different recipe.")::llvm::llvm_unreachable_internal("This instruction is handled by a different recipe." , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9370); | |||
9371 | case Instruction::UDiv: | |||
9372 | case Instruction::SDiv: | |||
9373 | case Instruction::SRem: | |||
9374 | case Instruction::URem: | |||
9375 | case Instruction::Add: | |||
9376 | case Instruction::FAdd: | |||
9377 | case Instruction::Sub: | |||
9378 | case Instruction::FSub: | |||
9379 | case Instruction::FNeg: | |||
9380 | case Instruction::Mul: | |||
9381 | case Instruction::FMul: | |||
9382 | case Instruction::FDiv: | |||
9383 | case Instruction::FRem: | |||
9384 | case Instruction::Shl: | |||
9385 | case Instruction::LShr: | |||
9386 | case Instruction::AShr: | |||
9387 | case Instruction::And: | |||
9388 | case Instruction::Or: | |||
9389 | case Instruction::Xor: { | |||
9390 | // Just widen unops and binops. | |||
9391 | State.ILV->setDebugLocFromInst(&I); | |||
9392 | ||||
9393 | for (unsigned Part = 0; Part < State.UF; ++Part) { | |||
9394 | SmallVector<Value *, 2> Ops; | |||
9395 | for (VPValue *VPOp : operands()) | |||
9396 | Ops.push_back(State.get(VPOp, Part)); | |||
9397 | ||||
9398 | Value *V = Builder.CreateNAryOp(I.getOpcode(), Ops); | |||
9399 | ||||
9400 | if (auto *VecOp = dyn_cast<Instruction>(V)) { | |||
9401 | VecOp->copyIRFlags(&I); | |||
9402 | ||||
9403 | // If the instruction is vectorized and was in a basic block that needed | |||
9404 | // predication, we can't propagate poison-generating flags (nuw/nsw, | |||
9405 | // exact, etc.). The control flow has been linearized and the | |||
9406 | // instruction is no longer guarded by the predicate, which could make | |||
9407 | // the flag properties to no longer hold. | |||
9408 | if (State.MayGeneratePoisonRecipes.contains(this)) | |||
9409 | VecOp->dropPoisonGeneratingFlags(); | |||
9410 | } | |||
9411 | ||||
9412 | // Use this vector value for all users of the original instruction. | |||
9413 | State.set(this, V, Part); | |||
9414 | State.ILV->addMetadata(V, &I); | |||
9415 | } | |||
9416 | ||||
9417 | break; | |||
9418 | } | |||
9419 | case Instruction::ICmp: | |||
9420 | case Instruction::FCmp: { | |||
9421 | // Widen compares. Generate vector compares. | |||
9422 | bool FCmp = (I.getOpcode() == Instruction::FCmp); | |||
9423 | auto *Cmp = cast<CmpInst>(&I); | |||
9424 | State.ILV->setDebugLocFromInst(Cmp); | |||
9425 | for (unsigned Part = 0; Part < State.UF; ++Part) { | |||
9426 | Value *A = State.get(getOperand(0), Part); | |||
9427 | Value *B = State.get(getOperand(1), Part); | |||
9428 | Value *C = nullptr; | |||
9429 | if (FCmp) { | |||
9430 | // Propagate fast math flags. | |||
9431 | IRBuilder<>::FastMathFlagGuard FMFG(Builder); | |||
9432 | Builder.setFastMathFlags(Cmp->getFastMathFlags()); | |||
9433 | C = Builder.CreateFCmp(Cmp->getPredicate(), A, B); | |||
9434 | } else { | |||
9435 | C = Builder.CreateICmp(Cmp->getPredicate(), A, B); | |||
9436 | } | |||
9437 | State.set(this, C, Part); | |||
9438 | State.ILV->addMetadata(C, &I); | |||
9439 | } | |||
9440 | ||||
9441 | break; | |||
9442 | } | |||
9443 | ||||
9444 | case Instruction::ZExt: | |||
9445 | case Instruction::SExt: | |||
9446 | case Instruction::FPToUI: | |||
9447 | case Instruction::FPToSI: | |||
9448 | case Instruction::FPExt: | |||
9449 | case Instruction::PtrToInt: | |||
9450 | case Instruction::IntToPtr: | |||
9451 | case Instruction::SIToFP: | |||
9452 | case Instruction::UIToFP: | |||
9453 | case Instruction::Trunc: | |||
9454 | case Instruction::FPTrunc: | |||
9455 | case Instruction::BitCast: { | |||
9456 | auto *CI = cast<CastInst>(&I); | |||
9457 | State.ILV->setDebugLocFromInst(CI); | |||
9458 | ||||
9459 | /// Vectorize casts. | |||
9460 | Type *DestTy = (State.VF.isScalar()) | |||
9461 | ? CI->getType() | |||
9462 | : VectorType::get(CI->getType(), State.VF); | |||
9463 | ||||
9464 | for (unsigned Part = 0; Part < State.UF; ++Part) { | |||
9465 | Value *A = State.get(getOperand(0), Part); | |||
9466 | Value *Cast = Builder.CreateCast(CI->getOpcode(), A, DestTy); | |||
9467 | State.set(this, Cast, Part); | |||
9468 | State.ILV->addMetadata(Cast, &I); | |||
9469 | } | |||
9470 | break; | |||
9471 | } | |||
9472 | default: | |||
9473 | // This instruction is not vectorized by simple widening. | |||
9474 | 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); | |||
9475 | llvm_unreachable("Unhandled instruction!")::llvm::llvm_unreachable_internal("Unhandled instruction!", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 9475); | |||
9476 | } // end of switch. | |||
9477 | } | |||
9478 | ||||
9479 | void VPWidenGEPRecipe::execute(VPTransformState &State) { | |||
9480 | auto *GEP = cast<GetElementPtrInst>(getUnderlyingInstr()); | |||
9481 | // Construct a vector GEP by widening the operands of the scalar GEP as | |||
9482 | // necessary. We mark the vector GEP 'inbounds' if appropriate. A GEP | |||
9483 | // results in a vector of pointers when at least one operand of the GEP | |||
9484 | // is vector-typed. Thus, to keep the representation compact, we only use | |||
9485 | // vector-typed operands for loop-varying values. | |||
9486 | ||||
9487 | if (State.VF.isVector() && IsPtrLoopInvariant && IsIndexLoopInvariant.all()) { | |||
9488 | // If we are vectorizing, but the GEP has only loop-invariant operands, | |||
9489 | // the GEP we build (by only using vector-typed operands for | |||
9490 | // loop-varying values) would be a scalar pointer. Thus, to ensure we | |||
9491 | // produce a vector of pointers, we need to either arbitrarily pick an | |||
9492 | // operand to broadcast, or broadcast a clone of the original GEP. | |||
9493 | // Here, we broadcast a clone of the original. | |||
9494 | // | |||
9495 | // TODO: If at some point we decide to scalarize instructions having | |||
9496 | // loop-invariant operands, this special case will no longer be | |||
9497 | // required. We would add the scalarization decision to | |||
9498 | // collectLoopScalars() and teach getVectorValue() to broadcast | |||
9499 | // the lane-zero scalar value. | |||
9500 | auto *Clone = State.Builder.Insert(GEP->clone()); | |||
9501 | for (unsigned Part = 0; Part < State.UF; ++Part) { | |||
9502 | Value *EntryPart = State.Builder.CreateVectorSplat(State.VF, Clone); | |||
9503 | State.set(this, EntryPart, Part); | |||
9504 | State.ILV->addMetadata(EntryPart, GEP); | |||
9505 | } | |||
9506 | } else { | |||
9507 | // If the GEP has at least one loop-varying operand, we are sure to | |||
9508 | // produce a vector of pointers. But if we are only unrolling, we want | |||
9509 | // to produce a scalar GEP for each unroll part. Thus, the GEP we | |||
9510 | // produce with the code below will be scalar (if VF == 1) or vector | |||
9511 | // (otherwise). Note that for the unroll-only case, we still maintain | |||
9512 | // values in the vector mapping with initVector, as we do for other | |||
9513 | // instructions. | |||
9514 | for (unsigned Part = 0; Part < State.UF; ++Part) { | |||
9515 | // The pointer operand of the new GEP. If it's loop-invariant, we | |||
9516 | // won't broadcast it. | |||
9517 | auto *Ptr = IsPtrLoopInvariant | |||
9518 | ? State.get(getOperand(0), VPIteration(0, 0)) | |||
9519 | : State.get(getOperand(0), Part); | |||
9520 | ||||
9521 | // Collect all the indices for the new GEP. If any index is | |||
9522 | // loop-invariant, we won't broadcast it. | |||
9523 | SmallVector<Value *, 4> Indices; | |||
9524 | for (unsigned I = 1, E = getNumOperands(); I < E; I++) { | |||
9525 | VPValue *Operand = getOperand(I); | |||
9526 | if (IsIndexLoopInvariant[I - 1]) | |||
9527 | Indices.push_back(State.get(Operand, VPIteration(0, 0))); | |||
9528 | else | |||
9529 | Indices.push_back(State.get(Operand, Part)); | |||
9530 | } | |||
9531 | ||||
9532 | // If the GEP instruction is vectorized and was in a basic block that | |||
9533 | // needed predication, we can't propagate the poison-generating 'inbounds' | |||
9534 | // flag. The control flow has been linearized and the GEP is no longer | |||
9535 | // guarded by the predicate, which could make the 'inbounds' properties to | |||
9536 | // no longer hold. | |||
9537 | bool IsInBounds = | |||
9538 | GEP->isInBounds() && State.MayGeneratePoisonRecipes.count(this) == 0; | |||
9539 | ||||
9540 | // Create the new GEP. Note that this GEP may be a scalar if VF == 1, | |||
9541 | // but it should be a vector, otherwise. | |||
9542 | auto *NewGEP = IsInBounds | |||
9543 | ? State.Builder.CreateInBoundsGEP( | |||
9544 | GEP->getSourceElementType(), Ptr, Indices) | |||
9545 | : State.Builder.CreateGEP(GEP->getSourceElementType(), | |||
9546 | Ptr, Indices); | |||
9547 | assert((State.VF.isScalar() || NewGEP->getType()->isVectorTy()) &&(static_cast <bool> ((State.VF.isScalar() || NewGEP-> getType()->isVectorTy()) && "NewGEP is not a pointer vector" ) ? void (0) : __assert_fail ("(State.VF.isScalar() || NewGEP->getType()->isVectorTy()) && \"NewGEP is not a pointer vector\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9548, __extension__ __PRETTY_FUNCTION__)) | |||
9548 | "NewGEP is not a pointer vector")(static_cast <bool> ((State.VF.isScalar() || NewGEP-> getType()->isVectorTy()) && "NewGEP is not a pointer vector" ) ? void (0) : __assert_fail ("(State.VF.isScalar() || NewGEP->getType()->isVectorTy()) && \"NewGEP is not a pointer vector\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9548, __extension__ __PRETTY_FUNCTION__)); | |||
9549 | State.set(this, NewGEP, Part); | |||
9550 | State.ILV->addMetadata(NewGEP, GEP); | |||
9551 | } | |||
9552 | } | |||
9553 | } | |||
9554 | ||||
9555 | void VPWidenIntOrFpInductionRecipe::execute(VPTransformState &State) { | |||
9556 | assert(!State.Instance && "Int or FP induction being replicated.")(static_cast <bool> (!State.Instance && "Int or FP induction being replicated." ) ? void (0) : __assert_fail ("!State.Instance && \"Int or FP induction being replicated.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9556, __extension__ __PRETTY_FUNCTION__)); | |||
9557 | ||||
9558 | Value *Start = getStartValue()->getLiveInIRValue(); | |||
9559 | const InductionDescriptor &ID = getInductionDescriptor(); | |||
9560 | TruncInst *Trunc = getTruncInst(); | |||
9561 | IRBuilderBase &Builder = State.Builder; | |||
9562 | assert(IV->getType() == ID.getStartValue()->getType() && "Types must match")(static_cast <bool> (IV->getType() == ID.getStartValue ()->getType() && "Types must match") ? void (0) : __assert_fail ("IV->getType() == ID.getStartValue()->getType() && \"Types must match\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9562, __extension__ __PRETTY_FUNCTION__)); | |||
9563 | assert(State.VF.isVector() && "must have vector VF")(static_cast <bool> (State.VF.isVector() && "must have vector VF" ) ? void (0) : __assert_fail ("State.VF.isVector() && \"must have vector VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9563, __extension__ __PRETTY_FUNCTION__)); | |||
9564 | ||||
9565 | // The value from the original loop to which we are mapping the new induction | |||
9566 | // variable. | |||
9567 | Instruction *EntryVal = Trunc ? cast<Instruction>(Trunc) : IV; | |||
9568 | ||||
9569 | auto &DL = EntryVal->getModule()->getDataLayout(); | |||
9570 | ||||
9571 | // Generate code for the induction step. Note that induction steps are | |||
9572 | // required to be loop-invariant | |||
9573 | auto CreateStepValue = [&](const SCEV *Step) -> Value * { | |||
9574 | if (SE.isSCEVable(IV->getType())) { | |||
9575 | SCEVExpander Exp(SE, DL, "induction"); | |||
9576 | return Exp.expandCodeFor(Step, Step->getType(), | |||
9577 | State.CFG.VectorPreHeader->getTerminator()); | |||
9578 | } | |||
9579 | return cast<SCEVUnknown>(Step)->getValue(); | |||
9580 | }; | |||
9581 | ||||
9582 | // Fast-math-flags propagate from the original induction instruction. | |||
9583 | IRBuilder<>::FastMathFlagGuard FMFG(Builder); | |||
9584 | if (ID.getInductionBinOp() && isa<FPMathOperator>(ID.getInductionBinOp())) | |||
9585 | Builder.setFastMathFlags(ID.getInductionBinOp()->getFastMathFlags()); | |||
9586 | ||||
9587 | // Now do the actual transformations, and start with creating the step value. | |||
9588 | Value *Step = CreateStepValue(ID.getStep()); | |||
9589 | ||||
9590 | assert((isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal)) &&(static_cast <bool> ((isa<PHINode>(EntryVal) || isa <TruncInst>(EntryVal)) && "Expected either an induction phi-node or a truncate of it!" ) ? void (0) : __assert_fail ("(isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal)) && \"Expected either an induction phi-node or a truncate of it!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9591, __extension__ __PRETTY_FUNCTION__)) | |||
9591 | "Expected either an induction phi-node or a truncate of it!")(static_cast <bool> ((isa<PHINode>(EntryVal) || isa <TruncInst>(EntryVal)) && "Expected either an induction phi-node or a truncate of it!" ) ? void (0) : __assert_fail ("(isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal)) && \"Expected either an induction phi-node or a truncate of it!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9591, __extension__ __PRETTY_FUNCTION__)); | |||
9592 | ||||
9593 | // Construct the initial value of the vector IV in the vector loop preheader | |||
9594 | auto CurrIP = Builder.saveIP(); | |||
9595 | Builder.SetInsertPoint(State.CFG.VectorPreHeader->getTerminator()); | |||
9596 | if (isa<TruncInst>(EntryVal)) { | |||
9597 | assert(Start->getType()->isIntegerTy() &&(static_cast <bool> (Start->getType()->isIntegerTy () && "Truncation requires an integer type") ? void ( 0) : __assert_fail ("Start->getType()->isIntegerTy() && \"Truncation requires an integer type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9598, __extension__ __PRETTY_FUNCTION__)) | |||
9598 | "Truncation requires an integer type")(static_cast <bool> (Start->getType()->isIntegerTy () && "Truncation requires an integer type") ? void ( 0) : __assert_fail ("Start->getType()->isIntegerTy() && \"Truncation requires an integer type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9598, __extension__ __PRETTY_FUNCTION__)); | |||
9599 | auto *TruncType = cast<IntegerType>(EntryVal->getType()); | |||
9600 | Step = Builder.CreateTrunc(Step, TruncType); | |||
9601 | Start = Builder.CreateCast(Instruction::Trunc, Start, TruncType); | |||
9602 | } | |||
9603 | ||||
9604 | Value *Zero = getSignedIntOrFpConstant(Start->getType(), 0); | |||
9605 | Value *SplatStart = Builder.CreateVectorSplat(State.VF, Start); | |||
9606 | Value *SteppedStart = getStepVector( | |||
9607 | SplatStart, Zero, Step, ID.getInductionOpcode(), State.VF, State.Builder); | |||
9608 | ||||
9609 | // We create vector phi nodes for both integer and floating-point induction | |||
9610 | // variables. Here, we determine the kind of arithmetic we will perform. | |||
9611 | Instruction::BinaryOps AddOp; | |||
9612 | Instruction::BinaryOps MulOp; | |||
9613 | if (Step->getType()->isIntegerTy()) { | |||
9614 | AddOp = Instruction::Add; | |||
9615 | MulOp = Instruction::Mul; | |||
9616 | } else { | |||
9617 | AddOp = ID.getInductionOpcode(); | |||
9618 | MulOp = Instruction::FMul; | |||
9619 | } | |||
9620 | ||||
9621 | // Multiply the vectorization factor by the step using integer or | |||
9622 | // floating-point arithmetic as appropriate. | |||
9623 | Type *StepType = Step->getType(); | |||
9624 | Value *RuntimeVF; | |||
9625 | if (Step->getType()->isFloatingPointTy()) | |||
9626 | RuntimeVF = getRuntimeVFAsFloat(Builder, StepType, State.VF); | |||
9627 | else | |||
9628 | RuntimeVF = getRuntimeVF(Builder, StepType, State.VF); | |||
9629 | Value *Mul = Builder.CreateBinOp(MulOp, Step, RuntimeVF); | |||
9630 | ||||
9631 | // Create a vector splat to use in the induction update. | |||
9632 | // | |||
9633 | // FIXME: If the step is non-constant, we create the vector splat with | |||
9634 | // IRBuilder. IRBuilder can constant-fold the multiply, but it doesn't | |||
9635 | // handle a constant vector splat. | |||
9636 | Value *SplatVF = isa<Constant>(Mul) | |||
9637 | ? ConstantVector::getSplat(State.VF, cast<Constant>(Mul)) | |||
9638 | : Builder.CreateVectorSplat(State.VF, Mul); | |||
9639 | Builder.restoreIP(CurrIP); | |||
9640 | ||||
9641 | // We may need to add the step a number of times, depending on the unroll | |||
9642 | // factor. The last of those goes into the PHI. | |||
9643 | PHINode *VecInd = PHINode::Create(SteppedStart->getType(), 2, "vec.ind", | |||
9644 | &*State.CFG.PrevBB->getFirstInsertionPt()); | |||
9645 | VecInd->setDebugLoc(EntryVal->getDebugLoc()); | |||
9646 | Instruction *LastInduction = VecInd; | |||
9647 | for (unsigned Part = 0; Part < State.UF; ++Part) { | |||
9648 | State.set(this, LastInduction, Part); | |||
9649 | ||||
9650 | if (isa<TruncInst>(EntryVal)) | |||
9651 | State.ILV->addMetadata(LastInduction, EntryVal); | |||
9652 | ||||
9653 | LastInduction = cast<Instruction>( | |||
9654 | Builder.CreateBinOp(AddOp, LastInduction, SplatVF, "step.add")); | |||
9655 | LastInduction->setDebugLoc(EntryVal->getDebugLoc()); | |||
9656 | } | |||
9657 | ||||
9658 | // Move the last step to the end of the latch block. This ensures consistent | |||
9659 | // placement of all induction updates. | |||
9660 | auto *LoopVectorLatch = | |||
9661 | State.LI->getLoopFor(State.CFG.PrevBB)->getLoopLatch(); | |||
9662 | auto *Br = cast<BranchInst>(LoopVectorLatch->getTerminator()); | |||
9663 | LastInduction->moveBefore(Br); | |||
9664 | LastInduction->setName("vec.ind.next"); | |||
9665 | ||||
9666 | VecInd->addIncoming(SteppedStart, State.CFG.VectorPreHeader); | |||
9667 | VecInd->addIncoming(LastInduction, LoopVectorLatch); | |||
9668 | } | |||
9669 | ||||
9670 | void VPScalarIVStepsRecipe::execute(VPTransformState &State) { | |||
9671 | assert(!State.Instance && "VPScalarIVStepsRecipe being replicated.")(static_cast <bool> (!State.Instance && "VPScalarIVStepsRecipe being replicated." ) ? void (0) : __assert_fail ("!State.Instance && \"VPScalarIVStepsRecipe being replicated.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9671, __extension__ __PRETTY_FUNCTION__)); | |||
9672 | ||||
9673 | // Fast-math-flags propagate from the original induction instruction. | |||
9674 | IRBuilder<>::FastMathFlagGuard FMFG(State.Builder); | |||
9675 | if (IndDesc.getInductionBinOp() && | |||
9676 | isa<FPMathOperator>(IndDesc.getInductionBinOp())) | |||
9677 | State.Builder.setFastMathFlags( | |||
9678 | IndDesc.getInductionBinOp()->getFastMathFlags()); | |||
9679 | ||||
9680 | Value *Step = State.get(getStepValue(), VPIteration(0, 0)); | |||
9681 | auto CreateScalarIV = [&](Value *&Step) -> Value * { | |||
9682 | Value *ScalarIV = State.get(getCanonicalIV(), VPIteration(0, 0)); | |||
9683 | auto *CanonicalIV = State.get(getParent()->getPlan()->getCanonicalIV(), 0); | |||
9684 | if (!isCanonical() || CanonicalIV->getType() != Ty) { | |||
9685 | ScalarIV = | |||
9686 | Ty->isIntegerTy() | |||
9687 | ? State.Builder.CreateSExtOrTrunc(ScalarIV, Ty) | |||
9688 | : State.Builder.CreateCast(Instruction::SIToFP, ScalarIV, Ty); | |||
9689 | ScalarIV = emitTransformedIndex(State.Builder, ScalarIV, | |||
9690 | getStartValue()->getLiveInIRValue(), Step, | |||
9691 | IndDesc); | |||
9692 | ScalarIV->setName("offset.idx"); | |||
9693 | } | |||
9694 | if (TruncToTy) { | |||
9695 | assert(Step->getType()->isIntegerTy() &&(static_cast <bool> (Step->getType()->isIntegerTy () && "Truncation requires an integer step") ? void ( 0) : __assert_fail ("Step->getType()->isIntegerTy() && \"Truncation requires an integer step\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9696, __extension__ __PRETTY_FUNCTION__)) | |||
9696 | "Truncation requires an integer step")(static_cast <bool> (Step->getType()->isIntegerTy () && "Truncation requires an integer step") ? void ( 0) : __assert_fail ("Step->getType()->isIntegerTy() && \"Truncation requires an integer step\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9696, __extension__ __PRETTY_FUNCTION__)); | |||
9697 | ScalarIV = State.Builder.CreateTrunc(ScalarIV, TruncToTy); | |||
9698 | Step = State.Builder.CreateTrunc(Step, TruncToTy); | |||
9699 | } | |||
9700 | return ScalarIV; | |||
9701 | }; | |||
9702 | ||||
9703 | Value *ScalarIV = CreateScalarIV(Step); | |||
9704 | if (State.VF.isVector()) { | |||
9705 | buildScalarSteps(ScalarIV, Step, IndDesc, this, State); | |||
9706 | return; | |||
9707 | } | |||
9708 | ||||
9709 | for (unsigned Part = 0; Part < State.UF; ++Part) { | |||
9710 | assert(!State.VF.isScalable() && "scalable vectors not yet supported.")(static_cast <bool> (!State.VF.isScalable() && "scalable vectors not yet supported." ) ? void (0) : __assert_fail ("!State.VF.isScalable() && \"scalable vectors not yet supported.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9710, __extension__ __PRETTY_FUNCTION__)); | |||
9711 | Value *EntryPart; | |||
9712 | if (Step->getType()->isFloatingPointTy()) { | |||
9713 | Value *StartIdx = | |||
9714 | getRuntimeVFAsFloat(State.Builder, Step->getType(), State.VF * Part); | |||
9715 | // Floating-point operations inherit FMF via the builder's flags. | |||
9716 | Value *MulOp = State.Builder.CreateFMul(StartIdx, Step); | |||
9717 | EntryPart = State.Builder.CreateBinOp(IndDesc.getInductionOpcode(), | |||
9718 | ScalarIV, MulOp); | |||
9719 | } else { | |||
9720 | Value *StartIdx = | |||
9721 | getRuntimeVF(State.Builder, Step->getType(), State.VF * Part); | |||
9722 | EntryPart = State.Builder.CreateAdd( | |||
9723 | ScalarIV, State.Builder.CreateMul(StartIdx, Step), "induction"); | |||
9724 | } | |||
9725 | State.set(this, EntryPart, Part); | |||
9726 | } | |||
9727 | } | |||
9728 | ||||
9729 | void VPWidenPHIRecipe::execute(VPTransformState &State) { | |||
9730 | State.ILV->widenPHIInstruction(cast<PHINode>(getUnderlyingValue()), this, | |||
9731 | State); | |||
9732 | } | |||
9733 | ||||
9734 | void VPBlendRecipe::execute(VPTransformState &State) { | |||
9735 | State.ILV->setDebugLocFromInst(Phi, &State.Builder); | |||
9736 | // We know that all PHIs in non-header blocks are converted into | |||
9737 | // selects, so we don't have to worry about the insertion order and we | |||
9738 | // can just use the builder. | |||
9739 | // At this point we generate the predication tree. There may be | |||
9740 | // duplications since this is a simple recursive scan, but future | |||
9741 | // optimizations will clean it up. | |||
9742 | ||||
9743 | unsigned NumIncoming = getNumIncomingValues(); | |||
9744 | ||||
9745 | // Generate a sequence of selects of the form: | |||
9746 | // SELECT(Mask3, In3, | |||
9747 | // SELECT(Mask2, In2, | |||
9748 | // SELECT(Mask1, In1, | |||
9749 | // In0))) | |||
9750 | // Note that Mask0 is never used: lanes for which no path reaches this phi and | |||
9751 | // are essentially undef are taken from In0. | |||
9752 | InnerLoopVectorizer::VectorParts Entry(State.UF); | |||
9753 | for (unsigned In = 0; In < NumIncoming; ++In) { | |||
9754 | for (unsigned Part = 0; Part < State.UF; ++Part) { | |||
9755 | // We might have single edge PHIs (blocks) - use an identity | |||
9756 | // 'select' for the first PHI operand. | |||
9757 | Value *In0 = State.get(getIncomingValue(In), Part); | |||
9758 | if (In == 0) | |||
9759 | Entry[Part] = In0; // Initialize with the first incoming value. | |||
9760 | else { | |||
9761 | // Select between the current value and the previous incoming edge | |||
9762 | // based on the incoming mask. | |||
9763 | Value *Cond = State.get(getMask(In), Part); | |||
9764 | Entry[Part] = | |||
9765 | State.Builder.CreateSelect(Cond, In0, Entry[Part], "predphi"); | |||
9766 | } | |||
9767 | } | |||
9768 | } | |||
9769 | for (unsigned Part = 0; Part < State.UF; ++Part) | |||
9770 | State.set(this, Entry[Part], Part); | |||
9771 | } | |||
9772 | ||||
9773 | void VPInterleaveRecipe::execute(VPTransformState &State) { | |||
9774 | assert(!State.Instance && "Interleave group being replicated.")(static_cast <bool> (!State.Instance && "Interleave group being replicated." ) ? void (0) : __assert_fail ("!State.Instance && \"Interleave group being replicated.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9774, __extension__ __PRETTY_FUNCTION__)); | |||
9775 | State.ILV->vectorizeInterleaveGroup(IG, definedValues(), State, getAddr(), | |||
9776 | getStoredValues(), getMask()); | |||
9777 | } | |||
9778 | ||||
9779 | void VPReductionRecipe::execute(VPTransformState &State) { | |||
9780 | assert(!State.Instance && "Reduction being replicated.")(static_cast <bool> (!State.Instance && "Reduction being replicated." ) ? void (0) : __assert_fail ("!State.Instance && \"Reduction being replicated.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9780, __extension__ __PRETTY_FUNCTION__)); | |||
9781 | Value *PrevInChain = State.get(getChainOp(), 0); | |||
9782 | RecurKind Kind = RdxDesc->getRecurrenceKind(); | |||
9783 | bool IsOrdered = State.ILV->useOrderedReductions(*RdxDesc); | |||
9784 | // Propagate the fast-math flags carried by the underlying instruction. | |||
9785 | IRBuilderBase::FastMathFlagGuard FMFGuard(State.Builder); | |||
9786 | State.Builder.setFastMathFlags(RdxDesc->getFastMathFlags()); | |||
9787 | for (unsigned Part = 0; Part < State.UF; ++Part) { | |||
9788 | Value *NewVecOp = State.get(getVecOp(), Part); | |||
9789 | if (VPValue *Cond = getCondOp()) { | |||
9790 | Value *NewCond = State.get(Cond, Part); | |||
9791 | VectorType *VecTy = cast<VectorType>(NewVecOp->getType()); | |||
9792 | Value *Iden = RdxDesc->getRecurrenceIdentity( | |||
9793 | Kind, VecTy->getElementType(), RdxDesc->getFastMathFlags()); | |||
9794 | Value *IdenVec = | |||
9795 | State.Builder.CreateVectorSplat(VecTy->getElementCount(), Iden); | |||
9796 | Value *Select = State.Builder.CreateSelect(NewCond, NewVecOp, IdenVec); | |||
9797 | NewVecOp = Select; | |||
9798 | } | |||
9799 | Value *NewRed; | |||
9800 | Value *NextInChain; | |||
9801 | if (IsOrdered) { | |||
9802 | if (State.VF.isVector()) | |||
9803 | NewRed = createOrderedReduction(State.Builder, *RdxDesc, NewVecOp, | |||
9804 | PrevInChain); | |||
9805 | else | |||
9806 | NewRed = State.Builder.CreateBinOp( | |||
9807 | (Instruction::BinaryOps)RdxDesc->getOpcode(Kind), PrevInChain, | |||
9808 | NewVecOp); | |||
9809 | PrevInChain = NewRed; | |||
9810 | } else { | |||
9811 | PrevInChain = State.get(getChainOp(), Part); | |||
9812 | NewRed = createTargetReduction(State.Builder, TTI, *RdxDesc, NewVecOp); | |||
9813 | } | |||
9814 | if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) { | |||
9815 | NextInChain = | |||
9816 | createMinMaxOp(State.Builder, RdxDesc->getRecurrenceKind(), | |||
9817 | NewRed, PrevInChain); | |||
9818 | } else if (IsOrdered) | |||
9819 | NextInChain = NewRed; | |||
9820 | else | |||
9821 | NextInChain = State.Builder.CreateBinOp( | |||
9822 | (Instruction::BinaryOps)RdxDesc->getOpcode(Kind), NewRed, | |||
9823 | PrevInChain); | |||
9824 | State.set(this, NextInChain, Part); | |||
9825 | } | |||
9826 | } | |||
9827 | ||||
9828 | void VPReplicateRecipe::execute(VPTransformState &State) { | |||
9829 | if (State.Instance) { // Generate a single instance. | |||
9830 | assert(!State.VF.isScalable() && "Can't scalarize a scalable vector")(static_cast <bool> (!State.VF.isScalable() && "Can't scalarize a scalable vector" ) ? void (0) : __assert_fail ("!State.VF.isScalable() && \"Can't scalarize a scalable vector\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9830, __extension__ __PRETTY_FUNCTION__)); | |||
9831 | State.ILV->scalarizeInstruction(getUnderlyingInstr(), this, *State.Instance, | |||
9832 | IsPredicated, State); | |||
9833 | // Insert scalar instance packing it into a vector. | |||
9834 | if (AlsoPack && State.VF.isVector()) { | |||
9835 | // If we're constructing lane 0, initialize to start from poison. | |||
9836 | if (State.Instance->Lane.isFirstLane()) { | |||
9837 | assert(!State.VF.isScalable() && "VF is assumed to be non scalable.")(static_cast <bool> (!State.VF.isScalable() && "VF is assumed to be non scalable." ) ? void (0) : __assert_fail ("!State.VF.isScalable() && \"VF is assumed to be non scalable.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9837, __extension__ __PRETTY_FUNCTION__)); | |||
9838 | Value *Poison = PoisonValue::get( | |||
9839 | VectorType::get(getUnderlyingValue()->getType(), State.VF)); | |||
9840 | State.set(this, Poison, State.Instance->Part); | |||
9841 | } | |||
9842 | State.ILV->packScalarIntoVectorValue(this, *State.Instance, State); | |||
9843 | } | |||
9844 | return; | |||
9845 | } | |||
9846 | ||||
9847 | // Generate scalar instances for all VF lanes of all UF parts, unless the | |||
9848 | // instruction is uniform inwhich case generate only the first lane for each | |||
9849 | // of the UF parts. | |||
9850 | unsigned EndLane = IsUniform ? 1 : State.VF.getKnownMinValue(); | |||
9851 | assert((!State.VF.isScalable() || IsUniform) &&(static_cast <bool> ((!State.VF.isScalable() || IsUniform ) && "Can't scalarize a scalable vector") ? void (0) : __assert_fail ("(!State.VF.isScalable() || IsUniform) && \"Can't scalarize a scalable vector\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9852, __extension__ __PRETTY_FUNCTION__)) | |||
9852 | "Can't scalarize a scalable vector")(static_cast <bool> ((!State.VF.isScalable() || IsUniform ) && "Can't scalarize a scalable vector") ? void (0) : __assert_fail ("(!State.VF.isScalable() || IsUniform) && \"Can't scalarize a scalable vector\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9852, __extension__ __PRETTY_FUNCTION__)); | |||
9853 | for (unsigned Part = 0; Part < State.UF; ++Part) | |||
9854 | for (unsigned Lane = 0; Lane < EndLane; ++Lane) | |||
9855 | State.ILV->scalarizeInstruction(getUnderlyingInstr(), this, | |||
9856 | VPIteration(Part, Lane), IsPredicated, | |||
9857 | State); | |||
9858 | } | |||
9859 | ||||
9860 | void VPBranchOnMaskRecipe::execute(VPTransformState &State) { | |||
9861 | assert(State.Instance && "Branch on Mask works only on single instance.")(static_cast <bool> (State.Instance && "Branch on Mask works only on single instance." ) ? void (0) : __assert_fail ("State.Instance && \"Branch on Mask works only on single instance.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9861, __extension__ __PRETTY_FUNCTION__)); | |||
9862 | ||||
9863 | unsigned Part = State.Instance->Part; | |||
9864 | unsigned Lane = State.Instance->Lane.getKnownLane(); | |||
9865 | ||||
9866 | Value *ConditionBit = nullptr; | |||
9867 | VPValue *BlockInMask = getMask(); | |||
9868 | if (BlockInMask) { | |||
9869 | ConditionBit = State.get(BlockInMask, Part); | |||
9870 | if (ConditionBit->getType()->isVectorTy()) | |||
9871 | ConditionBit = State.Builder.CreateExtractElement( | |||
9872 | ConditionBit, State.Builder.getInt32(Lane)); | |||
9873 | } else // Block in mask is all-one. | |||
9874 | ConditionBit = State.Builder.getTrue(); | |||
9875 | ||||
9876 | // Replace the temporary unreachable terminator with a new conditional branch, | |||
9877 | // whose two destinations will be set later when they are created. | |||
9878 | auto *CurrentTerminator = State.CFG.PrevBB->getTerminator(); | |||
9879 | assert(isa<UnreachableInst>(CurrentTerminator) &&(static_cast <bool> (isa<UnreachableInst>(CurrentTerminator ) && "Expected to replace unreachable terminator with conditional branch." ) ? void (0) : __assert_fail ("isa<UnreachableInst>(CurrentTerminator) && \"Expected to replace unreachable terminator with conditional branch.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9880, __extension__ __PRETTY_FUNCTION__)) | |||
9880 | "Expected to replace unreachable terminator with conditional branch.")(static_cast <bool> (isa<UnreachableInst>(CurrentTerminator ) && "Expected to replace unreachable terminator with conditional branch." ) ? void (0) : __assert_fail ("isa<UnreachableInst>(CurrentTerminator) && \"Expected to replace unreachable terminator with conditional branch.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9880, __extension__ __PRETTY_FUNCTION__)); | |||
9881 | auto *CondBr = BranchInst::Create(State.CFG.PrevBB, nullptr, ConditionBit); | |||
9882 | CondBr->setSuccessor(0, nullptr); | |||
9883 | ReplaceInstWithInst(CurrentTerminator, CondBr); | |||
9884 | } | |||
9885 | ||||
9886 | void VPPredInstPHIRecipe::execute(VPTransformState &State) { | |||
9887 | assert(State.Instance && "Predicated instruction PHI works per instance.")(static_cast <bool> (State.Instance && "Predicated instruction PHI works per instance." ) ? void (0) : __assert_fail ("State.Instance && \"Predicated instruction PHI works per instance.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9887, __extension__ __PRETTY_FUNCTION__)); | |||
9888 | Instruction *ScalarPredInst = | |||
9889 | cast<Instruction>(State.get(getOperand(0), *State.Instance)); | |||
9890 | BasicBlock *PredicatedBB = ScalarPredInst->getParent(); | |||
9891 | BasicBlock *PredicatingBB = PredicatedBB->getSinglePredecessor(); | |||
9892 | assert(PredicatingBB && "Predicated block has no single predecessor.")(static_cast <bool> (PredicatingBB && "Predicated block has no single predecessor." ) ? void (0) : __assert_fail ("PredicatingBB && \"Predicated block has no single predecessor.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9892, __extension__ __PRETTY_FUNCTION__)); | |||
9893 | assert(isa<VPReplicateRecipe>(getOperand(0)) &&(static_cast <bool> (isa<VPReplicateRecipe>(getOperand (0)) && "operand must be VPReplicateRecipe") ? void ( 0) : __assert_fail ("isa<VPReplicateRecipe>(getOperand(0)) && \"operand must be VPReplicateRecipe\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9894, __extension__ __PRETTY_FUNCTION__)) | |||
9894 | "operand must be VPReplicateRecipe")(static_cast <bool> (isa<VPReplicateRecipe>(getOperand (0)) && "operand must be VPReplicateRecipe") ? void ( 0) : __assert_fail ("isa<VPReplicateRecipe>(getOperand(0)) && \"operand must be VPReplicateRecipe\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9894, __extension__ __PRETTY_FUNCTION__)); | |||
9895 | ||||
9896 | // By current pack/unpack logic we need to generate only a single phi node: if | |||
9897 | // a vector value for the predicated instruction exists at this point it means | |||
9898 | // the instruction has vector users only, and a phi for the vector value is | |||
9899 | // needed. In this case the recipe of the predicated instruction is marked to | |||
9900 | // also do that packing, thereby "hoisting" the insert-element sequence. | |||
9901 | // Otherwise, a phi node for the scalar value is needed. | |||
9902 | unsigned Part = State.Instance->Part; | |||
9903 | if (State.hasVectorValue(getOperand(0), Part)) { | |||
9904 | Value *VectorValue = State.get(getOperand(0), Part); | |||
9905 | InsertElementInst *IEI = cast<InsertElementInst>(VectorValue); | |||
9906 | PHINode *VPhi = State.Builder.CreatePHI(IEI->getType(), 2); | |||
9907 | VPhi->addIncoming(IEI->getOperand(0), PredicatingBB); // Unmodified vector. | |||
9908 | VPhi->addIncoming(IEI, PredicatedBB); // New vector with inserted element. | |||
9909 | if (State.hasVectorValue(this, Part)) | |||
9910 | State.reset(this, VPhi, Part); | |||
9911 | else | |||
9912 | State.set(this, VPhi, Part); | |||
9913 | // NOTE: Currently we need to update the value of the operand, so the next | |||
9914 | // predicated iteration inserts its generated value in the correct vector. | |||
9915 | State.reset(getOperand(0), VPhi, Part); | |||
9916 | } else { | |||
9917 | Type *PredInstType = getOperand(0)->getUnderlyingValue()->getType(); | |||
9918 | PHINode *Phi = State.Builder.CreatePHI(PredInstType, 2); | |||
9919 | Phi->addIncoming(PoisonValue::get(ScalarPredInst->getType()), | |||
9920 | PredicatingBB); | |||
9921 | Phi->addIncoming(ScalarPredInst, PredicatedBB); | |||
9922 | if (State.hasScalarValue(this, *State.Instance)) | |||
9923 | State.reset(this, Phi, *State.Instance); | |||
9924 | else | |||
9925 | State.set(this, Phi, *State.Instance); | |||
9926 | // NOTE: Currently we need to update the value of the operand, so the next | |||
9927 | // predicated iteration inserts its generated value in the correct vector. | |||
9928 | State.reset(getOperand(0), Phi, *State.Instance); | |||
9929 | } | |||
9930 | } | |||
9931 | ||||
9932 | void VPWidenMemoryInstructionRecipe::execute(VPTransformState &State) { | |||
9933 | VPValue *StoredValue = isStore() ? getStoredValue() : nullptr; | |||
9934 | ||||
9935 | // Attempt to issue a wide load. | |||
9936 | LoadInst *LI = dyn_cast<LoadInst>(&Ingredient); | |||
9937 | StoreInst *SI = dyn_cast<StoreInst>(&Ingredient); | |||
9938 | ||||
9939 | assert((LI || SI) && "Invalid Load/Store instruction")(static_cast <bool> ((LI || SI) && "Invalid Load/Store instruction" ) ? void (0) : __assert_fail ("(LI || SI) && \"Invalid Load/Store instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9939, __extension__ __PRETTY_FUNCTION__)); | |||
9940 | assert((!SI || StoredValue) && "No stored value provided for widened store")(static_cast <bool> ((!SI || StoredValue) && "No stored value provided for widened store" ) ? void (0) : __assert_fail ("(!SI || StoredValue) && \"No stored value provided for widened store\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9940, __extension__ __PRETTY_FUNCTION__)); | |||
9941 | assert((!LI || !StoredValue) && "Stored value provided for widened load")(static_cast <bool> ((!LI || !StoredValue) && "Stored value provided for widened load" ) ? void (0) : __assert_fail ("(!LI || !StoredValue) && \"Stored value provided for widened load\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9941, __extension__ __PRETTY_FUNCTION__)); | |||
9942 | ||||
9943 | Type *ScalarDataTy = getLoadStoreType(&Ingredient); | |||
9944 | ||||
9945 | auto *DataTy = VectorType::get(ScalarDataTy, State.VF); | |||
9946 | const Align Alignment = getLoadStoreAlignment(&Ingredient); | |||
9947 | bool CreateGatherScatter = !Consecutive; | |||
9948 | ||||
9949 | auto &Builder = State.Builder; | |||
9950 | InnerLoopVectorizer::VectorParts BlockInMaskParts(State.UF); | |||
9951 | bool isMaskRequired = getMask(); | |||
9952 | if (isMaskRequired) | |||
9953 | for (unsigned Part = 0; Part < State.UF; ++Part) | |||
9954 | BlockInMaskParts[Part] = State.get(getMask(), Part); | |||
9955 | ||||
9956 | const auto CreateVecPtr = [&](unsigned Part, Value *Ptr) -> Value * { | |||
9957 | // Calculate the pointer for the specific unroll-part. | |||
9958 | GetElementPtrInst *PartPtr = nullptr; | |||
9959 | ||||
9960 | bool InBounds = false; | |||
9961 | if (auto *gep = dyn_cast<GetElementPtrInst>(Ptr->stripPointerCasts())) | |||
9962 | InBounds = gep->isInBounds(); | |||
9963 | if (Reverse) { | |||
9964 | // If the address is consecutive but reversed, then the | |||
9965 | // wide store needs to start at the last vector element. | |||
9966 | // RunTimeVF = VScale * VF.getKnownMinValue() | |||
9967 | // For fixed-width VScale is 1, then RunTimeVF = VF.getKnownMinValue() | |||
9968 | Value *RunTimeVF = getRuntimeVF(Builder, Builder.getInt32Ty(), State.VF); | |||
9969 | // NumElt = -Part * RunTimeVF | |||
9970 | Value *NumElt = Builder.CreateMul(Builder.getInt32(-Part), RunTimeVF); | |||
9971 | // LastLane = 1 - RunTimeVF | |||
9972 | Value *LastLane = Builder.CreateSub(Builder.getInt32(1), RunTimeVF); | |||
9973 | PartPtr = | |||
9974 | cast<GetElementPtrInst>(Builder.CreateGEP(ScalarDataTy, Ptr, NumElt)); | |||
9975 | PartPtr->setIsInBounds(InBounds); | |||
9976 | PartPtr = cast<GetElementPtrInst>( | |||
9977 | Builder.CreateGEP(ScalarDataTy, PartPtr, LastLane)); | |||
9978 | PartPtr->setIsInBounds(InBounds); | |||
9979 | if (isMaskRequired) // Reverse of a null all-one mask is a null mask. | |||
9980 | BlockInMaskParts[Part] = | |||
9981 | Builder.CreateVectorReverse(BlockInMaskParts[Part], "reverse"); | |||
9982 | } else { | |||
9983 | Value *Increment = | |||
9984 | createStepForVF(Builder, Builder.getInt32Ty(), State.VF, Part); | |||
9985 | PartPtr = cast<GetElementPtrInst>( | |||
9986 | Builder.CreateGEP(ScalarDataTy, Ptr, Increment)); | |||
9987 | PartPtr->setIsInBounds(InBounds); | |||
9988 | } | |||
9989 | ||||
9990 | unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace(); | |||
9991 | return Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace)); | |||
9992 | }; | |||
9993 | ||||
9994 | // Handle Stores: | |||
9995 | if (SI) { | |||
9996 | State.ILV->setDebugLocFromInst(SI); | |||
9997 | ||||
9998 | for (unsigned Part = 0; Part < State.UF; ++Part) { | |||
9999 | Instruction *NewSI = nullptr; | |||
10000 | Value *StoredVal = State.get(StoredValue, Part); | |||
10001 | if (CreateGatherScatter) { | |||
10002 | Value *MaskPart = isMaskRequired ? BlockInMaskParts[Part] : nullptr; | |||
10003 | Value *VectorGep = State.get(getAddr(), Part); | |||
10004 | NewSI = Builder.CreateMaskedScatter(StoredVal, VectorGep, Alignment, | |||
10005 | MaskPart); | |||
10006 | } else { | |||
10007 | if (Reverse) { | |||
10008 | // If we store to reverse consecutive memory locations, then we need | |||
10009 | // to reverse the order of elements in the stored value. | |||
10010 | StoredVal = Builder.CreateVectorReverse(StoredVal, "reverse"); | |||
10011 | // We don't want to update the value in the map as it might be used in | |||
10012 | // another expression. So don't call resetVectorValue(StoredVal). | |||
10013 | } | |||
10014 | auto *VecPtr = | |||
10015 | CreateVecPtr(Part, State.get(getAddr(), VPIteration(0, 0))); | |||
10016 | if (isMaskRequired) | |||
10017 | NewSI = Builder.CreateMaskedStore(StoredVal, VecPtr, Alignment, | |||
10018 | BlockInMaskParts[Part]); | |||
10019 | else | |||
10020 | NewSI = Builder.CreateAlignedStore(StoredVal, VecPtr, Alignment); | |||
10021 | } | |||
10022 | State.ILV->addMetadata(NewSI, SI); | |||
10023 | } | |||
10024 | return; | |||
10025 | } | |||
10026 | ||||
10027 | // Handle loads. | |||
10028 | assert(LI && "Must have a load instruction")(static_cast <bool> (LI && "Must have a load instruction" ) ? void (0) : __assert_fail ("LI && \"Must have a load instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 10028, __extension__ __PRETTY_FUNCTION__)); | |||
10029 | State.ILV->setDebugLocFromInst(LI); | |||
10030 | for (unsigned Part = 0; Part < State.UF; ++Part) { | |||
10031 | Value *NewLI; | |||
10032 | if (CreateGatherScatter) { | |||
10033 | Value *MaskPart = isMaskRequired ? BlockInMaskParts[Part] : nullptr; | |||
10034 | Value *VectorGep = State.get(getAddr(), Part); | |||
10035 | NewLI = Builder.CreateMaskedGather(DataTy, VectorGep, Alignment, MaskPart, | |||
10036 | nullptr, "wide.masked.gather"); | |||
10037 | State.ILV->addMetadata(NewLI, LI); | |||
10038 | } else { | |||
10039 | auto *VecPtr = | |||
10040 | CreateVecPtr(Part, State.get(getAddr(), VPIteration(0, 0))); | |||
10041 | if (isMaskRequired) | |||
10042 | NewLI = Builder.CreateMaskedLoad( | |||
10043 | DataTy, VecPtr, Alignment, BlockInMaskParts[Part], | |||
10044 | PoisonValue::get(DataTy), "wide.masked.load"); | |||
10045 | else | |||
10046 | NewLI = | |||
10047 | Builder.CreateAlignedLoad(DataTy, VecPtr, Alignment, "wide.load"); | |||
10048 | ||||
10049 | // Add metadata to the load, but setVectorValue to the reverse shuffle. | |||
10050 | State.ILV->addMetadata(NewLI, LI); | |||
10051 | if (Reverse) | |||
10052 | NewLI = Builder.CreateVectorReverse(NewLI, "reverse"); | |||
10053 | } | |||
10054 | ||||
10055 | State.set(this, NewLI, Part); | |||
10056 | } | |||
10057 | } | |||
10058 | ||||
10059 | // Determine how to lower the scalar epilogue, which depends on 1) optimising | |||
10060 | // for minimum code-size, 2) predicate compiler options, 3) loop hints forcing | |||
10061 | // predication, and 4) a TTI hook that analyses whether the loop is suitable | |||
10062 | // for predication. | |||
10063 | static ScalarEpilogueLowering getScalarEpilogueLowering( | |||
10064 | Function *F, Loop *L, LoopVectorizeHints &Hints, ProfileSummaryInfo *PSI, | |||
10065 | BlockFrequencyInfo *BFI, TargetTransformInfo *TTI, TargetLibraryInfo *TLI, | |||
10066 | AssumptionCache *AC, LoopInfo *LI, ScalarEvolution *SE, DominatorTree *DT, | |||
10067 | LoopVectorizationLegality &LVL) { | |||
10068 | // 1) OptSize takes precedence over all other options, i.e. if this is set, | |||
10069 | // don't look at hints or options, and don't request a scalar epilogue. | |||
10070 | // (For PGSO, as shouldOptimizeForSize isn't currently accessible from | |||
10071 | // LoopAccessInfo (due to code dependency and not being able to reliably get | |||
10072 | // PSI/BFI from a loop analysis under NPM), we cannot suppress the collection | |||
10073 | // of strides in LoopAccessInfo::analyzeLoop() and vectorize without | |||
10074 | // versioning when the vectorization is forced, unlike hasOptSize. So revert | |||
10075 | // back to the old way and vectorize with versioning when forced. See D81345.) | |||
10076 | if (F->hasOptSize() || (llvm::shouldOptimizeForSize(L->getHeader(), PSI, BFI, | |||
10077 | PGSOQueryType::IRPass) && | |||
10078 | Hints.getForce() != LoopVectorizeHints::FK_Enabled)) | |||
10079 | return CM_ScalarEpilogueNotAllowedOptSize; | |||
10080 | ||||
10081 | // 2) If set, obey the directives | |||
10082 | if (PreferPredicateOverEpilogue.getNumOccurrences()) { | |||
10083 | switch (PreferPredicateOverEpilogue) { | |||
10084 | case PreferPredicateTy::ScalarEpilogue: | |||
10085 | return CM_ScalarEpilogueAllowed; | |||
10086 | case PreferPredicateTy::PredicateElseScalarEpilogue: | |||
10087 | return CM_ScalarEpilogueNotNeededUsePredicate; | |||
10088 | case PreferPredicateTy::PredicateOrDontVectorize: | |||
10089 | return CM_ScalarEpilogueNotAllowedUsePredicate; | |||
10090 | }; | |||
10091 | } | |||
10092 | ||||
10093 | // 3) If set, obey the hints | |||
10094 | switch (Hints.getPredicate()) { | |||
10095 | case LoopVectorizeHints::FK_Enabled: | |||
10096 | return CM_ScalarEpilogueNotNeededUsePredicate; | |||
10097 | case LoopVectorizeHints::FK_Disabled: | |||
10098 | return CM_ScalarEpilogueAllowed; | |||
10099 | }; | |||
10100 | ||||
10101 | // 4) if the TTI hook indicates this is profitable, request predication. | |||
10102 | if (TTI->preferPredicateOverEpilogue(L, LI, *SE, *AC, TLI, DT, | |||
10103 | LVL.getLAI())) | |||
10104 | return CM_ScalarEpilogueNotNeededUsePredicate; | |||
10105 | ||||
10106 | return CM_ScalarEpilogueAllowed; | |||
10107 | } | |||
10108 | ||||
10109 | Value *VPTransformState::get(VPValue *Def, unsigned Part) { | |||
10110 | // If Values have been set for this Def return the one relevant for \p Part. | |||
10111 | if (hasVectorValue(Def, Part)) | |||
10112 | return Data.PerPartOutput[Def][Part]; | |||
10113 | ||||
10114 | if (!hasScalarValue(Def, {Part, 0})) { | |||
10115 | Value *IRV = Def->getLiveInIRValue(); | |||
10116 | Value *B = ILV->getBroadcastInstrs(IRV); | |||
10117 | set(Def, B, Part); | |||
10118 | return B; | |||
10119 | } | |||
10120 | ||||
10121 | Value *ScalarValue = get(Def, {Part, 0}); | |||
10122 | // If we aren't vectorizing, we can just copy the scalar map values over | |||
10123 | // to the vector map. | |||
10124 | if (VF.isScalar()) { | |||
10125 | set(Def, ScalarValue, Part); | |||
10126 | return ScalarValue; | |||
10127 | } | |||
10128 | ||||
10129 | auto *RepR = dyn_cast<VPReplicateRecipe>(Def); | |||
10130 | bool IsUniform = RepR && RepR->isUniform(); | |||
10131 | ||||
10132 | unsigned LastLane = IsUniform ? 0 : VF.getKnownMinValue() - 1; | |||
10133 | // Check if there is a scalar value for the selected lane. | |||
10134 | if (!hasScalarValue(Def, {Part, LastLane})) { | |||
10135 | // At the moment, VPWidenIntOrFpInductionRecipes can also be uniform. | |||
10136 | assert((isa<VPWidenIntOrFpInductionRecipe>(Def->getDef()) ||(static_cast <bool> ((isa<VPWidenIntOrFpInductionRecipe >(Def->getDef()) || isa<VPScalarIVStepsRecipe>(Def ->getDef())) && "unexpected recipe found to be invariant" ) ? void (0) : __assert_fail ("(isa<VPWidenIntOrFpInductionRecipe>(Def->getDef()) || isa<VPScalarIVStepsRecipe>(Def->getDef())) && \"unexpected recipe found to be invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 10138, __extension__ __PRETTY_FUNCTION__)) | |||
10137 | isa<VPScalarIVStepsRecipe>(Def->getDef())) &&(static_cast <bool> ((isa<VPWidenIntOrFpInductionRecipe >(Def->getDef()) || isa<VPScalarIVStepsRecipe>(Def ->getDef())) && "unexpected recipe found to be invariant" ) ? void (0) : __assert_fail ("(isa<VPWidenIntOrFpInductionRecipe>(Def->getDef()) || isa<VPScalarIVStepsRecipe>(Def->getDef())) && \"unexpected recipe found to be invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 10138, __extension__ __PRETTY_FUNCTION__)) | |||
10138 | "unexpected recipe found to be invariant")(static_cast <bool> ((isa<VPWidenIntOrFpInductionRecipe >(Def->getDef()) || isa<VPScalarIVStepsRecipe>(Def ->getDef())) && "unexpected recipe found to be invariant" ) ? void (0) : __assert_fail ("(isa<VPWidenIntOrFpInductionRecipe>(Def->getDef()) || isa<VPScalarIVStepsRecipe>(Def->getDef())) && \"unexpected recipe found to be invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 10138, __extension__ __PRETTY_FUNCTION__)); | |||
10139 | IsUniform = true; | |||
10140 | LastLane = 0; | |||
10141 | } | |||
10142 | ||||
10143 | auto *LastInst = cast<Instruction>(get(Def, {Part, LastLane})); | |||
10144 | // Set the insert point after the last scalarized instruction or after the | |||
10145 | // last PHI, if LastInst is a PHI. This ensures the insertelement sequence | |||
10146 | // will directly follow the scalar definitions. | |||
10147 | auto OldIP = Builder.saveIP(); | |||
10148 | auto NewIP = | |||
10149 | isa<PHINode>(LastInst) | |||
10150 | ? BasicBlock::iterator(LastInst->getParent()->getFirstNonPHI()) | |||
10151 | : std::next(BasicBlock::iterator(LastInst)); | |||
10152 | Builder.SetInsertPoint(&*NewIP); | |||
10153 | ||||
10154 | // However, if we are vectorizing, we need to construct the vector values. | |||
10155 | // If the value is known to be uniform after vectorization, we can just | |||
10156 | // broadcast the scalar value corresponding to lane zero for each unroll | |||
10157 | // iteration. Otherwise, we construct the vector values using | |||
10158 | // insertelement instructions. Since the resulting vectors are stored in | |||
10159 | // State, we will only generate the insertelements once. | |||
10160 | Value *VectorValue = nullptr; | |||
10161 | if (IsUniform) { | |||
10162 | VectorValue = ILV->getBroadcastInstrs(ScalarValue); | |||
10163 | set(Def, VectorValue, Part); | |||
10164 | } else { | |||
10165 | // Initialize packing with insertelements to start from undef. | |||
10166 | assert(!VF.isScalable() && "VF is assumed to be non scalable.")(static_cast <bool> (!VF.isScalable() && "VF is assumed to be non scalable." ) ? void (0) : __assert_fail ("!VF.isScalable() && \"VF is assumed to be non scalable.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 10166, __extension__ __PRETTY_FUNCTION__)); | |||
10167 | Value *Undef = PoisonValue::get(VectorType::get(LastInst->getType(), VF)); | |||
10168 | set(Def, Undef, Part); | |||
10169 | for (unsigned Lane = 0; Lane < VF.getKnownMinValue(); ++Lane) | |||
10170 | ILV->packScalarIntoVectorValue(Def, {Part, Lane}, *this); | |||
10171 | VectorValue = get(Def, Part); | |||
10172 | } | |||
10173 | Builder.restoreIP(OldIP); | |||
10174 | return VectorValue; | |||
10175 | } | |||
10176 | ||||
10177 | // Process the loop in the VPlan-native vectorization path. This path builds | |||
10178 | // VPlan upfront in the vectorization pipeline, which allows to apply | |||
10179 | // VPlan-to-VPlan transformations from the very beginning without modifying the | |||
10180 | // input LLVM IR. | |||
10181 | static bool processLoopInVPlanNativePath( | |||
10182 | Loop *L, PredicatedScalarEvolution &PSE, LoopInfo *LI, DominatorTree *DT, | |||
10183 | LoopVectorizationLegality *LVL, TargetTransformInfo *TTI, | |||
10184 | TargetLibraryInfo *TLI, DemandedBits *DB, AssumptionCache *AC, | |||
10185 | OptimizationRemarkEmitter *ORE, BlockFrequencyInfo *BFI, | |||
10186 | ProfileSummaryInfo *PSI, LoopVectorizeHints &Hints, | |||
10187 | LoopVectorizationRequirements &Requirements) { | |||
10188 | ||||
10189 | if (isa<SCEVCouldNotCompute>(PSE.getBackedgeTakenCount())) { | |||
10190 | LLVM_DEBUG(dbgs() << "LV: cannot compute the outer-loop trip count\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: cannot compute the outer-loop trip count\n" ; } } while (false); | |||
10191 | return false; | |||
10192 | } | |||
10193 | assert(EnableVPlanNativePath && "VPlan-native path is disabled.")(static_cast <bool> (EnableVPlanNativePath && "VPlan-native path is disabled." ) ? void (0) : __assert_fail ("EnableVPlanNativePath && \"VPlan-native path is disabled.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 10193, __extension__ __PRETTY_FUNCTION__)); | |||
10194 | Function *F = L->getHeader()->getParent(); | |||
10195 | InterleavedAccessInfo IAI(PSE, L, DT, LI, LVL->getLAI()); | |||
10196 | ||||
10197 | ScalarEpilogueLowering SEL = getScalarEpilogueLowering( | |||
10198 | F, L, Hints, PSI, BFI, TTI, TLI, AC, LI, PSE.getSE(), DT, *LVL); | |||
10199 | ||||
10200 | LoopVectorizationCostModel CM(SEL, L, PSE, LI, LVL, *TTI, TLI, DB, AC, ORE, F, | |||
10201 | &Hints, IAI); | |||
10202 | // Use the planner for outer loop vectorization. | |||
10203 | // TODO: CM is not used at this point inside the planner. Turn CM into an | |||
10204 | // optional argument if we don't need it in the future. | |||
10205 | LoopVectorizationPlanner LVP(L, LI, TLI, TTI, LVL, CM, IAI, PSE, Hints, | |||
10206 | Requirements, ORE); | |||
10207 | ||||
10208 | // Get user vectorization factor. | |||
10209 | ElementCount UserVF = Hints.getWidth(); | |||
10210 | ||||
10211 | CM.collectElementTypesForWidening(); | |||
10212 | ||||
10213 | // Plan how to best vectorize, return the best VF and its cost. | |||
10214 | const VectorizationFactor VF = LVP.planInVPlanNativePath(UserVF); | |||
10215 | ||||
10216 | // If we are stress testing VPlan builds, do not attempt to generate vector | |||
10217 | // code. Masked vector code generation support will follow soon. | |||
10218 | // Also, do not attempt to vectorize if no vector code will be produced. | |||
10219 | if (VPlanBuildStressTest || EnableVPlanPredication || | |||
10220 | VectorizationFactor::Disabled() == VF) | |||
10221 | return false; | |||
10222 | ||||
10223 | VPlan &BestPlan = LVP.getBestPlanFor(VF.Width); | |||
10224 | ||||
10225 | { | |||
10226 | GeneratedRTChecks Checks(*PSE.getSE(), DT, LI, | |||
10227 | F->getParent()->getDataLayout()); | |||
10228 | InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, VF.Width, 1, LVL, | |||
10229 | &CM, BFI, PSI, Checks); | |||
10230 | LLVM_DEBUG(dbgs() << "Vectorizing outer loop in \""do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Vectorizing outer loop in \"" << L->getHeader()->getParent()->getName() << "\"\n"; } } while (false) | |||
10231 | << L->getHeader()->getParent()->getName() << "\"\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Vectorizing outer loop in \"" << L->getHeader()->getParent()->getName() << "\"\n"; } } while (false); | |||
10232 | LVP.executePlan(VF.Width, 1, BestPlan, LB, DT); | |||
10233 | } | |||
10234 | ||||
10235 | // Mark the loop as already vectorized to avoid vectorizing again. | |||
10236 | Hints.setAlreadyVectorized(); | |||
10237 | assert(!verifyFunction(*L->getHeader()->getParent(), &dbgs()))(static_cast <bool> (!verifyFunction(*L->getHeader() ->getParent(), &dbgs())) ? void (0) : __assert_fail ("!verifyFunction(*L->getHeader()->getParent(), &dbgs())" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 10237, __extension__ __PRETTY_FUNCTION__)); | |||
10238 | return true; | |||
10239 | } | |||
10240 | ||||
10241 | // Emit a remark if there are stores to floats that required a floating point | |||
10242 | // extension. If the vectorized loop was generated with floating point there | |||
10243 | // will be a performance penalty from the conversion overhead and the change in | |||
10244 | // the vector width. | |||
10245 | static void checkMixedPrecision(Loop *L, OptimizationRemarkEmitter *ORE) { | |||
10246 | SmallVector<Instruction *, 4> Worklist; | |||
10247 | for (BasicBlock *BB : L->getBlocks()) { | |||
10248 | for (Instruction &Inst : *BB) { | |||
10249 | if (auto *S = dyn_cast<StoreInst>(&Inst)) { | |||
10250 | if (S->getValueOperand()->getType()->isFloatTy()) | |||
10251 | Worklist.push_back(S); | |||
10252 | } | |||
10253 | } | |||
10254 | } | |||
10255 | ||||
10256 | // Traverse the floating point stores upwards searching, for floating point | |||
10257 | // conversions. | |||
10258 | SmallPtrSet<const Instruction *, 4> Visited; | |||
10259 | SmallPtrSet<const Instruction *, 4> EmittedRemark; | |||
10260 | while (!Worklist.empty()) { | |||
10261 | auto *I = Worklist.pop_back_val(); | |||
10262 | if (!L->contains(I)) | |||
10263 | continue; | |||
10264 | if (!Visited.insert(I).second) | |||
10265 | continue; | |||
10266 | ||||
10267 | // Emit a remark if the floating point store required a floating | |||
10268 | // point conversion. | |||
10269 | // TODO: More work could be done to identify the root cause such as a | |||
10270 | // constant or a function return type and point the user to it. | |||
10271 | if (isa<FPExtInst>(I) && EmittedRemark.insert(I).second) | |||
10272 | ORE->emit([&]() { | |||
10273 | return OptimizationRemarkAnalysis(LV_NAME"loop-vectorize", "VectorMixedPrecision", | |||
10274 | I->getDebugLoc(), L->getHeader()) | |||
10275 | << "floating point conversion changes vector width. " | |||
10276 | << "Mixed floating point precision requires an up/down " | |||
10277 | << "cast that will negatively impact performance."; | |||
10278 | }); | |||
10279 | ||||
10280 | for (Use &Op : I->operands()) | |||
10281 | if (auto *OpI = dyn_cast<Instruction>(Op)) | |||
10282 | Worklist.push_back(OpI); | |||
10283 | } | |||
10284 | } | |||
10285 | ||||
10286 | LoopVectorizePass::LoopVectorizePass(LoopVectorizeOptions Opts) | |||
10287 | : InterleaveOnlyWhenForced(Opts.InterleaveOnlyWhenForced || | |||
10288 | !EnableLoopInterleaving), | |||
10289 | VectorizeOnlyWhenForced(Opts.VectorizeOnlyWhenForced || | |||
10290 | !EnableLoopVectorization) {} | |||
10291 | ||||
10292 | bool LoopVectorizePass::processLoop(Loop *L) { | |||
10293 | assert((EnableVPlanNativePath || L->isInnermost()) &&(static_cast <bool> ((EnableVPlanNativePath || L->isInnermost ()) && "VPlan-native path is not enabled. Only process inner loops." ) ? void (0) : __assert_fail ("(EnableVPlanNativePath || L->isInnermost()) && \"VPlan-native path is not enabled. Only process inner loops.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 10294, __extension__ __PRETTY_FUNCTION__)) | |||
10294 | "VPlan-native path is not enabled. Only process inner loops.")(static_cast <bool> ((EnableVPlanNativePath || L->isInnermost ()) && "VPlan-native path is not enabled. Only process inner loops." ) ? void (0) : __assert_fail ("(EnableVPlanNativePath || L->isInnermost()) && \"VPlan-native path is not enabled. Only process inner loops.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 10294, __extension__ __PRETTY_FUNCTION__)); | |||
10295 | ||||
10296 | #ifndef NDEBUG | |||
10297 | const std::string DebugLocStr = getDebugLocString(L); | |||
10298 | #endif /* NDEBUG */ | |||
10299 | ||||
10300 | LLVM_DEBUG(dbgs() << "\nLV: Checking a loop in '"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "\nLV: Checking a loop in '" << L->getHeader()->getParent()->getName() << "' from " << DebugLocStr << "\n"; } } while (false ) | |||
10301 | << L->getHeader()->getParent()->getName() << "' from "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "\nLV: Checking a loop in '" << L->getHeader()->getParent()->getName() << "' from " << DebugLocStr << "\n"; } } while (false ) | |||
10302 | << DebugLocStr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "\nLV: Checking a loop in '" << L->getHeader()->getParent()->getName() << "' from " << DebugLocStr << "\n"; } } while (false ); | |||
10303 | ||||
10304 | LoopVectorizeHints Hints(L, InterleaveOnlyWhenForced, *ORE, TTI); | |||
10305 | ||||
10306 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints:" << " force=" << (Hints.getForce() == LoopVectorizeHints:: FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints ::FK_Enabled ? "enabled" : "?")) << " width=" << Hints .getWidth() << " interleave=" << Hints.getInterleave () << "\n"; } } while (false) | |||
10307 | dbgs() << "LV: Loop hints:"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints:" << " force=" << (Hints.getForce() == LoopVectorizeHints:: FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints ::FK_Enabled ? "enabled" : "?")) << " width=" << Hints .getWidth() << " interleave=" << Hints.getInterleave () << "\n"; } } while (false) | |||
10308 | << " force="do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints:" << " force=" << (Hints.getForce() == LoopVectorizeHints:: FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints ::FK_Enabled ? "enabled" : "?")) << " width=" << Hints .getWidth() << " interleave=" << Hints.getInterleave () << "\n"; } } while (false) | |||
10309 | << (Hints.getForce() == LoopVectorizeHints::FK_Disableddo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints:" << " force=" << (Hints.getForce() == LoopVectorizeHints:: FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints ::FK_Enabled ? "enabled" : "?")) << " width=" << Hints .getWidth() << " interleave=" << Hints.getInterleave () << "\n"; } } while (false) | |||
10310 | ? "disabled"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints:" << " force=" << (Hints.getForce() == LoopVectorizeHints:: FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints ::FK_Enabled ? "enabled" : "?")) << " width=" << Hints .getWidth() << " interleave=" << Hints.getInterleave () << "\n"; } } while (false) | |||
10311 | : (Hints.getForce() == LoopVectorizeHints::FK_Enableddo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints:" << " force=" << (Hints.getForce() == LoopVectorizeHints:: FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints ::FK_Enabled ? "enabled" : "?")) << " width=" << Hints .getWidth() << " interleave=" << Hints.getInterleave () << "\n"; } } while (false) | |||
10312 | ? "enabled"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints:" << " force=" << (Hints.getForce() == LoopVectorizeHints:: FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints ::FK_Enabled ? "enabled" : "?")) << " width=" << Hints .getWidth() << " interleave=" << Hints.getInterleave () << "\n"; } } while (false) | |||
10313 | : "?"))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints:" << " force=" << (Hints.getForce() == LoopVectorizeHints:: FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints ::FK_Enabled ? "enabled" : "?")) << " width=" << Hints .getWidth() << " interleave=" << Hints.getInterleave () << "\n"; } } while (false) | |||
10314 | << " width=" << Hints.getWidth()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints:" << " force=" << (Hints.getForce() == LoopVectorizeHints:: FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints ::FK_Enabled ? "enabled" : "?")) << " width=" << Hints .getWidth() << " interleave=" << Hints.getInterleave () << "\n"; } } while (false) | |||
10315 | << " interleave=" << Hints.getInterleave() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints:" << " force=" << (Hints.getForce() == LoopVectorizeHints:: FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints ::FK_Enabled ? "enabled" : "?")) << " width=" << Hints .getWidth() << " interleave=" << Hints.getInterleave () << "\n"; } } while (false); | |||
10316 | ||||
10317 | // Function containing loop | |||
10318 | Function *F = L->getHeader()->getParent(); | |||
10319 | ||||
10320 | // Looking at the diagnostic output is the only way to determine if a loop | |||
10321 | // was vectorized (other than looking at the IR or machine code), so it | |||
10322 | // is important to generate an optimization remark for each loop. Most of | |||
10323 | // these messages are generated as OptimizationRemarkAnalysis. Remarks | |||
10324 | // generated as OptimizationRemark and OptimizationRemarkMissed are | |||
10325 | // less verbose reporting vectorized loops and unvectorized loops that may | |||
10326 | // benefit from vectorization, respectively. | |||
10327 | ||||
10328 | if (!Hints.allowVectorization(F, L, VectorizeOnlyWhenForced)) { | |||
10329 | LLVM_DEBUG(dbgs() << "LV: Loop hints prevent vectorization.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints prevent vectorization.\n" ; } } while (false); | |||
10330 | return false; | |||
10331 | } | |||
10332 | ||||
10333 | PredicatedScalarEvolution PSE(*SE, *L); | |||
10334 | ||||
10335 | // Check if it is legal to vectorize the loop. | |||
10336 | LoopVectorizationRequirements Requirements; | |||
10337 | LoopVectorizationLegality LVL(L, PSE, DT, TTI, TLI, AA, F, GetLAA, LI, ORE, | |||
10338 | &Requirements, &Hints, DB, AC, BFI, PSI); | |||
10339 | if (!LVL.canVectorize(EnableVPlanNativePath)) { | |||
10340 | LLVM_DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not vectorizing: Cannot prove legality.\n" ; } } while (false); | |||
10341 | Hints.emitRemarkWithHints(); | |||
10342 | return false; | |||
10343 | } | |||
10344 | ||||
10345 | // Check the function attributes and profiles to find out if this function | |||
10346 | // should be optimized for size. | |||
10347 | ScalarEpilogueLowering SEL = getScalarEpilogueLowering( | |||
10348 | F, L, Hints, PSI, BFI, TTI, TLI, AC, LI, PSE.getSE(), DT, LVL); | |||
10349 | ||||
10350 | // Entrance to the VPlan-native vectorization path. Outer loops are processed | |||
10351 | // here. They may require CFG and instruction level transformations before | |||
10352 | // even evaluating whether vectorization is profitable. Since we cannot modify | |||
10353 | // the incoming IR, we need to build VPlan upfront in the vectorization | |||
10354 | // pipeline. | |||
10355 | if (!L->isInnermost()) | |||
10356 | return processLoopInVPlanNativePath(L, PSE, LI, DT, &LVL, TTI, TLI, DB, AC, | |||
10357 | ORE, BFI, PSI, Hints, Requirements); | |||
10358 | ||||
10359 | assert(L->isInnermost() && "Inner loop expected.")(static_cast <bool> (L->isInnermost() && "Inner loop expected." ) ? void (0) : __assert_fail ("L->isInnermost() && \"Inner loop expected.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 10359, __extension__ __PRETTY_FUNCTION__)); | |||
10360 | ||||
10361 | // Check the loop for a trip count threshold: vectorize loops with a tiny trip | |||
10362 | // count by optimizing for size, to minimize overheads. | |||
10363 | auto ExpectedTC = getSmallBestKnownTC(*SE, L); | |||
10364 | if (ExpectedTC && *ExpectedTC < TinyTripCountVectorThreshold) { | |||
10365 | LLVM_DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found a loop with a very small trip count. " << "This loop is worth vectorizing only if no scalar " << "iteration overheads are incurred."; } } while (false ) | |||
10366 | << "This loop is worth vectorizing only if no scalar "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found a loop with a very small trip count. " << "This loop is worth vectorizing only if no scalar " << "iteration overheads are incurred."; } } while (false ) | |||
10367 | << "iteration overheads are incurred.")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found a loop with a very small trip count. " << "This loop is worth vectorizing only if no scalar " << "iteration overheads are incurred."; } } while (false ); | |||
10368 | if (Hints.getForce() == LoopVectorizeHints::FK_Enabled) | |||
10369 | LLVM_DEBUG(dbgs() << " But vectorizing was explicitly forced.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << " But vectorizing was explicitly forced.\n" ; } } while (false); | |||
10370 | else { | |||
10371 | LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "\n"; } } while (false); | |||
10372 | SEL = CM_ScalarEpilogueNotAllowedLowTripLoop; | |||
10373 | } | |||
10374 | } | |||
10375 | ||||
10376 | // Check the function attributes to see if implicit floats are allowed. | |||
10377 | // FIXME: This check doesn't seem possibly correct -- what if the loop is | |||
10378 | // an integer loop and the vector instructions selected are purely integer | |||
10379 | // vector instructions? | |||
10380 | if (F->hasFnAttribute(Attribute::NoImplicitFloat)) { | |||
10381 | reportVectorizationFailure( | |||
10382 | "Can't vectorize when the NoImplicitFloat attribute is used", | |||
10383 | "loop not vectorized due to NoImplicitFloat attribute", | |||
10384 | "NoImplicitFloat", ORE, L); | |||
10385 | Hints.emitRemarkWithHints(); | |||
10386 | return false; | |||
10387 | } | |||
10388 | ||||
10389 | // Check if the target supports potentially unsafe FP vectorization. | |||
10390 | // FIXME: Add a check for the type of safety issue (denormal, signaling) | |||
10391 | // for the target we're vectorizing for, to make sure none of the | |||
10392 | // additional fp-math flags can help. | |||
10393 | if (Hints.isPotentiallyUnsafe() && | |||
10394 | TTI->isFPVectorizationPotentiallyUnsafe()) { | |||
10395 | reportVectorizationFailure( | |||
10396 | "Potentially unsafe FP op prevents vectorization", | |||
10397 | "loop not vectorized due to unsafe FP support.", | |||
10398 | "UnsafeFP", ORE, L); | |||
10399 | Hints.emitRemarkWithHints(); | |||
10400 | return false; | |||
10401 | } | |||
10402 | ||||
10403 | bool AllowOrderedReductions; | |||
10404 | // If the flag is set, use that instead and override the TTI behaviour. | |||
10405 | if (ForceOrderedReductions.getNumOccurrences() > 0) | |||
10406 | AllowOrderedReductions = ForceOrderedReductions; | |||
10407 | else | |||
10408 | AllowOrderedReductions = TTI->enableOrderedReductions(); | |||
10409 | if (!LVL.canVectorizeFPMath(AllowOrderedReductions)) { | |||
10410 | ORE->emit([&]() { | |||
10411 | auto *ExactFPMathInst = Requirements.getExactFPInst(); | |||
10412 | return OptimizationRemarkAnalysisFPCommute(DEBUG_TYPE"loop-vectorize", "CantReorderFPOps", | |||
10413 | ExactFPMathInst->getDebugLoc(), | |||
10414 | ExactFPMathInst->getParent()) | |||
10415 | << "loop not vectorized: cannot prove it is safe to reorder " | |||
10416 | "floating-point operations"; | |||
10417 | }); | |||
10418 | LLVM_DEBUG(dbgs() << "LV: loop not vectorized: cannot prove it is safe to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: loop not vectorized: cannot prove it is safe to " "reorder floating-point operations\n"; } } while (false) | |||
10419 | "reorder floating-point operations\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: loop not vectorized: cannot prove it is safe to " "reorder floating-point operations\n"; } } while (false); | |||
10420 | Hints.emitRemarkWithHints(); | |||
10421 | return false; | |||
10422 | } | |||
10423 | ||||
10424 | bool UseInterleaved = TTI->enableInterleavedAccessVectorization(); | |||
10425 | InterleavedAccessInfo IAI(PSE, L, DT, LI, LVL.getLAI()); | |||
10426 | ||||
10427 | // If an override option has been passed in for interleaved accesses, use it. | |||
10428 | if (EnableInterleavedMemAccesses.getNumOccurrences() > 0) | |||
10429 | UseInterleaved = EnableInterleavedMemAccesses; | |||
10430 | ||||
10431 | // Analyze interleaved memory accesses. | |||
10432 | if (UseInterleaved) { | |||
10433 | IAI.analyzeInterleaving(useMaskedInterleavedAccesses(*TTI)); | |||
10434 | } | |||
10435 | ||||
10436 | // Use the cost model. | |||
10437 | LoopVectorizationCostModel CM(SEL, L, PSE, LI, &LVL, *TTI, TLI, DB, AC, ORE, | |||
10438 | F, &Hints, IAI); | |||
10439 | CM.collectValuesToIgnore(); | |||
10440 | CM.collectElementTypesForWidening(); | |||
10441 | ||||
10442 | // Use the planner for vectorization. | |||
10443 | LoopVectorizationPlanner LVP(L, LI, TLI, TTI, &LVL, CM, IAI, PSE, Hints, | |||
10444 | Requirements, ORE); | |||
10445 | ||||
10446 | // Get user vectorization factor and interleave count. | |||
10447 | ElementCount UserVF = Hints.getWidth(); | |||
10448 | unsigned UserIC = Hints.getInterleave(); | |||
10449 | ||||
10450 | // Plan how to best vectorize, return the best VF and its cost. | |||
10451 | Optional<VectorizationFactor> MaybeVF = LVP.plan(UserVF, UserIC); | |||
10452 | ||||
10453 | VectorizationFactor VF = VectorizationFactor::Disabled(); | |||
10454 | unsigned IC = 1; | |||
10455 | ||||
10456 | if (MaybeVF) { | |||
10457 | VF = *MaybeVF; | |||
10458 | // Select the interleave count. | |||
10459 | IC = CM.selectInterleaveCount(VF.Width, *VF.Cost.getValue()); | |||
10460 | } | |||
10461 | ||||
10462 | // Identify the diagnostic messages that should be produced. | |||
10463 | std::pair<StringRef, std::string> VecDiagMsg, IntDiagMsg; | |||
10464 | bool VectorizeLoop = true, InterleaveLoop = true; | |||
10465 | if (VF.Width.isScalar()) { | |||
10466 | LLVM_DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Vectorization is possible but not beneficial.\n" ; } } while (false); | |||
10467 | VecDiagMsg = std::make_pair( | |||
10468 | "VectorizationNotBeneficial", | |||
10469 | "the cost-model indicates that vectorization is not beneficial"); | |||
10470 | VectorizeLoop = false; | |||
10471 | } | |||
10472 | ||||
10473 | if (!MaybeVF && UserIC > 1) { | |||
10474 | // Tell the user interleaving was avoided up-front, despite being explicitly | |||
10475 | // requested. | |||
10476 | LLVM_DEBUG(dbgs() << "LV: Ignoring UserIC, because vectorization and "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Ignoring UserIC, because vectorization and " "interleaving should be avoided up front\n"; } } while (false ) | |||
10477 | "interleaving should be avoided up front\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Ignoring UserIC, because vectorization and " "interleaving should be avoided up front\n"; } } while (false ); | |||
10478 | IntDiagMsg = std::make_pair( | |||
10479 | "InterleavingAvoided", | |||
10480 | "Ignoring UserIC, because interleaving was avoided up front"); | |||
10481 | InterleaveLoop = false; | |||
10482 | } else if (IC == 1 && UserIC <= 1) { | |||
10483 | // Tell the user interleaving is not beneficial. | |||
10484 | LLVM_DEBUG(dbgs() << "LV: Interleaving is not beneficial.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving is not beneficial.\n" ; } } while (false); | |||
10485 | IntDiagMsg = std::make_pair( | |||
10486 | "InterleavingNotBeneficial", | |||
10487 | "the cost-model indicates that interleaving is not beneficial"); | |||
10488 | InterleaveLoop = false; | |||
10489 | if (UserIC == 1) { | |||
10490 | IntDiagMsg.first = "InterleavingNotBeneficialAndDisabled"; | |||
10491 | IntDiagMsg.second += | |||
10492 | " and is explicitly disabled or interleave count is set to 1"; | |||
10493 | } | |||
10494 | } else if (IC > 1 && UserIC == 1) { | |||
10495 | // Tell the user interleaving is beneficial, but it explicitly disabled. | |||
10496 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving is beneficial but is explicitly disabled." ; } } while (false) | |||
10497 | dbgs() << "LV: Interleaving is beneficial but is explicitly disabled.")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving is beneficial but is explicitly disabled." ; } } while (false); | |||
10498 | IntDiagMsg = std::make_pair( | |||
10499 | "InterleavingBeneficialButDisabled", | |||
10500 | "the cost-model indicates that interleaving is beneficial " | |||
10501 | "but is explicitly disabled or interleave count is set to 1"); | |||
10502 | InterleaveLoop = false; | |||
10503 | } | |||
10504 | ||||
10505 | // Override IC if user provided an interleave count. | |||
10506 | IC = UserIC > 0 ? UserIC : IC; | |||
10507 | ||||
10508 | // Emit diagnostic messages, if any. | |||
10509 | const char *VAPassName = Hints.vectorizeAnalysisPassName(); | |||
10510 | if (!VectorizeLoop && !InterleaveLoop) { | |||
10511 | // Do not vectorize or interleaving the loop. | |||
10512 | ORE->emit([&]() { | |||
10513 | return OptimizationRemarkMissed(VAPassName, VecDiagMsg.first, | |||
10514 | L->getStartLoc(), L->getHeader()) | |||
10515 | << VecDiagMsg.second; | |||
10516 | }); | |||
10517 | ORE->emit([&]() { | |||
10518 | return OptimizationRemarkMissed(LV_NAME"loop-vectorize", IntDiagMsg.first, | |||
10519 | L->getStartLoc(), L->getHeader()) | |||
10520 | << IntDiagMsg.second; | |||
10521 | }); | |||
10522 | return false; | |||
10523 | } else if (!VectorizeLoop && InterleaveLoop) { | |||
10524 | LLVM_DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleave Count is " << IC << '\n'; } } while (false); | |||
10525 | ORE->emit([&]() { | |||
10526 | return OptimizationRemarkAnalysis(VAPassName, VecDiagMsg.first, | |||
10527 | L->getStartLoc(), L->getHeader()) | |||
10528 | << VecDiagMsg.second; | |||
10529 | }); | |||
10530 | } else if (VectorizeLoop && !InterleaveLoop) { | |||
10531 | LLVM_DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Widthdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in " << DebugLocStr << '\n'; } } while (false) | |||
10532 | << ") in " << DebugLocStr << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in " << DebugLocStr << '\n'; } } while (false); | |||
10533 | ORE->emit([&]() { | |||
10534 | return OptimizationRemarkAnalysis(LV_NAME"loop-vectorize", IntDiagMsg.first, | |||
10535 | L->getStartLoc(), L->getHeader()) | |||
10536 | << IntDiagMsg.second; | |||
10537 | }); | |||
10538 | } else if (VectorizeLoop && InterleaveLoop) { | |||
10539 | LLVM_DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Widthdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in " << DebugLocStr << '\n'; } } while (false) | |||
10540 | << ") in " << DebugLocStr << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in " << DebugLocStr << '\n'; } } while (false); | |||
10541 | LLVM_DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleave Count is " << IC << '\n'; } } while (false); | |||
10542 | } | |||
10543 | ||||
10544 | bool DisableRuntimeUnroll = false; | |||
10545 | MDNode *OrigLoopID = L->getLoopID(); | |||
10546 | { | |||
10547 | // Optimistically generate runtime checks. Drop them if they turn out to not | |||
10548 | // be profitable. Limit the scope of Checks, so the cleanup happens | |||
10549 | // immediately after vector codegeneration is done. | |||
10550 | GeneratedRTChecks Checks(*PSE.getSE(), DT, LI, | |||
10551 | F->getParent()->getDataLayout()); | |||
10552 | if (!VF.Width.isScalar() || IC > 1) | |||
10553 | Checks.Create(L, *LVL.getLAI(), PSE.getPredicate()); | |||
10554 | ||||
10555 | using namespace ore; | |||
10556 | if (!VectorizeLoop) { | |||
10557 | assert(IC > 1 && "interleave count should not be 1 or 0")(static_cast <bool> (IC > 1 && "interleave count should not be 1 or 0" ) ? void (0) : __assert_fail ("IC > 1 && \"interleave count should not be 1 or 0\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 10557, __extension__ __PRETTY_FUNCTION__)); | |||
10558 | // If we decided that it is not legal to vectorize the loop, then | |||
10559 | // interleave it. | |||
10560 | InnerLoopUnroller Unroller(L, PSE, LI, DT, TLI, TTI, AC, ORE, IC, &LVL, | |||
10561 | &CM, BFI, PSI, Checks); | |||
10562 | ||||
10563 | VPlan &BestPlan = LVP.getBestPlanFor(VF.Width); | |||
10564 | LVP.executePlan(VF.Width, IC, BestPlan, Unroller, DT); | |||
10565 | ||||
10566 | ORE->emit([&]() { | |||
10567 | return OptimizationRemark(LV_NAME"loop-vectorize", "Interleaved", L->getStartLoc(), | |||
10568 | L->getHeader()) | |||
10569 | << "interleaved loop (interleaved count: " | |||
10570 | << NV("InterleaveCount", IC) << ")"; | |||
10571 | }); | |||
10572 | } else { | |||
10573 | // If we decided that it is *legal* to vectorize the loop, then do it. | |||
10574 | ||||
10575 | // Consider vectorizing the epilogue too if it's profitable. | |||
10576 | VectorizationFactor EpilogueVF = | |||
10577 | CM.selectEpilogueVectorizationFactor(VF.Width, LVP); | |||
10578 | if (EpilogueVF.Width.isVector()) { | |||
10579 | ||||
10580 | // The first pass vectorizes the main loop and creates a scalar epilogue | |||
10581 | // to be vectorized by executing the plan (potentially with a different | |||
10582 | // factor) again shortly afterwards. | |||
10583 | EpilogueLoopVectorizationInfo EPI(VF.Width, IC, EpilogueVF.Width, 1); | |||
10584 | EpilogueVectorizerMainLoop MainILV(L, PSE, LI, DT, TLI, TTI, AC, ORE, | |||
10585 | EPI, &LVL, &CM, BFI, PSI, Checks); | |||
10586 | ||||
10587 | VPlan &BestMainPlan = LVP.getBestPlanFor(EPI.MainLoopVF); | |||
10588 | LVP.executePlan(EPI.MainLoopVF, EPI.MainLoopUF, BestMainPlan, MainILV, | |||
10589 | DT); | |||
10590 | ++LoopsVectorized; | |||
10591 | ||||
10592 | simplifyLoop(L, DT, LI, SE, AC, nullptr, false /* PreserveLCSSA */); | |||
10593 | formLCSSARecursively(*L, *DT, LI, SE); | |||
10594 | ||||
10595 | // Second pass vectorizes the epilogue and adjusts the control flow | |||
10596 | // edges from the first pass. | |||
10597 | EPI.MainLoopVF = EPI.EpilogueVF; | |||
10598 | EPI.MainLoopUF = EPI.EpilogueUF; | |||
10599 | EpilogueVectorizerEpilogueLoop EpilogILV(L, PSE, LI, DT, TLI, TTI, AC, | |||
10600 | ORE, EPI, &LVL, &CM, BFI, PSI, | |||
10601 | Checks); | |||
10602 | ||||
10603 | VPlan &BestEpiPlan = LVP.getBestPlanFor(EPI.EpilogueVF); | |||
10604 | ||||
10605 | // Ensure that the start values for any VPReductionPHIRecipes are | |||
10606 | // updated before vectorising the epilogue loop. | |||
10607 | VPBasicBlock *Header = BestEpiPlan.getEntry()->getEntryBasicBlock(); | |||
10608 | for (VPRecipeBase &R : Header->phis()) { | |||
10609 | if (auto *ReductionPhi = dyn_cast<VPReductionPHIRecipe>(&R)) { | |||
10610 | if (auto *Resume = MainILV.getReductionResumeValue( | |||
10611 | ReductionPhi->getRecurrenceDescriptor())) { | |||
10612 | VPValue *StartVal = new VPValue(Resume); | |||
10613 | BestEpiPlan.addExternalDef(StartVal); | |||
10614 | ReductionPhi->setOperand(0, StartVal); | |||
10615 | } | |||
10616 | } | |||
10617 | } | |||
10618 | ||||
10619 | LVP.executePlan(EPI.EpilogueVF, EPI.EpilogueUF, BestEpiPlan, EpilogILV, | |||
10620 | DT); | |||
10621 | ++LoopsEpilogueVectorized; | |||
10622 | ||||
10623 | if (!MainILV.areSafetyChecksAdded()) | |||
10624 | DisableRuntimeUnroll = true; | |||
10625 | } else { | |||
10626 | InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, VF.Width, IC, | |||
10627 | &LVL, &CM, BFI, PSI, Checks); | |||
10628 | ||||
10629 | VPlan &BestPlan = LVP.getBestPlanFor(VF.Width); | |||
10630 | LVP.executePlan(VF.Width, IC, BestPlan, LB, DT); | |||
10631 | ++LoopsVectorized; | |||
10632 | ||||
10633 | // Add metadata to disable runtime unrolling a scalar loop when there | |||
10634 | // are no runtime checks about strides and memory. A scalar loop that is | |||
10635 | // rarely used is not worth unrolling. | |||
10636 | if (!LB.areSafetyChecksAdded()) | |||
10637 | DisableRuntimeUnroll = true; | |||
10638 | } | |||
10639 | // Report the vectorization decision. | |||
10640 | ORE->emit([&]() { | |||
10641 | return OptimizationRemark(LV_NAME"loop-vectorize", "Vectorized", L->getStartLoc(), | |||
10642 | L->getHeader()) | |||
10643 | << "vectorized loop (vectorization width: " | |||
10644 | << NV("VectorizationFactor", VF.Width) | |||
10645 | << ", interleaved count: " << NV("InterleaveCount", IC) << ")"; | |||
10646 | }); | |||
10647 | } | |||
10648 | ||||
10649 | if (ORE->allowExtraAnalysis(LV_NAME"loop-vectorize")) | |||
10650 | checkMixedPrecision(L, ORE); | |||
10651 | } | |||
10652 | ||||
10653 | Optional<MDNode *> RemainderLoopID = | |||
10654 | makeFollowupLoopID(OrigLoopID, {LLVMLoopVectorizeFollowupAll, | |||
10655 | LLVMLoopVectorizeFollowupEpilogue}); | |||
10656 | if (RemainderLoopID.hasValue()) { | |||
10657 | L->setLoopID(RemainderLoopID.getValue()); | |||
10658 | } else { | |||
10659 | if (DisableRuntimeUnroll) | |||
10660 | AddRuntimeUnrollDisableMetaData(L); | |||
10661 | ||||
10662 | // Mark the loop as already vectorized to avoid vectorizing again. | |||
10663 | Hints.setAlreadyVectorized(); | |||
10664 | } | |||
10665 | ||||
10666 | assert(!verifyFunction(*L->getHeader()->getParent(), &dbgs()))(static_cast <bool> (!verifyFunction(*L->getHeader() ->getParent(), &dbgs())) ? void (0) : __assert_fail ("!verifyFunction(*L->getHeader()->getParent(), &dbgs())" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 10666, __extension__ __PRETTY_FUNCTION__)); | |||
10667 | return true; | |||
10668 | } | |||
10669 | ||||
10670 | LoopVectorizeResult LoopVectorizePass::runImpl( | |||
10671 | Function &F, ScalarEvolution &SE_, LoopInfo &LI_, TargetTransformInfo &TTI_, | |||
10672 | DominatorTree &DT_, BlockFrequencyInfo &BFI_, TargetLibraryInfo *TLI_, | |||
10673 | DemandedBits &DB_, AAResults &AA_, AssumptionCache &AC_, | |||
10674 | std::function<const LoopAccessInfo &(Loop &)> &GetLAA_, | |||
10675 | OptimizationRemarkEmitter &ORE_, ProfileSummaryInfo *PSI_) { | |||
10676 | SE = &SE_; | |||
10677 | LI = &LI_; | |||
10678 | TTI = &TTI_; | |||
10679 | DT = &DT_; | |||
10680 | BFI = &BFI_; | |||
10681 | TLI = TLI_; | |||
10682 | AA = &AA_; | |||
10683 | AC = &AC_; | |||
10684 | GetLAA = &GetLAA_; | |||
10685 | DB = &DB_; | |||
10686 | ORE = &ORE_; | |||
10687 | PSI = PSI_; | |||
10688 | ||||
10689 | // Don't attempt if | |||
10690 | // 1. the target claims to have no vector registers, and | |||
10691 | // 2. interleaving won't help ILP. | |||
10692 | // | |||
10693 | // The second condition is necessary because, even if the target has no | |||
10694 | // vector registers, loop vectorization may still enable scalar | |||
10695 | // interleaving. | |||
10696 | if (!TTI->getNumberOfRegisters(TTI->getRegisterClassForType(true)) && | |||
10697 | TTI->getMaxInterleaveFactor(1) < 2) | |||
10698 | return LoopVectorizeResult(false, false); | |||
10699 | ||||
10700 | bool Changed = false, CFGChanged = false; | |||
10701 | ||||
10702 | // The vectorizer requires loops to be in simplified form. | |||
10703 | // Since simplification may add new inner loops, it has to run before the | |||
10704 | // legality and profitability checks. This means running the loop vectorizer | |||
10705 | // will simplify all loops, regardless of whether anything end up being | |||
10706 | // vectorized. | |||
10707 | for (auto &L : *LI) | |||
10708 | Changed |= CFGChanged |= | |||
10709 | simplifyLoop(L, DT, LI, SE, AC, nullptr, false /* PreserveLCSSA */); | |||
10710 | ||||
10711 | // Build up a worklist of inner-loops to vectorize. This is necessary as | |||
10712 | // the act of vectorizing or partially unrolling a loop creates new loops | |||
10713 | // and can invalidate iterators across the loops. | |||
10714 | SmallVector<Loop *, 8> Worklist; | |||
10715 | ||||
10716 | for (Loop *L : *LI) | |||
10717 | collectSupportedLoops(*L, LI, ORE, Worklist); | |||
10718 | ||||
10719 | LoopsAnalyzed += Worklist.size(); | |||
10720 | ||||
10721 | // Now walk the identified inner loops. | |||
10722 | while (!Worklist.empty()) { | |||
10723 | Loop *L = Worklist.pop_back_val(); | |||
10724 | ||||
10725 | // For the inner loops we actually process, form LCSSA to simplify the | |||
10726 | // transform. | |||
10727 | Changed |= formLCSSARecursively(*L, *DT, LI, SE); | |||
10728 | ||||
10729 | Changed |= CFGChanged |= processLoop(L); | |||
10730 | } | |||
10731 | ||||
10732 | // Process each loop nest in the function. | |||
10733 | return LoopVectorizeResult(Changed, CFGChanged); | |||
10734 | } | |||
10735 | ||||
10736 | PreservedAnalyses LoopVectorizePass::run(Function &F, | |||
10737 | FunctionAnalysisManager &AM) { | |||
10738 | auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F); | |||
10739 | auto &LI = AM.getResult<LoopAnalysis>(F); | |||
10740 | auto &TTI = AM.getResult<TargetIRAnalysis>(F); | |||
10741 | auto &DT = AM.getResult<DominatorTreeAnalysis>(F); | |||
10742 | auto &BFI = AM.getResult<BlockFrequencyAnalysis>(F); | |||
10743 | auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); | |||
10744 | auto &AA = AM.getResult<AAManager>(F); | |||
10745 | auto &AC = AM.getResult<AssumptionAnalysis>(F); | |||
10746 | auto &DB = AM.getResult<DemandedBitsAnalysis>(F); | |||
10747 | auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F); | |||
10748 | ||||
10749 | auto &LAM = AM.getResult<LoopAnalysisManagerFunctionProxy>(F).getManager(); | |||
10750 | std::function<const LoopAccessInfo &(Loop &)> GetLAA = | |||
10751 | [&](Loop &L) -> const LoopAccessInfo & { | |||
10752 | LoopStandardAnalysisResults AR = {AA, AC, DT, LI, SE, | |||
10753 | TLI, TTI, nullptr, nullptr, nullptr}; | |||
10754 | return LAM.getResult<LoopAccessAnalysis>(L, AR); | |||
10755 | }; | |||
10756 | auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F); | |||
10757 | ProfileSummaryInfo *PSI = | |||
10758 | MAMProxy.getCachedResult<ProfileSummaryAnalysis>(*F.getParent()); | |||
10759 | LoopVectorizeResult Result = | |||
10760 | runImpl(F, SE, LI, TTI, DT, BFI, &TLI, DB, AA, AC, GetLAA, ORE, PSI); | |||
10761 | if (!Result.MadeAnyChange) | |||
10762 | return PreservedAnalyses::all(); | |||
10763 | PreservedAnalyses PA; | |||
10764 | ||||
10765 | // We currently do not preserve loopinfo/dominator analyses with outer loop | |||
10766 | // vectorization. Until this is addressed, mark these analyses as preserved | |||
10767 | // only for non-VPlan-native path. | |||
10768 | // TODO: Preserve Loop and Dominator analyses for VPlan-native path. | |||
10769 | if (!EnableVPlanNativePath) { | |||
10770 | PA.preserve<LoopAnalysis>(); | |||
10771 | PA.preserve<DominatorTreeAnalysis>(); | |||
10772 | } | |||
10773 | ||||
10774 | if (Result.MadeCFGChange) { | |||
10775 | // Making CFG changes likely means a loop got vectorized. Indicate that | |||
10776 | // extra simplification passes should be run. | |||
10777 | // TODO: MadeCFGChanges is not a prefect proxy. Extra passes should only | |||
10778 | // be run if runtime checks have been added. | |||
10779 | AM.getResult<ShouldRunExtraVectorPasses>(F); | |||
10780 | PA.preserve<ShouldRunExtraVectorPasses>(); | |||
10781 | } else { | |||
10782 | PA.preserveSet<CFGAnalyses>(); | |||
10783 | } | |||
10784 | return PA; | |||
10785 | } | |||
10786 | ||||
10787 | void LoopVectorizePass::printPipeline( | |||
10788 | raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { | |||
10789 | static_cast<PassInfoMixin<LoopVectorizePass> *>(this)->printPipeline( | |||
10790 | OS, MapClassName2PassName); | |||
10791 | ||||
10792 | OS << "<"; | |||
10793 | OS << (InterleaveOnlyWhenForced ? "" : "no-") << "interleave-forced-only;"; | |||
10794 | OS << (VectorizeOnlyWhenForced ? "" : "no-") << "vectorize-forced-only;"; | |||
10795 | OS << ">"; | |||
10796 | } |